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/ScalarEvolution.h"
30 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
31 #include "llvm/Analysis/TargetTransformInfo.h"
32 #include "llvm/Analysis/ValueTracking.h"
33 #include "llvm/IR/Constants.h"
34 #include "llvm/IR/DataLayout.h"
35 #include "llvm/IR/DerivedTypes.h"
36 #include "llvm/IR/Dominators.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<DominatorTreeWrapperPass>().getDomTree();
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 (skipOptnoneFunction(BB))
393 if (!DT->isReachableFromEntry(&BB)) {
394 DEBUG(dbgs() << "BBV: skipping unreachable " << BB.getName() <<
395 " in " << BB.getParent()->getName() << "\n");
399 DEBUG(if (TTI) dbgs() << "BBV: using target information\n");
401 bool changed = false;
402 // Iterate a sufficient number of times to merge types of size 1 bit,
403 // then 2 bits, then 4, etc. up to half of the target vector width of the
404 // target vector register.
407 (TTI || v <= Config.VectorBits) &&
408 (!Config.MaxIter || n <= Config.MaxIter);
410 DEBUG(dbgs() << "BBV: fusing loop #" << n <<
411 " for " << BB.getName() << " in " <<
412 BB.getParent()->getName() << "...\n");
413 if (vectorizePairs(BB))
419 if (changed && !Pow2LenOnly) {
421 for (; !Config.MaxIter || n <= Config.MaxIter; ++n) {
422 DEBUG(dbgs() << "BBV: fusing for non-2^n-length vectors loop #: " <<
423 n << " for " << BB.getName() << " in " <<
424 BB.getParent()->getName() << "...\n");
425 if (!vectorizePairs(BB, true)) break;
429 DEBUG(dbgs() << "BBV: done!\n");
433 virtual bool runOnBasicBlock(BasicBlock &BB) {
434 // OptimizeNone check deferred to vectorizeBB().
436 AA = &getAnalysis<AliasAnalysis>();
437 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
438 SE = &getAnalysis<ScalarEvolution>();
439 TD = getAnalysisIfAvailable<DataLayout>();
440 TTI = IgnoreTargetInfo ? 0 : &getAnalysis<TargetTransformInfo>();
442 return vectorizeBB(BB);
445 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
446 BasicBlockPass::getAnalysisUsage(AU);
447 AU.addRequired<AliasAnalysis>();
448 AU.addRequired<DominatorTreeWrapperPass>();
449 AU.addRequired<ScalarEvolution>();
450 AU.addRequired<TargetTransformInfo>();
451 AU.addPreserved<AliasAnalysis>();
452 AU.addPreserved<DominatorTreeWrapperPass>();
453 AU.addPreserved<ScalarEvolution>();
454 AU.setPreservesCFG();
457 static inline VectorType *getVecTypeForPair(Type *ElemTy, Type *Elem2Ty) {
458 assert(ElemTy->getScalarType() == Elem2Ty->getScalarType() &&
459 "Cannot form vector from incompatible scalar types");
460 Type *STy = ElemTy->getScalarType();
463 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) {
464 numElem = VTy->getNumElements();
469 if (VectorType *VTy = dyn_cast<VectorType>(Elem2Ty)) {
470 numElem += VTy->getNumElements();
475 return VectorType::get(STy, numElem);
478 static inline void getInstructionTypes(Instruction *I,
479 Type *&T1, Type *&T2) {
480 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
481 // For stores, it is the value type, not the pointer type that matters
482 // because the value is what will come from a vector register.
484 Value *IVal = SI->getValueOperand();
485 T1 = IVal->getType();
490 if (CastInst *CI = dyn_cast<CastInst>(I))
495 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
496 T2 = SI->getCondition()->getType();
497 } else if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(I)) {
498 T2 = SI->getOperand(0)->getType();
499 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) {
500 T2 = CI->getOperand(0)->getType();
504 // Returns the weight associated with the provided value. A chain of
505 // candidate pairs has a length given by the sum of the weights of its
506 // members (one weight per pair; the weight of each member of the pair
507 // is assumed to be the same). This length is then compared to the
508 // chain-length threshold to determine if a given chain is significant
509 // enough to be vectorized. The length is also used in comparing
510 // candidate chains where longer chains are considered to be better.
511 // Note: when this function returns 0, the resulting instructions are
512 // not actually fused.
513 inline size_t getDepthFactor(Value *V) {
514 // InsertElement and ExtractElement have a depth factor of zero. This is
515 // for two reasons: First, they cannot be usefully fused. Second, because
516 // the pass generates a lot of these, they can confuse the simple metric
517 // used to compare the dags in the next iteration. Thus, giving them a
518 // weight of zero allows the pass to essentially ignore them in
519 // subsequent iterations when looking for vectorization opportunities
520 // while still tracking dependency chains that flow through those
522 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V))
525 // Give a load or store half of the required depth so that load/store
526 // pairs will vectorize.
527 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V)))
528 return Config.ReqChainDepth/2;
533 // Returns the cost of the provided instruction using TTI.
534 // This does not handle loads and stores.
535 unsigned getInstrCost(unsigned Opcode, Type *T1, Type *T2) {
538 case Instruction::GetElementPtr:
539 // We mark this instruction as zero-cost because scalar GEPs are usually
540 // lowered to the instruction addressing mode. At the moment we don't
541 // generate vector GEPs.
543 case Instruction::Br:
544 return TTI->getCFInstrCost(Opcode);
545 case Instruction::PHI:
547 case Instruction::Add:
548 case Instruction::FAdd:
549 case Instruction::Sub:
550 case Instruction::FSub:
551 case Instruction::Mul:
552 case Instruction::FMul:
553 case Instruction::UDiv:
554 case Instruction::SDiv:
555 case Instruction::FDiv:
556 case Instruction::URem:
557 case Instruction::SRem:
558 case Instruction::FRem:
559 case Instruction::Shl:
560 case Instruction::LShr:
561 case Instruction::AShr:
562 case Instruction::And:
563 case Instruction::Or:
564 case Instruction::Xor:
565 return TTI->getArithmeticInstrCost(Opcode, T1);
566 case Instruction::Select:
567 case Instruction::ICmp:
568 case Instruction::FCmp:
569 return TTI->getCmpSelInstrCost(Opcode, T1, T2);
570 case Instruction::ZExt:
571 case Instruction::SExt:
572 case Instruction::FPToUI:
573 case Instruction::FPToSI:
574 case Instruction::FPExt:
575 case Instruction::PtrToInt:
576 case Instruction::IntToPtr:
577 case Instruction::SIToFP:
578 case Instruction::UIToFP:
579 case Instruction::Trunc:
580 case Instruction::FPTrunc:
581 case Instruction::BitCast:
582 case Instruction::ShuffleVector:
583 return TTI->getCastInstrCost(Opcode, T1, T2);
589 // This determines the relative offset of two loads or stores, returning
590 // true if the offset could be determined to be some constant value.
591 // For example, if OffsetInElmts == 1, then J accesses the memory directly
592 // after I; if OffsetInElmts == -1 then I accesses the memory
594 bool getPairPtrInfo(Instruction *I, Instruction *J,
595 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment,
596 unsigned &IAddressSpace, unsigned &JAddressSpace,
597 int64_t &OffsetInElmts, bool ComputeOffset = true) {
599 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
600 LoadInst *LJ = cast<LoadInst>(J);
601 IPtr = LI->getPointerOperand();
602 JPtr = LJ->getPointerOperand();
603 IAlignment = LI->getAlignment();
604 JAlignment = LJ->getAlignment();
605 IAddressSpace = LI->getPointerAddressSpace();
606 JAddressSpace = LJ->getPointerAddressSpace();
608 StoreInst *SI = cast<StoreInst>(I), *SJ = cast<StoreInst>(J);
609 IPtr = SI->getPointerOperand();
610 JPtr = SJ->getPointerOperand();
611 IAlignment = SI->getAlignment();
612 JAlignment = SJ->getAlignment();
613 IAddressSpace = SI->getPointerAddressSpace();
614 JAddressSpace = SJ->getPointerAddressSpace();
620 const SCEV *IPtrSCEV = SE->getSCEV(IPtr);
621 const SCEV *JPtrSCEV = SE->getSCEV(JPtr);
623 // If this is a trivial offset, then we'll get something like
624 // 1*sizeof(type). With target data, which we need anyway, this will get
625 // constant folded into a number.
626 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV);
627 if (const SCEVConstant *ConstOffSCEV =
628 dyn_cast<SCEVConstant>(OffsetSCEV)) {
629 ConstantInt *IntOff = ConstOffSCEV->getValue();
630 int64_t Offset = IntOff->getSExtValue();
632 Type *VTy = IPtr->getType()->getPointerElementType();
633 int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy);
635 Type *VTy2 = JPtr->getType()->getPointerElementType();
636 if (VTy != VTy2 && Offset < 0) {
637 int64_t VTy2TSS = (int64_t) TD->getTypeStoreSize(VTy2);
638 OffsetInElmts = Offset/VTy2TSS;
639 return (abs64(Offset) % VTy2TSS) == 0;
642 OffsetInElmts = Offset/VTyTSS;
643 return (abs64(Offset) % VTyTSS) == 0;
649 // Returns true if the provided CallInst represents an intrinsic that can
651 bool isVectorizableIntrinsic(CallInst* I) {
652 Function *F = I->getCalledFunction();
653 if (!F) return false;
655 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID();
656 if (!IID) return false;
661 case Intrinsic::sqrt:
662 case Intrinsic::powi:
666 case Intrinsic::log2:
667 case Intrinsic::log10:
669 case Intrinsic::exp2:
671 return Config.VectorizeMath;
673 case Intrinsic::fmuladd:
674 return Config.VectorizeFMA;
678 bool isPureIEChain(InsertElementInst *IE) {
679 InsertElementInst *IENext = IE;
681 if (!isa<UndefValue>(IENext->getOperand(0)) &&
682 !isa<InsertElementInst>(IENext->getOperand(0))) {
686 dyn_cast<InsertElementInst>(IENext->getOperand(0))));
692 // This function implements one vectorization iteration on the provided
693 // basic block. It returns true if the block is changed.
694 bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) {
696 BasicBlock::iterator Start = BB.getFirstInsertionPt();
698 std::vector<Value *> AllPairableInsts;
699 DenseMap<Value *, Value *> AllChosenPairs;
700 DenseSet<ValuePair> AllFixedOrderPairs;
701 DenseMap<VPPair, unsigned> AllPairConnectionTypes;
702 DenseMap<ValuePair, std::vector<ValuePair> > AllConnectedPairs,
703 AllConnectedPairDeps;
706 std::vector<Value *> PairableInsts;
707 DenseMap<Value *, std::vector<Value *> > CandidatePairs;
708 DenseSet<ValuePair> FixedOrderPairs;
709 DenseMap<ValuePair, int> CandidatePairCostSavings;
710 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs,
712 CandidatePairCostSavings,
713 PairableInsts, NonPow2Len);
714 if (PairableInsts.empty()) continue;
716 // Build the candidate pair set for faster lookups.
717 DenseSet<ValuePair> CandidatePairsSet;
718 for (DenseMap<Value *, std::vector<Value *> >::iterator I =
719 CandidatePairs.begin(), E = CandidatePairs.end(); I != E; ++I)
720 for (std::vector<Value *>::iterator J = I->second.begin(),
721 JE = I->second.end(); J != JE; ++J)
722 CandidatePairsSet.insert(ValuePair(I->first, *J));
724 // Now we have a map of all of the pairable instructions and we need to
725 // select the best possible pairing. A good pairing is one such that the
726 // users of the pair are also paired. This defines a (directed) forest
727 // over the pairs such that two pairs are connected iff the second pair
730 // Note that it only matters that both members of the second pair use some
731 // element of the first pair (to allow for splatting).
