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 MaxPairs("bb-vectorize-max-pairs-per-group", cl::init(3000), cl::Hidden,
92 cl::desc("The maximum number of candidate instruction pairs per group"));
94 static cl::opt<unsigned>
95 MaxCandPairsForCycleCheck("bb-vectorize-max-cycle-check-pairs", cl::init(200),
96 cl::Hidden, cl::desc("The maximum number of candidate pairs with which to use"
97 " a full cycle check"));
100 NoBools("bb-vectorize-no-bools", cl::init(false), cl::Hidden,
101 cl::desc("Don't try to vectorize boolean (i1) values"));
104 NoInts("bb-vectorize-no-ints", cl::init(false), cl::Hidden,
105 cl::desc("Don't try to vectorize integer values"));
108 NoFloats("bb-vectorize-no-floats", cl::init(false), cl::Hidden,
109 cl::desc("Don't try to vectorize floating-point values"));
111 // FIXME: This should default to false once pointer vector support works.
113 NoPointers("bb-vectorize-no-pointers", cl::init(/*false*/ true), cl::Hidden,
114 cl::desc("Don't try to vectorize pointer values"));
117 NoCasts("bb-vectorize-no-casts", cl::init(false), cl::Hidden,
118 cl::desc("Don't try to vectorize casting (conversion) operations"));
121 NoMath("bb-vectorize-no-math", cl::init(false), cl::Hidden,
122 cl::desc("Don't try to vectorize floating-point math intrinsics"));
125 NoFMA("bb-vectorize-no-fma", cl::init(false), cl::Hidden,
126 cl::desc("Don't try to vectorize the fused-multiply-add intrinsic"));
129 NoSelect("bb-vectorize-no-select", cl::init(false), cl::Hidden,
130 cl::desc("Don't try to vectorize select instructions"));
133 NoCmp("bb-vectorize-no-cmp", cl::init(false), cl::Hidden,
134 cl::desc("Don't try to vectorize comparison instructions"));
137 NoGEP("bb-vectorize-no-gep", cl::init(false), cl::Hidden,
138 cl::desc("Don't try to vectorize getelementptr instructions"));
141 NoMemOps("bb-vectorize-no-mem-ops", cl::init(false), cl::Hidden,
142 cl::desc("Don't try to vectorize loads and stores"));
145 AlignedOnly("bb-vectorize-aligned-only", cl::init(false), cl::Hidden,
146 cl::desc("Only generate aligned loads and stores"));
149 NoMemOpBoost("bb-vectorize-no-mem-op-boost",
150 cl::init(false), cl::Hidden,
151 cl::desc("Don't boost the chain-depth contribution of loads and stores"));
154 FastDep("bb-vectorize-fast-dep", cl::init(false), cl::Hidden,
155 cl::desc("Use a fast instruction dependency analysis"));
159 DebugInstructionExamination("bb-vectorize-debug-instruction-examination",
160 cl::init(false), cl::Hidden,
161 cl::desc("When debugging is enabled, output information on the"
162 " instruction-examination process"));
164 DebugCandidateSelection("bb-vectorize-debug-candidate-selection",
165 cl::init(false), cl::Hidden,
166 cl::desc("When debugging is enabled, output information on the"
167 " candidate-selection process"));
169 DebugPairSelection("bb-vectorize-debug-pair-selection",
170 cl::init(false), cl::Hidden,
171 cl::desc("When debugging is enabled, output information on the"
172 " pair-selection process"));
174 DebugCycleCheck("bb-vectorize-debug-cycle-check",
175 cl::init(false), cl::Hidden,
176 cl::desc("When debugging is enabled, output information on the"
177 " cycle-checking process"));
180 PrintAfterEveryPair("bb-vectorize-debug-print-after-every-pair",
181 cl::init(false), cl::Hidden,
182 cl::desc("When debugging is enabled, dump the basic block after"
183 " every pair is fused"));
186 STATISTIC(NumFusedOps, "Number of operations fused by bb-vectorize");
189 struct BBVectorize : public BasicBlockPass {
190 static char ID; // Pass identification, replacement for typeid
192 const VectorizeConfig Config;
194 BBVectorize(const VectorizeConfig &C = VectorizeConfig())
195 : BasicBlockPass(ID), Config(C) {
196 initializeBBVectorizePass(*PassRegistry::getPassRegistry());
199 BBVectorize(Pass *P, const VectorizeConfig &C)
200 : BasicBlockPass(ID), Config(C) {
201 AA = &P->getAnalysis<AliasAnalysis>();
202 DT = &P->getAnalysis<DominatorTree>();
203 SE = &P->getAnalysis<ScalarEvolution>();
204 TD = P->getAnalysisIfAvailable<DataLayout>();
205 TTI = IgnoreTargetInfo ? 0 : &P->getAnalysis<TargetTransformInfo>();
208 typedef std::pair<Value *, Value *> ValuePair;
209 typedef std::pair<ValuePair, int> ValuePairWithCost;
210 typedef std::pair<ValuePair, size_t> ValuePairWithDepth;
211 typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair
212 typedef std::pair<VPPair, unsigned> VPPairWithType;
218 const TargetTransformInfo *TTI;
220 // FIXME: const correct?
222 bool vectorizePairs(BasicBlock &BB, bool NonPow2Len = false);
224 bool getCandidatePairs(BasicBlock &BB,
225 BasicBlock::iterator &Start,
226 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
227 DenseSet<ValuePair> &FixedOrderPairs,
228 DenseMap<ValuePair, int> &CandidatePairCostSavings,
229 std::vector<Value *> &PairableInsts, bool NonPow2Len);
231 // FIXME: The current implementation does not account for pairs that
232 // are connected in multiple ways. For example:
233 // C1 = A1 / A2; C2 = A2 / A1 (which may be both direct and a swap)
234 enum PairConnectionType {
235 PairConnectionDirect,
240 void computeConnectedPairs(
241 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
242 DenseSet<ValuePair> &CandidatePairsSet,
243 std::vector<Value *> &PairableInsts,
244 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
245 DenseMap<VPPair, unsigned> &PairConnectionTypes);
247 void buildDepMap(BasicBlock &BB,
248 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
249 std::vector<Value *> &PairableInsts,
250 DenseSet<ValuePair> &PairableInstUsers);
252 void choosePairs(DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
253 DenseSet<ValuePair> &CandidatePairsSet,
254 DenseMap<ValuePair, int> &CandidatePairCostSavings,
255 std::vector<Value *> &PairableInsts,
256 DenseSet<ValuePair> &FixedOrderPairs,
257 DenseMap<VPPair, unsigned> &PairConnectionTypes,
258 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
259 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
260 DenseSet<ValuePair> &PairableInstUsers,
261 DenseMap<Value *, Value *>& ChosenPairs);
263 void fuseChosenPairs(BasicBlock &BB,
264 std::vector<Value *> &PairableInsts,
265 DenseMap<Value *, Value *>& ChosenPairs,
266 DenseSet<ValuePair> &FixedOrderPairs,
267 DenseMap<VPPair, unsigned> &PairConnectionTypes,
268 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
269 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps);
272 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore);
274 bool areInstsCompatible(Instruction *I, Instruction *J,
275 bool IsSimpleLoadStore, bool NonPow2Len,
276 int &CostSavings, int &FixedOrder);
278 bool trackUsesOfI(DenseSet<Value *> &Users,
279 AliasSetTracker &WriteSet, Instruction *I,
280 Instruction *J, bool UpdateUsers = true,
281 DenseSet<ValuePair> *LoadMoveSetPairs = 0);
283 void computePairsConnectedTo(
284 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
285 DenseSet<ValuePair> &CandidatePairsSet,
286 std::vector<Value *> &PairableInsts,
287 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
288 DenseMap<VPPair, unsigned> &PairConnectionTypes,
291 bool pairsConflict(ValuePair P, ValuePair Q,
292 DenseSet<ValuePair> &PairableInstUsers,
293 DenseMap<ValuePair, std::vector<ValuePair> >
294 *PairableInstUserMap = 0,
295 DenseSet<VPPair> *PairableInstUserPairSet = 0);
297 bool pairWillFormCycle(ValuePair P,
298 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUsers,
299 DenseSet<ValuePair> &CurrentPairs);
302 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
303 std::vector<Value *> &PairableInsts,
304 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
305 DenseSet<ValuePair> &PairableInstUsers,
306 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
307 DenseSet<VPPair> &PairableInstUserPairSet,
308 DenseMap<Value *, Value *> &ChosenPairs,
309 DenseMap<ValuePair, size_t> &DAG,
310 DenseSet<ValuePair> &PrunedDAG, ValuePair J,
313 void buildInitialDAGFor(
314 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
315 DenseSet<ValuePair> &CandidatePairsSet,
316 std::vector<Value *> &PairableInsts,
317 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
318 DenseSet<ValuePair> &PairableInstUsers,
319 DenseMap<Value *, Value *> &ChosenPairs,
320 DenseMap<ValuePair, size_t> &DAG, ValuePair J);
323 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
324 DenseSet<ValuePair> &CandidatePairsSet,
325 DenseMap<ValuePair, int> &CandidatePairCostSavings,
326 std::vector<Value *> &PairableInsts,
327 DenseSet<ValuePair> &FixedOrderPairs,
328 DenseMap<VPPair, unsigned> &PairConnectionTypes,
329 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
330 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
331 DenseSet<ValuePair> &PairableInstUsers,
332 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
333 DenseSet<VPPair> &PairableInstUserPairSet,
334 DenseMap<Value *, Value *> &ChosenPairs,
335 DenseSet<ValuePair> &BestDAG, size_t &BestMaxDepth,
336 int &BestEffSize, Value *II, std::vector<Value *>&JJ,
339 Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I,
340 Instruction *J, unsigned o);
342 void fillNewShuffleMask(LLVMContext& Context, Instruction *J,
343 unsigned MaskOffset, unsigned NumInElem,
344 unsigned NumInElem1, unsigned IdxOffset,
345 std::vector<Constant*> &Mask);
347 Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I,
350 bool expandIEChain(LLVMContext& Context, Instruction *I, Instruction *J,
351 unsigned o, Value *&LOp, unsigned numElemL,
352 Type *ArgTypeL, Type *ArgTypeR, bool IBeforeJ,
353 unsigned IdxOff = 0);
355 Value *getReplacementInput(LLVMContext& Context, Instruction *I,
356 Instruction *J, unsigned o, bool IBeforeJ);
358 void getReplacementInputsForPair(LLVMContext& Context, Instruction *I,
359 Instruction *J, SmallVectorImpl<Value *> &ReplacedOperands,
362 void replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
363 Instruction *J, Instruction *K,
364 Instruction *&InsertionPt, Instruction *&K1,
367 void collectPairLoadMoveSet(BasicBlock &BB,
368 DenseMap<Value *, Value *> &ChosenPairs,
369 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
370 DenseSet<ValuePair> &LoadMoveSetPairs,
373 void collectLoadMoveSet(BasicBlock &BB,
374 std::vector<Value *> &PairableInsts,
375 DenseMap<Value *, Value *> &ChosenPairs,
376 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
377 DenseSet<ValuePair> &LoadMoveSetPairs);
379 bool canMoveUsesOfIAfterJ(BasicBlock &BB,
380 DenseSet<ValuePair> &LoadMoveSetPairs,
381 Instruction *I, Instruction *J);
383 void moveUsesOfIAfterJ(BasicBlock &BB,
384 DenseSet<ValuePair> &LoadMoveSetPairs,
385 Instruction *&InsertionPt,
386 Instruction *I, Instruction *J);
388 void combineMetadata(Instruction *K, const Instruction *J);
390 bool vectorizeBB(BasicBlock &BB) {
391 if (!DT->isReachableFromEntry(&BB)) {
392 DEBUG(dbgs() << "BBV: skipping unreachable " << BB.getName() <<
393 " in " << BB.getParent()->getName() << "\n");
397 DEBUG(if (TTI) dbgs() << "BBV: using target information\n");
399 bool changed = false;
400 // Iterate a sufficient number of times to merge types of size 1 bit,
401 // then 2 bits, then 4, etc. up to half of the target vector width of the
402 // target vector register.
405 (TTI || v <= Config.VectorBits) &&
406 (!Config.MaxIter || n <= Config.MaxIter);
408 DEBUG(dbgs() << "BBV: fusing loop #" << n <<
409 " for " << BB.getName() << " in " <<
410 BB.getParent()->getName() << "...\n");
411 if (vectorizePairs(BB))
417 if (changed && !Pow2LenOnly) {
419 for (; !Config.MaxIter || n <= Config.MaxIter; ++n) {
420 DEBUG(dbgs() << "BBV: fusing for non-2^n-length vectors loop #: " <<
421 n << " for " << BB.getName() << " in " <<
422 BB.getParent()->getName() << "...\n");
423 if (!vectorizePairs(BB, true)) break;
427 DEBUG(dbgs() << "BBV: done!\n");
431 virtual bool runOnBasicBlock(BasicBlock &BB) {
432 AA = &getAnalysis<AliasAnalysis>();
433 DT = &getAnalysis<DominatorTree>();
434 SE = &getAnalysis<ScalarEvolution>();
435 TD = getAnalysisIfAvailable<DataLayout>();
436 TTI = IgnoreTargetInfo ? 0 : &getAnalysis<TargetTransformInfo>();
438 return vectorizeBB(BB);
441 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
442 BasicBlockPass::getAnalysisUsage(AU);
443 AU.addRequired<AliasAnalysis>();
444 AU.addRequired<DominatorTree>();
445 AU.addRequired<ScalarEvolution>();
446 AU.addRequired<TargetTransformInfo>();
447 AU.addPreserved<AliasAnalysis>();
448 AU.addPreserved<DominatorTree>();
449 AU.addPreserved<ScalarEvolution>();
450 AU.setPreservesCFG();
453 static inline VectorType *getVecTypeForPair(Type *ElemTy, Type *Elem2Ty) {
454 assert(ElemTy->getScalarType() == Elem2Ty->getScalarType() &&
455 "Cannot form vector from incompatible scalar types");
456 Type *STy = ElemTy->getScalarType();
459 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) {
460 numElem = VTy->getNumElements();
465 if (VectorType *VTy = dyn_cast<VectorType>(Elem2Ty)) {
466 numElem += VTy->getNumElements();
471 return VectorType::get(STy, numElem);
474 static inline void getInstructionTypes(Instruction *I,
475 Type *&T1, Type *&T2) {
476 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
477 // For stores, it is the value type, not the pointer type that matters
478 // because the value is what will come from a vector register.
