1 //===- BBVectorize.cpp - A Basic-Block Vectorizer -------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements a basic-block vectorization pass. The algorithm was
11 // inspired by that used by the Vienna MAP Vectorizor by Franchetti and Kral,
12 // et al. It works by looking for chains of pairable operations and then
15 //===----------------------------------------------------------------------===//
17 #define BBV_NAME "bb-vectorize"
18 #define DEBUG_TYPE BBV_NAME
19 #include "llvm/Transforms/Vectorize.h"
20 #include "llvm/ADT/DenseMap.h"
21 #include "llvm/ADT/DenseSet.h"
22 #include "llvm/ADT/STLExtras.h"
23 #include "llvm/ADT/SmallSet.h"
24 #include "llvm/ADT/SmallVector.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/ADT/StringExtras.h"
27 #include "llvm/Analysis/AliasAnalysis.h"
28 #include "llvm/Analysis/AliasSetTracker.h"
29 #include "llvm/Analysis/Dominators.h"
30 #include "llvm/Analysis/ScalarEvolution.h"
31 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
32 #include "llvm/Analysis/TargetTransformInfo.h"
33 #include "llvm/Analysis/ValueTracking.h"
34 #include "llvm/IR/Constants.h"
35 #include "llvm/IR/DataLayout.h"
36 #include "llvm/IR/DerivedTypes.h"
37 #include "llvm/IR/Function.h"
38 #include "llvm/IR/Instructions.h"
39 #include "llvm/IR/IntrinsicInst.h"
40 #include "llvm/IR/Intrinsics.h"
41 #include "llvm/IR/LLVMContext.h"
42 #include "llvm/IR/Metadata.h"
43 #include "llvm/IR/Type.h"
44 #include "llvm/Pass.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/ValueHandle.h"
48 #include "llvm/Support/raw_ostream.h"
49 #include "llvm/Transforms/Utils/Local.h"
55 IgnoreTargetInfo("bb-vectorize-ignore-target-info", cl::init(false),
56 cl::Hidden, cl::desc("Ignore target information"));
58 static cl::opt<unsigned>
59 ReqChainDepth("bb-vectorize-req-chain-depth", cl::init(6), cl::Hidden,
60 cl::desc("The required chain depth for vectorization"));
63 UseChainDepthWithTI("bb-vectorize-use-chain-depth", cl::init(false),
64 cl::Hidden, cl::desc("Use the chain depth requirement with"
65 " target information"));
67 static cl::opt<unsigned>
68 SearchLimit("bb-vectorize-search-limit", cl::init(400), cl::Hidden,
69 cl::desc("The maximum search distance for instruction pairs"));
72 SplatBreaksChain("bb-vectorize-splat-breaks-chain", cl::init(false), cl::Hidden,
73 cl::desc("Replicating one element to a pair breaks the chain"));
75 static cl::opt<unsigned>
76 VectorBits("bb-vectorize-vector-bits", cl::init(128), cl::Hidden,
77 cl::desc("The size of the native vector registers"));
79 static cl::opt<unsigned>
80 MaxIter("bb-vectorize-max-iter", cl::init(0), cl::Hidden,
81 cl::desc("The maximum number of pairing iterations"));
84 Pow2LenOnly("bb-vectorize-pow2-len-only", cl::init(false), cl::Hidden,
85 cl::desc("Don't try to form non-2^n-length vectors"));
87 static cl::opt<unsigned>
88 MaxInsts("bb-vectorize-max-instr-per-group", cl::init(500), cl::Hidden,
89 cl::desc("The maximum number of pairable instructions per group"));
91 static cl::opt<unsigned>
92 MaxCandPairsForCycleCheck("bb-vectorize-max-cycle-check-pairs", cl::init(200),
93 cl::Hidden, cl::desc("The maximum number of candidate pairs with which to use"
94 " a full cycle check"));
97 NoBools("bb-vectorize-no-bools", cl::init(false), cl::Hidden,
98 cl::desc("Don't try to vectorize boolean (i1) values"));
101 NoInts("bb-vectorize-no-ints", cl::init(false), cl::Hidden,
102 cl::desc("Don't try to vectorize integer values"));
105 NoFloats("bb-vectorize-no-floats", cl::init(false), cl::Hidden,
106 cl::desc("Don't try to vectorize floating-point values"));
108 // FIXME: This should default to false once pointer vector support works.
110 NoPointers("bb-vectorize-no-pointers", cl::init(/*false*/ true), cl::Hidden,
111 cl::desc("Don't try to vectorize pointer values"));
114 NoCasts("bb-vectorize-no-casts", cl::init(false), cl::Hidden,
115 cl::desc("Don't try to vectorize casting (conversion) operations"));
118 NoMath("bb-vectorize-no-math", cl::init(false), cl::Hidden,
119 cl::desc("Don't try to vectorize floating-point math intrinsics"));
122 NoFMA("bb-vectorize-no-fma", cl::init(false), cl::Hidden,
123 cl::desc("Don't try to vectorize the fused-multiply-add intrinsic"));
126 NoSelect("bb-vectorize-no-select", cl::init(false), cl::Hidden,
127 cl::desc("Don't try to vectorize select instructions"));
130 NoCmp("bb-vectorize-no-cmp", cl::init(false), cl::Hidden,
131 cl::desc("Don't try to vectorize comparison instructions"));
134 NoGEP("bb-vectorize-no-gep", cl::init(false), cl::Hidden,
135 cl::desc("Don't try to vectorize getelementptr instructions"));
138 NoMemOps("bb-vectorize-no-mem-ops", cl::init(false), cl::Hidden,
139 cl::desc("Don't try to vectorize loads and stores"));
142 AlignedOnly("bb-vectorize-aligned-only", cl::init(false), cl::Hidden,
143 cl::desc("Only generate aligned loads and stores"));
146 NoMemOpBoost("bb-vectorize-no-mem-op-boost",
147 cl::init(false), cl::Hidden,
148 cl::desc("Don't boost the chain-depth contribution of loads and stores"));
151 FastDep("bb-vectorize-fast-dep", cl::init(false), cl::Hidden,
152 cl::desc("Use a fast instruction dependency analysis"));
156 DebugInstructionExamination("bb-vectorize-debug-instruction-examination",
157 cl::init(false), cl::Hidden,
158 cl::desc("When debugging is enabled, output information on the"
159 " instruction-examination process"));
161 DebugCandidateSelection("bb-vectorize-debug-candidate-selection",
162 cl::init(false), cl::Hidden,
163 cl::desc("When debugging is enabled, output information on the"
164 " candidate-selection process"));
166 DebugPairSelection("bb-vectorize-debug-pair-selection",
167 cl::init(false), cl::Hidden,
168 cl::desc("When debugging is enabled, output information on the"
169 " pair-selection process"));
171 DebugCycleCheck("bb-vectorize-debug-cycle-check",
172 cl::init(false), cl::Hidden,
173 cl::desc("When debugging is enabled, output information on the"
174 " cycle-checking process"));
177 PrintAfterEveryPair("bb-vectorize-debug-print-after-every-pair",
178 cl::init(false), cl::Hidden,
179 cl::desc("When debugging is enabled, dump the basic block after"
180 " every pair is fused"));
183 STATISTIC(NumFusedOps, "Number of operations fused by bb-vectorize");
186 struct BBVectorize : public BasicBlockPass {
187 static char ID; // Pass identification, replacement for typeid
189 const VectorizeConfig Config;
191 BBVectorize(const VectorizeConfig &C = VectorizeConfig())
192 : BasicBlockPass(ID), Config(C) {
193 initializeBBVectorizePass(*PassRegistry::getPassRegistry());
196 BBVectorize(Pass *P, const VectorizeConfig &C)
197 : BasicBlockPass(ID), Config(C) {
198 AA = &P->getAnalysis<AliasAnalysis>();
199 DT = &P->getAnalysis<DominatorTree>();
200 SE = &P->getAnalysis<ScalarEvolution>();
201 TD = P->getAnalysisIfAvailable<DataLayout>();
202 TTI = IgnoreTargetInfo ? 0 : &P->getAnalysis<TargetTransformInfo>();
205 typedef std::pair<Value *, Value *> ValuePair;
206 typedef std::pair<ValuePair, int> ValuePairWithCost;
207 typedef std::pair<ValuePair, size_t> ValuePairWithDepth;
208 typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair
209 typedef std::pair<VPPair, unsigned> VPPairWithType;
210 typedef std::pair<std::multimap<Value *, Value *>::iterator,
211 std::multimap<Value *, Value *>::iterator> VPIteratorPair;
212 typedef std::pair<std::multimap<ValuePair, ValuePair>::iterator,
213 std::multimap<ValuePair, ValuePair>::iterator>
220 const TargetTransformInfo *TTI;
222 // FIXME: const correct?
224 bool vectorizePairs(BasicBlock &BB, bool NonPow2Len = false);
226 bool getCandidatePairs(BasicBlock &BB,
227 BasicBlock::iterator &Start,
228 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
229 DenseSet<ValuePair> &FixedOrderPairs,
230 DenseMap<ValuePair, int> &CandidatePairCostSavings,
231 std::vector<Value *> &PairableInsts, bool NonPow2Len);
233 // FIXME: The current implementation does not account for pairs that
234 // are connected in multiple ways. For example:
235 // C1 = A1 / A2; C2 = A2 / A1 (which may be both direct and a swap)
236 enum PairConnectionType {
237 PairConnectionDirect,
242 void computeConnectedPairs(DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
243 DenseSet<ValuePair> &CandidatePairsSet,
244 std::vector<Value *> &PairableInsts,
245 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
246 DenseMap<VPPair, unsigned> &PairConnectionTypes);
248 void buildDepMap(BasicBlock &BB,
249 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
250 std::vector<Value *> &PairableInsts,
251 DenseSet<ValuePair> &PairableInstUsers);
253 void choosePairs(DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
254 DenseSet<ValuePair> &CandidatePairsSet,
255 DenseMap<ValuePair, int> &CandidatePairCostSavings,
256 std::vector<Value *> &PairableInsts,
257 DenseSet<ValuePair> &FixedOrderPairs,
258 DenseMap<VPPair, unsigned> &PairConnectionTypes,
259 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
260 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps,
261 DenseSet<ValuePair> &PairableInstUsers,
262 DenseMap<Value *, Value *>& ChosenPairs);
264 void fuseChosenPairs(BasicBlock &BB,
265 std::vector<Value *> &PairableInsts,
266 DenseMap<Value *, Value *>& ChosenPairs,
267 DenseSet<ValuePair> &FixedOrderPairs,
268 DenseMap<VPPair, unsigned> &PairConnectionTypes,
269 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
270 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps);
273 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore);
275 bool areInstsCompatible(Instruction *I, Instruction *J,
276 bool IsSimpleLoadStore, bool NonPow2Len,
277 int &CostSavings, int &FixedOrder);
279 bool trackUsesOfI(DenseSet<Value *> &Users,
280 AliasSetTracker &WriteSet, Instruction *I,
281 Instruction *J, bool UpdateUsers = true,
282 DenseSet<ValuePair> *LoadMoveSetPairs = 0);
284 void computePairsConnectedTo(
285 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
286 DenseSet<ValuePair> &CandidatePairsSet,
287 std::vector<Value *> &PairableInsts,
288 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
289 DenseMap<VPPair, unsigned> &PairConnectionTypes,
292 bool pairsConflict(ValuePair P, ValuePair Q,
293 DenseSet<ValuePair> &PairableInstUsers,
294 std::multimap<ValuePair, ValuePair> *PairableInstUserMap = 0,
295 DenseSet<VPPair> *PairableInstUserPairSet = 0);
297 bool pairWillFormCycle(ValuePair P,
298 std::multimap<ValuePair, ValuePair> &PairableInstUsers,
299 DenseSet<ValuePair> &CurrentPairs);
302 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
303 std::vector<Value *> &PairableInsts,
304 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
305 DenseSet<ValuePair> &PairableInstUsers,
306 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
307 DenseSet<VPPair> &PairableInstUserPairSet,
308 DenseMap<Value *, Value *> &ChosenPairs,
309 DenseMap<ValuePair, size_t> &Tree,
310 DenseSet<ValuePair> &PrunedTree, ValuePair J,
313 void buildInitialTreeFor(
314 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
315 DenseSet<ValuePair> &CandidatePairsSet,
316 std::vector<Value *> &PairableInsts,
317 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
318 DenseSet<ValuePair> &PairableInstUsers,
319 DenseMap<Value *, Value *> &ChosenPairs,
320 DenseMap<ValuePair, size_t> &Tree, ValuePair J);
322 void findBestTreeFor(
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 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
330 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps,
331 DenseSet<ValuePair> &PairableInstUsers,
332 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
333 DenseSet<VPPair> &PairableInstUserPairSet,
334 DenseMap<Value *, Value *> &ChosenPairs,
335 DenseSet<ValuePair> &BestTree, 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, SmallVector<Value *, 3> &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 std::multimap<Value *, Value *> &LoadMoveSet,
370 DenseSet<ValuePair> &LoadMoveSetPairs,
373 void collectLoadMoveSet(BasicBlock &BB,
374 std::vector<Value *> &PairableInsts,
375 DenseMap<Value *, Value *> &ChosenPairs,
376 std::multimap<Value *, Value *> &LoadMoveSet,
377 DenseSet<ValuePair> &LoadMoveSetPairs);
379 bool canMoveUsesOfIAfterJ(BasicBlock &BB,
380 DenseSet<ValuePair> &LoadMoveSetPairs,
381 Instruction *I, Instruction *J);
383 void moveUsesOfIAfterJ(BasicBlock &BB,
384 DenseSet<ValuePair> &LoadMoveSetPairs,
385 Instruction *&InsertionPt,
386 Instruction *I, Instruction *J);
388 void combineMetadata(Instruction *K, const Instruction *J);
390 bool vectorizeBB(BasicBlock &BB) {
391 if (!DT->isReachableFromEntry(&BB)) {
392 DEBUG(dbgs() << "BBV: skipping unreachable " << BB.getName() <<
393 " in " << BB.getParent()->getName() << "\n");
397 DEBUG(if (TTI) dbgs() << "BBV: using target information\n");
399 bool changed = false;
400 // Iterate a sufficient number of times to merge types of size 1 bit,
401 // then 2 bits, then 4, etc. up to half of the target vector width of the
402 // target vector register.