733 DenseMap<ValuePair, std::vector<ValuePair> > ConnectedPairs,
735 DenseMap<VPPair, unsigned> PairConnectionTypes;
736 computeConnectedPairs(CandidatePairs, CandidatePairsSet,
737 PairableInsts, ConnectedPairs, PairConnectionTypes);
738 if (ConnectedPairs.empty()) continue;
740 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator
741 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
743 for (std::vector<ValuePair>::iterator J = I->second.begin(),
744 JE = I->second.end(); J != JE; ++J)
745 ConnectedPairDeps[*J].push_back(I->first);
747 // Build the pairable-instruction dependency map
748 DenseSet<ValuePair> PairableInstUsers;
749 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers);
751 // There is now a graph of the connected pairs. For each variable, pick
752 // the pairing with the largest dag meeting the depth requirement on at
753 // least one branch. Then select all pairings that are part of that dag
754 // and remove them from the list of available pairings and pairable
757 DenseMap<Value *, Value *> ChosenPairs;
758 choosePairs(CandidatePairs, CandidatePairsSet,
759 CandidatePairCostSavings,
760 PairableInsts, FixedOrderPairs, PairConnectionTypes,
761 ConnectedPairs, ConnectedPairDeps,
762 PairableInstUsers, ChosenPairs);
764 if (ChosenPairs.empty()) continue;
765 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(),
766 PairableInsts.end());
767 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end());
769 // Only for the chosen pairs, propagate information on fixed-order pairs,
770 // pair connections, and their types to the data structures used by the
771 // pair fusion procedures.
772 for (DenseMap<Value *, Value *>::iterator I = ChosenPairs.begin(),
773 IE = ChosenPairs.end(); I != IE; ++I) {
774 if (FixedOrderPairs.count(*I))
775 AllFixedOrderPairs.insert(*I);
776 else if (FixedOrderPairs.count(ValuePair(I->second, I->first)))
777 AllFixedOrderPairs.insert(ValuePair(I->second, I->first));
779 for (DenseMap<Value *, Value *>::iterator J = ChosenPairs.begin();
781 DenseMap<VPPair, unsigned>::iterator K =
782 PairConnectionTypes.find(VPPair(*I, *J));
783 if (K != PairConnectionTypes.end()) {
784 AllPairConnectionTypes.insert(*K);
786 K = PairConnectionTypes.find(VPPair(*J, *I));
787 if (K != PairConnectionTypes.end())
788 AllPairConnectionTypes.insert(*K);
793 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator
794 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
796 for (std::vector<ValuePair>::iterator J = I->second.begin(),
797 JE = I->second.end(); J != JE; ++J)
798 if (AllPairConnectionTypes.count(VPPair(I->first, *J))) {
799 AllConnectedPairs[I->first].push_back(*J);
800 AllConnectedPairDeps[*J].push_back(I->first);
802 } while (ShouldContinue);
804 if (AllChosenPairs.empty()) return false;
805 NumFusedOps += AllChosenPairs.size();
807 // A set of pairs has now been selected. It is now necessary to replace the
808 // paired instructions with vector instructions. For this procedure each
809 // operand must be replaced with a vector operand. This vector is formed
810 // by using build_vector on the old operands. The replaced values are then
811 // replaced with a vector_extract on the result. Subsequent optimization
812 // passes should coalesce the build/extract combinations.
814 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs, AllFixedOrderPairs,
815 AllPairConnectionTypes,
816 AllConnectedPairs, AllConnectedPairDeps);
818 // It is important to cleanup here so that future iterations of this
819 // function have less work to do.
820 (void) SimplifyInstructionsInBlock(&BB, TD, AA->getTargetLibraryInfo());
824 // This function returns true if the provided instruction is capable of being
825 // fused into a vector instruction. This determination is based only on the
826 // type and other attributes of the instruction.
827 bool BBVectorize::isInstVectorizable(Instruction *I,
828 bool &IsSimpleLoadStore) {
829 IsSimpleLoadStore = false;
831 if (CallInst *C = dyn_cast<CallInst>(I)) {
832 if (!isVectorizableIntrinsic(C))
834 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) {
835 // Vectorize simple loads if possbile:
836 IsSimpleLoadStore = L->isSimple();
837 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
839 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) {
840 // Vectorize simple stores if possbile:
841 IsSimpleLoadStore = S->isSimple();
842 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
844 } else if (CastInst *C = dyn_cast<CastInst>(I)) {
845 // We can vectorize casts, but not casts of pointer types, etc.
846 if (!Config.VectorizeCasts)
849 Type *SrcTy = C->getSrcTy();
850 if (!SrcTy->isSingleValueType())
853 Type *DestTy = C->getDestTy();
854 if (!DestTy->isSingleValueType())
856 } else if (isa<SelectInst>(I)) {
857 if (!Config.VectorizeSelect)
859 } else if (isa<CmpInst>(I)) {
860 if (!Config.VectorizeCmp)
862 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) {
863 if (!Config.VectorizeGEP)
866 // Currently, vector GEPs exist only with one index.
867 if (G->getNumIndices() != 1)
869 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) ||
870 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) {
874 // We can't vectorize memory operations without target data
875 if (TD == 0 && IsSimpleLoadStore)
879 getInstructionTypes(I, T1, T2);
881 // Not every type can be vectorized...
882 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) ||
883 !(VectorType::isValidElementType(T2) || T2->isVectorTy()))
886 if (T1->getScalarSizeInBits() == 1) {
887 if (!Config.VectorizeBools)
890 if (!Config.VectorizeInts && T1->isIntOrIntVectorTy())
894 if (T2->getScalarSizeInBits() == 1) {
895 if (!Config.VectorizeBools)
898 if (!Config.VectorizeInts && T2->isIntOrIntVectorTy())
902 if (!Config.VectorizeFloats
903 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy()))
906 // Don't vectorize target-specific types.
907 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy())
909 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy())
912 if ((!Config.VectorizePointers || TD == 0) &&
913 (T1->getScalarType()->isPointerTy() ||
914 T2->getScalarType()->isPointerTy()))
917 if (!TTI && (T1->getPrimitiveSizeInBits() >= Config.VectorBits ||
918 T2->getPrimitiveSizeInBits() >= Config.VectorBits))
924 // This function returns true if the two provided instructions are compatible
925 // (meaning that they can be fused into a vector instruction). This assumes
926 // that I has already been determined to be vectorizable and that J is not
927 // in the use dag of I.
928 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J,
929 bool IsSimpleLoadStore, bool NonPow2Len,
930 int &CostSavings, int &FixedOrder) {
931 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I <<
932 " <-> " << *J << "\n");
937 // Loads and stores can be merged if they have different alignments,
938 // but are otherwise the same.
939 if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment |
940 (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0)))
943 Type *IT1, *IT2, *JT1, *JT2;
944 getInstructionTypes(I, IT1, IT2);
945 getInstructionTypes(J, JT1, JT2);
946 unsigned MaxTypeBits = std::max(
947 IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(),
948 IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits());
949 if (!TTI && MaxTypeBits > Config.VectorBits)
952 // FIXME: handle addsub-type operations!
954 if (IsSimpleLoadStore) {
956 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
957 int64_t OffsetInElmts = 0;
958 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
959 IAddressSpace, JAddressSpace,
960 OffsetInElmts) && abs64(OffsetInElmts) == 1) {
961 FixedOrder = (int) OffsetInElmts;
962 unsigned BottomAlignment = IAlignment;
963 if (OffsetInElmts < 0) BottomAlignment = JAlignment;
965 Type *aTypeI = isa<StoreInst>(I) ?
966 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType();
967 Type *aTypeJ = isa<StoreInst>(J) ?
968 cast<StoreInst>(J)->getValueOperand()->getType() : J->getType();
969 Type *VType = getVecTypeForPair(aTypeI, aTypeJ);
971 if (Config.AlignedOnly) {
972 // An aligned load or store is possible only if the instruction
973 // with the lower offset has an alignment suitable for the
976 unsigned VecAlignment = TD->getPrefTypeAlignment(VType);
977 if (BottomAlignment < VecAlignment)
982 unsigned ICost = TTI->getMemoryOpCost(I->getOpcode(), aTypeI,
983 IAlignment, IAddressSpace);
984 unsigned JCost = TTI->getMemoryOpCost(J->getOpcode(), aTypeJ,
985 JAlignment, JAddressSpace);
986 unsigned VCost = TTI->getMemoryOpCost(I->getOpcode(), VType,
990 ICost += TTI->getAddressComputationCost(aTypeI);
991 JCost += TTI->getAddressComputationCost(aTypeJ);
992 VCost += TTI->getAddressComputationCost(VType);
994 if (VCost > ICost + JCost)
997 // We don't want to fuse to a type that will be split, even
998 // if the two input types will also be split and there is no other
1000 unsigned VParts = TTI->getNumberOfParts(VType);
1003 else if (!VParts && VCost == ICost + JCost)
1006 CostSavings = ICost + JCost - VCost;
1012 unsigned ICost = getInstrCost(I->getOpcode(), IT1, IT2);
1013 unsigned JCost = getInstrCost(J->getOpcode(), JT1, JT2);
1014 Type *VT1 = getVecTypeForPair(IT1, JT1),
1015 *VT2 = getVecTypeForPair(IT2, JT2);
1017 // Note that this procedure is incorrect for insert and extract element
1018 // instructions (because combining these often results in a shuffle),
1019 // but this cost is ignored (because insert and extract element
1020 // instructions are assigned a zero depth factor and are not really
1021 // fused in general).
1022 unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2);
1024 if (VCost > ICost + JCost)
1027 // We don't want to fuse to a type that will be split, even
1028 // if the two input types will also be split and there is no other
1030 unsigned VParts1 = TTI->getNumberOfParts(VT1),
1031 VParts2 = TTI->getNumberOfParts(VT2);
1032 if (VParts1 > 1 || VParts2 > 1)
1034 else if ((!VParts1 || !VParts2) && VCost == ICost + JCost)
1037 CostSavings = ICost + JCost - VCost;
1040 // The powi intrinsic is special because only the first argument is
1041 // vectorized, the second arguments must be equal.
1042 CallInst *CI = dyn_cast<CallInst>(I);
1044 if (CI && (FI = CI->getCalledFunction())) {
1045 Intrinsic::ID IID = (Intrinsic::ID) FI->getIntrinsicID();
1046 if (IID == Intrinsic::powi) {
1047 Value *A1I = CI->getArgOperand(1),
1048 *A1J = cast<CallInst>(J)->getArgOperand(1);
1049 const SCEV *A1ISCEV = SE->getSCEV(A1I),
1050 *A1JSCEV = SE->getSCEV(A1J);
1051 return (A1ISCEV == A1JSCEV);
1055 SmallVector<Type*, 4> Tys;
1056 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
1057 Tys.push_back(CI->getArgOperand(i)->getType());
1058 unsigned ICost = TTI->getIntrinsicInstrCost(IID, IT1, Tys);
1061 CallInst *CJ = cast<CallInst>(J);
1062 for (unsigned i = 0, ie = CJ->getNumArgOperands(); i != ie; ++i)
1063 Tys.push_back(CJ->getArgOperand(i)->getType());
1064 unsigned JCost = TTI->getIntrinsicInstrCost(IID, JT1, Tys);
1067 assert(CI->getNumArgOperands() == CJ->getNumArgOperands() &&
1068 "Intrinsic argument counts differ");
1069 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
1070 if (IID == Intrinsic::powi && i == 1)
1071 Tys.push_back(CI->getArgOperand(i)->getType());
1073 Tys.push_back(getVecTypeForPair(CI->getArgOperand(i)->getType(),
1074 CJ->getArgOperand(i)->getType()));
1077 Type *RetTy = getVecTypeForPair(IT1, JT1);
1078 unsigned VCost = TTI->getIntrinsicInstrCost(IID, RetTy, Tys);
1080 if (VCost > ICost + JCost)
1083 // We don't want to fuse to a type that will be split, even
1084 // if the two input types will also be split and there is no other
1086 unsigned RetParts = TTI->getNumberOfParts(RetTy);
1089 else if (!RetParts && VCost == ICost + JCost)
1092 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
1093 if (!Tys[i]->isVectorTy())
1096 unsigned NumParts = TTI->getNumberOfParts(Tys[i]);
1099 else if (!NumParts && VCost == ICost + JCost)
1103 CostSavings = ICost + JCost - VCost;
1110 // Figure out whether or not J uses I and update the users and write-set
1111 // structures associated with I. Specifically, Users represents the set of
1112 // instructions that depend on I. WriteSet represents the set
1113 // of memory locations that are dependent on I. If UpdateUsers is true,
1114 // and J uses I, then Users is updated to contain J and WriteSet is updated
1115 // to contain any memory locations to which J writes. The function returns
1116 // true if J uses I. By default, alias analysis is used to determine
1117 // whether J reads from memory that overlaps with a location in WriteSet.