480 Value *IVal = SI->getValueOperand();
481 T1 = IVal->getType();
486 if (CastInst *CI = dyn_cast<CastInst>(I))
491 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
492 T2 = SI->getCondition()->getType();
493 } else if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(I)) {
494 T2 = SI->getOperand(0)->getType();
495 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) {
496 T2 = CI->getOperand(0)->getType();
500 // Returns the weight associated with the provided value. A chain of
501 // candidate pairs has a length given by the sum of the weights of its
502 // members (one weight per pair; the weight of each member of the pair
503 // is assumed to be the same). This length is then compared to the
504 // chain-length threshold to determine if a given chain is significant
505 // enough to be vectorized. The length is also used in comparing
506 // candidate chains where longer chains are considered to be better.
507 // Note: when this function returns 0, the resulting instructions are
508 // not actually fused.
509 inline size_t getDepthFactor(Value *V) {
510 // InsertElement and ExtractElement have a depth factor of zero. This is
511 // for two reasons: First, they cannot be usefully fused. Second, because
512 // the pass generates a lot of these, they can confuse the simple metric
513 // used to compare the dags in the next iteration. Thus, giving them a
514 // weight of zero allows the pass to essentially ignore them in
515 // subsequent iterations when looking for vectorization opportunities
516 // while still tracking dependency chains that flow through those
518 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V))
521 // Give a load or store half of the required depth so that load/store
522 // pairs will vectorize.
523 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V)))
524 return Config.ReqChainDepth/2;
529 // Returns the cost of the provided instruction using TTI.
530 // This does not handle loads and stores.
531 unsigned getInstrCost(unsigned Opcode, Type *T1, Type *T2) {
534 case Instruction::GetElementPtr:
535 // We mark this instruction as zero-cost because scalar GEPs are usually
536 // lowered to the intruction addressing mode. At the moment we don't
537 // generate vector GEPs.
539 case Instruction::Br:
540 return TTI->getCFInstrCost(Opcode);
541 case Instruction::PHI:
543 case Instruction::Add:
544 case Instruction::FAdd:
545 case Instruction::Sub:
546 case Instruction::FSub:
547 case Instruction::Mul:
548 case Instruction::FMul:
549 case Instruction::UDiv:
550 case Instruction::SDiv:
551 case Instruction::FDiv:
552 case Instruction::URem:
553 case Instruction::SRem:
554 case Instruction::FRem:
555 case Instruction::Shl:
556 case Instruction::LShr:
557 case Instruction::AShr:
558 case Instruction::And:
559 case Instruction::Or:
560 case Instruction::Xor:
561 return TTI->getArithmeticInstrCost(Opcode, T1);
562 case Instruction::Select:
563 case Instruction::ICmp:
564 case Instruction::FCmp:
565 return TTI->getCmpSelInstrCost(Opcode, T1, T2);
566 case Instruction::ZExt:
567 case Instruction::SExt:
568 case Instruction::FPToUI:
569 case Instruction::FPToSI:
570 case Instruction::FPExt:
571 case Instruction::PtrToInt:
572 case Instruction::IntToPtr:
573 case Instruction::SIToFP:
574 case Instruction::UIToFP:
575 case Instruction::Trunc:
576 case Instruction::FPTrunc:
577 case Instruction::BitCast:
578 case Instruction::ShuffleVector:
579 return TTI->getCastInstrCost(Opcode, T1, T2);
585 // This determines the relative offset of two loads or stores, returning
586 // true if the offset could be determined to be some constant value.
587 // For example, if OffsetInElmts == 1, then J accesses the memory directly
588 // after I; if OffsetInElmts == -1 then I accesses the memory
590 bool getPairPtrInfo(Instruction *I, Instruction *J,
591 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment,
592 unsigned &IAddressSpace, unsigned &JAddressSpace,
593 int64_t &OffsetInElmts, bool ComputeOffset = true) {
595 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
596 LoadInst *LJ = cast<LoadInst>(J);
597 IPtr = LI->getPointerOperand();
598 JPtr = LJ->getPointerOperand();
599 IAlignment = LI->getAlignment();
600 JAlignment = LJ->getAlignment();
601 IAddressSpace = LI->getPointerAddressSpace();
602 JAddressSpace = LJ->getPointerAddressSpace();
604 StoreInst *SI = cast<StoreInst>(I), *SJ = cast<StoreInst>(J);
605 IPtr = SI->getPointerOperand();
606 JPtr = SJ->getPointerOperand();
607 IAlignment = SI->getAlignment();
608 JAlignment = SJ->getAlignment();
609 IAddressSpace = SI->getPointerAddressSpace();
610 JAddressSpace = SJ->getPointerAddressSpace();
616 const SCEV *IPtrSCEV = SE->getSCEV(IPtr);
617 const SCEV *JPtrSCEV = SE->getSCEV(JPtr);
619 // If this is a trivial offset, then we'll get something like
620 // 1*sizeof(type). With target data, which we need anyway, this will get
621 // constant folded into a number.
622 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV);
623 if (const SCEVConstant *ConstOffSCEV =
624 dyn_cast<SCEVConstant>(OffsetSCEV)) {
625 ConstantInt *IntOff = ConstOffSCEV->getValue();
626 int64_t Offset = IntOff->getSExtValue();
628 Type *VTy = cast<PointerType>(IPtr->getType())->getElementType();
629 int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy);
631 Type *VTy2 = cast<PointerType>(JPtr->getType())->getElementType();
632 if (VTy != VTy2 && Offset < 0) {
633 int64_t VTy2TSS = (int64_t) TD->getTypeStoreSize(VTy2);
634 OffsetInElmts = Offset/VTy2TSS;
635 return (abs64(Offset) % VTy2TSS) == 0;
638 OffsetInElmts = Offset/VTyTSS;
639 return (abs64(Offset) % VTyTSS) == 0;
645 // Returns true if the provided CallInst represents an intrinsic that can
647 bool isVectorizableIntrinsic(CallInst* I) {
648 Function *F = I->getCalledFunction();
649 if (!F) return false;
651 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID();
652 if (!IID) return false;
657 case Intrinsic::sqrt:
658 case Intrinsic::powi:
662 case Intrinsic::log2:
663 case Intrinsic::log10:
665 case Intrinsic::exp2:
667 return Config.VectorizeMath;
669 case Intrinsic::fmuladd:
670 return Config.VectorizeFMA;
674 bool isPureIEChain(InsertElementInst *IE) {
675 InsertElementInst *IENext = IE;
677 if (!isa<UndefValue>(IENext->getOperand(0)) &&
678 !isa<InsertElementInst>(IENext->getOperand(0))) {
682 dyn_cast<InsertElementInst>(IENext->getOperand(0))));
688 // This function implements one vectorization iteration on the provided
689 // basic block. It returns true if the block is changed.
690 bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) {
692 BasicBlock::iterator Start = BB.getFirstInsertionPt();
694 std::vector<Value *> AllPairableInsts;
695 DenseMap<Value *, Value *> AllChosenPairs;
696 DenseSet<ValuePair> AllFixedOrderPairs;
697 DenseMap<VPPair, unsigned> AllPairConnectionTypes;
698 DenseMap<ValuePair, std::vector<ValuePair> > AllConnectedPairs,
699 AllConnectedPairDeps;
702 std::vector<Value *> PairableInsts;
703 DenseMap<Value *, std::vector<Value *> > CandidatePairs;
704 DenseSet<ValuePair> FixedOrderPairs;
705 DenseMap<ValuePair, int> CandidatePairCostSavings;
706 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs,
708 CandidatePairCostSavings,
709 PairableInsts, NonPow2Len);
710 if (PairableInsts.empty()) continue;
712 // Build the candidate pair set for faster lookups.
713 DenseSet<ValuePair> CandidatePairsSet;
714 for (DenseMap<Value *, std::vector<Value *> >::iterator I =
715 CandidatePairs.begin(), E = CandidatePairs.end(); I != E; ++I)
716 for (std::vector<Value *>::iterator J = I->second.begin(),
717 JE = I->second.end(); J != JE; ++J)
718 CandidatePairsSet.insert(ValuePair(I->first, *J));
720 // Now we have a map of all of the pairable instructions and we need to
721 // select the best possible pairing. A good pairing is one such that the
722 // users of the pair are also paired. This defines a (directed) forest
723 // over the pairs such that two pairs are connected iff the second pair
726 // Note that it only matters that both members of the second pair use some
727 // element of the first pair (to allow for splatting).
729 DenseMap<ValuePair, std::vector<ValuePair> > ConnectedPairs,
731 DenseMap<VPPair, unsigned> PairConnectionTypes;
732 computeConnectedPairs(CandidatePairs, CandidatePairsSet,
733 PairableInsts, ConnectedPairs, PairConnectionTypes);
734 if (ConnectedPairs.empty()) continue;
736 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator
737 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
739 for (std::vector<ValuePair>::iterator J = I->second.begin(),
740 JE = I->second.end(); J != JE; ++J)
741 ConnectedPairDeps[*J].push_back(I->first);
743 // Build the pairable-instruction dependency map
744 DenseSet<ValuePair> PairableInstUsers;
745 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers);
747 // There is now a graph of the connected pairs. For each variable, pick
748 // the pairing with the largest dag meeting the depth requirement on at
749 // least one branch. Then select all pairings that are part of that dag
750 // and remove them from the list of available pairings and pairable
753 DenseMap<Value *, Value *> ChosenPairs;
754 choosePairs(CandidatePairs, CandidatePairsSet,
755 CandidatePairCostSavings,
756 PairableInsts, FixedOrderPairs, PairConnectionTypes,
757 ConnectedPairs, ConnectedPairDeps,
758 PairableInstUsers, ChosenPairs);
760 if (ChosenPairs.empty()) continue;
761 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(),
762 PairableInsts.end());
763 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end());
765 // Only for the chosen pairs, propagate information on fixed-order pairs,
766 // pair connections, and their types to the data structures used by the
767 // pair fusion procedures.
768 for (DenseMap<Value *, Value *>::iterator I = ChosenPairs.begin(),
769 IE = ChosenPairs.end(); I != IE; ++I) {
770 if (FixedOrderPairs.count(*I))
771 AllFixedOrderPairs.insert(*I);
772 else if (FixedOrderPairs.count(ValuePair(I->second, I->first)))
773 AllFixedOrderPairs.insert(ValuePair(I->second, I->first));
775 for (DenseMap<Value *, Value *>::iterator J = ChosenPairs.begin();
777 DenseMap<VPPair, unsigned>::iterator K =
778 PairConnectionTypes.find(VPPair(*I, *J));
779 if (K != PairConnectionTypes.end()) {
780 AllPairConnectionTypes.insert(*K);
782 K = PairConnectionTypes.find(VPPair(*J, *I));
783 if (K != PairConnectionTypes.end())
784 AllPairConnectionTypes.insert(*K);
789 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator
790 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
792 for (std::vector<ValuePair>::iterator J = I->second.begin(),
793 JE = I->second.end(); J != JE; ++J)
794 if (AllPairConnectionTypes.count(VPPair(I->first, *J))) {
795 AllConnectedPairs[I->first].push_back(*J);
796 AllConnectedPairDeps[*J].push_back(I->first);
798 } while (ShouldContinue);
800 if (AllChosenPairs.empty()) return false;
801 NumFusedOps += AllChosenPairs.size();
803 // A set of pairs has now been selected. It is now necessary to replace the
804 // paired instructions with vector instructions. For this procedure each
805 // operand must be replaced with a vector operand. This vector is formed
806 // by using build_vector on the old operands. The replaced values are then
807 // replaced with a vector_extract on the result. Subsequent optimization
808 // passes should coalesce the build/extract combinations.
810 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs, AllFixedOrderPairs,
811 AllPairConnectionTypes,
812 AllConnectedPairs, AllConnectedPairDeps);
814 // It is important to cleanup here so that future iterations of this
815 // function have less work to do.
816 (void) SimplifyInstructionsInBlock(&BB, TD, AA->getTargetLibraryInfo());
820 // This function returns true if the provided instruction is capable of being
821 // fused into a vector instruction. This determination is based only on the
822 // type and other attributes of the instruction.