405 (TTI || v <= Config.VectorBits) &&
406 (!Config.MaxIter || n <= Config.MaxIter);
408 DEBUG(dbgs() << "BBV: fusing loop #" << n <<
409 " for " << BB.getName() << " in " <<
410 BB.getParent()->getName() << "...\n");
411 if (vectorizePairs(BB))
417 if (changed && !Pow2LenOnly) {
419 for (; !Config.MaxIter || n <= Config.MaxIter; ++n) {
420 DEBUG(dbgs() << "BBV: fusing for non-2^n-length vectors loop #: " <<
421 n << " for " << BB.getName() << " in " <<
422 BB.getParent()->getName() << "...\n");
423 if (!vectorizePairs(BB, true)) break;
427 DEBUG(dbgs() << "BBV: done!\n");
431 virtual bool runOnBasicBlock(BasicBlock &BB) {
432 AA = &getAnalysis<AliasAnalysis>();
433 DT = &getAnalysis<DominatorTree>();
434 SE = &getAnalysis<ScalarEvolution>();
435 TD = getAnalysisIfAvailable<DataLayout>();
436 TTI = IgnoreTargetInfo ? 0 : &getAnalysis<TargetTransformInfo>();
438 return vectorizeBB(BB);
441 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
442 BasicBlockPass::getAnalysisUsage(AU);
443 AU.addRequired<AliasAnalysis>();
444 AU.addRequired<DominatorTree>();
445 AU.addRequired<ScalarEvolution>();
446 AU.addRequired<TargetTransformInfo>();
447 AU.addPreserved<AliasAnalysis>();
448 AU.addPreserved<DominatorTree>();
449 AU.addPreserved<ScalarEvolution>();
450 AU.setPreservesCFG();
453 static inline VectorType *getVecTypeForPair(Type *ElemTy, Type *Elem2Ty) {
454 assert(ElemTy->getScalarType() == Elem2Ty->getScalarType() &&
455 "Cannot form vector from incompatible scalar types");
456 Type *STy = ElemTy->getScalarType();
459 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) {
460 numElem = VTy->getNumElements();
465 if (VectorType *VTy = dyn_cast<VectorType>(Elem2Ty)) {
466 numElem += VTy->getNumElements();
471 return VectorType::get(STy, numElem);
474 static inline void getInstructionTypes(Instruction *I,
475 Type *&T1, Type *&T2) {
476 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
477 // For stores, it is the value type, not the pointer type that matters
478 // because the value is what will come from a vector register.
480 Value *IVal = SI->getValueOperand();
481 T1 = IVal->getType();
486 if (CastInst *CI = dyn_cast<CastInst>(I))
491 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
492 T2 = SI->getCondition()->getType();
493 } else if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(I)) {
494 T2 = SI->getOperand(0)->getType();
495 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) {
496 T2 = CI->getOperand(0)->getType();
500 // Returns the weight associated with the provided value. A chain of
501 // candidate pairs has a length given by the sum of the weights of its
502 // members (one weight per pair; the weight of each member of the pair
503 // is assumed to be the same). This length is then compared to the
504 // chain-length threshold to determine if a given chain is significant
505 // enough to be vectorized. The length is also used in comparing
506 // candidate chains where longer chains are considered to be better.
507 // Note: when this function returns 0, the resulting instructions are
508 // not actually fused.
509 inline size_t getDepthFactor(Value *V) {
510 // InsertElement and ExtractElement have a depth factor of zero. This is
511 // for two reasons: First, they cannot be usefully fused. Second, because
512 // the pass generates a lot of these, they can confuse the simple metric
513 // used to compare the trees in the next iteration. Thus, giving them a
514 // weight of zero allows the pass to essentially ignore them in
515 // subsequent iterations when looking for vectorization opportunities
516 // while still tracking dependency chains that flow through those
518 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V))
521 // Give a load or store half of the required depth so that load/store
522 // pairs will vectorize.
523 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V)))
524 return Config.ReqChainDepth/2;
529 // Returns the cost of the provided instruction using TTI.
530 // This does not handle loads and stores.
531 unsigned getInstrCost(unsigned Opcode, Type *T1, Type *T2) {
534 case Instruction::GetElementPtr:
535 // We mark this instruction as zero-cost because scalar GEPs are usually
536 // lowered to the intruction addressing mode. At the moment we don't
537 // generate vector GEPs.
539 case Instruction::Br:
540 return TTI->getCFInstrCost(Opcode);
541 case Instruction::PHI:
543 case Instruction::Add:
544 case Instruction::FAdd:
545 case Instruction::Sub:
546 case Instruction::FSub:
547 case Instruction::Mul:
548 case Instruction::FMul:
549 case Instruction::UDiv:
550 case Instruction::SDiv:
551 case Instruction::FDiv:
552 case Instruction::URem:
553 case Instruction::SRem:
554 case Instruction::FRem:
555 case Instruction::Shl:
556 case Instruction::LShr:
557 case Instruction::AShr:
558 case Instruction::And:
559 case Instruction::Or:
560 case Instruction::Xor:
561 return TTI->getArithmeticInstrCost(Opcode, T1);
562 case Instruction::Select:
563 case Instruction::ICmp:
564 case Instruction::FCmp:
565 return TTI->getCmpSelInstrCost(Opcode, T1, T2);
566 case Instruction::ZExt:
567 case Instruction::SExt:
568 case Instruction::FPToUI:
569 case Instruction::FPToSI:
570 case Instruction::FPExt:
571 case Instruction::PtrToInt:
572 case Instruction::IntToPtr:
573 case Instruction::SIToFP:
574 case Instruction::UIToFP:
575 case Instruction::Trunc:
576 case Instruction::FPTrunc:
577 case Instruction::BitCast:
578 case Instruction::ShuffleVector:
579 return TTI->getCastInstrCost(Opcode, T1, T2);
585 // This determines the relative offset of two loads or stores, returning
586 // true if the offset could be determined to be some constant value.
587 // For example, if OffsetInElmts == 1, then J accesses the memory directly
588 // after I; if OffsetInElmts == -1 then I accesses the memory
590 bool getPairPtrInfo(Instruction *I, Instruction *J,
591 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment,
592 unsigned &IAddressSpace, unsigned &JAddressSpace,
593 int64_t &OffsetInElmts, bool ComputeOffset = true) {
595 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
596 LoadInst *LJ = cast<LoadInst>(J);
597 IPtr = LI->getPointerOperand();
598 JPtr = LJ->getPointerOperand();
599 IAlignment = LI->getAlignment();
600 JAlignment = LJ->getAlignment();
601 IAddressSpace = LI->getPointerAddressSpace();
602 JAddressSpace = LJ->getPointerAddressSpace();
604 StoreInst *SI = cast<StoreInst>(I), *SJ = cast<StoreInst>(J);
605 IPtr = SI->getPointerOperand();
606 JPtr = SJ->getPointerOperand();
607 IAlignment = SI->getAlignment();
608 JAlignment = SJ->getAlignment();
609 IAddressSpace = SI->getPointerAddressSpace();
610 JAddressSpace = SJ->getPointerAddressSpace();
616 const SCEV *IPtrSCEV = SE->getSCEV(IPtr);
617 const SCEV *JPtrSCEV = SE->getSCEV(JPtr);
619 // If this is a trivial offset, then we'll get something like
620 // 1*sizeof(type). With target data, which we need anyway, this will get
621 // constant folded into a number.
622 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV);
623 if (const SCEVConstant *ConstOffSCEV =
624 dyn_cast<SCEVConstant>(OffsetSCEV)) {
625 ConstantInt *IntOff = ConstOffSCEV->getValue();
626 int64_t Offset = IntOff->getSExtValue();
628 Type *VTy = cast<PointerType>(IPtr->getType())->getElementType();
629 int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy);
631 Type *VTy2 = cast<PointerType>(JPtr->getType())->getElementType();
632 if (VTy != VTy2 && Offset < 0) {
633 int64_t VTy2TSS = (int64_t) TD->getTypeStoreSize(VTy2);
634 OffsetInElmts = Offset/VTy2TSS;
635 return (abs64(Offset) % VTy2TSS) == 0;
638 OffsetInElmts = Offset/VTyTSS;
639 return (abs64(Offset) % VTyTSS) == 0;
645 // Returns true if the provided CallInst represents an intrinsic that can
647 bool isVectorizableIntrinsic(CallInst* I) {
648 Function *F = I->getCalledFunction();
649 if (!F) return false;
651 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID();
652 if (!IID) return false;
657 case Intrinsic::sqrt:
658 case Intrinsic::powi:
662 case Intrinsic::log2:
663 case Intrinsic::log10:
665 case Intrinsic::exp2:
667 return Config.VectorizeMath;
669 case Intrinsic::fmuladd:
670 return Config.VectorizeFMA;
674 bool isPureIEChain(InsertElementInst *IE) {
675 InsertElementInst *IENext = IE;
677 if (!isa<UndefValue>(IENext->getOperand(0)) &&
678 !isa<InsertElementInst>(IENext->getOperand(0))) {
682 dyn_cast<InsertElementInst>(IENext->getOperand(0))));
688 // This function implements one vectorization iteration on the provided
689 // basic block. It returns true if the block is changed.
690 bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) {
692 BasicBlock::iterator Start = BB.getFirstInsertionPt();
694 std::vector<Value *> AllPairableInsts;
695 DenseMap<Value *, Value *> AllChosenPairs;
696 DenseSet<ValuePair> AllFixedOrderPairs;
697 DenseMap<VPPair, unsigned> AllPairConnectionTypes;
698 std::multimap<ValuePair, ValuePair> AllConnectedPairs, AllConnectedPairDeps;
701 std::vector<Value *> PairableInsts;
702 DenseMap<Value *, std::vector<Value *> > CandidatePairs;
703 DenseSet<ValuePair> FixedOrderPairs;
704 DenseMap<ValuePair, int> CandidatePairCostSavings;
705 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs,
707 CandidatePairCostSavings,
708 PairableInsts, NonPow2Len);
709 if (PairableInsts.empty()) continue;
711 // Build the candidate pair set for faster lookups.
712 DenseSet<ValuePair> CandidatePairsSet;
713 for (DenseMap<Value *, std::vector<Value *> >::iterator I =
714 CandidatePairs.begin(), E = CandidatePairs.end(); I != E; ++I)
715 for (std::vector<Value *>::iterator J = I->second.begin(),
716 JE = I->second.end(); J != JE; ++J)
717 CandidatePairsSet.insert(ValuePair(I->first, *J));
719 // Now we have a map of all of the pairable instructions and we need to
720 // select the best possible pairing. A good pairing is one such that the
721 // users of the pair are also paired. This defines a (directed) forest
722 // over the pairs such that two pairs are connected iff the second pair
725 // Note that it only matters that both members of the second pair use some
726 // element of the first pair (to allow for splatting).
728 std::multimap<ValuePair, ValuePair> ConnectedPairs, ConnectedPairDeps;
729 DenseMap<VPPair, unsigned> PairConnectionTypes;
730 computeConnectedPairs(CandidatePairs, CandidatePairsSet,
731 PairableInsts, ConnectedPairs, PairConnectionTypes);
732 if (ConnectedPairs.empty()) continue;
734 for (std::multimap<ValuePair, ValuePair>::iterator
735 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
737 ConnectedPairDeps.insert(VPPair(I->second, I->first));
740 // Build the pairable-instruction dependency map
741 DenseSet<ValuePair> PairableInstUsers;
742 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers);
744 // There is now a graph of the connected pairs. For each variable, pick
745 // the pairing with the largest tree meeting the depth requirement on at
746 // least one branch. Then select all pairings that are part of that tree
747 // and remove them from the list of available pairings and pairable
750 DenseMap<Value *, Value *> ChosenPairs;
751 choosePairs(CandidatePairs, CandidatePairsSet,
752 CandidatePairCostSavings,
753 PairableInsts, FixedOrderPairs, PairConnectionTypes,
754 ConnectedPairs, ConnectedPairDeps,
755 PairableInstUsers, ChosenPairs);
757 if (ChosenPairs.empty()) continue;
758 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(),
759 PairableInsts.end());
760 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end());
762 // Only for the chosen pairs, propagate information on fixed-order pairs,
763 // pair connections, and their types to the data structures used by the
764 // pair fusion procedures.