1118 // If LoadMoveSet is not null, then it is a previously-computed map
1119 // where the key is the memory-based user instruction and the value is
1120 // the instruction to be compared with I. So, if LoadMoveSet is provided,
1121 // then the alias analysis is not used. This is necessary because this
1122 // function is called during the process of moving instructions during
1123 // vectorization and the results of the alias analysis are not stable during
1125 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users,
1126 AliasSetTracker &WriteSet, Instruction *I,
1127 Instruction *J, bool UpdateUsers,
1128 DenseSet<ValuePair> *LoadMoveSetPairs) {
1131 // This instruction may already be marked as a user due, for example, to
1132 // being a member of a selected pair.
1137 for (User::op_iterator JU = J->op_begin(), JE = J->op_end();
1140 if (I == V || Users.count(V)) {
1145 if (!UsesI && J->mayReadFromMemory()) {
1146 if (LoadMoveSetPairs) {
1147 UsesI = LoadMoveSetPairs->count(ValuePair(J, I));
1149 for (AliasSetTracker::iterator W = WriteSet.begin(),
1150 WE = WriteSet.end(); W != WE; ++W) {
1151 if (W->aliasesUnknownInst(J, *AA)) {
1159 if (UsesI && UpdateUsers) {
1160 if (J->mayWriteToMemory()) WriteSet.add(J);
1167 // This function iterates over all instruction pairs in the provided
1168 // basic block and collects all candidate pairs for vectorization.
1169 bool BBVectorize::getCandidatePairs(BasicBlock &BB,
1170 BasicBlock::iterator &Start,
1171 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1172 DenseSet<ValuePair> &FixedOrderPairs,
1173 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1174 std::vector<Value *> &PairableInsts, bool NonPow2Len) {
1175 size_t TotalPairs = 0;
1176 BasicBlock::iterator E = BB.end();
1177 if (Start == E) return false;
1179 bool ShouldContinue = false, IAfterStart = false;
1180 for (BasicBlock::iterator I = Start++; I != E; ++I) {
1181 if (I == Start) IAfterStart = true;
1183 bool IsSimpleLoadStore;
1184 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue;
1186 // Look for an instruction with which to pair instruction *I...
1187 DenseSet<Value *> Users;
1188 AliasSetTracker WriteSet(*AA);
1189 if (I->mayWriteToMemory()) WriteSet.add(I);
1191 bool JAfterStart = IAfterStart;
1192 BasicBlock::iterator J = llvm::next(I);
1193 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) {
1194 if (J == Start) JAfterStart = true;
1196 // Determine if J uses I, if so, exit the loop.
1197 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep);
1198 if (Config.FastDep) {
1199 // Note: For this heuristic to be effective, independent operations
1200 // must tend to be intermixed. This is likely to be true from some
1201 // kinds of grouped loop unrolling (but not the generic LLVM pass),
1202 // but otherwise may require some kind of reordering pass.
1204 // When using fast dependency analysis,
1205 // stop searching after first use:
1208 if (UsesI) continue;
1211 // J does not use I, and comes before the first use of I, so it can be
1212 // merged with I if the instructions are compatible.
1213 int CostSavings, FixedOrder;
1214 if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len,
1215 CostSavings, FixedOrder)) continue;
1217 // J is a candidate for merging with I.
1218 if (!PairableInsts.size() ||
1219 PairableInsts[PairableInsts.size()-1] != I) {
1220 PairableInsts.push_back(I);
1223 CandidatePairs[I].push_back(J);
1226 CandidatePairCostSavings.insert(ValuePairWithCost(ValuePair(I, J),
1229 if (FixedOrder == 1)
1230 FixedOrderPairs.insert(ValuePair(I, J));
1231 else if (FixedOrder == -1)
1232 FixedOrderPairs.insert(ValuePair(J, I));
1234 // The next call to this function must start after the last instruction
1235 // selected during this invocation.
1237 Start = llvm::next(J);
1238 IAfterStart = JAfterStart = false;
1241 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair "
1242 << *I << " <-> " << *J << " (cost savings: " <<
1243 CostSavings << ")\n");
1245 // If we have already found too many pairs, break here and this function
1246 // will be called again starting after the last instruction selected
1247 // during this invocation.
1248 if (PairableInsts.size() >= Config.MaxInsts ||
1249 TotalPairs >= Config.MaxPairs) {
1250 ShouldContinue = true;
1259 DEBUG(dbgs() << "BBV: found " << PairableInsts.size()
1260 << " instructions with candidate pairs\n");
1262 return ShouldContinue;
1265 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that
1266 // it looks for pairs such that both members have an input which is an
1267 // output of PI or PJ.
1268 void BBVectorize::computePairsConnectedTo(
1269 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1270 DenseSet<ValuePair> &CandidatePairsSet,
1271 std::vector<Value *> &PairableInsts,
1272 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1273 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1277 // For each possible pairing for this variable, look at the uses of
1278 // the first value...
1279 for (Value::use_iterator I = P.first->use_begin(),
1280 E = P.first->use_end(); I != E; ++I) {
1281 if (isa<LoadInst>(*I)) {
1282 // A pair cannot be connected to a load because the load only takes one
1283 // operand (the address) and it is a scalar even after vectorization.
1285 } else if ((SI = dyn_cast<StoreInst>(*I)) &&
1286 P.first == SI->getPointerOperand()) {
1287 // Similarly, a pair cannot be connected to a store through its
1292 // For each use of the first variable, look for uses of the second
1294 for (Value::use_iterator J = P.second->use_begin(),
1295 E2 = P.second->use_end(); J != E2; ++J) {
1296 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1297 P.second == SJ->getPointerOperand())
1301 if (CandidatePairsSet.count(ValuePair(*I, *J))) {
1302 VPPair VP(P, ValuePair(*I, *J));
1303 ConnectedPairs[VP.first].push_back(VP.second);
1304 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionDirect));
1308 if (CandidatePairsSet.count(ValuePair(*J, *I))) {
1309 VPPair VP(P, ValuePair(*J, *I));
1310 ConnectedPairs[VP.first].push_back(VP.second);
1311 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSwap));
1315 if (Config.SplatBreaksChain) continue;
1316 // Look for cases where just the first value in the pair is used by
1317 // both members of another pair (splatting).
1318 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) {
1319 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1320 P.first == SJ->getPointerOperand())
1323 if (CandidatePairsSet.count(ValuePair(*I, *J))) {
1324 VPPair VP(P, ValuePair(*I, *J));
1325 ConnectedPairs[VP.first].push_back(VP.second);
1326 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1331 if (Config.SplatBreaksChain) return;
1332 // Look for cases where just the second value in the pair is used by
1333 // both members of another pair (splatting).
1334 for (Value::use_iterator I = P.second->use_begin(),
1335 E = P.second->use_end(); I != E; ++I) {
1336 if (isa<LoadInst>(*I))
1338 else if ((SI = dyn_cast<StoreInst>(*I)) &&
1339 P.second == SI->getPointerOperand())
1342 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) {
1343 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1344 P.second == SJ->getPointerOperand())
1347 if (CandidatePairsSet.count(ValuePair(*I, *J))) {
1348 VPPair VP(P, ValuePair(*I, *J));
1349 ConnectedPairs[VP.first].push_back(VP.second);
1350 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1356 // This function figures out which pairs are connected. Two pairs are
1357 // connected if some output of the first pair forms an input to both members
1358 // of the second pair.
1359 void BBVectorize::computeConnectedPairs(
1360 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1361 DenseSet<ValuePair> &CandidatePairsSet,
1362 std::vector<Value *> &PairableInsts,
1363 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1364 DenseMap<VPPair, unsigned> &PairConnectionTypes) {
1365 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
1366 PE = PairableInsts.end(); PI != PE; ++PI) {
1367 DenseMap<Value *, std::vector<Value *> >::iterator PP =
1368 CandidatePairs.find(*PI);
1369 if (PP == CandidatePairs.end())
1372 for (std::vector<Value *>::iterator P = PP->second.begin(),
1373 E = PP->second.end(); P != E; ++P)
1374 computePairsConnectedTo(CandidatePairs, CandidatePairsSet,
1375 PairableInsts, ConnectedPairs,
1376 PairConnectionTypes, ValuePair(*PI, *P));
1379 DEBUG(size_t TotalPairs = 0;
1380 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator I =
1381 ConnectedPairs.begin(), IE = ConnectedPairs.end(); I != IE; ++I)
1382 TotalPairs += I->second.size();
1383 dbgs() << "BBV: found " << TotalPairs
1384 << " pair connections.\n");
1387 // This function builds a set of use tuples such that <A, B> is in the set
1388 // if B is in the use dag of A. If B is in the use dag of A, then B
1389 // depends on the output of A.
1390 void BBVectorize::buildDepMap(
1392 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1393 std::vector<Value *> &PairableInsts,
1394 DenseSet<ValuePair> &PairableInstUsers) {
1395 DenseSet<Value *> IsInPair;
1396 for (DenseMap<Value *, std::vector<Value *> >::iterator C =
1397 CandidatePairs.begin(), E = CandidatePairs.end(); C != E; ++C) {
1398 IsInPair.insert(C->first);
1399 IsInPair.insert(C->second.begin(), C->second.end());
1402 // Iterate through the basic block, recording all users of each
1403 // pairable instruction.
1405 BasicBlock::iterator E = BB.end(), EL =
1406 BasicBlock::iterator(cast<Instruction>(PairableInsts.back()));
1407 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) {
1408 if (IsInPair.find(I) == IsInPair.end()) continue;
1410 DenseSet<Value *> Users;
1411 AliasSetTracker WriteSet(*AA);
1412 if (I->mayWriteToMemory()) WriteSet.add(I);
1414 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J) {
1415 (void) trackUsesOfI(Users, WriteSet, I, J);
1421 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end();
1423 if (IsInPair.find(*U) == IsInPair.end()) continue;
1424 PairableInstUsers.insert(ValuePair(I, *U));
1432 // Returns true if an input to pair P is an output of pair Q and also an
1433 // input of pair Q is an output of pair P. If this is the case, then these
1434 // two pairs cannot be simultaneously fused.
1435 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q,
1436 DenseSet<ValuePair> &PairableInstUsers,
1437 DenseMap<ValuePair, std::vector<ValuePair> > *PairableInstUserMap,
1438 DenseSet<VPPair> *PairableInstUserPairSet) {
1439 // Two pairs are in conflict if they are mutual Users of eachother.
1440 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) ||
1441 PairableInstUsers.count(ValuePair(P.first, Q.second)) ||
1442 PairableInstUsers.count(ValuePair(P.second, Q.first)) ||
1443 PairableInstUsers.count(ValuePair(P.second, Q.second));
1444 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) ||
1445 PairableInstUsers.count(ValuePair(Q.first, P.second)) ||
1446 PairableInstUsers.count(ValuePair(Q.second, P.first)) ||
1447 PairableInstUsers.count(ValuePair(Q.second, P.second));
1448 if (PairableInstUserMap) {
1449 // FIXME: The expensive part of the cycle check is not so much the cycle
1450 // check itself but this edge insertion procedure. This needs some
1451 // profiling and probably a different data structure.