823 bool BBVectorize::isInstVectorizable(Instruction *I,
824 bool &IsSimpleLoadStore) {
825 IsSimpleLoadStore = false;
827 if (CallInst *C = dyn_cast<CallInst>(I)) {
828 if (!isVectorizableIntrinsic(C))
830 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) {
831 // Vectorize simple loads if possbile:
832 IsSimpleLoadStore = L->isSimple();
833 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
835 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) {
836 // Vectorize simple stores if possbile:
837 IsSimpleLoadStore = S->isSimple();
838 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
840 } else if (CastInst *C = dyn_cast<CastInst>(I)) {
841 // We can vectorize casts, but not casts of pointer types, etc.
842 if (!Config.VectorizeCasts)
845 Type *SrcTy = C->getSrcTy();
846 if (!SrcTy->isSingleValueType())
849 Type *DestTy = C->getDestTy();
850 if (!DestTy->isSingleValueType())
852 } else if (isa<SelectInst>(I)) {
853 if (!Config.VectorizeSelect)
855 } else if (isa<CmpInst>(I)) {
856 if (!Config.VectorizeCmp)
858 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) {
859 if (!Config.VectorizeGEP)
862 // Currently, vector GEPs exist only with one index.
863 if (G->getNumIndices() != 1)
865 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) ||
866 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) {
870 // We can't vectorize memory operations without target data
871 if (TD == 0 && IsSimpleLoadStore)
875 getInstructionTypes(I, T1, T2);
877 // Not every type can be vectorized...
878 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) ||
879 !(VectorType::isValidElementType(T2) || T2->isVectorTy()))
882 if (T1->getScalarSizeInBits() == 1) {
883 if (!Config.VectorizeBools)
886 if (!Config.VectorizeInts && T1->isIntOrIntVectorTy())
890 if (T2->getScalarSizeInBits() == 1) {
891 if (!Config.VectorizeBools)
894 if (!Config.VectorizeInts && T2->isIntOrIntVectorTy())
898 if (!Config.VectorizeFloats
899 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy()))
902 // Don't vectorize target-specific types.
903 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy())
905 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy())
908 if ((!Config.VectorizePointers || TD == 0) &&
909 (T1->getScalarType()->isPointerTy() ||
910 T2->getScalarType()->isPointerTy()))
913 if (!TTI && (T1->getPrimitiveSizeInBits() >= Config.VectorBits ||
914 T2->getPrimitiveSizeInBits() >= Config.VectorBits))
920 // This function returns true if the two provided instructions are compatible
921 // (meaning that they can be fused into a vector instruction). This assumes
922 // that I has already been determined to be vectorizable and that J is not
923 // in the use dag of I.
924 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J,
925 bool IsSimpleLoadStore, bool NonPow2Len,
926 int &CostSavings, int &FixedOrder) {
927 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I <<
928 " <-> " << *J << "\n");
933 // Loads and stores can be merged if they have different alignments,
934 // but are otherwise the same.
935 if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment |
936 (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0)))
939 Type *IT1, *IT2, *JT1, *JT2;
940 getInstructionTypes(I, IT1, IT2);
941 getInstructionTypes(J, JT1, JT2);
942 unsigned MaxTypeBits = std::max(
943 IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(),
944 IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits());
945 if (!TTI && MaxTypeBits > Config.VectorBits)
948 // FIXME: handle addsub-type operations!
950 if (IsSimpleLoadStore) {
952 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
953 int64_t OffsetInElmts = 0;
954 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
955 IAddressSpace, JAddressSpace,
956 OffsetInElmts) && abs64(OffsetInElmts) == 1) {
957 FixedOrder = (int) OffsetInElmts;
958 unsigned BottomAlignment = IAlignment;
959 if (OffsetInElmts < 0) BottomAlignment = JAlignment;
961 Type *aTypeI = isa<StoreInst>(I) ?
962 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType();
963 Type *aTypeJ = isa<StoreInst>(J) ?
964 cast<StoreInst>(J)->getValueOperand()->getType() : J->getType();
965 Type *VType = getVecTypeForPair(aTypeI, aTypeJ);
967 if (Config.AlignedOnly) {
968 // An aligned load or store is possible only if the instruction
969 // with the lower offset has an alignment suitable for the
972 unsigned VecAlignment = TD->getPrefTypeAlignment(VType);
973 if (BottomAlignment < VecAlignment)
978 unsigned ICost = TTI->getMemoryOpCost(I->getOpcode(), aTypeI,
979 IAlignment, IAddressSpace);
980 unsigned JCost = TTI->getMemoryOpCost(J->getOpcode(), aTypeJ,
981 JAlignment, JAddressSpace);
982 unsigned VCost = TTI->getMemoryOpCost(I->getOpcode(), VType,
986 ICost += TTI->getAddressComputationCost(aTypeI);
987 JCost += TTI->getAddressComputationCost(aTypeJ);
988 VCost += TTI->getAddressComputationCost(VType);
990 if (VCost > ICost + JCost)
993 // We don't want to fuse to a type that will be split, even
994 // if the two input types will also be split and there is no other
996 unsigned VParts = TTI->getNumberOfParts(VType);
999 else if (!VParts && VCost == ICost + JCost)
1002 CostSavings = ICost + JCost - VCost;
1008 unsigned ICost = getInstrCost(I->getOpcode(), IT1, IT2);
1009 unsigned JCost = getInstrCost(J->getOpcode(), JT1, JT2);
1010 Type *VT1 = getVecTypeForPair(IT1, JT1),
1011 *VT2 = getVecTypeForPair(IT2, JT2);
1013 // Note that this procedure is incorrect for insert and extract element
1014 // instructions (because combining these often results in a shuffle),
1015 // but this cost is ignored (because insert and extract element
1016 // instructions are assigned a zero depth factor and are not really
1017 // fused in general).
1018 unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2);
1020 if (VCost > ICost + JCost)
1023 // We don't want to fuse to a type that will be split, even
1024 // if the two input types will also be split and there is no other
1026 unsigned VParts1 = TTI->getNumberOfParts(VT1),
1027 VParts2 = TTI->getNumberOfParts(VT2);
1028 if (VParts1 > 1 || VParts2 > 1)
1030 else if ((!VParts1 || !VParts2) && VCost == ICost + JCost)
1033 CostSavings = ICost + JCost - VCost;
1036 // The powi intrinsic is special because only the first argument is
1037 // vectorized, the second arguments must be equal.
1038 CallInst *CI = dyn_cast<CallInst>(I);
1040 if (CI && (FI = CI->getCalledFunction())) {
1041 Intrinsic::ID IID = (Intrinsic::ID) FI->getIntrinsicID();
1042 if (IID == Intrinsic::powi) {
1043 Value *A1I = CI->getArgOperand(1),
1044 *A1J = cast<CallInst>(J)->getArgOperand(1);
1045 const SCEV *A1ISCEV = SE->getSCEV(A1I),
1046 *A1JSCEV = SE->getSCEV(A1J);
1047 return (A1ISCEV == A1JSCEV);
1051 SmallVector<Type*, 4> Tys;
1052 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
1053 Tys.push_back(CI->getArgOperand(i)->getType());
1054 unsigned ICost = TTI->getIntrinsicInstrCost(IID, IT1, Tys);
1057 CallInst *CJ = cast<CallInst>(J);
1058 for (unsigned i = 0, ie = CJ->getNumArgOperands(); i != ie; ++i)
1059 Tys.push_back(CJ->getArgOperand(i)->getType());
1060 unsigned JCost = TTI->getIntrinsicInstrCost(IID, JT1, Tys);
1063 assert(CI->getNumArgOperands() == CJ->getNumArgOperands() &&
1064 "Intrinsic argument counts differ");
1065 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
1066 if (IID == Intrinsic::powi && i == 1)
1067 Tys.push_back(CI->getArgOperand(i)->getType());
1069 Tys.push_back(getVecTypeForPair(CI->getArgOperand(i)->getType(),
1070 CJ->getArgOperand(i)->getType()));
1073 Type *RetTy = getVecTypeForPair(IT1, JT1);
1074 unsigned VCost = TTI->getIntrinsicInstrCost(IID, RetTy, Tys);
1076 if (VCost > ICost + JCost)
1079 // We don't want to fuse to a type that will be split, even
1080 // if the two input types will also be split and there is no other
1082 unsigned RetParts = TTI->getNumberOfParts(RetTy);
1085 else if (!RetParts && VCost == ICost + JCost)
1088 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
1089 if (!Tys[i]->isVectorTy())
1092 unsigned NumParts = TTI->getNumberOfParts(Tys[i]);
1095 else if (!NumParts && VCost == ICost + JCost)
1099 CostSavings = ICost + JCost - VCost;
1106 // Figure out whether or not J uses I and update the users and write-set
1107 // structures associated with I. Specifically, Users represents the set of
1108 // instructions that depend on I. WriteSet represents the set
1109 // of memory locations that are dependent on I. If UpdateUsers is true,
1110 // and J uses I, then Users is updated to contain J and WriteSet is updated
1111 // to contain any memory locations to which J writes. The function returns
1112 // true if J uses I. By default, alias analysis is used to determine
1113 // whether J reads from memory that overlaps with a location in WriteSet.
1114 // If LoadMoveSet is not null, then it is a previously-computed map
1115 // where the key is the memory-based user instruction and the value is
1116 // the instruction to be compared with I. So, if LoadMoveSet is provided,
1117 // then the alias analysis is not used. This is necessary because this
1118 // function is called during the process of moving instructions during
1119 // vectorization and the results of the alias analysis are not stable during
1121 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users,
1122 AliasSetTracker &WriteSet, Instruction *I,
1123 Instruction *J, bool UpdateUsers,
1124 DenseSet<ValuePair> *LoadMoveSetPairs) {
1127 // This instruction may already be marked as a user due, for example, to
1128 // being a member of a selected pair.
1133 for (User::op_iterator JU = J->op_begin(), JE = J->op_end();
1136 if (I == V || Users.count(V)) {
1141 if (!UsesI && J->mayReadFromMemory()) {
1142 if (LoadMoveSetPairs) {
1143 UsesI = LoadMoveSetPairs->count(ValuePair(J, I));
1145 for (AliasSetTracker::iterator W = WriteSet.begin(),
1146 WE = WriteSet.end(); W != WE; ++W) {
1147 if (W->aliasesUnknownInst(J, *AA)) {
1155 if (UsesI && UpdateUsers) {
1156 if (J->mayWriteToMemory()) WriteSet.add(J);
1163 // This function iterates over all instruction pairs in the provided
1164 // basic block and collects all candidate pairs for vectorization.
1165 bool BBVectorize::getCandidatePairs(BasicBlock &BB,
1166 BasicBlock::iterator &Start,
1167 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1168 DenseSet<ValuePair> &FixedOrderPairs,
1169 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1170 std::vector<Value *> &PairableInsts, bool NonPow2Len) {
1171 size_t TotalPairs = 0;
1172 BasicBlock::iterator E = BB.end();
1173 if (Start == E) return false;
1175 bool ShouldContinue = false, IAfterStart = false;
1176 for (BasicBlock::iterator I = Start++; I != E; ++I) {
1177 if (I == Start) IAfterStart = true;
1179 bool IsSimpleLoadStore;
1180 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue;
1182 // Look for an instruction with which to pair instruction *I...
1183 DenseSet<Value *> Users;
1184 AliasSetTracker WriteSet(*AA);
1185 if (I->mayWriteToMemory()) WriteSet.add(I);
1187 bool JAfterStart = IAfterStart;
1188 BasicBlock::iterator J = llvm::next(I);
1189 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) {
1190 if (J == Start) JAfterStart = true;
1192 // Determine if J uses I, if so, exit the loop.
1193 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep);
1194 if (Config.FastDep) {
1195 // Note: For this heuristic to be effective, independent operations
1196 // must tend to be intermixed. This is likely to be true from some
1197 // kinds of grouped loop unrolling (but not the generic LLVM pass),
1198 // but otherwise may require some kind of reordering pass.
1200 // When using fast dependency analysis,
1201 // stop searching after first use:
1204 if (UsesI) continue;
1207 // J does not use I, and comes before the first use of I, so it can be
1208 // merged with I if the instructions are compatible.
1209 int CostSavings, FixedOrder;
1210 if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len,
1211 CostSavings, FixedOrder)) continue;
1213 // J is a candidate for merging with I.
1214 if (!PairableInsts.size() ||
1215 PairableInsts[PairableInsts.size()-1] != I) {
1216 PairableInsts.push_back(I);
1219 CandidatePairs[I].push_back(J);
1222 CandidatePairCostSavings.insert(ValuePairWithCost(ValuePair(I, J),
1225 if (FixedOrder == 1)
1226 FixedOrderPairs.insert(ValuePair(I, J));
1227 else if (FixedOrder == -1)
1228 FixedOrderPairs.insert(ValuePair(J, I));
1230 // The next call to this function must start after the last instruction
1231 // selected during this invocation.
1233 Start = llvm::next(J);
1234 IAfterStart = JAfterStart = false;
1237 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair "
1238 << *I << " <-> " << *J << " (cost savings: " <<
1239 CostSavings << ")\n");
1241 // If we have already found too many pairs, break here and this function
1242 // will be called again starting after the last instruction selected
1243 // during this invocation.