765 for (DenseMap<Value *, Value *>::iterator I = ChosenPairs.begin(),
766 IE = ChosenPairs.end(); I != IE; ++I) {
767 if (FixedOrderPairs.count(*I))
768 AllFixedOrderPairs.insert(*I);
769 else if (FixedOrderPairs.count(ValuePair(I->second, I->first)))
770 AllFixedOrderPairs.insert(ValuePair(I->second, I->first));
772 for (DenseMap<Value *, Value *>::iterator J = ChosenPairs.begin();
774 DenseMap<VPPair, unsigned>::iterator K =
775 PairConnectionTypes.find(VPPair(*I, *J));
776 if (K != PairConnectionTypes.end()) {
777 AllPairConnectionTypes.insert(*K);
779 K = PairConnectionTypes.find(VPPair(*J, *I));
780 if (K != PairConnectionTypes.end())
781 AllPairConnectionTypes.insert(*K);
786 for (std::multimap<ValuePair, ValuePair>::iterator
787 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
789 if (AllPairConnectionTypes.count(*I)) {
790 AllConnectedPairs.insert(*I);
791 AllConnectedPairDeps.insert(VPPair(I->second, I->first));
794 } while (ShouldContinue);
796 if (AllChosenPairs.empty()) return false;
797 NumFusedOps += AllChosenPairs.size();
799 // A set of pairs has now been selected. It is now necessary to replace the
800 // paired instructions with vector instructions. For this procedure each
801 // operand must be replaced with a vector operand. This vector is formed
802 // by using build_vector on the old operands. The replaced values are then
803 // replaced with a vector_extract on the result. Subsequent optimization
804 // passes should coalesce the build/extract combinations.
806 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs, AllFixedOrderPairs,
807 AllPairConnectionTypes,
808 AllConnectedPairs, AllConnectedPairDeps);
810 // It is important to cleanup here so that future iterations of this
811 // function have less work to do.
812 (void) SimplifyInstructionsInBlock(&BB, TD, AA->getTargetLibraryInfo());
816 // This function returns true if the provided instruction is capable of being
817 // fused into a vector instruction. This determination is based only on the
818 // type and other attributes of the instruction.
819 bool BBVectorize::isInstVectorizable(Instruction *I,
820 bool &IsSimpleLoadStore) {
821 IsSimpleLoadStore = false;
823 if (CallInst *C = dyn_cast<CallInst>(I)) {
824 if (!isVectorizableIntrinsic(C))
826 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) {
827 // Vectorize simple loads if possbile:
828 IsSimpleLoadStore = L->isSimple();
829 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
831 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) {
832 // Vectorize simple stores if possbile:
833 IsSimpleLoadStore = S->isSimple();
834 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
836 } else if (CastInst *C = dyn_cast<CastInst>(I)) {
837 // We can vectorize casts, but not casts of pointer types, etc.
838 if (!Config.VectorizeCasts)
841 Type *SrcTy = C->getSrcTy();
842 if (!SrcTy->isSingleValueType())
845 Type *DestTy = C->getDestTy();
846 if (!DestTy->isSingleValueType())
848 } else if (isa<SelectInst>(I)) {
849 if (!Config.VectorizeSelect)
851 } else if (isa<CmpInst>(I)) {
852 if (!Config.VectorizeCmp)
854 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) {
855 if (!Config.VectorizeGEP)
858 // Currently, vector GEPs exist only with one index.
859 if (G->getNumIndices() != 1)
861 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) ||
862 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) {
866 // We can't vectorize memory operations without target data
867 if (TD == 0 && IsSimpleLoadStore)
871 getInstructionTypes(I, T1, T2);
873 // Not every type can be vectorized...
874 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) ||
875 !(VectorType::isValidElementType(T2) || T2->isVectorTy()))
878 if (T1->getScalarSizeInBits() == 1) {
879 if (!Config.VectorizeBools)
882 if (!Config.VectorizeInts && T1->isIntOrIntVectorTy())
886 if (T2->getScalarSizeInBits() == 1) {
887 if (!Config.VectorizeBools)
890 if (!Config.VectorizeInts && T2->isIntOrIntVectorTy())
894 if (!Config.VectorizeFloats
895 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy()))
898 // Don't vectorize target-specific types.
899 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy())
901 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy())
904 if ((!Config.VectorizePointers || TD == 0) &&
905 (T1->getScalarType()->isPointerTy() ||
906 T2->getScalarType()->isPointerTy()))
909 if (!TTI && (T1->getPrimitiveSizeInBits() >= Config.VectorBits ||
910 T2->getPrimitiveSizeInBits() >= Config.VectorBits))
916 // This function returns true if the two provided instructions are compatible
917 // (meaning that they can be fused into a vector instruction). This assumes
918 // that I has already been determined to be vectorizable and that J is not
919 // in the use tree of I.
920 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J,
921 bool IsSimpleLoadStore, bool NonPow2Len,
922 int &CostSavings, int &FixedOrder) {
923 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I <<
924 " <-> " << *J << "\n");
929 // Loads and stores can be merged if they have different alignments,
930 // but are otherwise the same.
931 if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment |
932 (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0)))
935 Type *IT1, *IT2, *JT1, *JT2;
936 getInstructionTypes(I, IT1, IT2);
937 getInstructionTypes(J, JT1, JT2);
938 unsigned MaxTypeBits = std::max(
939 IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(),
940 IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits());
941 if (!TTI && MaxTypeBits > Config.VectorBits)
944 // FIXME: handle addsub-type operations!
946 if (IsSimpleLoadStore) {
948 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
949 int64_t OffsetInElmts = 0;
950 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
951 IAddressSpace, JAddressSpace,
952 OffsetInElmts) && abs64(OffsetInElmts) == 1) {
953 FixedOrder = (int) OffsetInElmts;
954 unsigned BottomAlignment = IAlignment;
955 if (OffsetInElmts < 0) BottomAlignment = JAlignment;
957 Type *aTypeI = isa<StoreInst>(I) ?
958 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType();
959 Type *aTypeJ = isa<StoreInst>(J) ?
960 cast<StoreInst>(J)->getValueOperand()->getType() : J->getType();
961 Type *VType = getVecTypeForPair(aTypeI, aTypeJ);
963 if (Config.AlignedOnly) {
964 // An aligned load or store is possible only if the instruction
965 // with the lower offset has an alignment suitable for the
968 unsigned VecAlignment = TD->getPrefTypeAlignment(VType);
969 if (BottomAlignment < VecAlignment)
974 unsigned ICost = TTI->getMemoryOpCost(I->getOpcode(), aTypeI,
975 IAlignment, IAddressSpace);
976 unsigned JCost = TTI->getMemoryOpCost(J->getOpcode(), aTypeJ,
977 JAlignment, JAddressSpace);
978 unsigned VCost = TTI->getMemoryOpCost(I->getOpcode(), VType,
982 ICost += TTI->getAddressComputationCost(aTypeI);
983 JCost += TTI->getAddressComputationCost(aTypeJ);
984 VCost += TTI->getAddressComputationCost(VType);
986 if (VCost > ICost + JCost)
989 // We don't want to fuse to a type that will be split, even
990 // if the two input types will also be split and there is no other
992 unsigned VParts = TTI->getNumberOfParts(VType);
995 else if (!VParts && VCost == ICost + JCost)
998 CostSavings = ICost + JCost - VCost;
1004 unsigned ICost = getInstrCost(I->getOpcode(), IT1, IT2);
1005 unsigned JCost = getInstrCost(J->getOpcode(), JT1, JT2);
1006 Type *VT1 = getVecTypeForPair(IT1, JT1),
1007 *VT2 = getVecTypeForPair(IT2, JT2);
1009 // Note that this procedure is incorrect for insert and extract element
1010 // instructions (because combining these often results in a shuffle),
1011 // but this cost is ignored (because insert and extract element
1012 // instructions are assigned a zero depth factor and are not really
1013 // fused in general).
1014 unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2);
1016 if (VCost > ICost + JCost)
1019 // We don't want to fuse to a type that will be split, even
1020 // if the two input types will also be split and there is no other
1022 unsigned VParts1 = TTI->getNumberOfParts(VT1),
1023 VParts2 = TTI->getNumberOfParts(VT2);
1024 if (VParts1 > 1 || VParts2 > 1)
1026 else if ((!VParts1 || !VParts2) && VCost == ICost + JCost)
1029 CostSavings = ICost + JCost - VCost;
1032 // The powi intrinsic is special because only the first argument is
1033 // vectorized, the second arguments must be equal.
1034 CallInst *CI = dyn_cast<CallInst>(I);
1036 if (CI && (FI = CI->getCalledFunction())) {
1037 Intrinsic::ID IID = (Intrinsic::ID) FI->getIntrinsicID();
1038 if (IID == Intrinsic::powi) {
1039 Value *A1I = CI->getArgOperand(1),
1040 *A1J = cast<CallInst>(J)->getArgOperand(1);
1041 const SCEV *A1ISCEV = SE->getSCEV(A1I),
1042 *A1JSCEV = SE->getSCEV(A1J);
1043 return (A1ISCEV == A1JSCEV);
1047 SmallVector<Type*, 4> Tys;
1048 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
1049 Tys.push_back(CI->getArgOperand(i)->getType());
1050 unsigned ICost = TTI->getIntrinsicInstrCost(IID, IT1, Tys);
1053 CallInst *CJ = cast<CallInst>(J);
1054 for (unsigned i = 0, ie = CJ->getNumArgOperands(); i != ie; ++i)
1055 Tys.push_back(CJ->getArgOperand(i)->getType());
1056 unsigned JCost = TTI->getIntrinsicInstrCost(IID, JT1, Tys);
1059 assert(CI->getNumArgOperands() == CJ->getNumArgOperands() &&
1060 "Intrinsic argument counts differ");
1061 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
1062 if (IID == Intrinsic::powi && i == 1)
1063 Tys.push_back(CI->getArgOperand(i)->getType());
1065 Tys.push_back(getVecTypeForPair(CI->getArgOperand(i)->getType(),
1066 CJ->getArgOperand(i)->getType()));
1069 Type *RetTy = getVecTypeForPair(IT1, JT1);
1070 unsigned VCost = TTI->getIntrinsicInstrCost(IID, RetTy, Tys);
1072 if (VCost > ICost + JCost)
1075 // We don't want to fuse to a type that will be split, even
1076 // if the two input types will also be split and there is no other
1078 unsigned RetParts = TTI->getNumberOfParts(RetTy);
1081 else if (!RetParts && VCost == ICost + JCost)
1084 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
1085 if (!Tys[i]->isVectorTy())
1088 unsigned NumParts = TTI->getNumberOfParts(Tys[i]);
1091 else if (!NumParts && VCost == ICost + JCost)
1095 CostSavings = ICost + JCost - VCost;
1102 // Figure out whether or not J uses I and update the users and write-set
1103 // structures associated with I. Specifically, Users represents the set of
1104 // instructions that depend on I. WriteSet represents the set
1105 // of memory locations that are dependent on I. If UpdateUsers is true,
1106 // and J uses I, then Users is updated to contain J and WriteSet is updated
1107 // to contain any memory locations to which J writes. The function returns
1108 // true if J uses I. By default, alias analysis is used to determine
1109 // whether J reads from memory that overlaps with a location in WriteSet.
1110 // If LoadMoveSet is not null, then it is a previously-computed multimap
1111 // where the key is the memory-based user instruction and the value is
1112 // the instruction to be compared with I. So, if LoadMoveSet is provided,
1113 // then the alias analysis is not used. This is necessary because this
1114 // function is called during the process of moving instructions during
1115 // vectorization and the results of the alias analysis are not stable during
1117 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users,
1118 AliasSetTracker &WriteSet, Instruction *I,
1119 Instruction *J, bool UpdateUsers,
1120 DenseSet<ValuePair> *LoadMoveSetPairs) {
1123 // This instruction may already be marked as a user due, for example, to
1124 // being a member of a selected pair.
1129 for (User::op_iterator JU = J->op_begin(), JE = J->op_end();
1132 if (I == V || Users.count(V)) {
1137 if (!UsesI && J->mayReadFromMemory()) {
1138 if (LoadMoveSetPairs) {
1139 UsesI = LoadMoveSetPairs->count(ValuePair(J, I));
1141 for (AliasSetTracker::iterator W = WriteSet.begin(),
1142 WE = WriteSet.end(); W != WE; ++W) {
1143 if (W->aliasesUnknownInst(J, *AA)) {
1151 if (UsesI && UpdateUsers) {
1152 if (J->mayWriteToMemory()) WriteSet.add(J);
1159 // This function iterates over all instruction pairs in the provided
1160 // basic block and collects all candidate pairs for vectorization.