1453 if (PairableInstUserPairSet->insert(VPPair(Q, P)).second)
1454 (*PairableInstUserMap)[Q].push_back(P);
1457 if (PairableInstUserPairSet->insert(VPPair(P, Q)).second)
1458 (*PairableInstUserMap)[P].push_back(Q);
1462 return (QUsesP && PUsesQ);
1465 // This function walks the use graph of current pairs to see if, starting
1466 // from P, the walk returns to P.
1467 bool BBVectorize::pairWillFormCycle(ValuePair P,
1468 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
1469 DenseSet<ValuePair> &CurrentPairs) {
1470 DEBUG(if (DebugCycleCheck)
1471 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> "
1472 << *P.second << "\n");
1473 // A lookup table of visisted pairs is kept because the PairableInstUserMap
1474 // contains non-direct associations.
1475 DenseSet<ValuePair> Visited;
1476 SmallVector<ValuePair, 32> Q;
1477 // General depth-first post-order traversal:
1480 ValuePair QTop = Q.pop_back_val();
1481 Visited.insert(QTop);
1483 DEBUG(if (DebugCycleCheck)
1484 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> "
1485 << *QTop.second << "\n");
1486 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
1487 PairableInstUserMap.find(QTop);
1488 if (QQ == PairableInstUserMap.end())
1491 for (std::vector<ValuePair>::iterator C = QQ->second.begin(),
1492 CE = QQ->second.end(); C != CE; ++C) {
1495 << "BBV: rejected to prevent non-trivial cycle formation: "
1496 << QTop.first << " <-> " << C->second << "\n");
1500 if (CurrentPairs.count(*C) && !Visited.count(*C))
1503 } while (!Q.empty());
1508 // This function builds the initial dag of connected pairs with the
1509 // pair J at the root.
1510 void BBVectorize::buildInitialDAGFor(
1511 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1512 DenseSet<ValuePair> &CandidatePairsSet,
1513 std::vector<Value *> &PairableInsts,
1514 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1515 DenseSet<ValuePair> &PairableInstUsers,
1516 DenseMap<Value *, Value *> &ChosenPairs,
1517 DenseMap<ValuePair, size_t> &DAG, ValuePair J) {
1518 // Each of these pairs is viewed as the root node of a DAG. The DAG
1519 // is then walked (depth-first). As this happens, we keep track of
1520 // the pairs that compose the DAG and the maximum depth of the DAG.
1521 SmallVector<ValuePairWithDepth, 32> Q;
1522 // General depth-first post-order traversal:
1523 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1525 ValuePairWithDepth QTop = Q.back();
1527 // Push each child onto the queue:
1528 bool MoreChildren = false;
1529 size_t MaxChildDepth = QTop.second;
1530 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
1531 ConnectedPairs.find(QTop.first);
1532 if (QQ != ConnectedPairs.end())
1533 for (std::vector<ValuePair>::iterator k = QQ->second.begin(),
1534 ke = QQ->second.end(); k != ke; ++k) {
1535 // Make sure that this child pair is still a candidate:
1536 if (CandidatePairsSet.count(*k)) {
1537 DenseMap<ValuePair, size_t>::iterator C = DAG.find(*k);
1538 if (C == DAG.end()) {
1539 size_t d = getDepthFactor(k->first);
1540 Q.push_back(ValuePairWithDepth(*k, QTop.second+d));
1541 MoreChildren = true;
1543 MaxChildDepth = std::max(MaxChildDepth, C->second);
1548 if (!MoreChildren) {
1549 // Record the current pair as part of the DAG:
1550 DAG.insert(ValuePairWithDepth(QTop.first, MaxChildDepth));
1553 } while (!Q.empty());
1556 // Given some initial dag, prune it by removing conflicting pairs (pairs
1557 // that cannot be simultaneously chosen for vectorization).
1558 void BBVectorize::pruneDAGFor(
1559 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1560 std::vector<Value *> &PairableInsts,
1561 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1562 DenseSet<ValuePair> &PairableInstUsers,
1563 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
1564 DenseSet<VPPair> &PairableInstUserPairSet,
1565 DenseMap<Value *, Value *> &ChosenPairs,
1566 DenseMap<ValuePair, size_t> &DAG,
1567 DenseSet<ValuePair> &PrunedDAG, ValuePair J,
1568 bool UseCycleCheck) {
1569 SmallVector<ValuePairWithDepth, 32> Q;
1570 // General depth-first post-order traversal:
1571 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1573 ValuePairWithDepth QTop = Q.pop_back_val();
1574 PrunedDAG.insert(QTop.first);
1576 // Visit each child, pruning as necessary...
1577 SmallVector<ValuePairWithDepth, 8> BestChildren;
1578 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
1579 ConnectedPairs.find(QTop.first);
1580 if (QQ == ConnectedPairs.end())
1583 for (std::vector<ValuePair>::iterator K = QQ->second.begin(),
1584 KE = QQ->second.end(); K != KE; ++K) {
1585 DenseMap<ValuePair, size_t>::iterator C = DAG.find(*K);
1586 if (C == DAG.end()) continue;
1588 // This child is in the DAG, now we need to make sure it is the
1589 // best of any conflicting children. There could be multiple
1590 // conflicting children, so first, determine if we're keeping
1591 // this child, then delete conflicting children as necessary.
1593 // It is also necessary to guard against pairing-induced
1594 // dependencies. Consider instructions a .. x .. y .. b
1595 // such that (a,b) are to be fused and (x,y) are to be fused
1596 // but a is an input to x and b is an output from y. This
1597 // means that y cannot be moved after b but x must be moved
1598 // after b for (a,b) to be fused. In other words, after
1599 // fusing (a,b) we have y .. a/b .. x where y is an input
1600 // to a/b and x is an output to a/b: x and y can no longer
1601 // be legally fused. To prevent this condition, we must
1602 // make sure that a child pair added to the DAG is not
1603 // both an input and output of an already-selected pair.
1605 // Pairing-induced dependencies can also form from more complicated
1606 // cycles. The pair vs. pair conflicts are easy to check, and so
1607 // that is done explicitly for "fast rejection", and because for
1608 // child vs. child conflicts, we may prefer to keep the current
1609 // pair in preference to the already-selected child.
1610 DenseSet<ValuePair> CurrentPairs;
1613 for (SmallVectorImpl<ValuePairWithDepth>::iterator C2
1614 = BestChildren.begin(), E2 = BestChildren.end();
1616 if (C2->first.first == C->first.first ||
1617 C2->first.first == C->first.second ||
1618 C2->first.second == C->first.first ||
1619 C2->first.second == C->first.second ||
1620 pairsConflict(C2->first, C->first, PairableInstUsers,
1621 UseCycleCheck ? &PairableInstUserMap : 0,
1622 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1623 if (C2->second >= C->second) {
1628 CurrentPairs.insert(C2->first);
1631 if (!CanAdd) continue;
1633 // Even worse, this child could conflict with another node already
1634 // selected for the DAG. If that is the case, ignore this child.
1635 for (DenseSet<ValuePair>::iterator T = PrunedDAG.begin(),
1636 E2 = PrunedDAG.end(); T != E2; ++T) {
1637 if (T->first == C->first.first ||
1638 T->first == C->first.second ||
1639 T->second == C->first.first ||
1640 T->second == C->first.second ||
1641 pairsConflict(*T, C->first, PairableInstUsers,
1642 UseCycleCheck ? &PairableInstUserMap : 0,
1643 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1648 CurrentPairs.insert(*T);
1650 if (!CanAdd) continue;
1652 // And check the queue too...
1653 for (SmallVectorImpl<ValuePairWithDepth>::iterator C2 = Q.begin(),
1654 E2 = Q.end(); C2 != E2; ++C2) {
1655 if (C2->first.first == C->first.first ||
1656 C2->first.first == C->first.second ||
1657 C2->first.second == C->first.first ||
1658 C2->first.second == C->first.second ||
1659 pairsConflict(C2->first, C->first, PairableInstUsers,
1660 UseCycleCheck ? &PairableInstUserMap : 0,
1661 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1666 CurrentPairs.insert(C2->first);
1668 if (!CanAdd) continue;
1670 // Last but not least, check for a conflict with any of the
1671 // already-chosen pairs.
1672 for (DenseMap<Value *, Value *>::iterator C2 =
1673 ChosenPairs.begin(), E2 = ChosenPairs.end();
1675 if (pairsConflict(*C2, C->first, PairableInstUsers,
1676 UseCycleCheck ? &PairableInstUserMap : 0,
1677 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1682 CurrentPairs.insert(*C2);
1684 if (!CanAdd) continue;
1686 // To check for non-trivial cycles formed by the addition of the
1687 // current pair we've formed a list of all relevant pairs, now use a
1688 // graph walk to check for a cycle. We start from the current pair and
1689 // walk the use dag to see if we again reach the current pair. If we
1690 // do, then the current pair is rejected.
1692 // FIXME: It may be more efficient to use a topological-ordering
1693 // algorithm to improve the cycle check. This should be investigated.
1694 if (UseCycleCheck &&
1695 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs))
1698 // This child can be added, but we may have chosen it in preference
1699 // to an already-selected child. Check for this here, and if a
1700 // conflict is found, then remove the previously-selected child
1701 // before adding this one in its place.
1702 for (SmallVectorImpl<ValuePairWithDepth>::iterator C2
1703 = BestChildren.begin(); C2 != BestChildren.end();) {
1704 if (C2->first.first == C->first.first ||
1705 C2->first.first == C->first.second ||
1706 C2->first.second == C->first.first ||
1707 C2->first.second == C->first.second ||
1708 pairsConflict(C2->first, C->first, PairableInstUsers))
1709 C2 = BestChildren.erase(C2);
1714 BestChildren.push_back(ValuePairWithDepth(C->first, C->second));
1717 for (SmallVectorImpl<ValuePairWithDepth>::iterator C
1718 = BestChildren.begin(), E2 = BestChildren.end();
1720 size_t DepthF = getDepthFactor(C->first.first);
1721 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF));
1723 } while (!Q.empty());
1726 // This function finds the best dag of mututally-compatible connected
1727 // pairs, given the choice of root pairs as an iterator range.
1728 void BBVectorize::findBestDAGFor(
1729 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1730 DenseSet<ValuePair> &CandidatePairsSet,
1731 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1732 std::vector<Value *> &PairableInsts,
1733 DenseSet<ValuePair> &FixedOrderPairs,
1734 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1735 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1736 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
1737 DenseSet<ValuePair> &PairableInstUsers,
1738 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
1739 DenseSet<VPPair> &PairableInstUserPairSet,
1740 DenseMap<Value *, Value *> &ChosenPairs,
1741 DenseSet<ValuePair> &BestDAG, size_t &BestMaxDepth,
1742 int &BestEffSize, Value *II, std::vector<Value *>&JJ,
1743 bool UseCycleCheck) {
1744 for (std::vector<Value *>::iterator J = JJ.begin(), JE = JJ.end();
1746 ValuePair IJ(II, *J);
1747 if (!CandidatePairsSet.count(IJ))
1750 // Before going any further, make sure that this pair does not
1751 // conflict with any already-selected pairs (see comment below
1752 // near the DAG pruning for more details).
1753 DenseSet<ValuePair> ChosenPairSet;
1754 bool DoesConflict = false;
1755 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(),
1756 E = ChosenPairs.end(); C != E; ++C) {
1757 if (pairsConflict(*C, IJ, PairableInstUsers,
1758 UseCycleCheck ? &PairableInstUserMap : 0,
1759 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1760 DoesConflict = true;
1764 ChosenPairSet.insert(*C);
1766 if (DoesConflict) continue;
1768 if (UseCycleCheck &&
1769 pairWillFormCycle(IJ, PairableInstUserMap, ChosenPairSet))
1772 DenseMap<ValuePair, size_t> DAG;
1773 buildInitialDAGFor(CandidatePairs, CandidatePairsSet,
1774 PairableInsts, ConnectedPairs,
1775 PairableInstUsers, ChosenPairs, DAG, IJ);
1777 // Because we'll keep the child with the largest depth, the largest
1778 // depth is still the same in the unpruned DAG.