1244 if (PairableInsts.size() >= Config.MaxInsts ||
1245 TotalPairs >= Config.MaxPairs) {
1246 ShouldContinue = true;
1255 DEBUG(dbgs() << "BBV: found " << PairableInsts.size()
1256 << " instructions with candidate pairs\n");
1258 return ShouldContinue;
1261 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that
1262 // it looks for pairs such that both members have an input which is an
1263 // output of PI or PJ.
1264 void BBVectorize::computePairsConnectedTo(
1265 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1266 DenseSet<ValuePair> &CandidatePairsSet,
1267 std::vector<Value *> &PairableInsts,
1268 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1269 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1273 // For each possible pairing for this variable, look at the uses of
1274 // the first value...
1275 for (Value::use_iterator I = P.first->use_begin(),
1276 E = P.first->use_end(); I != E; ++I) {
1277 if (isa<LoadInst>(*I)) {
1278 // A pair cannot be connected to a load because the load only takes one
1279 // operand (the address) and it is a scalar even after vectorization.
1281 } else if ((SI = dyn_cast<StoreInst>(*I)) &&
1282 P.first == SI->getPointerOperand()) {
1283 // Similarly, a pair cannot be connected to a store through its
1288 // For each use of the first variable, look for uses of the second
1290 for (Value::use_iterator J = P.second->use_begin(),
1291 E2 = P.second->use_end(); J != E2; ++J) {
1292 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1293 P.second == SJ->getPointerOperand())
1297 if (CandidatePairsSet.count(ValuePair(*I, *J))) {
1298 VPPair VP(P, ValuePair(*I, *J));
1299 ConnectedPairs[VP.first].push_back(VP.second);
1300 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionDirect));
1304 if (CandidatePairsSet.count(ValuePair(*J, *I))) {
1305 VPPair VP(P, ValuePair(*J, *I));
1306 ConnectedPairs[VP.first].push_back(VP.second);
1307 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSwap));
1311 if (Config.SplatBreaksChain) continue;
1312 // Look for cases where just the first value in the pair is used by
1313 // both members of another pair (splatting).
1314 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) {
1315 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1316 P.first == SJ->getPointerOperand())
1319 if (CandidatePairsSet.count(ValuePair(*I, *J))) {
1320 VPPair VP(P, ValuePair(*I, *J));
1321 ConnectedPairs[VP.first].push_back(VP.second);
1322 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1327 if (Config.SplatBreaksChain) return;
1328 // Look for cases where just the second value in the pair is used by
1329 // both members of another pair (splatting).
1330 for (Value::use_iterator I = P.second->use_begin(),
1331 E = P.second->use_end(); I != E; ++I) {
1332 if (isa<LoadInst>(*I))
1334 else if ((SI = dyn_cast<StoreInst>(*I)) &&
1335 P.second == SI->getPointerOperand())
1338 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) {
1339 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1340 P.second == SJ->getPointerOperand())
1343 if (CandidatePairsSet.count(ValuePair(*I, *J))) {
1344 VPPair VP(P, ValuePair(*I, *J));
1345 ConnectedPairs[VP.first].push_back(VP.second);
1346 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1352 // This function figures out which pairs are connected. Two pairs are
1353 // connected if some output of the first pair forms an input to both members
1354 // of the second pair.
1355 void BBVectorize::computeConnectedPairs(
1356 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1357 DenseSet<ValuePair> &CandidatePairsSet,
1358 std::vector<Value *> &PairableInsts,
1359 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1360 DenseMap<VPPair, unsigned> &PairConnectionTypes) {
1361 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
1362 PE = PairableInsts.end(); PI != PE; ++PI) {
1363 DenseMap<Value *, std::vector<Value *> >::iterator PP =
1364 CandidatePairs.find(*PI);
1365 if (PP == CandidatePairs.end())
1368 for (std::vector<Value *>::iterator P = PP->second.begin(),
1369 E = PP->second.end(); P != E; ++P)
1370 computePairsConnectedTo(CandidatePairs, CandidatePairsSet,
1371 PairableInsts, ConnectedPairs,
1372 PairConnectionTypes, ValuePair(*PI, *P));
1375 DEBUG(size_t TotalPairs = 0;
1376 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator I =
1377 ConnectedPairs.begin(), IE = ConnectedPairs.end(); I != IE; ++I)
1378 TotalPairs += I->second.size();
1379 dbgs() << "BBV: found " << TotalPairs
1380 << " pair connections.\n");
1383 // This function builds a set of use tuples such that <A, B> is in the set
1384 // if B is in the use dag of A. If B is in the use dag of A, then B
1385 // depends on the output of A.
1386 void BBVectorize::buildDepMap(
1388 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1389 std::vector<Value *> &PairableInsts,
1390 DenseSet<ValuePair> &PairableInstUsers) {
1391 DenseSet<Value *> IsInPair;
1392 for (DenseMap<Value *, std::vector<Value *> >::iterator C =
1393 CandidatePairs.begin(), E = CandidatePairs.end(); C != E; ++C) {
1394 IsInPair.insert(C->first);
1395 IsInPair.insert(C->second.begin(), C->second.end());
1398 // Iterate through the basic block, recording all users of each
1399 // pairable instruction.
1401 BasicBlock::iterator E = BB.end(), EL =
1402 BasicBlock::iterator(cast<Instruction>(PairableInsts.back()));
1403 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) {
1404 if (IsInPair.find(I) == IsInPair.end()) continue;
1406 DenseSet<Value *> Users;
1407 AliasSetTracker WriteSet(*AA);
1408 if (I->mayWriteToMemory()) WriteSet.add(I);
1410 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J) {
1411 (void) trackUsesOfI(Users, WriteSet, I, J);
1417 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end();
1419 if (IsInPair.find(*U) == IsInPair.end()) continue;
1420 PairableInstUsers.insert(ValuePair(I, *U));
1428 // Returns true if an input to pair P is an output of pair Q and also an
1429 // input of pair Q is an output of pair P. If this is the case, then these
1430 // two pairs cannot be simultaneously fused.
1431 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q,
1432 DenseSet<ValuePair> &PairableInstUsers,
1433 DenseMap<ValuePair, std::vector<ValuePair> > *PairableInstUserMap,
1434 DenseSet<VPPair> *PairableInstUserPairSet) {
1435 // Two pairs are in conflict if they are mutual Users of eachother.
1436 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) ||
1437 PairableInstUsers.count(ValuePair(P.first, Q.second)) ||
1438 PairableInstUsers.count(ValuePair(P.second, Q.first)) ||
1439 PairableInstUsers.count(ValuePair(P.second, Q.second));
1440 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) ||
1441 PairableInstUsers.count(ValuePair(Q.first, P.second)) ||
1442 PairableInstUsers.count(ValuePair(Q.second, P.first)) ||
1443 PairableInstUsers.count(ValuePair(Q.second, P.second));
1444 if (PairableInstUserMap) {
1445 // FIXME: The expensive part of the cycle check is not so much the cycle
1446 // check itself but this edge insertion procedure. This needs some
1447 // profiling and probably a different data structure.
1449 if (PairableInstUserPairSet->insert(VPPair(Q, P)).second)
1450 (*PairableInstUserMap)[Q].push_back(P);
1453 if (PairableInstUserPairSet->insert(VPPair(P, Q)).second)
1454 (*PairableInstUserMap)[P].push_back(Q);
1458 return (QUsesP && PUsesQ);
1461 // This function walks the use graph of current pairs to see if, starting
1462 // from P, the walk returns to P.
1463 bool BBVectorize::pairWillFormCycle(ValuePair P,
1464 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
1465 DenseSet<ValuePair> &CurrentPairs) {
1466 DEBUG(if (DebugCycleCheck)
1467 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> "
1468 << *P.second << "\n");
1469 // A lookup table of visisted pairs is kept because the PairableInstUserMap
1470 // contains non-direct associations.
1471 DenseSet<ValuePair> Visited;
1472 SmallVector<ValuePair, 32> Q;
1473 // General depth-first post-order traversal:
1476 ValuePair QTop = Q.pop_back_val();
1477 Visited.insert(QTop);
1479 DEBUG(if (DebugCycleCheck)
1480 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> "
1481 << *QTop.second << "\n");
1482 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
1483 PairableInstUserMap.find(QTop);
1484 if (QQ == PairableInstUserMap.end())
1487 for (std::vector<ValuePair>::iterator C = QQ->second.begin(),
1488 CE = QQ->second.end(); C != CE; ++C) {
1491 << "BBV: rejected to prevent non-trivial cycle formation: "
1492 << QTop.first << " <-> " << C->second << "\n");
1496 if (CurrentPairs.count(*C) && !Visited.count(*C))
1499 } while (!Q.empty());
1504 // This function builds the initial dag of connected pairs with the
1505 // pair J at the root.
1506 void BBVectorize::buildInitialDAGFor(
1507 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1508 DenseSet<ValuePair> &CandidatePairsSet,
1509 std::vector<Value *> &PairableInsts,
1510 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1511 DenseSet<ValuePair> &PairableInstUsers,
1512 DenseMap<Value *, Value *> &ChosenPairs,
1513 DenseMap<ValuePair, size_t> &DAG, ValuePair J) {
1514 // Each of these pairs is viewed as the root node of a DAG. The DAG
1515 // is then walked (depth-first). As this happens, we keep track of
1516 // the pairs that compose the DAG and the maximum depth of the DAG.
1517 SmallVector<ValuePairWithDepth, 32> Q;
1518 // General depth-first post-order traversal:
1519 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1521 ValuePairWithDepth QTop = Q.back();
1523 // Push each child onto the queue:
1524 bool MoreChildren = false;
1525 size_t MaxChildDepth = QTop.second;
1526 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
1527 ConnectedPairs.find(QTop.first);
1528 if (QQ != ConnectedPairs.end())
1529 for (std::vector<ValuePair>::iterator k = QQ->second.begin(),
1530 ke = QQ->second.end(); k != ke; ++k) {
1531 // Make sure that this child pair is still a candidate:
1532 if (CandidatePairsSet.count(*k)) {
1533 DenseMap<ValuePair, size_t>::iterator C = DAG.find(*k);
1534 if (C == DAG.end()) {
1535 size_t d = getDepthFactor(k->first);
1536 Q.push_back(ValuePairWithDepth(*k, QTop.second+d));
1537 MoreChildren = true;
1539 MaxChildDepth = std::max(MaxChildDepth, C->second);
1544 if (!MoreChildren) {
1545 // Record the current pair as part of the DAG:
1546 DAG.insert(ValuePairWithDepth(QTop.first, MaxChildDepth));
1549 } while (!Q.empty());
1552 // Given some initial dag, prune it by removing conflicting pairs (pairs
1553 // that cannot be simultaneously chosen for vectorization).
1554 void BBVectorize::pruneDAGFor(
1555 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1556 std::vector<Value *> &PairableInsts,
1557 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1558 DenseSet<ValuePair> &PairableInstUsers,
1559 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
1560 DenseSet<VPPair> &PairableInstUserPairSet,
1561 DenseMap<Value *, Value *> &ChosenPairs,
1562 DenseMap<ValuePair, size_t> &DAG,
1563 DenseSet<ValuePair> &PrunedDAG, ValuePair J,
1564 bool UseCycleCheck) {
1565 SmallVector<ValuePairWithDepth, 32> Q;
1566 // General depth-first post-order traversal:
1567 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1569 ValuePairWithDepth QTop = Q.pop_back_val();
1570 PrunedDAG.insert(QTop.first);
1572 // Visit each child, pruning as necessary...
1573 SmallVector<ValuePairWithDepth, 8> BestChildren;
1574 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
1575 ConnectedPairs.find(QTop.first);
1576 if (QQ == ConnectedPairs.end())
1579 for (std::vector<ValuePair>::iterator K = QQ->second.begin(),
1580 KE = QQ->second.end(); K != KE; ++K) {
1581 DenseMap<ValuePair, size_t>::iterator C = DAG.find(*K);
1582 if (C == DAG.end()) continue;
1584 // This child is in the DAG, now we need to make sure it is the
1585 // best of any conflicting children. There could be multiple
1586 // conflicting children, so first, determine if we're keeping
1587 // this child, then delete conflicting children as necessary.
1589 // It is also necessary to guard against pairing-induced
1590 // dependencies. Consider instructions a .. x .. y .. b
1591 // such that (a,b) are to be fused and (x,y) are to be fused
1592 // but a is an input to x and b is an output from y. This
1593 // means that y cannot be moved after b but x must be moved
1594 // after b for (a,b) to be fused. In other words, after
1595 // fusing (a,b) we have y .. a/b .. x where y is an input
1596 // to a/b and x is an output to a/b: x and y can no longer
1597 // be legally fused. To prevent this condition, we must
1598 // make sure that a child pair added to the DAG is not
1599 // both an input and output of an already-selected pair.
1601 // Pairing-induced dependencies can also form from more complicated
1602 // cycles. The pair vs. pair conflicts are easy to check, and so
1603 // that is done explicitly for "fast rejection", and because for
1604 // child vs. child conflicts, we may prefer to keep the current
1605 // pair in preference to the already-selected child.