1161 bool BBVectorize::getCandidatePairs(BasicBlock &BB,
1162 BasicBlock::iterator &Start,
1163 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1164 DenseSet<ValuePair> &FixedOrderPairs,
1165 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1166 std::vector<Value *> &PairableInsts, bool NonPow2Len) {
1167 BasicBlock::iterator E = BB.end();
1168 if (Start == E) return false;
1170 bool ShouldContinue = false, IAfterStart = false;
1171 for (BasicBlock::iterator I = Start++; I != E; ++I) {
1172 if (I == Start) IAfterStart = true;
1174 bool IsSimpleLoadStore;
1175 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue;
1177 // Look for an instruction with which to pair instruction *I...
1178 DenseSet<Value *> Users;
1179 AliasSetTracker WriteSet(*AA);
1180 bool JAfterStart = IAfterStart;
1181 BasicBlock::iterator J = llvm::next(I);
1182 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) {
1183 if (J == Start) JAfterStart = true;
1185 // Determine if J uses I, if so, exit the loop.
1186 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep);
1187 if (Config.FastDep) {
1188 // Note: For this heuristic to be effective, independent operations
1189 // must tend to be intermixed. This is likely to be true from some
1190 // kinds of grouped loop unrolling (but not the generic LLVM pass),
1191 // but otherwise may require some kind of reordering pass.
1193 // When using fast dependency analysis,
1194 // stop searching after first use:
1197 if (UsesI) continue;
1200 // J does not use I, and comes before the first use of I, so it can be
1201 // merged with I if the instructions are compatible.
1202 int CostSavings, FixedOrder;
1203 if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len,
1204 CostSavings, FixedOrder)) continue;
1206 // J is a candidate for merging with I.
1207 if (!PairableInsts.size() ||
1208 PairableInsts[PairableInsts.size()-1] != I) {
1209 PairableInsts.push_back(I);
1212 CandidatePairs[I].push_back(J);
1214 CandidatePairCostSavings.insert(ValuePairWithCost(ValuePair(I, J),
1217 if (FixedOrder == 1)
1218 FixedOrderPairs.insert(ValuePair(I, J));
1219 else if (FixedOrder == -1)
1220 FixedOrderPairs.insert(ValuePair(J, I));
1222 // The next call to this function must start after the last instruction
1223 // selected during this invocation.
1225 Start = llvm::next(J);
1226 IAfterStart = JAfterStart = false;
1229 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair "
1230 << *I << " <-> " << *J << " (cost savings: " <<
1231 CostSavings << ")\n");
1233 // If we have already found too many pairs, break here and this function
1234 // will be called again starting after the last instruction selected
1235 // during this invocation.
1236 if (PairableInsts.size() >= Config.MaxInsts) {
1237 ShouldContinue = true;
1246 DEBUG(dbgs() << "BBV: found " << PairableInsts.size()
1247 << " instructions with candidate pairs\n");
1249 return ShouldContinue;
1252 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that
1253 // it looks for pairs such that both members have an input which is an
1254 // output of PI or PJ.
1255 void BBVectorize::computePairsConnectedTo(
1256 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1257 DenseSet<ValuePair> &CandidatePairsSet,
1258 std::vector<Value *> &PairableInsts,
1259 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1260 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1264 // For each possible pairing for this variable, look at the uses of
1265 // the first value...
1266 for (Value::use_iterator I = P.first->use_begin(),
1267 E = P.first->use_end(); I != E; ++I) {
1268 if (isa<LoadInst>(*I)) {
1269 // A pair cannot be connected to a load because the load only takes one
1270 // operand (the address) and it is a scalar even after vectorization.
1272 } else if ((SI = dyn_cast<StoreInst>(*I)) &&
1273 P.first == SI->getPointerOperand()) {
1274 // Similarly, a pair cannot be connected to a store through its
1279 // For each use of the first variable, look for uses of the second
1281 for (Value::use_iterator J = P.second->use_begin(),
1282 E2 = P.second->use_end(); J != E2; ++J) {
1283 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1284 P.second == SJ->getPointerOperand())
1288 if (CandidatePairsSet.count(ValuePair(*I, *J))) {
1289 VPPair VP(P, ValuePair(*I, *J));
1290 ConnectedPairs.insert(VP);
1291 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionDirect));
1295 if (CandidatePairsSet.count(ValuePair(*J, *I))) {
1296 VPPair VP(P, ValuePair(*J, *I));
1297 ConnectedPairs.insert(VP);
1298 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSwap));
1302 if (Config.SplatBreaksChain) continue;
1303 // Look for cases where just the first value in the pair is used by
1304 // both members of another pair (splatting).
1305 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) {
1306 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1307 P.first == SJ->getPointerOperand())
1310 if (CandidatePairsSet.count(ValuePair(*I, *J))) {
1311 VPPair VP(P, ValuePair(*I, *J));
1312 ConnectedPairs.insert(VP);
1313 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1318 if (Config.SplatBreaksChain) return;
1319 // Look for cases where just the second value in the pair is used by
1320 // both members of another pair (splatting).
1321 for (Value::use_iterator I = P.second->use_begin(),
1322 E = P.second->use_end(); I != E; ++I) {
1323 if (isa<LoadInst>(*I))
1325 else if ((SI = dyn_cast<StoreInst>(*I)) &&
1326 P.second == SI->getPointerOperand())
1329 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) {
1330 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1331 P.second == SJ->getPointerOperand())
1334 if (CandidatePairsSet.count(ValuePair(*I, *J))) {
1335 VPPair VP(P, ValuePair(*I, *J));
1336 ConnectedPairs.insert(VP);
1337 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1343 // This function figures out which pairs are connected. Two pairs are
1344 // connected if some output of the first pair forms an input to both members
1345 // of the second pair.
1346 void BBVectorize::computeConnectedPairs(
1347 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1348 DenseSet<ValuePair> &CandidatePairsSet,
1349 std::vector<Value *> &PairableInsts,
1350 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1351 DenseMap<VPPair, unsigned> &PairConnectionTypes) {
1352 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
1353 PE = PairableInsts.end(); PI != PE; ++PI) {
1354 DenseMap<Value *, std::vector<Value *> >::iterator PP =
1355 CandidatePairs.find(*PI);
1356 if (PP == CandidatePairs.end())
1359 for (std::vector<Value *>::iterator P = PP->second.begin(),
1360 E = PP->second.end(); P != E; ++P)
1361 computePairsConnectedTo(CandidatePairs, CandidatePairsSet,
1362 PairableInsts, ConnectedPairs,
1363 PairConnectionTypes, ValuePair(*PI, *P));
1366 DEBUG(dbgs() << "BBV: found " << ConnectedPairs.size()
1367 << " pair connections.\n");
1370 // This function builds a set of use tuples such that <A, B> is in the set
1371 // if B is in the use tree of A. If B is in the use tree of A, then B
1372 // depends on the output of A.
1373 void BBVectorize::buildDepMap(
1375 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1376 std::vector<Value *> &PairableInsts,
1377 DenseSet<ValuePair> &PairableInstUsers) {
1378 DenseSet<Value *> IsInPair;
1379 for (DenseMap<Value *, std::vector<Value *> >::iterator C =
1380 CandidatePairs.begin(), E = CandidatePairs.end(); C != E; ++C) {
1381 IsInPair.insert(C->first);
1382 IsInPair.insert(C->second.begin(), C->second.end());
1385 // Iterate through the basic block, recording all users of each
1386 // pairable instruction.
1388 BasicBlock::iterator E = BB.end(), EL =
1389 BasicBlock::iterator(cast<Instruction>(PairableInsts.back()));
1390 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) {
1391 if (IsInPair.find(I) == IsInPair.end()) continue;
1393 DenseSet<Value *> Users;
1394 AliasSetTracker WriteSet(*AA);
1395 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J) {
1396 (void) trackUsesOfI(Users, WriteSet, I, J);
1402 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end();
1404 if (IsInPair.find(*U) == IsInPair.end()) continue;
1405 PairableInstUsers.insert(ValuePair(I, *U));
1413 // Returns true if an input to pair P is an output of pair Q and also an
1414 // input of pair Q is an output of pair P. If this is the case, then these
1415 // two pairs cannot be simultaneously fused.
1416 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q,
1417 DenseSet<ValuePair> &PairableInstUsers,
1418 std::multimap<ValuePair, ValuePair> *PairableInstUserMap,
1419 DenseSet<VPPair> *PairableInstUserPairSet) {
1420 // Two pairs are in conflict if they are mutual Users of eachother.
1421 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) ||
1422 PairableInstUsers.count(ValuePair(P.first, Q.second)) ||
1423 PairableInstUsers.count(ValuePair(P.second, Q.first)) ||
1424 PairableInstUsers.count(ValuePair(P.second, Q.second));
1425 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) ||
1426 PairableInstUsers.count(ValuePair(Q.first, P.second)) ||
1427 PairableInstUsers.count(ValuePair(Q.second, P.first)) ||
1428 PairableInstUsers.count(ValuePair(Q.second, P.second));
1429 if (PairableInstUserMap) {
1430 // FIXME: The expensive part of the cycle check is not so much the cycle
1431 // check itself but this edge insertion procedure. This needs some
1432 // profiling and probably a different data structure (same is true of
1433 // most uses of std::multimap).
1435 if (PairableInstUserPairSet->insert(VPPair(Q, P)).second)
1436 PairableInstUserMap->insert(VPPair(Q, P));
1439 if (PairableInstUserPairSet->insert(VPPair(P, Q)).second)
1440 PairableInstUserMap->insert(VPPair(P, Q));
1444 return (QUsesP && PUsesQ);
1447 // This function walks the use graph of current pairs to see if, starting
1448 // from P, the walk returns to P.
1449 bool BBVectorize::pairWillFormCycle(ValuePair P,
1450 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1451 DenseSet<ValuePair> &CurrentPairs) {
1452 DEBUG(if (DebugCycleCheck)
1453 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> "
1454 << *P.second << "\n");
1455 // A lookup table of visisted pairs is kept because the PairableInstUserMap
1456 // contains non-direct associations.
1457 DenseSet<ValuePair> Visited;
1458 SmallVector<ValuePair, 32> Q;
1459 // General depth-first post-order traversal:
1462 ValuePair QTop = Q.pop_back_val();
1463 Visited.insert(QTop);
1465 DEBUG(if (DebugCycleCheck)
1466 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> "
1467 << *QTop.second << "\n");
1468 VPPIteratorPair QPairRange = PairableInstUserMap.equal_range(QTop);
1469 for (std::multimap<ValuePair, ValuePair>::iterator C = QPairRange.first;
1470 C != QPairRange.second; ++C) {
1471 if (C->second == P) {
1473 << "BBV: rejected to prevent non-trivial cycle formation: "
1474 << *C->first.first << " <-> " << *C->first.second << "\n");
1478 if (CurrentPairs.count(C->second) && !Visited.count(C->second))
1479 Q.push_back(C->second);
1481 } while (!Q.empty());
1486 // This function builds the initial tree of connected pairs with the
1487 // pair J at the root.
1488 void BBVectorize::buildInitialTreeFor(
1489 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1490 DenseSet<ValuePair> &CandidatePairsSet,
1491 std::vector<Value *> &PairableInsts,
1492 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1493 DenseSet<ValuePair> &PairableInstUsers,
1494 DenseMap<Value *, Value *> &ChosenPairs,
1495 DenseMap<ValuePair, size_t> &Tree, ValuePair J) {
1496 // Each of these pairs is viewed as the root node of a Tree. The Tree
1497 // is then walked (depth-first). As this happens, we keep track of
1498 // the pairs that compose the Tree and the maximum depth of the Tree.
1499 SmallVector<ValuePairWithDepth, 32> Q;
1500 // General depth-first post-order traversal:
1501 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1503 ValuePairWithDepth QTop = Q.back();
1505 // Push each child onto the queue:
1506 bool MoreChildren = false;
1507 size_t MaxChildDepth = QTop.second;
1508 VPPIteratorPair qtRange = ConnectedPairs.equal_range(QTop.first);
1509 for (std::multimap<ValuePair, ValuePair>::iterator k = qtRange.first;
1510 k != qtRange.second; ++k) {
1511 // Make sure that this child pair is still a candidate:
1512 if (CandidatePairsSet.count(ValuePair(k->second))) {
1513 DenseMap<ValuePair, size_t>::iterator C = Tree.find(k->second);
1514 if (C == Tree.end()) {
1515 size_t d = getDepthFactor(k->second.first);
1516 Q.push_back(ValuePairWithDepth(k->second, QTop.second+d));
1517 MoreChildren = true;
1519 MaxChildDepth = std::max(MaxChildDepth, C->second);
1524 if (!MoreChildren) {
1525 // Record the current pair as part of the Tree:
1526 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth));
1529 } while (!Q.empty());
1532 // Given some initial tree, prune it by removing conflicting pairs (pairs
1533 // that cannot be simultaneously chosen for vectorization).
1534 void BBVectorize::pruneTreeFor(
1535 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1536 std::vector<Value *> &PairableInsts,
1537 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1538 DenseSet<ValuePair> &PairableInstUsers,
1539 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1540 DenseSet<VPPair> &PairableInstUserPairSet,
1541 DenseMap<Value *, Value *> &ChosenPairs,
1542 DenseMap<ValuePair, size_t> &Tree,
1543 DenseSet<ValuePair> &PrunedTree, ValuePair J,
1544 bool UseCycleCheck) {
1545 SmallVector<ValuePairWithDepth, 32> Q;
1546 // General depth-first post-order traversal:
1547 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1549 ValuePairWithDepth QTop = Q.pop_back_val();
1550 PrunedTree.insert(QTop.first);
1552 // Visit each child, pruning as necessary...