1779 size_t MaxDepth = DAG.lookup(IJ);
1781 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found DAG for pair {"
1782 << *IJ.first << " <-> " << *IJ.second << "} of depth " <<
1783 MaxDepth << " and size " << DAG.size() << "\n");
1785 // At this point the DAG has been constructed, but, may contain
1786 // contradictory children (meaning that different children of
1787 // some dag node may be attempting to fuse the same instruction).
1788 // So now we walk the dag again, in the case of a conflict,
1789 // keep only the child with the largest depth. To break a tie,
1790 // favor the first child.
1792 DenseSet<ValuePair> PrunedDAG;
1793 pruneDAGFor(CandidatePairs, PairableInsts, ConnectedPairs,
1794 PairableInstUsers, PairableInstUserMap,
1795 PairableInstUserPairSet,
1796 ChosenPairs, DAG, PrunedDAG, IJ, UseCycleCheck);
1800 DenseSet<Value *> PrunedDAGInstrs;
1801 for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(),
1802 E = PrunedDAG.end(); S != E; ++S) {
1803 PrunedDAGInstrs.insert(S->first);
1804 PrunedDAGInstrs.insert(S->second);
1807 // The set of pairs that have already contributed to the total cost.
1808 DenseSet<ValuePair> IncomingPairs;
1810 // If the cost model were perfect, this might not be necessary; but we
1811 // need to make sure that we don't get stuck vectorizing our own
1813 bool HasNontrivialInsts = false;
1815 // The node weights represent the cost savings associated with
1816 // fusing the pair of instructions.
1817 for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(),
1818 E = PrunedDAG.end(); S != E; ++S) {
1819 if (!isa<ShuffleVectorInst>(S->first) &&
1820 !isa<InsertElementInst>(S->first) &&
1821 !isa<ExtractElementInst>(S->first))
1822 HasNontrivialInsts = true;
1824 bool FlipOrder = false;
1826 if (getDepthFactor(S->first)) {
1827 int ESContrib = CandidatePairCostSavings.find(*S)->second;
1828 DEBUG(if (DebugPairSelection) dbgs() << "\tweight {"
1829 << *S->first << " <-> " << *S->second << "} = " <<
1831 EffSize += ESContrib;
1834 // The edge weights contribute in a negative sense: they represent
1835 // the cost of shuffles.
1836 DenseMap<ValuePair, std::vector<ValuePair> >::iterator SS =
1837 ConnectedPairDeps.find(*S);
1838 if (SS != ConnectedPairDeps.end()) {
1839 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
1840 for (std::vector<ValuePair>::iterator T = SS->second.begin(),
1841 TE = SS->second.end(); T != TE; ++T) {
1843 if (!PrunedDAG.count(Q.second))
1845 DenseMap<VPPair, unsigned>::iterator R =
1846 PairConnectionTypes.find(VPPair(Q.second, Q.first));
1847 assert(R != PairConnectionTypes.end() &&
1848 "Cannot find pair connection type");
1849 if (R->second == PairConnectionDirect)
1851 else if (R->second == PairConnectionSwap)
1855 // If there are more swaps than direct connections, then
1856 // the pair order will be flipped during fusion. So the real
1857 // number of swaps is the minimum number.
1858 FlipOrder = !FixedOrderPairs.count(*S) &&
1859 ((NumDepsSwap > NumDepsDirect) ||
1860 FixedOrderPairs.count(ValuePair(S->second, S->first)));
1862 for (std::vector<ValuePair>::iterator T = SS->second.begin(),
1863 TE = SS->second.end(); T != TE; ++T) {
1865 if (!PrunedDAG.count(Q.second))
1867 DenseMap<VPPair, unsigned>::iterator R =
1868 PairConnectionTypes.find(VPPair(Q.second, Q.first));
1869 assert(R != PairConnectionTypes.end() &&
1870 "Cannot find pair connection type");
1871 Type *Ty1 = Q.second.first->getType(),
1872 *Ty2 = Q.second.second->getType();
1873 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1874 if ((R->second == PairConnectionDirect && FlipOrder) ||
1875 (R->second == PairConnectionSwap && !FlipOrder) ||
1876 R->second == PairConnectionSplat) {
1877 int ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1880 if (VTy->getVectorNumElements() == 2) {
1881 if (R->second == PairConnectionSplat)
1882 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1883 TargetTransformInfo::SK_Broadcast, VTy));
1885 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1886 TargetTransformInfo::SK_Reverse, VTy));
1889 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1890 *Q.second.first << " <-> " << *Q.second.second <<
1892 *S->first << " <-> " << *S->second << "} = " <<
1894 EffSize -= ESContrib;
1899 // Compute the cost of outgoing edges. We assume that edges outgoing
1900 // to shuffles, inserts or extracts can be merged, and so contribute
1901 // no additional cost.
1902 if (!S->first->getType()->isVoidTy()) {
1903 Type *Ty1 = S->first->getType(),
1904 *Ty2 = S->second->getType();
1905 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1907 bool NeedsExtraction = false;
1908 for (Value::use_iterator I = S->first->use_begin(),
1909 IE = S->first->use_end(); I != IE; ++I) {
1910 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) {
1911 // Shuffle can be folded if it has no other input
1912 if (isa<UndefValue>(SI->getOperand(1)))
1915 if (isa<ExtractElementInst>(*I))
1917 if (PrunedDAGInstrs.count(*I))
1919 NeedsExtraction = true;
1923 if (NeedsExtraction) {
1925 if (Ty1->isVectorTy()) {
1926 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1928 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1929 TargetTransformInfo::SK_ExtractSubvector, VTy, 0, Ty1));
1931 ESContrib = (int) TTI->getVectorInstrCost(
1932 Instruction::ExtractElement, VTy, 0);
1934 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1935 *S->first << "} = " << ESContrib << "\n");
1936 EffSize -= ESContrib;
1939 NeedsExtraction = false;
1940 for (Value::use_iterator I = S->second->use_begin(),
1941 IE = S->second->use_end(); I != IE; ++I) {
1942 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) {
1943 // Shuffle can be folded if it has no other input
1944 if (isa<UndefValue>(SI->getOperand(1)))
1947 if (isa<ExtractElementInst>(*I))
1949 if (PrunedDAGInstrs.count(*I))
1951 NeedsExtraction = true;
1955 if (NeedsExtraction) {
1957 if (Ty2->isVectorTy()) {
1958 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1960 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1961 TargetTransformInfo::SK_ExtractSubvector, VTy,
1962 Ty1->isVectorTy() ? Ty1->getVectorNumElements() : 1, Ty2));
1964 ESContrib = (int) TTI->getVectorInstrCost(
1965 Instruction::ExtractElement, VTy, 1);
1966 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1967 *S->second << "} = " << ESContrib << "\n");
1968 EffSize -= ESContrib;
1972 // Compute the cost of incoming edges.
1973 if (!isa<LoadInst>(S->first) && !isa<StoreInst>(S->first)) {
1974 Instruction *S1 = cast<Instruction>(S->first),
1975 *S2 = cast<Instruction>(S->second);
1976 for (unsigned o = 0; o < S1->getNumOperands(); ++o) {
1977 Value *O1 = S1->getOperand(o), *O2 = S2->getOperand(o);
1979 // Combining constants into vector constants (or small vector
1980 // constants into larger ones are assumed free).
1981 if (isa<Constant>(O1) && isa<Constant>(O2))
1987 ValuePair VP = ValuePair(O1, O2);
1988 ValuePair VPR = ValuePair(O2, O1);
1990 // Internal edges are not handled here.
1991 if (PrunedDAG.count(VP) || PrunedDAG.count(VPR))
1994 Type *Ty1 = O1->getType(),
1995 *Ty2 = O2->getType();
1996 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1998 // Combining vector operations of the same type is also assumed
1999 // folded with other operations.
2001 // If both are insert elements, then both can be widened.
2002 InsertElementInst *IEO1 = dyn_cast<InsertElementInst>(O1),
2003 *IEO2 = dyn_cast<InsertElementInst>(O2);
2004 if (IEO1 && IEO2 && isPureIEChain(IEO1) && isPureIEChain(IEO2))
2006 // If both are extract elements, and both have the same input
2007 // type, then they can be replaced with a shuffle
2008 ExtractElementInst *EIO1 = dyn_cast<ExtractElementInst>(O1),
2009 *EIO2 = dyn_cast<ExtractElementInst>(O2);
2011 EIO1->getOperand(0)->getType() ==
2012 EIO2->getOperand(0)->getType())
2014 // If both are a shuffle with equal operand types and only two
2015 // unqiue operands, then they can be replaced with a single
2017 ShuffleVectorInst *SIO1 = dyn_cast<ShuffleVectorInst>(O1),
2018 *SIO2 = dyn_cast<ShuffleVectorInst>(O2);
2020 SIO1->getOperand(0)->getType() ==
2021 SIO2->getOperand(0)->getType()) {
2022 SmallSet<Value *, 4> SIOps;
2023 SIOps.insert(SIO1->getOperand(0));
2024 SIOps.insert(SIO1->getOperand(1));
2025 SIOps.insert(SIO2->getOperand(0));
2026 SIOps.insert(SIO2->getOperand(1));
2027 if (SIOps.size() <= 2)
2033 // This pair has already been formed.
2034 if (IncomingPairs.count(VP)) {
2036 } else if (IncomingPairs.count(VPR)) {
2037 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2040 if (VTy->getVectorNumElements() == 2)
2041 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
2042 TargetTransformInfo::SK_Reverse, VTy));
2043 } else if (!Ty1->isVectorTy() && !Ty2->isVectorTy()) {
2044 ESContrib = (int) TTI->getVectorInstrCost(
2045 Instruction::InsertElement, VTy, 0);
2046 ESContrib += (int) TTI->getVectorInstrCost(
2047 Instruction::InsertElement, VTy, 1);
2048 } else if (!Ty1->isVectorTy()) {
2049 // O1 needs to be inserted into a vector of size O2, and then
2050 // both need to be shuffled together.
2051 ESContrib = (int) TTI->getVectorInstrCost(
2052 Instruction::InsertElement, Ty2, 0);
2053 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2055 } else if (!Ty2->isVectorTy()) {
2056 // O2 needs to be inserted into a vector of size O1, and then
2057 // both need to be shuffled together.
2058 ESContrib = (int) TTI->getVectorInstrCost(
2059 Instruction::InsertElement, Ty1, 0);
2060 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2063 Type *TyBig = Ty1, *TySmall = Ty2;
2064 if (Ty2->getVectorNumElements() > Ty1->getVectorNumElements())
2065 std::swap(TyBig, TySmall);
2067 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2069 if (TyBig != TySmall)
2070 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2074 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {"
2075 << *O1 << " <-> " << *O2 << "} = " <<
2077 EffSize -= ESContrib;
2078 IncomingPairs.insert(VP);
2083 if (!HasNontrivialInsts) {
2084 DEBUG(if (DebugPairSelection) dbgs() <<
2085 "\tNo non-trivial instructions in DAG;"
2086 " override to zero effective size\n");
2090 for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(),
2091 E = PrunedDAG.end(); S != E; ++S)
2092 EffSize += (int) getDepthFactor(S->first);
2095 DEBUG(if (DebugPairSelection)
2096 dbgs() << "BBV: found pruned DAG for pair {"
2097 << *IJ.first << " <-> " << *IJ.second << "} of depth " <<
2098 MaxDepth << " and size " << PrunedDAG.size() <<
2099 " (effective size: " << EffSize << ")\n");
2100 if (((TTI && !UseChainDepthWithTI) ||
2101 MaxDepth >= Config.ReqChainDepth) &&
2102 EffSize > 0 && EffSize > BestEffSize) {
2103 BestMaxDepth = MaxDepth;
2104 BestEffSize = EffSize;
2105 BestDAG = PrunedDAG;
2110 // Given the list of candidate pairs, this function selects those
2111 // that will be fused into vector instructions.