1606 DenseSet<ValuePair> CurrentPairs;
1609 for (SmallVectorImpl<ValuePairWithDepth>::iterator C2
1610 = BestChildren.begin(), E2 = BestChildren.end();
1612 if (C2->first.first == C->first.first ||
1613 C2->first.first == C->first.second ||
1614 C2->first.second == C->first.first ||
1615 C2->first.second == C->first.second ||
1616 pairsConflict(C2->first, C->first, PairableInstUsers,
1617 UseCycleCheck ? &PairableInstUserMap : 0,
1618 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1619 if (C2->second >= C->second) {
1624 CurrentPairs.insert(C2->first);
1627 if (!CanAdd) continue;
1629 // Even worse, this child could conflict with another node already
1630 // selected for the DAG. If that is the case, ignore this child.
1631 for (DenseSet<ValuePair>::iterator T = PrunedDAG.begin(),
1632 E2 = PrunedDAG.end(); T != E2; ++T) {
1633 if (T->first == C->first.first ||
1634 T->first == C->first.second ||
1635 T->second == C->first.first ||
1636 T->second == C->first.second ||
1637 pairsConflict(*T, C->first, PairableInstUsers,
1638 UseCycleCheck ? &PairableInstUserMap : 0,
1639 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1644 CurrentPairs.insert(*T);
1646 if (!CanAdd) continue;
1648 // And check the queue too...
1649 for (SmallVectorImpl<ValuePairWithDepth>::iterator C2 = Q.begin(),
1650 E2 = Q.end(); C2 != E2; ++C2) {
1651 if (C2->first.first == C->first.first ||
1652 C2->first.first == C->first.second ||
1653 C2->first.second == C->first.first ||
1654 C2->first.second == C->first.second ||
1655 pairsConflict(C2->first, C->first, PairableInstUsers,
1656 UseCycleCheck ? &PairableInstUserMap : 0,
1657 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1662 CurrentPairs.insert(C2->first);
1664 if (!CanAdd) continue;
1666 // Last but not least, check for a conflict with any of the
1667 // already-chosen pairs.
1668 for (DenseMap<Value *, Value *>::iterator C2 =
1669 ChosenPairs.begin(), E2 = ChosenPairs.end();
1671 if (pairsConflict(*C2, C->first, PairableInstUsers,
1672 UseCycleCheck ? &PairableInstUserMap : 0,
1673 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1678 CurrentPairs.insert(*C2);
1680 if (!CanAdd) continue;
1682 // To check for non-trivial cycles formed by the addition of the
1683 // current pair we've formed a list of all relevant pairs, now use a
1684 // graph walk to check for a cycle. We start from the current pair and
1685 // walk the use dag to see if we again reach the current pair. If we
1686 // do, then the current pair is rejected.
1688 // FIXME: It may be more efficient to use a topological-ordering
1689 // algorithm to improve the cycle check. This should be investigated.
1690 if (UseCycleCheck &&
1691 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs))
1694 // This child can be added, but we may have chosen it in preference
1695 // to an already-selected child. Check for this here, and if a
1696 // conflict is found, then remove the previously-selected child
1697 // before adding this one in its place.
1698 for (SmallVectorImpl<ValuePairWithDepth>::iterator C2
1699 = BestChildren.begin(); C2 != BestChildren.end();) {
1700 if (C2->first.first == C->first.first ||
1701 C2->first.first == C->first.second ||
1702 C2->first.second == C->first.first ||
1703 C2->first.second == C->first.second ||
1704 pairsConflict(C2->first, C->first, PairableInstUsers))
1705 C2 = BestChildren.erase(C2);
1710 BestChildren.push_back(ValuePairWithDepth(C->first, C->second));
1713 for (SmallVectorImpl<ValuePairWithDepth>::iterator C
1714 = BestChildren.begin(), E2 = BestChildren.end();
1716 size_t DepthF = getDepthFactor(C->first.first);
1717 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF));
1719 } while (!Q.empty());
1722 // This function finds the best dag of mututally-compatible connected
1723 // pairs, given the choice of root pairs as an iterator range.
1724 void BBVectorize::findBestDAGFor(
1725 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1726 DenseSet<ValuePair> &CandidatePairsSet,
1727 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1728 std::vector<Value *> &PairableInsts,
1729 DenseSet<ValuePair> &FixedOrderPairs,
1730 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1731 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1732 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
1733 DenseSet<ValuePair> &PairableInstUsers,
1734 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
1735 DenseSet<VPPair> &PairableInstUserPairSet,
1736 DenseMap<Value *, Value *> &ChosenPairs,
1737 DenseSet<ValuePair> &BestDAG, size_t &BestMaxDepth,
1738 int &BestEffSize, Value *II, std::vector<Value *>&JJ,
1739 bool UseCycleCheck) {
1740 for (std::vector<Value *>::iterator J = JJ.begin(), JE = JJ.end();
1742 ValuePair IJ(II, *J);
1743 if (!CandidatePairsSet.count(IJ))
1746 // Before going any further, make sure that this pair does not
1747 // conflict with any already-selected pairs (see comment below
1748 // near the DAG pruning for more details).
1749 DenseSet<ValuePair> ChosenPairSet;
1750 bool DoesConflict = false;
1751 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(),
1752 E = ChosenPairs.end(); C != E; ++C) {
1753 if (pairsConflict(*C, IJ, PairableInstUsers,
1754 UseCycleCheck ? &PairableInstUserMap : 0,
1755 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1756 DoesConflict = true;
1760 ChosenPairSet.insert(*C);
1762 if (DoesConflict) continue;
1764 if (UseCycleCheck &&
1765 pairWillFormCycle(IJ, PairableInstUserMap, ChosenPairSet))
1768 DenseMap<ValuePair, size_t> DAG;
1769 buildInitialDAGFor(CandidatePairs, CandidatePairsSet,
1770 PairableInsts, ConnectedPairs,
1771 PairableInstUsers, ChosenPairs, DAG, IJ);
1773 // Because we'll keep the child with the largest depth, the largest
1774 // depth is still the same in the unpruned DAG.
1775 size_t MaxDepth = DAG.lookup(IJ);
1777 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found DAG for pair {"
1778 << *IJ.first << " <-> " << *IJ.second << "} of depth " <<
1779 MaxDepth << " and size " << DAG.size() << "\n");
1781 // At this point the DAG has been constructed, but, may contain
1782 // contradictory children (meaning that different children of
1783 // some dag node may be attempting to fuse the same instruction).
1784 // So now we walk the dag again, in the case of a conflict,
1785 // keep only the child with the largest depth. To break a tie,
1786 // favor the first child.
1788 DenseSet<ValuePair> PrunedDAG;
1789 pruneDAGFor(CandidatePairs, PairableInsts, ConnectedPairs,
1790 PairableInstUsers, PairableInstUserMap,
1791 PairableInstUserPairSet,
1792 ChosenPairs, DAG, PrunedDAG, IJ, UseCycleCheck);
1796 DenseSet<Value *> PrunedDAGInstrs;
1797 for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(),
1798 E = PrunedDAG.end(); S != E; ++S) {
1799 PrunedDAGInstrs.insert(S->first);
1800 PrunedDAGInstrs.insert(S->second);
1803 // The set of pairs that have already contributed to the total cost.
1804 DenseSet<ValuePair> IncomingPairs;
1806 // If the cost model were perfect, this might not be necessary; but we
1807 // need to make sure that we don't get stuck vectorizing our own
1809 bool HasNontrivialInsts = false;
1811 // The node weights represent the cost savings associated with
1812 // fusing the pair of instructions.
1813 for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(),
1814 E = PrunedDAG.end(); S != E; ++S) {
1815 if (!isa<ShuffleVectorInst>(S->first) &&
1816 !isa<InsertElementInst>(S->first) &&
1817 !isa<ExtractElementInst>(S->first))
1818 HasNontrivialInsts = true;
1820 bool FlipOrder = false;
1822 if (getDepthFactor(S->first)) {
1823 int ESContrib = CandidatePairCostSavings.find(*S)->second;
1824 DEBUG(if (DebugPairSelection) dbgs() << "\tweight {"
1825 << *S->first << " <-> " << *S->second << "} = " <<
1827 EffSize += ESContrib;
1830 // The edge weights contribute in a negative sense: they represent
1831 // the cost of shuffles.
1832 DenseMap<ValuePair, std::vector<ValuePair> >::iterator SS =
1833 ConnectedPairDeps.find(*S);
1834 if (SS != ConnectedPairDeps.end()) {
1835 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
1836 for (std::vector<ValuePair>::iterator T = SS->second.begin(),
1837 TE = SS->second.end(); T != TE; ++T) {
1839 if (!PrunedDAG.count(Q.second))
1841 DenseMap<VPPair, unsigned>::iterator R =
1842 PairConnectionTypes.find(VPPair(Q.second, Q.first));
1843 assert(R != PairConnectionTypes.end() &&
1844 "Cannot find pair connection type");
1845 if (R->second == PairConnectionDirect)
1847 else if (R->second == PairConnectionSwap)
1851 // If there are more swaps than direct connections, then
1852 // the pair order will be flipped during fusion. So the real
1853 // number of swaps is the minimum number.
1854 FlipOrder = !FixedOrderPairs.count(*S) &&
1855 ((NumDepsSwap > NumDepsDirect) ||
1856 FixedOrderPairs.count(ValuePair(S->second, S->first)));
1858 for (std::vector<ValuePair>::iterator T = SS->second.begin(),
1859 TE = SS->second.end(); T != TE; ++T) {
1861 if (!PrunedDAG.count(Q.second))
1863 DenseMap<VPPair, unsigned>::iterator R =
1864 PairConnectionTypes.find(VPPair(Q.second, Q.first));
1865 assert(R != PairConnectionTypes.end() &&
1866 "Cannot find pair connection type");
1867 Type *Ty1 = Q.second.first->getType(),
1868 *Ty2 = Q.second.second->getType();
1869 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1870 if ((R->second == PairConnectionDirect && FlipOrder) ||
1871 (R->second == PairConnectionSwap && !FlipOrder) ||
1872 R->second == PairConnectionSplat) {
1873 int ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1876 if (VTy->getVectorNumElements() == 2) {
1877 if (R->second == PairConnectionSplat)
1878 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1879 TargetTransformInfo::SK_Broadcast, VTy));
1881 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1882 TargetTransformInfo::SK_Reverse, VTy));
1885 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1886 *Q.second.first << " <-> " << *Q.second.second <<
1888 *S->first << " <-> " << *S->second << "} = " <<
1890 EffSize -= ESContrib;
1895 // Compute the cost of outgoing edges. We assume that edges outgoing
1896 // to shuffles, inserts or extracts can be merged, and so contribute
1897 // no additional cost.
1898 if (!S->first->getType()->isVoidTy()) {
1899 Type *Ty1 = S->first->getType(),
1900 *Ty2 = S->second->getType();
1901 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1903 bool NeedsExtraction = false;
1904 for (Value::use_iterator I = S->first->use_begin(),
1905 IE = S->first->use_end(); I != IE; ++I) {
1906 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) {
1907 // Shuffle can be folded if it has no other input
1908 if (isa<UndefValue>(SI->getOperand(1)))
1911 if (isa<ExtractElementInst>(*I))
1913 if (PrunedDAGInstrs.count(*I))
1915 NeedsExtraction = true;
1919 if (NeedsExtraction) {
1921 if (Ty1->isVectorTy()) {
1922 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1924 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1925 TargetTransformInfo::SK_ExtractSubvector, VTy, 0, Ty1));
1927 ESContrib = (int) TTI->getVectorInstrCost(
1928 Instruction::ExtractElement, VTy, 0);
1930 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1931 *S->first << "} = " << ESContrib << "\n");
1932 EffSize -= ESContrib;
1935 NeedsExtraction = false;
1936 for (Value::use_iterator I = S->second->use_begin(),
1937 IE = S->second->use_end(); I != IE; ++I) {
1938 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) {
1939 // Shuffle can be folded if it has no other input
1940 if (isa<UndefValue>(SI->getOperand(1)))
1943 if (isa<ExtractElementInst>(*I))
1945 if (PrunedDAGInstrs.count(*I))
1947 NeedsExtraction = true;
1951 if (NeedsExtraction) {
1953 if (Ty2->isVectorTy()) {
1954 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1956 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1957 TargetTransformInfo::SK_ExtractSubvector, VTy,
1958 Ty1->isVectorTy() ? Ty1->getVectorNumElements() : 1, Ty2));
1960 ESContrib = (int) TTI->getVectorInstrCost(
1961 Instruction::ExtractElement, VTy, 1);
1962 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1963 *S->second << "} = " << ESContrib << "\n");
1964 EffSize -= ESContrib;
1968 // Compute the cost of incoming edges.
1969 if (!isa<LoadInst>(S->first) && !isa<StoreInst>(S->first)) {
1970 Instruction *S1 = cast<Instruction>(S->first),
1971 *S2 = cast<Instruction>(S->second);
1972 for (unsigned o = 0; o < S1->getNumOperands(); ++o) {
1973 Value *O1 = S1->getOperand(o), *O2 = S2->getOperand(o);
1975 // Combining constants into vector constants (or small vector
1976 // constants into larger ones are assumed free).
1977 if (isa<Constant>(O1) && isa<Constant>(O2))
1983 ValuePair VP = ValuePair(O1, O2);
1984 ValuePair VPR = ValuePair(O2, O1);
1986 // Internal edges are not handled here.
1987 if (PrunedDAG.count(VP) || PrunedDAG.count(VPR))
1990 Type *Ty1 = O1->getType(),
1991 *Ty2 = O2->getType();
1992 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1994 // Combining vector operations of the same type is also assumed
1995 // folded with other operations.