1553 SmallVector<ValuePairWithDepth, 8> BestChildren;
1554 VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first);
1555 for (std::multimap<ValuePair, ValuePair>::iterator K = QTopRange.first;
1556 K != QTopRange.second; ++K) {
1557 DenseMap<ValuePair, size_t>::iterator C = Tree.find(K->second);
1558 if (C == Tree.end()) continue;
1560 // This child is in the Tree, now we need to make sure it is the
1561 // best of any conflicting children. There could be multiple
1562 // conflicting children, so first, determine if we're keeping
1563 // this child, then delete conflicting children as necessary.
1565 // It is also necessary to guard against pairing-induced
1566 // dependencies. Consider instructions a .. x .. y .. b
1567 // such that (a,b) are to be fused and (x,y) are to be fused
1568 // but a is an input to x and b is an output from y. This
1569 // means that y cannot be moved after b but x must be moved
1570 // after b for (a,b) to be fused. In other words, after
1571 // fusing (a,b) we have y .. a/b .. x where y is an input
1572 // to a/b and x is an output to a/b: x and y can no longer
1573 // be legally fused. To prevent this condition, we must
1574 // make sure that a child pair added to the Tree is not
1575 // both an input and output of an already-selected pair.
1577 // Pairing-induced dependencies can also form from more complicated
1578 // cycles. The pair vs. pair conflicts are easy to check, and so
1579 // that is done explicitly for "fast rejection", and because for
1580 // child vs. child conflicts, we may prefer to keep the current
1581 // pair in preference to the already-selected child.
1582 DenseSet<ValuePair> CurrentPairs;
1585 for (SmallVector<ValuePairWithDepth, 8>::iterator C2
1586 = BestChildren.begin(), E2 = BestChildren.end();
1588 if (C2->first.first == C->first.first ||
1589 C2->first.first == C->first.second ||
1590 C2->first.second == C->first.first ||
1591 C2->first.second == C->first.second ||
1592 pairsConflict(C2->first, C->first, PairableInstUsers,
1593 UseCycleCheck ? &PairableInstUserMap : 0,
1594 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1595 if (C2->second >= C->second) {
1600 CurrentPairs.insert(C2->first);
1603 if (!CanAdd) continue;
1605 // Even worse, this child could conflict with another node already
1606 // selected for the Tree. If that is the case, ignore this child.
1607 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(),
1608 E2 = PrunedTree.end(); T != E2; ++T) {
1609 if (T->first == C->first.first ||
1610 T->first == C->first.second ||
1611 T->second == C->first.first ||
1612 T->second == C->first.second ||
1613 pairsConflict(*T, C->first, PairableInstUsers,
1614 UseCycleCheck ? &PairableInstUserMap : 0,
1615 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1620 CurrentPairs.insert(*T);
1622 if (!CanAdd) continue;
1624 // And check the queue too...
1625 for (SmallVector<ValuePairWithDepth, 32>::iterator C2 = Q.begin(),
1626 E2 = Q.end(); C2 != E2; ++C2) {
1627 if (C2->first.first == C->first.first ||
1628 C2->first.first == C->first.second ||
1629 C2->first.second == C->first.first ||
1630 C2->first.second == C->first.second ||
1631 pairsConflict(C2->first, C->first, PairableInstUsers,
1632 UseCycleCheck ? &PairableInstUserMap : 0,
1633 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1638 CurrentPairs.insert(C2->first);
1640 if (!CanAdd) continue;
1642 // Last but not least, check for a conflict with any of the
1643 // already-chosen pairs.
1644 for (DenseMap<Value *, Value *>::iterator C2 =
1645 ChosenPairs.begin(), E2 = ChosenPairs.end();
1647 if (pairsConflict(*C2, C->first, PairableInstUsers,
1648 UseCycleCheck ? &PairableInstUserMap : 0,
1649 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1654 CurrentPairs.insert(*C2);
1656 if (!CanAdd) continue;
1658 // To check for non-trivial cycles formed by the addition of the
1659 // current pair we've formed a list of all relevant pairs, now use a
1660 // graph walk to check for a cycle. We start from the current pair and
1661 // walk the use tree to see if we again reach the current pair. If we
1662 // do, then the current pair is rejected.
1664 // FIXME: It may be more efficient to use a topological-ordering
1665 // algorithm to improve the cycle check. This should be investigated.
1666 if (UseCycleCheck &&
1667 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs))
1670 // This child can be added, but we may have chosen it in preference
1671 // to an already-selected child. Check for this here, and if a
1672 // conflict is found, then remove the previously-selected child
1673 // before adding this one in its place.
1674 for (SmallVector<ValuePairWithDepth, 8>::iterator C2
1675 = BestChildren.begin(); C2 != BestChildren.end();) {
1676 if (C2->first.first == C->first.first ||
1677 C2->first.first == C->first.second ||
1678 C2->first.second == C->first.first ||
1679 C2->first.second == C->first.second ||
1680 pairsConflict(C2->first, C->first, PairableInstUsers))
1681 C2 = BestChildren.erase(C2);
1686 BestChildren.push_back(ValuePairWithDepth(C->first, C->second));
1689 for (SmallVector<ValuePairWithDepth, 8>::iterator C
1690 = BestChildren.begin(), E2 = BestChildren.end();
1692 size_t DepthF = getDepthFactor(C->first.first);
1693 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF));
1695 } while (!Q.empty());
1698 // This function finds the best tree of mututally-compatible connected
1699 // pairs, given the choice of root pairs as an iterator range.
1700 void BBVectorize::findBestTreeFor(
1701 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1702 DenseSet<ValuePair> &CandidatePairsSet,
1703 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1704 std::vector<Value *> &PairableInsts,
1705 DenseSet<ValuePair> &FixedOrderPairs,
1706 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1707 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1708 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps,
1709 DenseSet<ValuePair> &PairableInstUsers,
1710 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1711 DenseSet<VPPair> &PairableInstUserPairSet,
1712 DenseMap<Value *, Value *> &ChosenPairs,
1713 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth,
1714 int &BestEffSize, Value *II, std::vector<Value *>&JJ,
1715 bool UseCycleCheck) {
1716 for (std::vector<Value *>::iterator J = JJ.begin(), JE = JJ.end();
1718 ValuePair IJ(II, *J);
1719 if (!CandidatePairsSet.count(IJ))
1722 // Before going any further, make sure that this pair does not
1723 // conflict with any already-selected pairs (see comment below
1724 // near the Tree pruning for more details).
1725 DenseSet<ValuePair> ChosenPairSet;
1726 bool DoesConflict = false;
1727 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(),
1728 E = ChosenPairs.end(); C != E; ++C) {
1729 if (pairsConflict(*C, IJ, PairableInstUsers,
1730 UseCycleCheck ? &PairableInstUserMap : 0,
1731 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1732 DoesConflict = true;
1736 ChosenPairSet.insert(*C);
1738 if (DoesConflict) continue;
1740 if (UseCycleCheck &&
1741 pairWillFormCycle(IJ, PairableInstUserMap, ChosenPairSet))
1744 DenseMap<ValuePair, size_t> Tree;
1745 buildInitialTreeFor(CandidatePairs, CandidatePairsSet,
1746 PairableInsts, ConnectedPairs,
1747 PairableInstUsers, ChosenPairs, Tree, IJ);
1749 // Because we'll keep the child with the largest depth, the largest
1750 // depth is still the same in the unpruned Tree.
1751 size_t MaxDepth = Tree.lookup(IJ);
1753 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found Tree for pair {"
1754 << IJ.first << " <-> " << IJ.second << "} of depth " <<
1755 MaxDepth << " and size " << Tree.size() << "\n");
1757 // At this point the Tree has been constructed, but, may contain
1758 // contradictory children (meaning that different children of
1759 // some tree node may be attempting to fuse the same instruction).
1760 // So now we walk the tree again, in the case of a conflict,
1761 // keep only the child with the largest depth. To break a tie,
1762 // favor the first child.
1764 DenseSet<ValuePair> PrunedTree;
1765 pruneTreeFor(CandidatePairs, PairableInsts, ConnectedPairs,
1766 PairableInstUsers, PairableInstUserMap,
1767 PairableInstUserPairSet,
1768 ChosenPairs, Tree, PrunedTree, IJ, UseCycleCheck);
1772 DenseSet<Value *> PrunedTreeInstrs;
1773 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
1774 E = PrunedTree.end(); S != E; ++S) {
1775 PrunedTreeInstrs.insert(S->first);
1776 PrunedTreeInstrs.insert(S->second);
1779 // The set of pairs that have already contributed to the total cost.
1780 DenseSet<ValuePair> IncomingPairs;
1782 // If the cost model were perfect, this might not be necessary; but we
1783 // need to make sure that we don't get stuck vectorizing our own
1785 bool HasNontrivialInsts = false;
1787 // The node weights represent the cost savings associated with
1788 // fusing the pair of instructions.
1789 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
1790 E = PrunedTree.end(); S != E; ++S) {
1791 if (!isa<ShuffleVectorInst>(S->first) &&
1792 !isa<InsertElementInst>(S->first) &&
1793 !isa<ExtractElementInst>(S->first))
1794 HasNontrivialInsts = true;
1796 bool FlipOrder = false;
1798 if (getDepthFactor(S->first)) {
1799 int ESContrib = CandidatePairCostSavings.find(*S)->second;
1800 DEBUG(if (DebugPairSelection) dbgs() << "\tweight {"
1801 << *S->first << " <-> " << *S->second << "} = " <<
1803 EffSize += ESContrib;
1806 // The edge weights contribute in a negative sense: they represent
1807 // the cost of shuffles.
1808 VPPIteratorPair IP = ConnectedPairDeps.equal_range(*S);
1809 if (IP.first != ConnectedPairDeps.end()) {
1810 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
1811 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first;
1812 Q != IP.second; ++Q) {
1813 if (!PrunedTree.count(Q->second))
1815 DenseMap<VPPair, unsigned>::iterator R =
1816 PairConnectionTypes.find(VPPair(Q->second, Q->first));
1817 assert(R != PairConnectionTypes.end() &&
1818 "Cannot find pair connection type");
1819 if (R->second == PairConnectionDirect)
1821 else if (R->second == PairConnectionSwap)
1825 // If there are more swaps than direct connections, then
1826 // the pair order will be flipped during fusion. So the real
1827 // number of swaps is the minimum number.
1828 FlipOrder = !FixedOrderPairs.count(*S) &&
1829 ((NumDepsSwap > NumDepsDirect) ||
1830 FixedOrderPairs.count(ValuePair(S->second, S->first)));
1832 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first;
1833 Q != IP.second; ++Q) {
1834 if (!PrunedTree.count(Q->second))
1836 DenseMap<VPPair, unsigned>::iterator R =
1837 PairConnectionTypes.find(VPPair(Q->second, Q->first));
1838 assert(R != PairConnectionTypes.end() &&
1839 "Cannot find pair connection type");
1840 Type *Ty1 = Q->second.first->getType(),
1841 *Ty2 = Q->second.second->getType();
1842 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1843 if ((R->second == PairConnectionDirect && FlipOrder) ||
1844 (R->second == PairConnectionSwap && !FlipOrder) ||
1845 R->second == PairConnectionSplat) {
1846 int ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1849 if (VTy->getVectorNumElements() == 2) {
1850 if (R->second == PairConnectionSplat)
1851 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1852 TargetTransformInfo::SK_Broadcast, VTy));
1854 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1855 TargetTransformInfo::SK_Reverse, VTy));
1858 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1859 *Q->second.first << " <-> " << *Q->second.second <<
1861 *S->first << " <-> " << *S->second << "} = " <<
1863 EffSize -= ESContrib;
1868 // Compute the cost of outgoing edges. We assume that edges outgoing
1869 // to shuffles, inserts or extracts can be merged, and so contribute
1870 // no additional cost.
1871 if (!S->first->getType()->isVoidTy()) {
1872 Type *Ty1 = S->first->getType(),
1873 *Ty2 = S->second->getType();
1874 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1876 bool NeedsExtraction = false;
1877 for (Value::use_iterator I = S->first->use_begin(),
1878 IE = S->first->use_end(); I != IE; ++I) {
1879 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) {
1880 // Shuffle can be folded if it has no other input
1881 if (isa<UndefValue>(SI->getOperand(1)))
1884 if (isa<ExtractElementInst>(*I))
1886 if (PrunedTreeInstrs.count(*I))
1888 NeedsExtraction = true;
1892 if (NeedsExtraction) {
1894 if (Ty1->isVectorTy()) {
1895 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1897 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1898 TargetTransformInfo::SK_ExtractSubvector, VTy, 0, Ty1));
1900 ESContrib = (int) TTI->getVectorInstrCost(
1901 Instruction::ExtractElement, VTy, 0);
1903 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1904 *S->first << "} = " << ESContrib << "\n");
1905 EffSize -= ESContrib;
1908 NeedsExtraction = false;
1909 for (Value::use_iterator I = S->second->use_begin(),
1910 IE = S->second->use_end(); I != IE; ++I) {
1911 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) {
1912 // Shuffle can be folded if it has no other input
1913 if (isa<UndefValue>(SI->getOperand(1)))
1916 if (isa<ExtractElementInst>(*I))
1918 if (PrunedTreeInstrs.count(*I))
1920 NeedsExtraction = true;
1924 if (NeedsExtraction) {
1926 if (Ty2->isVectorTy()) {
1927 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1929 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1930 TargetTransformInfo::SK_ExtractSubvector, VTy,
1931 Ty1->isVectorTy() ? Ty1->getVectorNumElements() : 1, Ty2));
1933 ESContrib = (int) TTI->getVectorInstrCost(
1934 Instruction::ExtractElement, VTy, 1);
1935 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1936 *S->second << "} = " << ESContrib << "\n");
1937 EffSize -= ESContrib;
1941 // Compute the cost of incoming edges.