2112 void BBVectorize::choosePairs(
2113 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
2114 DenseSet<ValuePair> &CandidatePairsSet,
2115 DenseMap<ValuePair, int> &CandidatePairCostSavings,
2116 std::vector<Value *> &PairableInsts,
2117 DenseSet<ValuePair> &FixedOrderPairs,
2118 DenseMap<VPPair, unsigned> &PairConnectionTypes,
2119 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
2120 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
2121 DenseSet<ValuePair> &PairableInstUsers,
2122 DenseMap<Value *, Value *>& ChosenPairs) {
2123 bool UseCycleCheck =
2124 CandidatePairsSet.size() <= Config.MaxCandPairsForCycleCheck;
2126 DenseMap<Value *, std::vector<Value *> > CandidatePairs2;
2127 for (DenseSet<ValuePair>::iterator I = CandidatePairsSet.begin(),
2128 E = CandidatePairsSet.end(); I != E; ++I) {
2129 std::vector<Value *> &JJ = CandidatePairs2[I->second];
2130 if (JJ.empty()) JJ.reserve(32);
2131 JJ.push_back(I->first);
2134 DenseMap<ValuePair, std::vector<ValuePair> > PairableInstUserMap;
2135 DenseSet<VPPair> PairableInstUserPairSet;
2136 for (std::vector<Value *>::iterator I = PairableInsts.begin(),
2137 E = PairableInsts.end(); I != E; ++I) {
2138 // The number of possible pairings for this variable:
2139 size_t NumChoices = CandidatePairs.lookup(*I).size();
2140 if (!NumChoices) continue;
2142 std::vector<Value *> &JJ = CandidatePairs[*I];
2144 // The best pair to choose and its dag:
2145 size_t BestMaxDepth = 0;
2146 int BestEffSize = 0;
2147 DenseSet<ValuePair> BestDAG;
2148 findBestDAGFor(CandidatePairs, CandidatePairsSet,
2149 CandidatePairCostSavings,
2150 PairableInsts, FixedOrderPairs, PairConnectionTypes,
2151 ConnectedPairs, ConnectedPairDeps,
2152 PairableInstUsers, PairableInstUserMap,
2153 PairableInstUserPairSet, ChosenPairs,
2154 BestDAG, BestMaxDepth, BestEffSize, *I, JJ,
2157 if (BestDAG.empty())
2160 // A dag has been chosen (or not) at this point. If no dag was
2161 // chosen, then this instruction, I, cannot be paired (and is no longer
2164 DEBUG(dbgs() << "BBV: selected pairs in the best DAG for: "
2165 << *cast<Instruction>(*I) << "\n");
2167 for (DenseSet<ValuePair>::iterator S = BestDAG.begin(),
2168 SE2 = BestDAG.end(); S != SE2; ++S) {
2169 // Insert the members of this dag into the list of chosen pairs.
2170 ChosenPairs.insert(ValuePair(S->first, S->second));
2171 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " <<
2172 *S->second << "\n");
2174 // Remove all candidate pairs that have values in the chosen dag.
2175 std::vector<Value *> &KK = CandidatePairs[S->first];
2176 for (std::vector<Value *>::iterator K = KK.begin(), KE = KK.end();
2178 if (*K == S->second)
2181 CandidatePairsSet.erase(ValuePair(S->first, *K));
2184 std::vector<Value *> &LL = CandidatePairs2[S->second];
2185 for (std::vector<Value *>::iterator L = LL.begin(), LE = LL.end();
2190 CandidatePairsSet.erase(ValuePair(*L, S->second));
2193 std::vector<Value *> &MM = CandidatePairs[S->second];
2194 for (std::vector<Value *>::iterator M = MM.begin(), ME = MM.end();
2196 assert(*M != S->first && "Flipped pair in candidate list?");
2197 CandidatePairsSet.erase(ValuePair(S->second, *M));
2200 std::vector<Value *> &NN = CandidatePairs2[S->first];
2201 for (std::vector<Value *>::iterator N = NN.begin(), NE = NN.end();
2203 assert(*N != S->second && "Flipped pair in candidate list?");
2204 CandidatePairsSet.erase(ValuePair(*N, S->first));
2209 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n");
2212 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o,
2217 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) +
2218 (n > 0 ? "." + utostr(n) : "")).str();
2221 // Returns the value that is to be used as the pointer input to the vector
2222 // instruction that fuses I with J.
2223 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context,
2224 Instruction *I, Instruction *J, unsigned o) {
2226 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
2227 int64_t OffsetInElmts;
2229 // Note: the analysis might fail here, that is why the pair order has
2230 // been precomputed (OffsetInElmts must be unused here).
2231 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
2232 IAddressSpace, JAddressSpace,
2233 OffsetInElmts, false);
2235 // The pointer value is taken to be the one with the lowest offset.
2238 Type *ArgTypeI = IPtr->getType()->getPointerElementType();
2239 Type *ArgTypeJ = JPtr->getType()->getPointerElementType();
2240 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2242 = PointerType::get(VArgType,
2243 IPtr->getType()->getPointerAddressSpace());
2244 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o),
2245 /* insert before */ I);
2248 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J,
2249 unsigned MaskOffset, unsigned NumInElem,
2250 unsigned NumInElem1, unsigned IdxOffset,
2251 std::vector<Constant*> &Mask) {
2252 unsigned NumElem1 = J->getType()->getVectorNumElements();
2253 for (unsigned v = 0; v < NumElem1; ++v) {
2254 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v);
2256 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context));
2258 unsigned mm = m + (int) IdxOffset;
2259 if (m >= (int) NumInElem1)
2260 mm += (int) NumInElem;
2262 Mask[v+MaskOffset] =
2263 ConstantInt::get(Type::getInt32Ty(Context), mm);
2268 // Returns the value that is to be used as the vector-shuffle mask to the
2269 // vector instruction that fuses I with J.
2270 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context,
2271 Instruction *I, Instruction *J) {
2272 // This is the shuffle mask. We need to append the second
2273 // mask to the first, and the numbers need to be adjusted.
2275 Type *ArgTypeI = I->getType();
2276 Type *ArgTypeJ = J->getType();
2277 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2279 unsigned NumElemI = ArgTypeI->getVectorNumElements();
2281 // Get the total number of elements in the fused vector type.
2282 // By definition, this must equal the number of elements in
2284 unsigned NumElem = VArgType->getVectorNumElements();
2285 std::vector<Constant*> Mask(NumElem);
2287 Type *OpTypeI = I->getOperand(0)->getType();
2288 unsigned NumInElemI = OpTypeI->getVectorNumElements();
2289 Type *OpTypeJ = J->getOperand(0)->getType();
2290 unsigned NumInElemJ = OpTypeJ->getVectorNumElements();
2292 // The fused vector will be:
2293 // -----------------------------------------------------
2294 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ |
2295 // -----------------------------------------------------
2296 // from which we'll extract NumElem total elements (where the first NumElemI
2297 // of them come from the mask in I and the remainder come from the mask
2300 // For the mask from the first pair...
2301 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI,
2304 // For the mask from the second pair...
2305 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ,
2308 return ConstantVector::get(Mask);
2311 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I,
2312 Instruction *J, unsigned o, Value *&LOp,
2314 Type *ArgTypeL, Type *ArgTypeH,
2315 bool IBeforeJ, unsigned IdxOff) {
2316 bool ExpandedIEChain = false;
2317 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) {
2318 // If we have a pure insertelement chain, then this can be rewritten
2319 // into a chain that directly builds the larger type.
2320 if (isPureIEChain(LIE)) {
2321 SmallVector<Value *, 8> VectElemts(numElemL,
2322 UndefValue::get(ArgTypeL->getScalarType()));
2323 InsertElementInst *LIENext = LIE;
2326 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue();
2327 VectElemts[Idx] = LIENext->getOperand(1);
2329 dyn_cast<InsertElementInst>(LIENext->getOperand(0))));
2332 Value *LIEPrev = UndefValue::get(ArgTypeH);
2333 for (unsigned i = 0; i < numElemL; ++i) {
2334 if (isa<UndefValue>(VectElemts[i])) continue;
2335 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i],
2336 ConstantInt::get(Type::getInt32Ty(Context),
2338 getReplacementName(IBeforeJ ? I : J,
2340 LIENext->insertBefore(IBeforeJ ? J : I);
2344 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH);
2345 ExpandedIEChain = true;
2349 return ExpandedIEChain;
2352 static unsigned getNumScalarElements(Type *Ty) {
2353 if (VectorType *VecTy = dyn_cast<VectorType>(Ty))
2354 return VecTy->getNumElements();
2358 // Returns the value to be used as the specified operand of the vector
2359 // instruction that fuses I with J.
2360 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I,
2361 Instruction *J, unsigned o, bool IBeforeJ) {
2362 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2363 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1);
2365 // Compute the fused vector type for this operand
2366 Type *ArgTypeI = I->getOperand(o)->getType();
2367 Type *ArgTypeJ = J->getOperand(o)->getType();
2368 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2370 Instruction *L = I, *H = J;
2371 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ;
2373 unsigned numElemL = getNumScalarElements(ArgTypeL);
2374 unsigned numElemH = getNumScalarElements(ArgTypeH);
2376 Value *LOp = L->getOperand(o);
2377 Value *HOp = H->getOperand(o);
2378 unsigned numElem = VArgType->getNumElements();
2380 // First, we check if we can reuse the "original" vector outputs (if these
2381 // exist). We might need a shuffle.
2382 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp);
2383 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp);
2384 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp);
2385 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp);
2387 // FIXME: If we're fusing shuffle instructions, then we can't apply this
2388 // optimization. The input vectors to the shuffle might be a different
2389 // length from the shuffle outputs. Unfortunately, the replacement
2390 // shuffle mask has already been formed, and the mask entries are sensitive
2391 // to the sizes of the inputs.
2392 bool IsSizeChangeShuffle =
2393 isa<ShuffleVectorInst>(L) &&
2394 (LOp->getType() != L->getType() || HOp->getType() != H->getType());
2396 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) {
2397 // We can have at most two unique vector inputs.
2398 bool CanUseInputs = true;
2401 I1 = LEE->getOperand(0);
2403 I1 = LSV->getOperand(0);
2404 I2 = LSV->getOperand(1);
2405 if (I2 == I1 || isa<UndefValue>(I2))
2410 Value *I3 = HEE->getOperand(0);
2411 if (!I2 && I3 != I1)
2413 else if (I3 != I1 && I3 != I2)
2414 CanUseInputs = false;
2416 Value *I3 = HSV->getOperand(0);
2417 if (!I2 && I3 != I1)
2419 else if (I3 != I1 && I3 != I2)
2420 CanUseInputs = false;
2423 Value *I4 = HSV->getOperand(1);
2424 if (!isa<UndefValue>(I4)) {
2425 if (!I2 && I4 != I1)
2427 else if (I4 != I1 && I4 != I2)
2428 CanUseInputs = false;
2435 cast<Instruction>(LOp)->getOperand(0)->getType()
2436 ->getVectorNumElements();
2439 cast<Instruction>(HOp)->getOperand(0)->getType()
2440 ->getVectorNumElements();
2442 // We have one or two input vectors. We need to map each index of the
2443 // operands to the index of the original vector.
2444 SmallVector<std::pair<int, int>, 8> II(numElem);
2445 for (unsigned i = 0; i < numElemL; ++i) {
2449 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue();
2450 INum = LEE->getOperand(0) == I1 ? 0 : 1;
2452 Idx = LSV->getMaskValue(i);
2453 if (Idx < (int) LOpElem) {
2454 INum = LSV->getOperand(0) == I1 ? 0 : 1;
2457 INum = LSV->getOperand(1) == I1 ? 0 : 1;
2461 II[i] = std::pair<int, int>(Idx, INum);
2463 for (unsigned i = 0; i < numElemH; ++i) {
2467 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue();
2468 INum = HEE->getOperand(0) == I1 ? 0 : 1;
2470 Idx = HSV->getMaskValue(i);
2471 if (Idx < (int) HOpElem) {
2472 INum = HSV->getOperand(0) == I1 ? 0 : 1;
2475 INum = HSV->getOperand(1) == I1 ? 0 : 1;
2479 II[i + numElemL] = std::pair<int, int>(Idx, INum);
2482 // We now have an array which tells us from which index of which
2483 // input vector each element of the operand comes.