1997 // If both are insert elements, then both can be widened.
1998 InsertElementInst *IEO1 = dyn_cast<InsertElementInst>(O1),
1999 *IEO2 = dyn_cast<InsertElementInst>(O2);
2000 if (IEO1 && IEO2 && isPureIEChain(IEO1) && isPureIEChain(IEO2))
2002 // If both are extract elements, and both have the same input
2003 // type, then they can be replaced with a shuffle
2004 ExtractElementInst *EIO1 = dyn_cast<ExtractElementInst>(O1),
2005 *EIO2 = dyn_cast<ExtractElementInst>(O2);
2007 EIO1->getOperand(0)->getType() ==
2008 EIO2->getOperand(0)->getType())
2010 // If both are a shuffle with equal operand types and only two
2011 // unqiue operands, then they can be replaced with a single
2013 ShuffleVectorInst *SIO1 = dyn_cast<ShuffleVectorInst>(O1),
2014 *SIO2 = dyn_cast<ShuffleVectorInst>(O2);
2016 SIO1->getOperand(0)->getType() ==
2017 SIO2->getOperand(0)->getType()) {
2018 SmallSet<Value *, 4> SIOps;
2019 SIOps.insert(SIO1->getOperand(0));
2020 SIOps.insert(SIO1->getOperand(1));
2021 SIOps.insert(SIO2->getOperand(0));
2022 SIOps.insert(SIO2->getOperand(1));
2023 if (SIOps.size() <= 2)
2029 // This pair has already been formed.
2030 if (IncomingPairs.count(VP)) {
2032 } else if (IncomingPairs.count(VPR)) {
2033 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2036 if (VTy->getVectorNumElements() == 2)
2037 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
2038 TargetTransformInfo::SK_Reverse, VTy));
2039 } else if (!Ty1->isVectorTy() && !Ty2->isVectorTy()) {
2040 ESContrib = (int) TTI->getVectorInstrCost(
2041 Instruction::InsertElement, VTy, 0);
2042 ESContrib += (int) TTI->getVectorInstrCost(
2043 Instruction::InsertElement, VTy, 1);
2044 } else if (!Ty1->isVectorTy()) {
2045 // O1 needs to be inserted into a vector of size O2, and then
2046 // both need to be shuffled together.
2047 ESContrib = (int) TTI->getVectorInstrCost(
2048 Instruction::InsertElement, Ty2, 0);
2049 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2051 } else if (!Ty2->isVectorTy()) {
2052 // O2 needs to be inserted into a vector of size O1, and then
2053 // both need to be shuffled together.
2054 ESContrib = (int) TTI->getVectorInstrCost(
2055 Instruction::InsertElement, Ty1, 0);
2056 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2059 Type *TyBig = Ty1, *TySmall = Ty2;
2060 if (Ty2->getVectorNumElements() > Ty1->getVectorNumElements())
2061 std::swap(TyBig, TySmall);
2063 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2065 if (TyBig != TySmall)
2066 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2070 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {"
2071 << *O1 << " <-> " << *O2 << "} = " <<
2073 EffSize -= ESContrib;
2074 IncomingPairs.insert(VP);
2079 if (!HasNontrivialInsts) {
2080 DEBUG(if (DebugPairSelection) dbgs() <<
2081 "\tNo non-trivial instructions in DAG;"
2082 " override to zero effective size\n");
2086 for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(),
2087 E = PrunedDAG.end(); S != E; ++S)
2088 EffSize += (int) getDepthFactor(S->first);
2091 DEBUG(if (DebugPairSelection)
2092 dbgs() << "BBV: found pruned DAG for pair {"
2093 << *IJ.first << " <-> " << *IJ.second << "} of depth " <<
2094 MaxDepth << " and size " << PrunedDAG.size() <<
2095 " (effective size: " << EffSize << ")\n");
2096 if (((TTI && !UseChainDepthWithTI) ||
2097 MaxDepth >= Config.ReqChainDepth) &&
2098 EffSize > 0 && EffSize > BestEffSize) {
2099 BestMaxDepth = MaxDepth;
2100 BestEffSize = EffSize;
2101 BestDAG = PrunedDAG;
2106 // Given the list of candidate pairs, this function selects those
2107 // that will be fused into vector instructions.
2108 void BBVectorize::choosePairs(
2109 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
2110 DenseSet<ValuePair> &CandidatePairsSet,
2111 DenseMap<ValuePair, int> &CandidatePairCostSavings,
2112 std::vector<Value *> &PairableInsts,
2113 DenseSet<ValuePair> &FixedOrderPairs,
2114 DenseMap<VPPair, unsigned> &PairConnectionTypes,
2115 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
2116 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
2117 DenseSet<ValuePair> &PairableInstUsers,
2118 DenseMap<Value *, Value *>& ChosenPairs) {
2119 bool UseCycleCheck =
2120 CandidatePairsSet.size() <= Config.MaxCandPairsForCycleCheck;
2122 DenseMap<Value *, std::vector<Value *> > CandidatePairs2;
2123 for (DenseSet<ValuePair>::iterator I = CandidatePairsSet.begin(),
2124 E = CandidatePairsSet.end(); I != E; ++I) {
2125 std::vector<Value *> &JJ = CandidatePairs2[I->second];
2126 if (JJ.empty()) JJ.reserve(32);
2127 JJ.push_back(I->first);
2130 DenseMap<ValuePair, std::vector<ValuePair> > PairableInstUserMap;
2131 DenseSet<VPPair> PairableInstUserPairSet;
2132 for (std::vector<Value *>::iterator I = PairableInsts.begin(),
2133 E = PairableInsts.end(); I != E; ++I) {
2134 // The number of possible pairings for this variable:
2135 size_t NumChoices = CandidatePairs.lookup(*I).size();
2136 if (!NumChoices) continue;
2138 std::vector<Value *> &JJ = CandidatePairs[*I];
2140 // The best pair to choose and its dag:
2141 size_t BestMaxDepth = 0;
2142 int BestEffSize = 0;
2143 DenseSet<ValuePair> BestDAG;
2144 findBestDAGFor(CandidatePairs, CandidatePairsSet,
2145 CandidatePairCostSavings,
2146 PairableInsts, FixedOrderPairs, PairConnectionTypes,
2147 ConnectedPairs, ConnectedPairDeps,
2148 PairableInstUsers, PairableInstUserMap,
2149 PairableInstUserPairSet, ChosenPairs,
2150 BestDAG, BestMaxDepth, BestEffSize, *I, JJ,
2153 if (BestDAG.empty())
2156 // A dag has been chosen (or not) at this point. If no dag was
2157 // chosen, then this instruction, I, cannot be paired (and is no longer
2160 DEBUG(dbgs() << "BBV: selected pairs in the best DAG for: "
2161 << *cast<Instruction>(*I) << "\n");
2163 for (DenseSet<ValuePair>::iterator S = BestDAG.begin(),
2164 SE2 = BestDAG.end(); S != SE2; ++S) {
2165 // Insert the members of this dag into the list of chosen pairs.
2166 ChosenPairs.insert(ValuePair(S->first, S->second));
2167 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " <<
2168 *S->second << "\n");
2170 // Remove all candidate pairs that have values in the chosen dag.
2171 std::vector<Value *> &KK = CandidatePairs[S->first];
2172 for (std::vector<Value *>::iterator K = KK.begin(), KE = KK.end();
2174 if (*K == S->second)
2177 CandidatePairsSet.erase(ValuePair(S->first, *K));
2180 std::vector<Value *> &LL = CandidatePairs2[S->second];
2181 for (std::vector<Value *>::iterator L = LL.begin(), LE = LL.end();
2186 CandidatePairsSet.erase(ValuePair(*L, S->second));
2189 std::vector<Value *> &MM = CandidatePairs[S->second];
2190 for (std::vector<Value *>::iterator M = MM.begin(), ME = MM.end();
2192 assert(*M != S->first && "Flipped pair in candidate list?");
2193 CandidatePairsSet.erase(ValuePair(S->second, *M));
2196 std::vector<Value *> &NN = CandidatePairs2[S->first];
2197 for (std::vector<Value *>::iterator N = NN.begin(), NE = NN.end();
2199 assert(*N != S->second && "Flipped pair in candidate list?");
2200 CandidatePairsSet.erase(ValuePair(*N, S->first));
2205 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n");
2208 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o,
2213 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) +
2214 (n > 0 ? "." + utostr(n) : "")).str();
2217 // Returns the value that is to be used as the pointer input to the vector
2218 // instruction that fuses I with J.
2219 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context,
2220 Instruction *I, Instruction *J, unsigned o) {
2222 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
2223 int64_t OffsetInElmts;
2225 // Note: the analysis might fail here, that is why the pair order has
2226 // been precomputed (OffsetInElmts must be unused here).
2227 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
2228 IAddressSpace, JAddressSpace,
2229 OffsetInElmts, false);
2231 // The pointer value is taken to be the one with the lowest offset.
2234 Type *ArgTypeI = cast<PointerType>(IPtr->getType())->getElementType();
2235 Type *ArgTypeJ = cast<PointerType>(JPtr->getType())->getElementType();
2236 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2237 Type *VArgPtrType = PointerType::get(VArgType,
2238 cast<PointerType>(IPtr->getType())->getAddressSpace());
2239 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o),
2240 /* insert before */ I);
2243 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J,
2244 unsigned MaskOffset, unsigned NumInElem,
2245 unsigned NumInElem1, unsigned IdxOffset,
2246 std::vector<Constant*> &Mask) {
2247 unsigned NumElem1 = cast<VectorType>(J->getType())->getNumElements();
2248 for (unsigned v = 0; v < NumElem1; ++v) {
2249 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v);
2251 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context));
2253 unsigned mm = m + (int) IdxOffset;
2254 if (m >= (int) NumInElem1)
2255 mm += (int) NumInElem;
2257 Mask[v+MaskOffset] =
2258 ConstantInt::get(Type::getInt32Ty(Context), mm);
2263 // Returns the value that is to be used as the vector-shuffle mask to the
2264 // vector instruction that fuses I with J.
2265 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context,
2266 Instruction *I, Instruction *J) {
2267 // This is the shuffle mask. We need to append the second
2268 // mask to the first, and the numbers need to be adjusted.
2270 Type *ArgTypeI = I->getType();
2271 Type *ArgTypeJ = J->getType();
2272 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2274 unsigned NumElemI = cast<VectorType>(ArgTypeI)->getNumElements();
2276 // Get the total number of elements in the fused vector type.
2277 // By definition, this must equal the number of elements in
2279 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements();
2280 std::vector<Constant*> Mask(NumElem);
2282 Type *OpTypeI = I->getOperand(0)->getType();
2283 unsigned NumInElemI = cast<VectorType>(OpTypeI)->getNumElements();
2284 Type *OpTypeJ = J->getOperand(0)->getType();
2285 unsigned NumInElemJ = cast<VectorType>(OpTypeJ)->getNumElements();
2287 // The fused vector will be:
2288 // -----------------------------------------------------
2289 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ |
2290 // -----------------------------------------------------
2291 // from which we'll extract NumElem total elements (where the first NumElemI
2292 // of them come from the mask in I and the remainder come from the mask
2295 // For the mask from the first pair...
2296 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI,
2299 // For the mask from the second pair...
2300 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ,
2303 return ConstantVector::get(Mask);
2306 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I,
2307 Instruction *J, unsigned o, Value *&LOp,
2309 Type *ArgTypeL, Type *ArgTypeH,
2310 bool IBeforeJ, unsigned IdxOff) {
2311 bool ExpandedIEChain = false;
2312 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) {
2313 // If we have a pure insertelement chain, then this can be rewritten
2314 // into a chain that directly builds the larger type.
2315 if (isPureIEChain(LIE)) {
2316 SmallVector<Value *, 8> VectElemts(numElemL,
2317 UndefValue::get(ArgTypeL->getScalarType()));
2318 InsertElementInst *LIENext = LIE;
2321 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue();
2322 VectElemts[Idx] = LIENext->getOperand(1);
2324 dyn_cast<InsertElementInst>(LIENext->getOperand(0))));
2327 Value *LIEPrev = UndefValue::get(ArgTypeH);
2328 for (unsigned i = 0; i < numElemL; ++i) {
2329 if (isa<UndefValue>(VectElemts[i])) continue;
2330 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i],
2331 ConstantInt::get(Type::getInt32Ty(Context),
2333 getReplacementName(IBeforeJ ? I : J,
2335 LIENext->insertBefore(IBeforeJ ? J : I);
2339 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH);
2340 ExpandedIEChain = true;
2344 return ExpandedIEChain;
2347 // Returns the value to be used as the specified operand of the vector
2348 // instruction that fuses I with J.
2349 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I,
2350 Instruction *J, unsigned o, bool IBeforeJ) {
2351 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2352 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1);
2354 // Compute the fused vector type for this operand
2355 Type *ArgTypeI = I->getOperand(o)->getType();
2356 Type *ArgTypeJ = J->getOperand(o)->getType();
2357 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2359 Instruction *L = I, *H = J;
2360 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ;
2363 if (ArgTypeL->isVectorTy())
2364 numElemL = cast<VectorType>(ArgTypeL)->getNumElements();
2369 if (ArgTypeH->isVectorTy())
2370 numElemH = cast<VectorType>(ArgTypeH)->getNumElements();
2374 Value *LOp = L->getOperand(o);
2375 Value *HOp = H->getOperand(o);
2376 unsigned numElem = VArgType->getNumElements();
2378 // First, we check if we can reuse the "original" vector outputs (if these
2379 // exist). We might need a shuffle.