1942 if (!isa<LoadInst>(S->first) && !isa<StoreInst>(S->first)) {
1943 Instruction *S1 = cast<Instruction>(S->first),
1944 *S2 = cast<Instruction>(S->second);
1945 for (unsigned o = 0; o < S1->getNumOperands(); ++o) {
1946 Value *O1 = S1->getOperand(o), *O2 = S2->getOperand(o);
1948 // Combining constants into vector constants (or small vector
1949 // constants into larger ones are assumed free).
1950 if (isa<Constant>(O1) && isa<Constant>(O2))
1956 ValuePair VP = ValuePair(O1, O2);
1957 ValuePair VPR = ValuePair(O2, O1);
1959 // Internal edges are not handled here.
1960 if (PrunedTree.count(VP) || PrunedTree.count(VPR))
1963 Type *Ty1 = O1->getType(),
1964 *Ty2 = O2->getType();
1965 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1967 // Combining vector operations of the same type is also assumed
1968 // folded with other operations.
1970 // If both are insert elements, then both can be widened.
1971 InsertElementInst *IEO1 = dyn_cast<InsertElementInst>(O1),
1972 *IEO2 = dyn_cast<InsertElementInst>(O2);
1973 if (IEO1 && IEO2 && isPureIEChain(IEO1) && isPureIEChain(IEO2))
1975 // If both are extract elements, and both have the same input
1976 // type, then they can be replaced with a shuffle
1977 ExtractElementInst *EIO1 = dyn_cast<ExtractElementInst>(O1),
1978 *EIO2 = dyn_cast<ExtractElementInst>(O2);
1980 EIO1->getOperand(0)->getType() ==
1981 EIO2->getOperand(0)->getType())
1983 // If both are a shuffle with equal operand types and only two
1984 // unqiue operands, then they can be replaced with a single
1986 ShuffleVectorInst *SIO1 = dyn_cast<ShuffleVectorInst>(O1),
1987 *SIO2 = dyn_cast<ShuffleVectorInst>(O2);
1989 SIO1->getOperand(0)->getType() ==
1990 SIO2->getOperand(0)->getType()) {
1991 SmallSet<Value *, 4> SIOps;
1992 SIOps.insert(SIO1->getOperand(0));
1993 SIOps.insert(SIO1->getOperand(1));
1994 SIOps.insert(SIO2->getOperand(0));
1995 SIOps.insert(SIO2->getOperand(1));
1996 if (SIOps.size() <= 2)
2002 // This pair has already been formed.
2003 if (IncomingPairs.count(VP)) {
2005 } else if (IncomingPairs.count(VPR)) {
2006 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2009 if (VTy->getVectorNumElements() == 2)
2010 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
2011 TargetTransformInfo::SK_Reverse, VTy));
2012 } else if (!Ty1->isVectorTy() && !Ty2->isVectorTy()) {
2013 ESContrib = (int) TTI->getVectorInstrCost(
2014 Instruction::InsertElement, VTy, 0);
2015 ESContrib += (int) TTI->getVectorInstrCost(
2016 Instruction::InsertElement, VTy, 1);
2017 } else if (!Ty1->isVectorTy()) {
2018 // O1 needs to be inserted into a vector of size O2, and then
2019 // both need to be shuffled together.
2020 ESContrib = (int) TTI->getVectorInstrCost(
2021 Instruction::InsertElement, Ty2, 0);
2022 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2024 } else if (!Ty2->isVectorTy()) {
2025 // O2 needs to be inserted into a vector of size O1, and then
2026 // both need to be shuffled together.
2027 ESContrib = (int) TTI->getVectorInstrCost(
2028 Instruction::InsertElement, Ty1, 0);
2029 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2032 Type *TyBig = Ty1, *TySmall = Ty2;
2033 if (Ty2->getVectorNumElements() > Ty1->getVectorNumElements())
2034 std::swap(TyBig, TySmall);
2036 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2038 if (TyBig != TySmall)
2039 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2043 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {"
2044 << *O1 << " <-> " << *O2 << "} = " <<
2046 EffSize -= ESContrib;
2047 IncomingPairs.insert(VP);
2052 if (!HasNontrivialInsts) {
2053 DEBUG(if (DebugPairSelection) dbgs() <<
2054 "\tNo non-trivial instructions in tree;"
2055 " override to zero effective size\n");
2059 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
2060 E = PrunedTree.end(); S != E; ++S)
2061 EffSize += (int) getDepthFactor(S->first);
2064 DEBUG(if (DebugPairSelection)
2065 dbgs() << "BBV: found pruned Tree for pair {"
2066 << IJ.first << " <-> " << IJ.second << "} of depth " <<
2067 MaxDepth << " and size " << PrunedTree.size() <<
2068 " (effective size: " << EffSize << ")\n");
2069 if (((TTI && !UseChainDepthWithTI) ||
2070 MaxDepth >= Config.ReqChainDepth) &&
2071 EffSize > 0 && EffSize > BestEffSize) {
2072 BestMaxDepth = MaxDepth;
2073 BestEffSize = EffSize;
2074 BestTree = PrunedTree;
2079 // Given the list of candidate pairs, this function selects those
2080 // that will be fused into vector instructions.
2081 void BBVectorize::choosePairs(
2082 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
2083 DenseSet<ValuePair> &CandidatePairsSet,
2084 DenseMap<ValuePair, int> &CandidatePairCostSavings,
2085 std::vector<Value *> &PairableInsts,
2086 DenseSet<ValuePair> &FixedOrderPairs,
2087 DenseMap<VPPair, unsigned> &PairConnectionTypes,
2088 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
2089 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps,
2090 DenseSet<ValuePair> &PairableInstUsers,
2091 DenseMap<Value *, Value *>& ChosenPairs) {
2092 bool UseCycleCheck =
2093 CandidatePairsSet.size() <= Config.MaxCandPairsForCycleCheck;
2095 DenseMap<Value *, std::vector<Value *> > CandidatePairs2;
2096 for (DenseSet<ValuePair>::iterator I = CandidatePairsSet.begin(),
2097 E = CandidatePairsSet.end(); I != E; ++I) {
2098 std::vector<Value *> &JJ = CandidatePairs2[I->second];
2099 if (JJ.empty()) JJ.reserve(32);
2100 JJ.push_back(I->first);
2103 std::multimap<ValuePair, ValuePair> PairableInstUserMap;
2104 DenseSet<VPPair> PairableInstUserPairSet;
2105 for (std::vector<Value *>::iterator I = PairableInsts.begin(),
2106 E = PairableInsts.end(); I != E; ++I) {
2107 // The number of possible pairings for this variable:
2108 size_t NumChoices = CandidatePairs.lookup(*I).size();
2109 if (!NumChoices) continue;
2111 std::vector<Value *> &JJ = CandidatePairs[*I];
2113 // The best pair to choose and its tree:
2114 size_t BestMaxDepth = 0;
2115 int BestEffSize = 0;
2116 DenseSet<ValuePair> BestTree;
2117 findBestTreeFor(CandidatePairs, CandidatePairsSet,
2118 CandidatePairCostSavings,
2119 PairableInsts, FixedOrderPairs, PairConnectionTypes,
2120 ConnectedPairs, ConnectedPairDeps,
2121 PairableInstUsers, PairableInstUserMap,
2122 PairableInstUserPairSet, ChosenPairs,
2123 BestTree, BestMaxDepth, BestEffSize, *I, JJ,
2126 if (BestTree.empty())
2129 // A tree has been chosen (or not) at this point. If no tree was
2130 // chosen, then this instruction, I, cannot be paired (and is no longer
2133 DEBUG(dbgs() << "BBV: selected pairs in the best tree for: "
2134 << *cast<Instruction>(*I) << "\n");
2136 for (DenseSet<ValuePair>::iterator S = BestTree.begin(),
2137 SE2 = BestTree.end(); S != SE2; ++S) {
2138 // Insert the members of this tree into the list of chosen pairs.
2139 ChosenPairs.insert(ValuePair(S->first, S->second));
2140 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " <<
2141 *S->second << "\n");
2143 // Remove all candidate pairs that have values in the chosen tree.
2144 std::vector<Value *> &KK = CandidatePairs[S->first],
2145 &LL = CandidatePairs2[S->second],
2146 &MM = CandidatePairs[S->second],
2147 &NN = CandidatePairs2[S->first];
2148 for (std::vector<Value *>::iterator K = KK.begin(), KE = KK.end();
2150 if (*K == S->second)
2153 CandidatePairsSet.erase(ValuePair(S->first, *K));
2155 for (std::vector<Value *>::iterator L = LL.begin(), LE = LL.end();
2160 CandidatePairsSet.erase(ValuePair(*L, S->second));
2162 for (std::vector<Value *>::iterator M = MM.begin(), ME = MM.end();
2164 assert(*M != S->first && "Flipped pair in candidate list?");
2165 CandidatePairsSet.erase(ValuePair(S->second, *M));
2167 for (std::vector<Value *>::iterator N = NN.begin(), NE = NN.end();
2169 assert(*N != S->second && "Flipped pair in candidate list?");
2170 CandidatePairsSet.erase(ValuePair(*N, S->first));
2175 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n");
2178 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o,
2183 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) +
2184 (n > 0 ? "." + utostr(n) : "")).str();
2187 // Returns the value that is to be used as the pointer input to the vector
2188 // instruction that fuses I with J.
2189 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context,
2190 Instruction *I, Instruction *J, unsigned o) {
2192 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
2193 int64_t OffsetInElmts;
2195 // Note: the analysis might fail here, that is why the pair order has
2196 // been precomputed (OffsetInElmts must be unused here).
2197 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
2198 IAddressSpace, JAddressSpace,
2199 OffsetInElmts, false);
2201 // The pointer value is taken to be the one with the lowest offset.
2204 Type *ArgTypeI = cast<PointerType>(IPtr->getType())->getElementType();
2205 Type *ArgTypeJ = cast<PointerType>(JPtr->getType())->getElementType();
2206 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2207 Type *VArgPtrType = PointerType::get(VArgType,
2208 cast<PointerType>(IPtr->getType())->getAddressSpace());
2209 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o),
2210 /* insert before */ I);
2213 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J,
2214 unsigned MaskOffset, unsigned NumInElem,
2215 unsigned NumInElem1, unsigned IdxOffset,
2216 std::vector<Constant*> &Mask) {
2217 unsigned NumElem1 = cast<VectorType>(J->getType())->getNumElements();
2218 for (unsigned v = 0; v < NumElem1; ++v) {
2219 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v);
2221 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context));
2223 unsigned mm = m + (int) IdxOffset;
2224 if (m >= (int) NumInElem1)
2225 mm += (int) NumInElem;
2227 Mask[v+MaskOffset] =
2228 ConstantInt::get(Type::getInt32Ty(Context), mm);
2233 // Returns the value that is to be used as the vector-shuffle mask to the
2234 // vector instruction that fuses I with J.
2235 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context,
2236 Instruction *I, Instruction *J) {
2237 // This is the shuffle mask. We need to append the second
2238 // mask to the first, and the numbers need to be adjusted.
2240 Type *ArgTypeI = I->getType();
2241 Type *ArgTypeJ = J->getType();
2242 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2244 unsigned NumElemI = cast<VectorType>(ArgTypeI)->getNumElements();
2246 // Get the total number of elements in the fused vector type.
2247 // By definition, this must equal the number of elements in
2249 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements();
2250 std::vector<Constant*> Mask(NumElem);
2252 Type *OpTypeI = I->getOperand(0)->getType();
2253 unsigned NumInElemI = cast<VectorType>(OpTypeI)->getNumElements();
2254 Type *OpTypeJ = J->getOperand(0)->getType();
2255 unsigned NumInElemJ = cast<VectorType>(OpTypeJ)->getNumElements();
2257 // The fused vector will be:
2258 // -----------------------------------------------------
2259 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ |
2260 // -----------------------------------------------------
2261 // from which we'll extract NumElem total elements (where the first NumElemI
2262 // of them come from the mask in I and the remainder come from the mask
2265 // For the mask from the first pair...
2266 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI,
2269 // For the mask from the second pair...