2484 VectorType *I1T = cast<VectorType>(I1->getType());
2485 unsigned I1Elem = I1T->getNumElements();
2488 // In this case there is only one underlying vector input. Check for
2489 // the trivial case where we can use the input directly.
2490 if (I1Elem == numElem) {
2491 bool ElemInOrder = true;
2492 for (unsigned i = 0; i < numElem; ++i) {
2493 if (II[i].first != (int) i && II[i].first != -1) {
2494 ElemInOrder = false;
2503 // A shuffle is needed.
2504 std::vector<Constant *> Mask(numElem);
2505 for (unsigned i = 0; i < numElem; ++i) {
2506 int Idx = II[i].first;
2508 Mask[i] = UndefValue::get(Type::getInt32Ty(Context));
2510 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2514 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2515 ConstantVector::get(Mask),
2516 getReplacementName(IBeforeJ ? I : J,
2518 S->insertBefore(IBeforeJ ? J : I);
2522 VectorType *I2T = cast<VectorType>(I2->getType());
2523 unsigned I2Elem = I2T->getNumElements();
2525 // This input comes from two distinct vectors. The first step is to
2526 // make sure that both vectors are the same length. If not, the
2527 // smaller one will need to grow before they can be shuffled together.
2528 if (I1Elem < I2Elem) {
2529 std::vector<Constant *> Mask(I2Elem);
2531 for (; v < I1Elem; ++v)
2532 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2533 for (; v < I2Elem; ++v)
2534 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2536 Instruction *NewI1 =
2537 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2538 ConstantVector::get(Mask),
2539 getReplacementName(IBeforeJ ? I : J,
2541 NewI1->insertBefore(IBeforeJ ? J : I);
2545 } else if (I1Elem > I2Elem) {
2546 std::vector<Constant *> Mask(I1Elem);
2548 for (; v < I2Elem; ++v)
2549 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2550 for (; v < I1Elem; ++v)
2551 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2553 Instruction *NewI2 =
2554 new ShuffleVectorInst(I2, UndefValue::get(I2T),
2555 ConstantVector::get(Mask),
2556 getReplacementName(IBeforeJ ? I : J,
2558 NewI2->insertBefore(IBeforeJ ? J : I);
2564 // Now that both I1 and I2 are the same length we can shuffle them
2565 // together (and use the result).
2566 std::vector<Constant *> Mask(numElem);
2567 for (unsigned v = 0; v < numElem; ++v) {
2568 if (II[v].first == -1) {
2569 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2571 int Idx = II[v].first + II[v].second * I1Elem;
2572 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2576 Instruction *NewOp =
2577 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask),
2578 getReplacementName(IBeforeJ ? I : J, true, o));
2579 NewOp->insertBefore(IBeforeJ ? J : I);
2584 Type *ArgType = ArgTypeL;
2585 if (numElemL < numElemH) {
2586 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH,
2587 ArgTypeL, VArgType, IBeforeJ, 1)) {
2588 // This is another short-circuit case: we're combining a scalar into
2589 // a vector that is formed by an IE chain. We've just expanded the IE
2590 // chain, now insert the scalar and we're done.
2592 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0,
2593 getReplacementName(IBeforeJ ? I : J, true, o));
2594 S->insertBefore(IBeforeJ ? J : I);
2596 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL,
2597 ArgTypeH, IBeforeJ)) {
2598 // The two vector inputs to the shuffle must be the same length,
2599 // so extend the smaller vector to be the same length as the larger one.
2603 std::vector<Constant *> Mask(numElemH);
2605 for (; v < numElemL; ++v)
2606 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2607 for (; v < numElemH; ++v)
2608 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2610 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL),
2611 ConstantVector::get(Mask),
2612 getReplacementName(IBeforeJ ? I : J,
2615 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0,
2616 getReplacementName(IBeforeJ ? I : J,
2620 NLOp->insertBefore(IBeforeJ ? J : I);
2625 } else if (numElemL > numElemH) {
2626 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL,
2627 ArgTypeH, VArgType, IBeforeJ)) {
2629 InsertElementInst::Create(LOp, HOp,
2630 ConstantInt::get(Type::getInt32Ty(Context),
2632 getReplacementName(IBeforeJ ? I : J,
2634 S->insertBefore(IBeforeJ ? J : I);
2636 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH,
2637 ArgTypeL, IBeforeJ)) {
2640 std::vector<Constant *> Mask(numElemL);
2642 for (; v < numElemH; ++v)
2643 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2644 for (; v < numElemL; ++v)
2645 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2647 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH),
2648 ConstantVector::get(Mask),
2649 getReplacementName(IBeforeJ ? I : J,
2652 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0,
2653 getReplacementName(IBeforeJ ? I : J,
2657 NHOp->insertBefore(IBeforeJ ? J : I);
2662 if (ArgType->isVectorTy()) {
2663 unsigned numElem = VArgType->getVectorNumElements();
2664 std::vector<Constant*> Mask(numElem);
2665 for (unsigned v = 0; v < numElem; ++v) {
2667 // If the low vector was expanded, we need to skip the extra
2668 // undefined entries.
2669 if (v >= numElemL && numElemH > numElemL)
2670 Idx += (numElemH - numElemL);
2671 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2674 Instruction *BV = new ShuffleVectorInst(LOp, HOp,
2675 ConstantVector::get(Mask),
2676 getReplacementName(IBeforeJ ? I : J, true, o));
2677 BV->insertBefore(IBeforeJ ? J : I);
2681 Instruction *BV1 = InsertElementInst::Create(
2682 UndefValue::get(VArgType), LOp, CV0,
2683 getReplacementName(IBeforeJ ? I : J,
2685 BV1->insertBefore(IBeforeJ ? J : I);
2686 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1,
2687 getReplacementName(IBeforeJ ? I : J,
2689 BV2->insertBefore(IBeforeJ ? J : I);
2693 // This function creates an array of values that will be used as the inputs
2694 // to the vector instruction that fuses I with J.
2695 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context,
2696 Instruction *I, Instruction *J,
2697 SmallVectorImpl<Value *> &ReplacedOperands,
2699 unsigned NumOperands = I->getNumOperands();
2701 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) {
2702 // Iterate backward so that we look at the store pointer
2703 // first and know whether or not we need to flip the inputs.
2705 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) {
2706 // This is the pointer for a load/store instruction.
2707 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o);
2709 } else if (isa<CallInst>(I)) {
2710 Function *F = cast<CallInst>(I)->getCalledFunction();
2711 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID();
2712 if (o == NumOperands-1) {
2713 BasicBlock &BB = *I->getParent();
2715 Module *M = BB.getParent()->getParent();
2716 Type *ArgTypeI = I->getType();
2717 Type *ArgTypeJ = J->getType();
2718 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2720 ReplacedOperands[o] = Intrinsic::getDeclaration(M, IID, VArgType);
2722 } else if (IID == Intrinsic::powi && o == 1) {
2723 // The second argument of powi is a single integer and we've already
2724 // checked that both arguments are equal. As a result, we just keep
2725 // I's second argument.
2726 ReplacedOperands[o] = I->getOperand(o);
2729 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) {
2730 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J);
2734 ReplacedOperands[o] = getReplacementInput(Context, I, J, o, IBeforeJ);
2738 // This function creates two values that represent the outputs of the
2739 // original I and J instructions. These are generally vector shuffles
2740 // or extracts. In many cases, these will end up being unused and, thus,
2741 // eliminated by later passes.
2742 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
2743 Instruction *J, Instruction *K,
2744 Instruction *&InsertionPt,
2745 Instruction *&K1, Instruction *&K2) {
2746 if (isa<StoreInst>(I)) {
2747 AA->replaceWithNewValue(I, K);
2748 AA->replaceWithNewValue(J, K);
2750 Type *IType = I->getType();
2751 Type *JType = J->getType();
2753 VectorType *VType = getVecTypeForPair(IType, JType);
2754 unsigned numElem = VType->getNumElements();
2756 unsigned numElemI = getNumScalarElements(IType);
2757 unsigned numElemJ = getNumScalarElements(JType);
2759 if (IType->isVectorTy()) {
2760 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI);
2761 for (unsigned v = 0; v < numElemI; ++v) {
2762 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2763 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v);
2766 K1 = new ShuffleVectorInst(K, UndefValue::get(VType),
2767 ConstantVector::get( Mask1),
2768 getReplacementName(K, false, 1));
2770 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2771 K1 = ExtractElementInst::Create(K, CV0,
2772 getReplacementName(K, false, 1));
2775 if (JType->isVectorTy()) {
2776 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ);
2777 for (unsigned v = 0; v < numElemJ; ++v) {
2778 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2779 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v);
2782 K2 = new ShuffleVectorInst(K, UndefValue::get(VType),
2783 ConstantVector::get( Mask2),
2784 getReplacementName(K, false, 2));
2786 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1);
2787 K2 = ExtractElementInst::Create(K, CV1,
2788 getReplacementName(K, false, 2));
2792 K2->insertAfter(K1);
2797 // Move all uses of the function I (including pairing-induced uses) after J.
2798 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB,
2799 DenseSet<ValuePair> &LoadMoveSetPairs,
2800 Instruction *I, Instruction *J) {
2801 // Skip to the first instruction past I.
2802 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2804 DenseSet<Value *> Users;
2805 AliasSetTracker WriteSet(*AA);
2806 if (I->mayWriteToMemory()) WriteSet.add(I);
2808 for (; cast<Instruction>(L) != J; ++L)
2809 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs);
2811 assert(cast<Instruction>(L) == J &&
2812 "Tracking has not proceeded far enough to check for dependencies");
2813 // If J is now in the use set of I, then trackUsesOfI will return true
2814 // and we have a dependency cycle (and the fusing operation must abort).
2815 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSetPairs);
2818 // Move all uses of the function I (including pairing-induced uses) after J.
2819 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB,
2820 DenseSet<ValuePair> &LoadMoveSetPairs,
2821 Instruction *&InsertionPt,
2822 Instruction *I, Instruction *J) {
2823 // Skip to the first instruction past I.
2824 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2826 DenseSet<Value *> Users;
2827 AliasSetTracker WriteSet(*AA);
2828 if (I->mayWriteToMemory()) WriteSet.add(I);
2830 for (; cast<Instruction>(L) != J;) {
2831 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs)) {
2832 // Move this instruction
2833 Instruction *InstToMove = L; ++L;
2835 DEBUG(dbgs() << "BBV: moving: " << *InstToMove <<
2836 " to after " << *InsertionPt << "\n");
2837 InstToMove->removeFromParent();
2838 InstToMove->insertAfter(InsertionPt);
2839 InsertionPt = InstToMove;
2846 // Collect all load instruction that are in the move set of a given first
2847 // pair member. These loads depend on the first instruction, I, and so need
2848 // to be moved after J (the second instruction) when the pair is fused.
2849 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB,
2850 DenseMap<Value *, Value *> &ChosenPairs,
2851 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
2852 DenseSet<ValuePair> &LoadMoveSetPairs,
2854 // Skip to the first instruction past I.
2855 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2857 DenseSet<Value *> Users;
2858 AliasSetTracker WriteSet(*AA);
2859 if (I->mayWriteToMemory()) WriteSet.add(I);
2861 // Note: We cannot end the loop when we reach J because J could be moved
2862 // farther down the use chain by another instruction pairing. Also, J
2863 // could be before I if this is an inverted input.