2380 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp);
2381 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp);
2382 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp);
2383 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp);
2385 // FIXME: If we're fusing shuffle instructions, then we can't apply this
2386 // optimization. The input vectors to the shuffle might be a different
2387 // length from the shuffle outputs. Unfortunately, the replacement
2388 // shuffle mask has already been formed, and the mask entries are sensitive
2389 // to the sizes of the inputs.
2390 bool IsSizeChangeShuffle =
2391 isa<ShuffleVectorInst>(L) &&
2392 (LOp->getType() != L->getType() || HOp->getType() != H->getType());
2394 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) {
2395 // We can have at most two unique vector inputs.
2396 bool CanUseInputs = true;
2399 I1 = LEE->getOperand(0);
2401 I1 = LSV->getOperand(0);
2402 I2 = LSV->getOperand(1);
2403 if (I2 == I1 || isa<UndefValue>(I2))
2408 Value *I3 = HEE->getOperand(0);
2409 if (!I2 && I3 != I1)
2411 else if (I3 != I1 && I3 != I2)
2412 CanUseInputs = false;
2414 Value *I3 = HSV->getOperand(0);
2415 if (!I2 && I3 != I1)
2417 else if (I3 != I1 && I3 != I2)
2418 CanUseInputs = false;
2421 Value *I4 = HSV->getOperand(1);
2422 if (!isa<UndefValue>(I4)) {
2423 if (!I2 && I4 != I1)
2425 else if (I4 != I1 && I4 != I2)
2426 CanUseInputs = false;
2433 cast<VectorType>(cast<Instruction>(LOp)->getOperand(0)->getType())
2436 cast<VectorType>(cast<Instruction>(HOp)->getOperand(0)->getType())
2439 // We have one or two input vectors. We need to map each index of the
2440 // operands to the index of the original vector.
2441 SmallVector<std::pair<int, int>, 8> II(numElem);
2442 for (unsigned i = 0; i < numElemL; ++i) {
2446 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue();
2447 INum = LEE->getOperand(0) == I1 ? 0 : 1;
2449 Idx = LSV->getMaskValue(i);
2450 if (Idx < (int) LOpElem) {
2451 INum = LSV->getOperand(0) == I1 ? 0 : 1;
2454 INum = LSV->getOperand(1) == I1 ? 0 : 1;
2458 II[i] = std::pair<int, int>(Idx, INum);
2460 for (unsigned i = 0; i < numElemH; ++i) {
2464 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue();
2465 INum = HEE->getOperand(0) == I1 ? 0 : 1;
2467 Idx = HSV->getMaskValue(i);
2468 if (Idx < (int) HOpElem) {
2469 INum = HSV->getOperand(0) == I1 ? 0 : 1;
2472 INum = HSV->getOperand(1) == I1 ? 0 : 1;
2476 II[i + numElemL] = std::pair<int, int>(Idx, INum);
2479 // We now have an array which tells us from which index of which
2480 // input vector each element of the operand comes.
2481 VectorType *I1T = cast<VectorType>(I1->getType());
2482 unsigned I1Elem = I1T->getNumElements();
2485 // In this case there is only one underlying vector input. Check for
2486 // the trivial case where we can use the input directly.
2487 if (I1Elem == numElem) {
2488 bool ElemInOrder = true;
2489 for (unsigned i = 0; i < numElem; ++i) {
2490 if (II[i].first != (int) i && II[i].first != -1) {
2491 ElemInOrder = false;
2500 // A shuffle is needed.
2501 std::vector<Constant *> Mask(numElem);
2502 for (unsigned i = 0; i < numElem; ++i) {
2503 int Idx = II[i].first;
2505 Mask[i] = UndefValue::get(Type::getInt32Ty(Context));
2507 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2511 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2512 ConstantVector::get(Mask),
2513 getReplacementName(IBeforeJ ? I : J,
2515 S->insertBefore(IBeforeJ ? J : I);
2519 VectorType *I2T = cast<VectorType>(I2->getType());
2520 unsigned I2Elem = I2T->getNumElements();
2522 // This input comes from two distinct vectors. The first step is to
2523 // make sure that both vectors are the same length. If not, the
2524 // smaller one will need to grow before they can be shuffled together.
2525 if (I1Elem < I2Elem) {
2526 std::vector<Constant *> Mask(I2Elem);
2528 for (; v < I1Elem; ++v)
2529 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2530 for (; v < I2Elem; ++v)
2531 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2533 Instruction *NewI1 =
2534 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2535 ConstantVector::get(Mask),
2536 getReplacementName(IBeforeJ ? I : J,
2538 NewI1->insertBefore(IBeforeJ ? J : I);
2542 } else if (I1Elem > I2Elem) {
2543 std::vector<Constant *> Mask(I1Elem);
2545 for (; v < I2Elem; ++v)
2546 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2547 for (; v < I1Elem; ++v)
2548 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2550 Instruction *NewI2 =
2551 new ShuffleVectorInst(I2, UndefValue::get(I2T),
2552 ConstantVector::get(Mask),
2553 getReplacementName(IBeforeJ ? I : J,
2555 NewI2->insertBefore(IBeforeJ ? J : I);
2561 // Now that both I1 and I2 are the same length we can shuffle them
2562 // together (and use the result).
2563 std::vector<Constant *> Mask(numElem);
2564 for (unsigned v = 0; v < numElem; ++v) {
2565 if (II[v].first == -1) {
2566 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2568 int Idx = II[v].first + II[v].second * I1Elem;
2569 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2573 Instruction *NewOp =
2574 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask),
2575 getReplacementName(IBeforeJ ? I : J, true, o));
2576 NewOp->insertBefore(IBeforeJ ? J : I);
2581 Type *ArgType = ArgTypeL;
2582 if (numElemL < numElemH) {
2583 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH,
2584 ArgTypeL, VArgType, IBeforeJ, 1)) {
2585 // This is another short-circuit case: we're combining a scalar into
2586 // a vector that is formed by an IE chain. We've just expanded the IE
2587 // chain, now insert the scalar and we're done.
2589 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0,
2590 getReplacementName(IBeforeJ ? I : J, true, o));
2591 S->insertBefore(IBeforeJ ? J : I);
2593 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL,
2594 ArgTypeH, IBeforeJ)) {
2595 // The two vector inputs to the shuffle must be the same length,
2596 // so extend the smaller vector to be the same length as the larger one.
2600 std::vector<Constant *> Mask(numElemH);
2602 for (; v < numElemL; ++v)
2603 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2604 for (; v < numElemH; ++v)
2605 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2607 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL),
2608 ConstantVector::get(Mask),
2609 getReplacementName(IBeforeJ ? I : J,
2612 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0,
2613 getReplacementName(IBeforeJ ? I : J,
2617 NLOp->insertBefore(IBeforeJ ? J : I);
2622 } else if (numElemL > numElemH) {
2623 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL,
2624 ArgTypeH, VArgType, IBeforeJ)) {
2626 InsertElementInst::Create(LOp, HOp,
2627 ConstantInt::get(Type::getInt32Ty(Context),
2629 getReplacementName(IBeforeJ ? I : J,
2631 S->insertBefore(IBeforeJ ? J : I);
2633 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH,
2634 ArgTypeL, IBeforeJ)) {
2637 std::vector<Constant *> Mask(numElemL);
2639 for (; v < numElemH; ++v)
2640 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2641 for (; v < numElemL; ++v)
2642 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2644 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH),
2645 ConstantVector::get(Mask),
2646 getReplacementName(IBeforeJ ? I : J,
2649 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0,
2650 getReplacementName(IBeforeJ ? I : J,
2654 NHOp->insertBefore(IBeforeJ ? J : I);
2659 if (ArgType->isVectorTy()) {
2660 unsigned numElem = cast<VectorType>(VArgType)->getNumElements();
2661 std::vector<Constant*> Mask(numElem);
2662 for (unsigned v = 0; v < numElem; ++v) {
2664 // If the low vector was expanded, we need to skip the extra
2665 // undefined entries.
2666 if (v >= numElemL && numElemH > numElemL)
2667 Idx += (numElemH - numElemL);
2668 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2671 Instruction *BV = new ShuffleVectorInst(LOp, HOp,
2672 ConstantVector::get(Mask),
2673 getReplacementName(IBeforeJ ? I : J, true, o));
2674 BV->insertBefore(IBeforeJ ? J : I);
2678 Instruction *BV1 = InsertElementInst::Create(
2679 UndefValue::get(VArgType), LOp, CV0,
2680 getReplacementName(IBeforeJ ? I : J,
2682 BV1->insertBefore(IBeforeJ ? J : I);
2683 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1,
2684 getReplacementName(IBeforeJ ? I : J,
2686 BV2->insertBefore(IBeforeJ ? J : I);
2690 // This function creates an array of values that will be used as the inputs
2691 // to the vector instruction that fuses I with J.
2692 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context,
2693 Instruction *I, Instruction *J,
2694 SmallVectorImpl<Value *> &ReplacedOperands,
2696 unsigned NumOperands = I->getNumOperands();
2698 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) {
2699 // Iterate backward so that we look at the store pointer
2700 // first and know whether or not we need to flip the inputs.
2702 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) {
2703 // This is the pointer for a load/store instruction.
2704 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o);
2706 } else if (isa<CallInst>(I)) {
2707 Function *F = cast<CallInst>(I)->getCalledFunction();
2708 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID();
2709 if (o == NumOperands-1) {
2710 BasicBlock &BB = *I->getParent();
2712 Module *M = BB.getParent()->getParent();
2713 Type *ArgTypeI = I->getType();
2714 Type *ArgTypeJ = J->getType();
2715 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2717 ReplacedOperands[o] = Intrinsic::getDeclaration(M, IID, VArgType);
2719 } else if (IID == Intrinsic::powi && o == 1) {
2720 // The second argument of powi is a single integer and we've already
2721 // checked that both arguments are equal. As a result, we just keep
2722 // I's second argument.
2723 ReplacedOperands[o] = I->getOperand(o);
2726 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) {
2727 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J);
2731 ReplacedOperands[o] = getReplacementInput(Context, I, J, o, IBeforeJ);
2735 // This function creates two values that represent the outputs of the
2736 // original I and J instructions. These are generally vector shuffles
2737 // or extracts. In many cases, these will end up being unused and, thus,
2738 // eliminated by later passes.
2739 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
2740 Instruction *J, Instruction *K,
2741 Instruction *&InsertionPt,
2742 Instruction *&K1, Instruction *&K2) {
2743 if (isa<StoreInst>(I)) {
2744 AA->replaceWithNewValue(I, K);
2745 AA->replaceWithNewValue(J, K);
2747 Type *IType = I->getType();
2748 Type *JType = J->getType();
2750 VectorType *VType = getVecTypeForPair(IType, JType);
2751 unsigned numElem = VType->getNumElements();
2753 unsigned numElemI, numElemJ;
2754 if (IType->isVectorTy())
2755 numElemI = cast<VectorType>(IType)->getNumElements();
2759 if (JType->isVectorTy())
2760 numElemJ = cast<VectorType>(JType)->getNumElements();
2764 if (IType->isVectorTy()) {
2765 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI);
2766 for (unsigned v = 0; v < numElemI; ++v) {
2767 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2768 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v);
2771 K1 = new ShuffleVectorInst(K, UndefValue::get(VType),
2772 ConstantVector::get( Mask1),
2773 getReplacementName(K, false, 1));
2775 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2776 K1 = ExtractElementInst::Create(K, CV0,
2777 getReplacementName(K, false, 1));
2780 if (JType->isVectorTy()) {
2781 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ);
2782 for (unsigned v = 0; v < numElemJ; ++v) {
2783 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2784 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v);
2787 K2 = new ShuffleVectorInst(K, UndefValue::get(VType),
2788 ConstantVector::get( Mask2),
2789 getReplacementName(K, false, 2));
2791 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1);
2792 K2 = ExtractElementInst::Create(K, CV1,
2793 getReplacementName(K, false, 2));
2797 K2->insertAfter(K1);
2802 // Move all uses of the function I (including pairing-induced uses) after J.
2803 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB,
2804 DenseSet<ValuePair> &LoadMoveSetPairs,
2805 Instruction *I, Instruction *J) {
2806 // Skip to the first instruction past I.
2807 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2809 DenseSet<Value *> Users;
2810 AliasSetTracker WriteSet(*AA);
2811 if (I->mayWriteToMemory()) WriteSet.add(I);
2813 for (; cast<Instruction>(L) != J; ++L)
2814 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs);
2816 assert(cast<Instruction>(L) == J &&
2817 "Tracking has not proceeded far enough to check for dependencies");
2818 // If J is now in the use set of I, then trackUsesOfI will return true
2819 // and we have a dependency cycle (and the fusing operation must abort).
2820 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSetPairs);
2823 // Move all uses of the function I (including pairing-induced uses) after J.
2824 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB,
2825 DenseSet<ValuePair> &LoadMoveSetPairs,
2826 Instruction *&InsertionPt,
2827 Instruction *I, Instruction *J) {
2828 // Skip to the first instruction past I.