2270 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ,
2273 return ConstantVector::get(Mask);
2276 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I,
2277 Instruction *J, unsigned o, Value *&LOp,
2279 Type *ArgTypeL, Type *ArgTypeH,
2280 bool IBeforeJ, unsigned IdxOff) {
2281 bool ExpandedIEChain = false;
2282 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) {
2283 // If we have a pure insertelement chain, then this can be rewritten
2284 // into a chain that directly builds the larger type.
2285 if (isPureIEChain(LIE)) {
2286 SmallVector<Value *, 8> VectElemts(numElemL,
2287 UndefValue::get(ArgTypeL->getScalarType()));
2288 InsertElementInst *LIENext = LIE;
2291 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue();
2292 VectElemts[Idx] = LIENext->getOperand(1);
2294 dyn_cast<InsertElementInst>(LIENext->getOperand(0))));
2297 Value *LIEPrev = UndefValue::get(ArgTypeH);
2298 for (unsigned i = 0; i < numElemL; ++i) {
2299 if (isa<UndefValue>(VectElemts[i])) continue;
2300 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i],
2301 ConstantInt::get(Type::getInt32Ty(Context),
2303 getReplacementName(IBeforeJ ? I : J,
2305 LIENext->insertBefore(IBeforeJ ? J : I);
2309 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH);
2310 ExpandedIEChain = true;
2314 return ExpandedIEChain;
2317 // Returns the value to be used as the specified operand of the vector
2318 // instruction that fuses I with J.
2319 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I,
2320 Instruction *J, unsigned o, bool IBeforeJ) {
2321 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2322 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1);
2324 // Compute the fused vector type for this operand
2325 Type *ArgTypeI = I->getOperand(o)->getType();
2326 Type *ArgTypeJ = J->getOperand(o)->getType();
2327 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2329 Instruction *L = I, *H = J;
2330 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ;
2333 if (ArgTypeL->isVectorTy())
2334 numElemL = cast<VectorType>(ArgTypeL)->getNumElements();
2339 if (ArgTypeH->isVectorTy())
2340 numElemH = cast<VectorType>(ArgTypeH)->getNumElements();
2344 Value *LOp = L->getOperand(o);
2345 Value *HOp = H->getOperand(o);
2346 unsigned numElem = VArgType->getNumElements();
2348 // First, we check if we can reuse the "original" vector outputs (if these
2349 // exist). We might need a shuffle.
2350 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp);
2351 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp);
2352 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp);
2353 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp);
2355 // FIXME: If we're fusing shuffle instructions, then we can't apply this
2356 // optimization. The input vectors to the shuffle might be a different
2357 // length from the shuffle outputs. Unfortunately, the replacement
2358 // shuffle mask has already been formed, and the mask entries are sensitive
2359 // to the sizes of the inputs.
2360 bool IsSizeChangeShuffle =
2361 isa<ShuffleVectorInst>(L) &&
2362 (LOp->getType() != L->getType() || HOp->getType() != H->getType());
2364 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) {
2365 // We can have at most two unique vector inputs.
2366 bool CanUseInputs = true;
2369 I1 = LEE->getOperand(0);
2371 I1 = LSV->getOperand(0);
2372 I2 = LSV->getOperand(1);
2373 if (I2 == I1 || isa<UndefValue>(I2))
2378 Value *I3 = HEE->getOperand(0);
2379 if (!I2 && I3 != I1)
2381 else if (I3 != I1 && I3 != I2)
2382 CanUseInputs = false;
2384 Value *I3 = HSV->getOperand(0);
2385 if (!I2 && I3 != I1)
2387 else if (I3 != I1 && I3 != I2)
2388 CanUseInputs = false;
2391 Value *I4 = HSV->getOperand(1);
2392 if (!isa<UndefValue>(I4)) {
2393 if (!I2 && I4 != I1)
2395 else if (I4 != I1 && I4 != I2)
2396 CanUseInputs = false;
2403 cast<VectorType>(cast<Instruction>(LOp)->getOperand(0)->getType())
2406 cast<VectorType>(cast<Instruction>(HOp)->getOperand(0)->getType())
2409 // We have one or two input vectors. We need to map each index of the
2410 // operands to the index of the original vector.
2411 SmallVector<std::pair<int, int>, 8> II(numElem);
2412 for (unsigned i = 0; i < numElemL; ++i) {
2416 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue();
2417 INum = LEE->getOperand(0) == I1 ? 0 : 1;
2419 Idx = LSV->getMaskValue(i);
2420 if (Idx < (int) LOpElem) {
2421 INum = LSV->getOperand(0) == I1 ? 0 : 1;
2424 INum = LSV->getOperand(1) == I1 ? 0 : 1;
2428 II[i] = std::pair<int, int>(Idx, INum);
2430 for (unsigned i = 0; i < numElemH; ++i) {
2434 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue();
2435 INum = HEE->getOperand(0) == I1 ? 0 : 1;
2437 Idx = HSV->getMaskValue(i);
2438 if (Idx < (int) HOpElem) {
2439 INum = HSV->getOperand(0) == I1 ? 0 : 1;
2442 INum = HSV->getOperand(1) == I1 ? 0 : 1;
2446 II[i + numElemL] = std::pair<int, int>(Idx, INum);
2449 // We now have an array which tells us from which index of which
2450 // input vector each element of the operand comes.
2451 VectorType *I1T = cast<VectorType>(I1->getType());
2452 unsigned I1Elem = I1T->getNumElements();
2455 // In this case there is only one underlying vector input. Check for
2456 // the trivial case where we can use the input directly.
2457 if (I1Elem == numElem) {
2458 bool ElemInOrder = true;
2459 for (unsigned i = 0; i < numElem; ++i) {
2460 if (II[i].first != (int) i && II[i].first != -1) {
2461 ElemInOrder = false;
2470 // A shuffle is needed.
2471 std::vector<Constant *> Mask(numElem);
2472 for (unsigned i = 0; i < numElem; ++i) {
2473 int Idx = II[i].first;
2475 Mask[i] = UndefValue::get(Type::getInt32Ty(Context));
2477 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2481 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2482 ConstantVector::get(Mask),
2483 getReplacementName(IBeforeJ ? I : J,
2485 S->insertBefore(IBeforeJ ? J : I);
2489 VectorType *I2T = cast<VectorType>(I2->getType());
2490 unsigned I2Elem = I2T->getNumElements();
2492 // This input comes from two distinct vectors. The first step is to
2493 // make sure that both vectors are the same length. If not, the
2494 // smaller one will need to grow before they can be shuffled together.
2495 if (I1Elem < I2Elem) {
2496 std::vector<Constant *> Mask(I2Elem);
2498 for (; v < I1Elem; ++v)
2499 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2500 for (; v < I2Elem; ++v)
2501 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2503 Instruction *NewI1 =
2504 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2505 ConstantVector::get(Mask),
2506 getReplacementName(IBeforeJ ? I : J,
2508 NewI1->insertBefore(IBeforeJ ? J : I);
2512 } else if (I1Elem > I2Elem) {
2513 std::vector<Constant *> Mask(I1Elem);
2515 for (; v < I2Elem; ++v)
2516 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2517 for (; v < I1Elem; ++v)
2518 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2520 Instruction *NewI2 =
2521 new ShuffleVectorInst(I2, UndefValue::get(I2T),
2522 ConstantVector::get(Mask),
2523 getReplacementName(IBeforeJ ? I : J,
2525 NewI2->insertBefore(IBeforeJ ? J : I);
2531 // Now that both I1 and I2 are the same length we can shuffle them
2532 // together (and use the result).
2533 std::vector<Constant *> Mask(numElem);
2534 for (unsigned v = 0; v < numElem; ++v) {
2535 if (II[v].first == -1) {
2536 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2538 int Idx = II[v].first + II[v].second * I1Elem;
2539 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2543 Instruction *NewOp =
2544 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask),
2545 getReplacementName(IBeforeJ ? I : J, true, o));
2546 NewOp->insertBefore(IBeforeJ ? J : I);
2551 Type *ArgType = ArgTypeL;
2552 if (numElemL < numElemH) {
2553 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH,
2554 ArgTypeL, VArgType, IBeforeJ, 1)) {
2555 // This is another short-circuit case: we're combining a scalar into
2556 // a vector that is formed by an IE chain. We've just expanded the IE
2557 // chain, now insert the scalar and we're done.
2559 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0,
2560 getReplacementName(IBeforeJ ? I : J, true, o));
2561 S->insertBefore(IBeforeJ ? J : I);
2563 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL,
2564 ArgTypeH, IBeforeJ)) {
2565 // The two vector inputs to the shuffle must be the same length,
2566 // so extend the smaller vector to be the same length as the larger one.
2570 std::vector<Constant *> Mask(numElemH);
2572 for (; v < numElemL; ++v)
2573 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2574 for (; v < numElemH; ++v)
2575 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2577 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL),
2578 ConstantVector::get(Mask),
2579 getReplacementName(IBeforeJ ? I : J,
2582 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0,
2583 getReplacementName(IBeforeJ ? I : J,
2587 NLOp->insertBefore(IBeforeJ ? J : I);
2592 } else if (numElemL > numElemH) {
2593 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL,
2594 ArgTypeH, VArgType, IBeforeJ)) {
2596 InsertElementInst::Create(LOp, HOp,
2597 ConstantInt::get(Type::getInt32Ty(Context),
2599 getReplacementName(IBeforeJ ? I : J,
2601 S->insertBefore(IBeforeJ ? J : I);
2603 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH,
2604 ArgTypeL, IBeforeJ)) {
2607 std::vector<Constant *> Mask(numElemL);
2609 for (; v < numElemH; ++v)
2610 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2611 for (; v < numElemL; ++v)
2612 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2614 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH),
2615 ConstantVector::get(Mask),
2616 getReplacementName(IBeforeJ ? I : J,
2619 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0,
2620 getReplacementName(IBeforeJ ? I : J,
2624 NHOp->insertBefore(IBeforeJ ? J : I);
2629 if (ArgType->isVectorTy()) {
2630 unsigned numElem = cast<VectorType>(VArgType)->getNumElements();
2631 std::vector<Constant*> Mask(numElem);
2632 for (unsigned v = 0; v < numElem; ++v) {
2634 // If the low vector was expanded, we need to skip the extra
2635 // undefined entries.
2636 if (v >= numElemL && numElemH > numElemL)
2637 Idx += (numElemH - numElemL);
2638 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2641 Instruction *BV = new ShuffleVectorInst(LOp, HOp,
2642 ConstantVector::get(Mask),
2643 getReplacementName(IBeforeJ ? I : J, true, o));
2644 BV->insertBefore(IBeforeJ ? J : I);
2648 Instruction *BV1 = InsertElementInst::Create(
2649 UndefValue::get(VArgType), LOp, CV0,
2650 getReplacementName(IBeforeJ ? I : J,
2652 BV1->insertBefore(IBeforeJ ? J : I);
2653 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1,
2654 getReplacementName(IBeforeJ ? I : J,
2656 BV2->insertBefore(IBeforeJ ? J : I);
2660 // This function creates an array of values that will be used as the inputs
2661 // to the vector instruction that fuses I with J.
2662 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context,
2663 Instruction *I, Instruction *J,
2664 SmallVector<Value *, 3> &ReplacedOperands,
2666 unsigned NumOperands = I->getNumOperands();
2668 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) {
2669 // Iterate backward so that we look at the store pointer
2670 // first and know whether or not we need to flip the inputs.
2672 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) {
2673 // This is the pointer for a load/store instruction.
2674 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o);
2676 } else if (isa<CallInst>(I)) {
2677 Function *F = cast<CallInst>(I)->getCalledFunction();
2678 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID();
2679 if (o == NumOperands-1) {
2680 BasicBlock &BB = *I->getParent();
2682 Module *M = BB.getParent()->getParent();
2683 Type *ArgTypeI = I->getType();
2684 Type *ArgTypeJ = J->getType();
2685 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2687 ReplacedOperands[o] = Intrinsic::getDeclaration(M, IID, VArgType);
2689 } else if (IID == Intrinsic::powi && o == 1) {
2690 // The second argument of powi is a single integer and we've already
2691 // checked that both arguments are equal. As a result, we just keep
2692 // I's second argument.
2693 ReplacedOperands[o] = I->getOperand(o);
2696 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) {
2697 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J);
2701 ReplacedOperands[o] = getReplacementInput(Context, I, J, o, IBeforeJ);
2705 // This function creates two values that represent the outputs of the
2706 // original I and J instructions. These are generally vector shuffles
2707 // or extracts. In many cases, these will end up being unused and, thus,
2708 // eliminated by later passes.