2864 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) {
2865 if (trackUsesOfI(Users, WriteSet, I, L)) {
2866 if (L->mayReadFromMemory()) {
2867 LoadMoveSet[L].push_back(I);
2868 LoadMoveSetPairs.insert(ValuePair(L, I));
2874 // In cases where both load/stores and the computation of their pointers
2875 // are chosen for vectorization, we can end up in a situation where the
2876 // aliasing analysis starts returning different query results as the
2877 // process of fusing instruction pairs continues. Because the algorithm
2878 // relies on finding the same use dags here as were found earlier, we'll
2879 // need to precompute the necessary aliasing information here and then
2880 // manually update it during the fusion process.
2881 void BBVectorize::collectLoadMoveSet(BasicBlock &BB,
2882 std::vector<Value *> &PairableInsts,
2883 DenseMap<Value *, Value *> &ChosenPairs,
2884 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
2885 DenseSet<ValuePair> &LoadMoveSetPairs) {
2886 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
2887 PIE = PairableInsts.end(); PI != PIE; ++PI) {
2888 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI);
2889 if (P == ChosenPairs.end()) continue;
2891 Instruction *I = cast<Instruction>(P->first);
2892 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet,
2893 LoadMoveSetPairs, I);
2897 // When the first instruction in each pair is cloned, it will inherit its
2898 // parent's metadata. This metadata must be combined with that of the other
2899 // instruction in a safe way.
2900 void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) {
2901 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata;
2902 K->getAllMetadataOtherThanDebugLoc(Metadata);
2903 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) {
2904 unsigned Kind = Metadata[i].first;
2905 MDNode *JMD = J->getMetadata(Kind);
2906 MDNode *KMD = Metadata[i].second;
2910 K->setMetadata(Kind, 0); // Remove unknown metadata
2912 case LLVMContext::MD_tbaa:
2913 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD));
2915 case LLVMContext::MD_fpmath:
2916 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD));
2922 // This function fuses the chosen instruction pairs into vector instructions,
2923 // taking care preserve any needed scalar outputs and, then, it reorders the
2924 // remaining instructions as needed (users of the first member of the pair
2925 // need to be moved to after the location of the second member of the pair
2926 // because the vector instruction is inserted in the location of the pair's
2928 void BBVectorize::fuseChosenPairs(BasicBlock &BB,
2929 std::vector<Value *> &PairableInsts,
2930 DenseMap<Value *, Value *> &ChosenPairs,
2931 DenseSet<ValuePair> &FixedOrderPairs,
2932 DenseMap<VPPair, unsigned> &PairConnectionTypes,
2933 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
2934 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps) {
2935 LLVMContext& Context = BB.getContext();
2937 // During the vectorization process, the order of the pairs to be fused
2938 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs
2939 // list. After a pair is fused, the flipped pair is removed from the list.
2940 DenseSet<ValuePair> FlippedPairs;
2941 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(),
2942 E = ChosenPairs.end(); P != E; ++P)
2943 FlippedPairs.insert(ValuePair(P->second, P->first));
2944 for (DenseSet<ValuePair>::iterator P = FlippedPairs.begin(),
2945 E = FlippedPairs.end(); P != E; ++P)
2946 ChosenPairs.insert(*P);
2948 DenseMap<Value *, std::vector<Value *> > LoadMoveSet;
2949 DenseSet<ValuePair> LoadMoveSetPairs;
2950 collectLoadMoveSet(BB, PairableInsts, ChosenPairs,
2951 LoadMoveSet, LoadMoveSetPairs);
2953 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n");
2955 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) {
2956 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI);
2957 if (P == ChosenPairs.end()) {
2962 if (getDepthFactor(P->first) == 0) {
2963 // These instructions are not really fused, but are tracked as though
2964 // they are. Any case in which it would be interesting to fuse them
2965 // will be taken care of by InstCombine.
2971 Instruction *I = cast<Instruction>(P->first),
2972 *J = cast<Instruction>(P->second);
2974 DEBUG(dbgs() << "BBV: fusing: " << *I <<
2975 " <-> " << *J << "\n");
2977 // Remove the pair and flipped pair from the list.
2978 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second);
2979 assert(FP != ChosenPairs.end() && "Flipped pair not found in list");
2980 ChosenPairs.erase(FP);
2981 ChosenPairs.erase(P);
2983 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSetPairs, I, J)) {
2984 DEBUG(dbgs() << "BBV: fusion of: " << *I <<
2986 " aborted because of non-trivial dependency cycle\n");
2992 // If the pair must have the other order, then flip it.
2993 bool FlipPairOrder = FixedOrderPairs.count(ValuePair(J, I));
2994 if (!FlipPairOrder && !FixedOrderPairs.count(ValuePair(I, J))) {
2995 // This pair does not have a fixed order, and so we might want to
2996 // flip it if that will yield fewer shuffles. We count the number
2997 // of dependencies connected via swaps, and those directly connected,
2998 // and flip the order if the number of swaps is greater.
2999 bool OrigOrder = true;
3000 DenseMap<ValuePair, std::vector<ValuePair> >::iterator IJ =
3001 ConnectedPairDeps.find(ValuePair(I, J));
3002 if (IJ == ConnectedPairDeps.end()) {
3003 IJ = ConnectedPairDeps.find(ValuePair(J, I));
3007 if (IJ != ConnectedPairDeps.end()) {
3008 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
3009 for (std::vector<ValuePair>::iterator T = IJ->second.begin(),
3010 TE = IJ->second.end(); T != TE; ++T) {
3011 VPPair Q(IJ->first, *T);
3012 DenseMap<VPPair, unsigned>::iterator R =
3013 PairConnectionTypes.find(VPPair(Q.second, Q.first));
3014 assert(R != PairConnectionTypes.end() &&
3015 "Cannot find pair connection type");
3016 if (R->second == PairConnectionDirect)
3018 else if (R->second == PairConnectionSwap)
3023 std::swap(NumDepsDirect, NumDepsSwap);
3025 if (NumDepsSwap > NumDepsDirect) {
3026 FlipPairOrder = true;
3027 DEBUG(dbgs() << "BBV: reordering pair: " << *I <<
3028 " <-> " << *J << "\n");
3033 Instruction *L = I, *H = J;
3037 // If the pair being fused uses the opposite order from that in the pair
3038 // connection map, then we need to flip the types.
3039 DenseMap<ValuePair, std::vector<ValuePair> >::iterator HL =
3040 ConnectedPairs.find(ValuePair(H, L));
3041 if (HL != ConnectedPairs.end())
3042 for (std::vector<ValuePair>::iterator T = HL->second.begin(),
3043 TE = HL->second.end(); T != TE; ++T) {
3044 VPPair Q(HL->first, *T);
3045 DenseMap<VPPair, unsigned>::iterator R = PairConnectionTypes.find(Q);
3046 assert(R != PairConnectionTypes.end() &&
3047 "Cannot find pair connection type");
3048 if (R->second == PairConnectionDirect)
3049 R->second = PairConnectionSwap;
3050 else if (R->second == PairConnectionSwap)
3051 R->second = PairConnectionDirect;
3054 bool LBeforeH = !FlipPairOrder;
3055 unsigned NumOperands = I->getNumOperands();
3056 SmallVector<Value *, 3> ReplacedOperands(NumOperands);
3057 getReplacementInputsForPair(Context, L, H, ReplacedOperands,
3060 // Make a copy of the original operation, change its type to the vector
3061 // type and replace its operands with the vector operands.
3062 Instruction *K = L->clone();
3065 else if (H->hasName())
3068 if (!isa<StoreInst>(K))
3069 K->mutateType(getVecTypeForPair(L->getType(), H->getType()));
3071 combineMetadata(K, H);
3072 K->intersectOptionalDataWith(H);
3074 for (unsigned o = 0; o < NumOperands; ++o)
3075 K->setOperand(o, ReplacedOperands[o]);
3079 // Instruction insertion point:
3080 Instruction *InsertionPt = K;
3081 Instruction *K1 = 0, *K2 = 0;
3082 replaceOutputsOfPair(Context, L, H, K, InsertionPt, K1, K2);
3084 // The use dag of the first original instruction must be moved to after
3085 // the location of the second instruction. The entire use dag of the
3086 // first instruction is disjoint from the input dag of the second
3087 // (by definition), and so commutes with it.
3089 moveUsesOfIAfterJ(BB, LoadMoveSetPairs, InsertionPt, I, J);
3091 if (!isa<StoreInst>(I)) {
3092 L->replaceAllUsesWith(K1);
3093 H->replaceAllUsesWith(K2);
3094 AA->replaceWithNewValue(L, K1);
3095 AA->replaceWithNewValue(H, K2);
3098 // Instructions that may read from memory may be in the load move set.
3099 // Once an instruction is fused, we no longer need its move set, and so
3100 // the values of the map never need to be updated. However, when a load
3101 // is fused, we need to merge the entries from both instructions in the
3102 // pair in case those instructions were in the move set of some other
3103 // yet-to-be-fused pair. The loads in question are the keys of the map.
3104 if (I->mayReadFromMemory()) {
3105 std::vector<ValuePair> NewSetMembers;
3106 DenseMap<Value *, std::vector<Value *> >::iterator II =
3107 LoadMoveSet.find(I);
3108 if (II != LoadMoveSet.end())
3109 for (std::vector<Value *>::iterator N = II->second.begin(),
3110 NE = II->second.end(); N != NE; ++N)
3111 NewSetMembers.push_back(ValuePair(K, *N));
3112 DenseMap<Value *, std::vector<Value *> >::iterator JJ =
3113 LoadMoveSet.find(J);
3114 if (JJ != LoadMoveSet.end())
3115 for (std::vector<Value *>::iterator N = JJ->second.begin(),
3116 NE = JJ->second.end(); N != NE; ++N)
3117 NewSetMembers.push_back(ValuePair(K, *N));
3118 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(),
3119 AE = NewSetMembers.end(); A != AE; ++A) {
3120 LoadMoveSet[A->first].push_back(A->second);
3121 LoadMoveSetPairs.insert(*A);
3125 // Before removing I, set the iterator to the next instruction.
3126 PI = llvm::next(BasicBlock::iterator(I));
3127 if (cast<Instruction>(PI) == J)
3132 I->eraseFromParent();
3133 J->eraseFromParent();
3135 DEBUG(if (PrintAfterEveryPair) dbgs() << "BBV: block is now: \n" <<
3139 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n");
3143 char BBVectorize::ID = 0;
3144 static const char bb_vectorize_name[] = "Basic-Block Vectorization";
3145 INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
3146 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
3147 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
3148 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
3149 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3150 INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
3152 BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) {
3153 return new BBVectorize(C);
3157 llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) {
3158 BBVectorize BBVectorizer(P, C);
3159 return BBVectorizer.vectorizeBB(BB);
3162 //===----------------------------------------------------------------------===//
3163 VectorizeConfig::VectorizeConfig() {
3164 VectorBits = ::VectorBits;
3165 VectorizeBools = !::NoBools;
3166 VectorizeInts = !::NoInts;
3167 VectorizeFloats = !::NoFloats;
3168 VectorizePointers = !::NoPointers;
3169 VectorizeCasts = !::NoCasts;
3170 VectorizeMath = !::NoMath;
3171 VectorizeFMA = !::NoFMA;
3172 VectorizeSelect = !::NoSelect;
3173 VectorizeCmp = !::NoCmp;
3174 VectorizeGEP = !::NoGEP;
3175 VectorizeMemOps = !::NoMemOps;
3176 AlignedOnly = ::AlignedOnly;
3177 ReqChainDepth= ::ReqChainDepth;
3178 SearchLimit = ::SearchLimit;
3179 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck;
3180 SplatBreaksChain = ::SplatBreaksChain;
3181 MaxInsts = ::MaxInsts;
3182 MaxPairs = ::MaxPairs;
3183 MaxIter = ::MaxIter;
3184 Pow2LenOnly = ::Pow2LenOnly;
3185 NoMemOpBoost = ::NoMemOpBoost;
3186 FastDep = ::FastDep;