2829 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2831 DenseSet<Value *> Users;
2832 AliasSetTracker WriteSet(*AA);
2833 if (I->mayWriteToMemory()) WriteSet.add(I);
2835 for (; cast<Instruction>(L) != J;) {
2836 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs)) {
2837 // Move this instruction
2838 Instruction *InstToMove = L; ++L;
2840 DEBUG(dbgs() << "BBV: moving: " << *InstToMove <<
2841 " to after " << *InsertionPt << "\n");
2842 InstToMove->removeFromParent();
2843 InstToMove->insertAfter(InsertionPt);
2844 InsertionPt = InstToMove;
2851 // Collect all load instruction that are in the move set of a given first
2852 // pair member. These loads depend on the first instruction, I, and so need
2853 // to be moved after J (the second instruction) when the pair is fused.
2854 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB,
2855 DenseMap<Value *, Value *> &ChosenPairs,
2856 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
2857 DenseSet<ValuePair> &LoadMoveSetPairs,
2859 // Skip to the first instruction past I.
2860 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2862 DenseSet<Value *> Users;
2863 AliasSetTracker WriteSet(*AA);
2864 if (I->mayWriteToMemory()) WriteSet.add(I);
2866 // Note: We cannot end the loop when we reach J because J could be moved
2867 // farther down the use chain by another instruction pairing. Also, J
2868 // could be before I if this is an inverted input.
2869 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) {
2870 if (trackUsesOfI(Users, WriteSet, I, L)) {
2871 if (L->mayReadFromMemory()) {
2872 LoadMoveSet[L].push_back(I);
2873 LoadMoveSetPairs.insert(ValuePair(L, I));
2879 // In cases where both load/stores and the computation of their pointers
2880 // are chosen for vectorization, we can end up in a situation where the
2881 // aliasing analysis starts returning different query results as the
2882 // process of fusing instruction pairs continues. Because the algorithm
2883 // relies on finding the same use dags here as were found earlier, we'll
2884 // need to precompute the necessary aliasing information here and then
2885 // manually update it during the fusion process.
2886 void BBVectorize::collectLoadMoveSet(BasicBlock &BB,
2887 std::vector<Value *> &PairableInsts,
2888 DenseMap<Value *, Value *> &ChosenPairs,
2889 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
2890 DenseSet<ValuePair> &LoadMoveSetPairs) {
2891 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
2892 PIE = PairableInsts.end(); PI != PIE; ++PI) {
2893 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI);
2894 if (P == ChosenPairs.end()) continue;
2896 Instruction *I = cast<Instruction>(P->first);
2897 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet,
2898 LoadMoveSetPairs, I);
2902 // When the first instruction in each pair is cloned, it will inherit its
2903 // parent's metadata. This metadata must be combined with that of the other
2904 // instruction in a safe way.
2905 void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) {
2906 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata;
2907 K->getAllMetadataOtherThanDebugLoc(Metadata);
2908 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) {
2909 unsigned Kind = Metadata[i].first;
2910 MDNode *JMD = J->getMetadata(Kind);
2911 MDNode *KMD = Metadata[i].second;
2915 K->setMetadata(Kind, 0); // Remove unknown metadata
2917 case LLVMContext::MD_tbaa:
2918 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD));
2920 case LLVMContext::MD_fpmath:
2921 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD));
2927 // This function fuses the chosen instruction pairs into vector instructions,
2928 // taking care preserve any needed scalar outputs and, then, it reorders the
2929 // remaining instructions as needed (users of the first member of the pair
2930 // need to be moved to after the location of the second member of the pair
2931 // because the vector instruction is inserted in the location of the pair's
2933 void BBVectorize::fuseChosenPairs(BasicBlock &BB,
2934 std::vector<Value *> &PairableInsts,
2935 DenseMap<Value *, Value *> &ChosenPairs,
2936 DenseSet<ValuePair> &FixedOrderPairs,
2937 DenseMap<VPPair, unsigned> &PairConnectionTypes,
2938 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
2939 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps) {
2940 LLVMContext& Context = BB.getContext();
2942 // During the vectorization process, the order of the pairs to be fused
2943 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs
2944 // list. After a pair is fused, the flipped pair is removed from the list.
2945 DenseSet<ValuePair> FlippedPairs;
2946 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(),
2947 E = ChosenPairs.end(); P != E; ++P)
2948 FlippedPairs.insert(ValuePair(P->second, P->first));
2949 for (DenseSet<ValuePair>::iterator P = FlippedPairs.begin(),
2950 E = FlippedPairs.end(); P != E; ++P)
2951 ChosenPairs.insert(*P);
2953 DenseMap<Value *, std::vector<Value *> > LoadMoveSet;
2954 DenseSet<ValuePair> LoadMoveSetPairs;
2955 collectLoadMoveSet(BB, PairableInsts, ChosenPairs,
2956 LoadMoveSet, LoadMoveSetPairs);
2958 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n");
2960 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) {
2961 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI);
2962 if (P == ChosenPairs.end()) {
2967 if (getDepthFactor(P->first) == 0) {
2968 // These instructions are not really fused, but are tracked as though
2969 // they are. Any case in which it would be interesting to fuse them
2970 // will be taken care of by InstCombine.
2976 Instruction *I = cast<Instruction>(P->first),
2977 *J = cast<Instruction>(P->second);
2979 DEBUG(dbgs() << "BBV: fusing: " << *I <<
2980 " <-> " << *J << "\n");
2982 // Remove the pair and flipped pair from the list.
2983 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second);
2984 assert(FP != ChosenPairs.end() && "Flipped pair not found in list");
2985 ChosenPairs.erase(FP);
2986 ChosenPairs.erase(P);
2988 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSetPairs, I, J)) {
2989 DEBUG(dbgs() << "BBV: fusion of: " << *I <<
2991 " aborted because of non-trivial dependency cycle\n");
2997 // If the pair must have the other order, then flip it.
2998 bool FlipPairOrder = FixedOrderPairs.count(ValuePair(J, I));
2999 if (!FlipPairOrder && !FixedOrderPairs.count(ValuePair(I, J))) {
3000 // This pair does not have a fixed order, and so we might want to
3001 // flip it if that will yield fewer shuffles. We count the number
3002 // of dependencies connected via swaps, and those directly connected,
3003 // and flip the order if the number of swaps is greater.
3004 bool OrigOrder = true;
3005 DenseMap<ValuePair, std::vector<ValuePair> >::iterator IJ =
3006 ConnectedPairDeps.find(ValuePair(I, J));
3007 if (IJ == ConnectedPairDeps.end()) {
3008 IJ = ConnectedPairDeps.find(ValuePair(J, I));
3012 if (IJ != ConnectedPairDeps.end()) {
3013 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
3014 for (std::vector<ValuePair>::iterator T = IJ->second.begin(),
3015 TE = IJ->second.end(); T != TE; ++T) {
3016 VPPair Q(IJ->first, *T);
3017 DenseMap<VPPair, unsigned>::iterator R =
3018 PairConnectionTypes.find(VPPair(Q.second, Q.first));
3019 assert(R != PairConnectionTypes.end() &&
3020 "Cannot find pair connection type");
3021 if (R->second == PairConnectionDirect)
3023 else if (R->second == PairConnectionSwap)
3028 std::swap(NumDepsDirect, NumDepsSwap);
3030 if (NumDepsSwap > NumDepsDirect) {
3031 FlipPairOrder = true;
3032 DEBUG(dbgs() << "BBV: reordering pair: " << *I <<
3033 " <-> " << *J << "\n");
3038 Instruction *L = I, *H = J;
3042 // If the pair being fused uses the opposite order from that in the pair
3043 // connection map, then we need to flip the types.
3044 DenseMap<ValuePair, std::vector<ValuePair> >::iterator HL =
3045 ConnectedPairs.find(ValuePair(H, L));
3046 if (HL != ConnectedPairs.end())
3047 for (std::vector<ValuePair>::iterator T = HL->second.begin(),
3048 TE = HL->second.end(); T != TE; ++T) {
3049 VPPair Q(HL->first, *T);
3050 DenseMap<VPPair, unsigned>::iterator R = PairConnectionTypes.find(Q);
3051 assert(R != PairConnectionTypes.end() &&
3052 "Cannot find pair connection type");
3053 if (R->second == PairConnectionDirect)
3054 R->second = PairConnectionSwap;
3055 else if (R->second == PairConnectionSwap)
3056 R->second = PairConnectionDirect;
3059 bool LBeforeH = !FlipPairOrder;
3060 unsigned NumOperands = I->getNumOperands();
3061 SmallVector<Value *, 3> ReplacedOperands(NumOperands);
3062 getReplacementInputsForPair(Context, L, H, ReplacedOperands,
3065 // Make a copy of the original operation, change its type to the vector
3066 // type and replace its operands with the vector operands.
3067 Instruction *K = L->clone();
3070 else if (H->hasName())
3073 if (!isa<StoreInst>(K))
3074 K->mutateType(getVecTypeForPair(L->getType(), H->getType()));
3076 combineMetadata(K, H);
3077 K->intersectOptionalDataWith(H);
3079 for (unsigned o = 0; o < NumOperands; ++o)
3080 K->setOperand(o, ReplacedOperands[o]);
3084 // Instruction insertion point:
3085 Instruction *InsertionPt = K;
3086 Instruction *K1 = 0, *K2 = 0;
3087 replaceOutputsOfPair(Context, L, H, K, InsertionPt, K1, K2);
3089 // The use dag of the first original instruction must be moved to after
3090 // the location of the second instruction. The entire use dag of the
3091 // first instruction is disjoint from the input dag of the second
3092 // (by definition), and so commutes with it.
3094 moveUsesOfIAfterJ(BB, LoadMoveSetPairs, InsertionPt, I, J);
3096 if (!isa<StoreInst>(I)) {
3097 L->replaceAllUsesWith(K1);
3098 H->replaceAllUsesWith(K2);
3099 AA->replaceWithNewValue(L, K1);
3100 AA->replaceWithNewValue(H, K2);
3103 // Instructions that may read from memory may be in the load move set.
3104 // Once an instruction is fused, we no longer need its move set, and so
3105 // the values of the map never need to be updated. However, when a load
3106 // is fused, we need to merge the entries from both instructions in the
3107 // pair in case those instructions were in the move set of some other
3108 // yet-to-be-fused pair. The loads in question are the keys of the map.
3109 if (I->mayReadFromMemory()) {
3110 std::vector<ValuePair> NewSetMembers;
3111 DenseMap<Value *, std::vector<Value *> >::iterator II =
3112 LoadMoveSet.find(I);
3113 if (II != LoadMoveSet.end())
3114 for (std::vector<Value *>::iterator N = II->second.begin(),
3115 NE = II->second.end(); N != NE; ++N)
3116 NewSetMembers.push_back(ValuePair(K, *N));
3117 DenseMap<Value *, std::vector<Value *> >::iterator JJ =
3118 LoadMoveSet.find(J);
3119 if (JJ != LoadMoveSet.end())
3120 for (std::vector<Value *>::iterator N = JJ->second.begin(),
3121 NE = JJ->second.end(); N != NE; ++N)
3122 NewSetMembers.push_back(ValuePair(K, *N));
3123 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(),
3124 AE = NewSetMembers.end(); A != AE; ++A) {
3125 LoadMoveSet[A->first].push_back(A->second);
3126 LoadMoveSetPairs.insert(*A);
3130 // Before removing I, set the iterator to the next instruction.
3131 PI = llvm::next(BasicBlock::iterator(I));
3132 if (cast<Instruction>(PI) == J)
3137 I->eraseFromParent();
3138 J->eraseFromParent();
3140 DEBUG(if (PrintAfterEveryPair) dbgs() << "BBV: block is now: \n" <<
3144 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n");
3148 char BBVectorize::ID = 0;
3149 static const char bb_vectorize_name[] = "Basic-Block Vectorization";
3150 INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
3151 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
3152 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
3153 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
3154 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3155 INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
3157 BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) {
3158 return new BBVectorize(C);
3162 llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) {
3163 BBVectorize BBVectorizer(P, C);
3164 return BBVectorizer.vectorizeBB(BB);
3167 //===----------------------------------------------------------------------===//
3168 VectorizeConfig::VectorizeConfig() {
3169 VectorBits = ::VectorBits;
3170 VectorizeBools = !::NoBools;
3171 VectorizeInts = !::NoInts;
3172 VectorizeFloats = !::NoFloats;
3173 VectorizePointers = !::NoPointers;
3174 VectorizeCasts = !::NoCasts;
3175 VectorizeMath = !::NoMath;
3176 VectorizeFMA = !::NoFMA;
3177 VectorizeSelect = !::NoSelect;
3178 VectorizeCmp = !::NoCmp;
3179 VectorizeGEP = !::NoGEP;
3180 VectorizeMemOps = !::NoMemOps;
3181 AlignedOnly = ::AlignedOnly;
3182 ReqChainDepth= ::ReqChainDepth;
3183 SearchLimit = ::SearchLimit;
3184 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck;
3185 SplatBreaksChain = ::SplatBreaksChain;
3186 MaxInsts = ::MaxInsts;
3187 MaxPairs = ::MaxPairs;
3188 MaxIter = ::MaxIter;
3189 Pow2LenOnly = ::Pow2LenOnly;
3190 NoMemOpBoost = ::NoMemOpBoost;
3191 FastDep = ::FastDep;