2709 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
2710 Instruction *J, Instruction *K,
2711 Instruction *&InsertionPt,
2712 Instruction *&K1, Instruction *&K2) {
2713 if (isa<StoreInst>(I)) {
2714 AA->replaceWithNewValue(I, K);
2715 AA->replaceWithNewValue(J, K);
2717 Type *IType = I->getType();
2718 Type *JType = J->getType();
2720 VectorType *VType = getVecTypeForPair(IType, JType);
2721 unsigned numElem = VType->getNumElements();
2723 unsigned numElemI, numElemJ;
2724 if (IType->isVectorTy())
2725 numElemI = cast<VectorType>(IType)->getNumElements();
2729 if (JType->isVectorTy())
2730 numElemJ = cast<VectorType>(JType)->getNumElements();
2734 if (IType->isVectorTy()) {
2735 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI);
2736 for (unsigned v = 0; v < numElemI; ++v) {
2737 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2738 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v);
2741 K1 = new ShuffleVectorInst(K, UndefValue::get(VType),
2742 ConstantVector::get( Mask1),
2743 getReplacementName(K, false, 1));
2745 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2746 K1 = ExtractElementInst::Create(K, CV0,
2747 getReplacementName(K, false, 1));
2750 if (JType->isVectorTy()) {
2751 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ);
2752 for (unsigned v = 0; v < numElemJ; ++v) {
2753 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2754 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v);
2757 K2 = new ShuffleVectorInst(K, UndefValue::get(VType),
2758 ConstantVector::get( Mask2),
2759 getReplacementName(K, false, 2));
2761 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1);
2762 K2 = ExtractElementInst::Create(K, CV1,
2763 getReplacementName(K, false, 2));
2767 K2->insertAfter(K1);
2772 // Move all uses of the function I (including pairing-induced uses) after J.
2773 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB,
2774 DenseSet<ValuePair> &LoadMoveSetPairs,
2775 Instruction *I, Instruction *J) {
2776 // Skip to the first instruction past I.
2777 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2779 DenseSet<Value *> Users;
2780 AliasSetTracker WriteSet(*AA);
2781 for (; cast<Instruction>(L) != J; ++L)
2782 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs);
2784 assert(cast<Instruction>(L) == J &&
2785 "Tracking has not proceeded far enough to check for dependencies");
2786 // If J is now in the use set of I, then trackUsesOfI will return true
2787 // and we have a dependency cycle (and the fusing operation must abort).
2788 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSetPairs);
2791 // Move all uses of the function I (including pairing-induced uses) after J.
2792 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB,
2793 DenseSet<ValuePair> &LoadMoveSetPairs,
2794 Instruction *&InsertionPt,
2795 Instruction *I, Instruction *J) {
2796 // Skip to the first instruction past I.
2797 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2799 DenseSet<Value *> Users;
2800 AliasSetTracker WriteSet(*AA);
2801 for (; cast<Instruction>(L) != J;) {
2802 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs)) {
2803 // Move this instruction
2804 Instruction *InstToMove = L; ++L;
2806 DEBUG(dbgs() << "BBV: moving: " << *InstToMove <<
2807 " to after " << *InsertionPt << "\n");
2808 InstToMove->removeFromParent();
2809 InstToMove->insertAfter(InsertionPt);
2810 InsertionPt = InstToMove;
2817 // Collect all load instruction that are in the move set of a given first
2818 // pair member. These loads depend on the first instruction, I, and so need
2819 // to be moved after J (the second instruction) when the pair is fused.
2820 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB,
2821 DenseMap<Value *, Value *> &ChosenPairs,
2822 std::multimap<Value *, Value *> &LoadMoveSet,
2823 DenseSet<ValuePair> &LoadMoveSetPairs,
2825 // Skip to the first instruction past I.
2826 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2828 DenseSet<Value *> Users;
2829 AliasSetTracker WriteSet(*AA);
2831 // Note: We cannot end the loop when we reach J because J could be moved
2832 // farther down the use chain by another instruction pairing. Also, J
2833 // could be before I if this is an inverted input.
2834 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) {
2835 if (trackUsesOfI(Users, WriteSet, I, L)) {
2836 if (L->mayReadFromMemory()) {
2837 LoadMoveSet.insert(ValuePair(L, I));
2838 LoadMoveSetPairs.insert(ValuePair(L, I));
2844 // In cases where both load/stores and the computation of their pointers
2845 // are chosen for vectorization, we can end up in a situation where the
2846 // aliasing analysis starts returning different query results as the
2847 // process of fusing instruction pairs continues. Because the algorithm
2848 // relies on finding the same use trees here as were found earlier, we'll
2849 // need to precompute the necessary aliasing information here and then
2850 // manually update it during the fusion process.
2851 void BBVectorize::collectLoadMoveSet(BasicBlock &BB,
2852 std::vector<Value *> &PairableInsts,
2853 DenseMap<Value *, Value *> &ChosenPairs,
2854 std::multimap<Value *, Value *> &LoadMoveSet,
2855 DenseSet<ValuePair> &LoadMoveSetPairs) {
2856 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
2857 PIE = PairableInsts.end(); PI != PIE; ++PI) {
2858 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI);
2859 if (P == ChosenPairs.end()) continue;
2861 Instruction *I = cast<Instruction>(P->first);
2862 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet,
2863 LoadMoveSetPairs, I);
2867 // When the first instruction in each pair is cloned, it will inherit its
2868 // parent's metadata. This metadata must be combined with that of the other
2869 // instruction in a safe way.
2870 void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) {
2871 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata;
2872 K->getAllMetadataOtherThanDebugLoc(Metadata);
2873 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) {
2874 unsigned Kind = Metadata[i].first;
2875 MDNode *JMD = J->getMetadata(Kind);
2876 MDNode *KMD = Metadata[i].second;
2880 K->setMetadata(Kind, 0); // Remove unknown metadata
2882 case LLVMContext::MD_tbaa:
2883 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD));
2885 case LLVMContext::MD_fpmath:
2886 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD));
2892 // This function fuses the chosen instruction pairs into vector instructions,
2893 // taking care preserve any needed scalar outputs and, then, it reorders the
2894 // remaining instructions as needed (users of the first member of the pair
2895 // need to be moved to after the location of the second member of the pair
2896 // because the vector instruction is inserted in the location of the pair's
2898 void BBVectorize::fuseChosenPairs(BasicBlock &BB,
2899 std::vector<Value *> &PairableInsts,
2900 DenseMap<Value *, Value *> &ChosenPairs,
2901 DenseSet<ValuePair> &FixedOrderPairs,
2902 DenseMap<VPPair, unsigned> &PairConnectionTypes,
2903 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
2904 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps) {
2905 LLVMContext& Context = BB.getContext();
2907 // During the vectorization process, the order of the pairs to be fused
2908 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs
2909 // list. After a pair is fused, the flipped pair is removed from the list.
2910 DenseSet<ValuePair> FlippedPairs;
2911 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(),
2912 E = ChosenPairs.end(); P != E; ++P)
2913 FlippedPairs.insert(ValuePair(P->second, P->first));
2914 for (DenseSet<ValuePair>::iterator P = FlippedPairs.begin(),
2915 E = FlippedPairs.end(); P != E; ++P)
2916 ChosenPairs.insert(*P);
2918 std::multimap<Value *, Value *> LoadMoveSet;
2919 DenseSet<ValuePair> LoadMoveSetPairs;
2920 collectLoadMoveSet(BB, PairableInsts, ChosenPairs,
2921 LoadMoveSet, LoadMoveSetPairs);
2923 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n");
2925 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) {
2926 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI);
2927 if (P == ChosenPairs.end()) {
2932 if (getDepthFactor(P->first) == 0) {
2933 // These instructions are not really fused, but are tracked as though
2934 // they are. Any case in which it would be interesting to fuse them
2935 // will be taken care of by InstCombine.
2941 Instruction *I = cast<Instruction>(P->first),
2942 *J = cast<Instruction>(P->second);
2944 DEBUG(dbgs() << "BBV: fusing: " << *I <<
2945 " <-> " << *J << "\n");
2947 // Remove the pair and flipped pair from the list.
2948 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second);
2949 assert(FP != ChosenPairs.end() && "Flipped pair not found in list");
2950 ChosenPairs.erase(FP);
2951 ChosenPairs.erase(P);
2953 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSetPairs, I, J)) {
2954 DEBUG(dbgs() << "BBV: fusion of: " << *I <<
2956 " aborted because of non-trivial dependency cycle\n");
2962 // If the pair must have the other order, then flip it.
2963 bool FlipPairOrder = FixedOrderPairs.count(ValuePair(J, I));
2964 if (!FlipPairOrder && !FixedOrderPairs.count(ValuePair(I, J))) {
2965 // This pair does not have a fixed order, and so we might want to
2966 // flip it if that will yield fewer shuffles. We count the number
2967 // of dependencies connected via swaps, and those directly connected,
2968 // and flip the order if the number of swaps is greater.
2969 bool OrigOrder = true;
2970 VPPIteratorPair IP = ConnectedPairDeps.equal_range(ValuePair(I, J));
2971 if (IP.first == ConnectedPairDeps.end()) {
2972 IP = ConnectedPairDeps.equal_range(ValuePair(J, I));
2976 if (IP.first != ConnectedPairDeps.end()) {
2977 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
2978 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first;
2979 Q != IP.second; ++Q) {
2980 DenseMap<VPPair, unsigned>::iterator R =
2981 PairConnectionTypes.find(VPPair(Q->second, Q->first));
2982 assert(R != PairConnectionTypes.end() &&
2983 "Cannot find pair connection type");
2984 if (R->second == PairConnectionDirect)
2986 else if (R->second == PairConnectionSwap)
2991 std::swap(NumDepsDirect, NumDepsSwap);
2993 if (NumDepsSwap > NumDepsDirect) {
2994 FlipPairOrder = true;
2995 DEBUG(dbgs() << "BBV: reordering pair: " << *I <<
2996 " <-> " << *J << "\n");
3001 Instruction *L = I, *H = J;
3005 // If the pair being fused uses the opposite order from that in the pair
3006 // connection map, then we need to flip the types.
3007 VPPIteratorPair IP = ConnectedPairs.equal_range(ValuePair(H, L));
3008 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first;
3009 Q != IP.second; ++Q) {
3010 DenseMap<VPPair, unsigned>::iterator R = PairConnectionTypes.find(*Q);
3011 assert(R != PairConnectionTypes.end() &&
3012 "Cannot find pair connection type");
3013 if (R->second == PairConnectionDirect)
3014 R->second = PairConnectionSwap;
3015 else if (R->second == PairConnectionSwap)
3016 R->second = PairConnectionDirect;
3019 bool LBeforeH = !FlipPairOrder;
3020 unsigned NumOperands = I->getNumOperands();
3021 SmallVector<Value *, 3> ReplacedOperands(NumOperands);
3022 getReplacementInputsForPair(Context, L, H, ReplacedOperands,
3025 // Make a copy of the original operation, change its type to the vector
3026 // type and replace its operands with the vector operands.
3027 Instruction *K = L->clone();
3030 else if (H->hasName())
3033 if (!isa<StoreInst>(K))
3034 K->mutateType(getVecTypeForPair(L->getType(), H->getType()));
3036 combineMetadata(K, H);
3037 K->intersectOptionalDataWith(H);
3039 for (unsigned o = 0; o < NumOperands; ++o)
3040 K->setOperand(o, ReplacedOperands[o]);
3044 // Instruction insertion point:
3045 Instruction *InsertionPt = K;
3046 Instruction *K1 = 0, *K2 = 0;
3047 replaceOutputsOfPair(Context, L, H, K, InsertionPt, K1, K2);
3049 // The use tree of the first original instruction must be moved to after
3050 // the location of the second instruction. The entire use tree of the
3051 // first instruction is disjoint from the input tree of the second
3052 // (by definition), and so commutes with it.
3054 moveUsesOfIAfterJ(BB, LoadMoveSetPairs, InsertionPt, I, J);
3056 if (!isa<StoreInst>(I)) {
3057 L->replaceAllUsesWith(K1);
3058 H->replaceAllUsesWith(K2);
3059 AA->replaceWithNewValue(L, K1);
3060 AA->replaceWithNewValue(H, K2);
3063 // Instructions that may read from memory may be in the load move set.
3064 // Once an instruction is fused, we no longer need its move set, and so
3065 // the values of the map never need to be updated. However, when a load
3066 // is fused, we need to merge the entries from both instructions in the
3067 // pair in case those instructions were in the move set of some other
3068 // yet-to-be-fused pair. The loads in question are the keys of the map.
3069 if (I->mayReadFromMemory()) {
3070 std::vector<ValuePair> NewSetMembers;
3071 VPIteratorPair IPairRange = LoadMoveSet.equal_range(I);
3072 VPIteratorPair JPairRange = LoadMoveSet.equal_range(J);
3073 for (std::multimap<Value *, Value *>::iterator N = IPairRange.first;
3074 N != IPairRange.second; ++N)
3075 NewSetMembers.push_back(ValuePair(K, N->second));
3076 for (std::multimap<Value *, Value *>::iterator N = JPairRange.first;
3077 N != JPairRange.second; ++N)
3078 NewSetMembers.push_back(ValuePair(K, N->second));
3079 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(),
3080 AE = NewSetMembers.end(); A != AE; ++A) {
3081 LoadMoveSet.insert(*A);
3082 LoadMoveSetPairs.insert(*A);
3086 // Before removing I, set the iterator to the next instruction.
3087 PI = llvm::next(BasicBlock::iterator(I));
3088 if (cast<Instruction>(PI) == J)
3093 I->eraseFromParent();
3094 J->eraseFromParent();
3096 DEBUG(if (PrintAfterEveryPair) dbgs() << "BBV: block is now: \n" <<
3100 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n");
3104 char BBVectorize::ID = 0;
3105 static const char bb_vectorize_name[] = "Basic-Block Vectorization";
3106 INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
3107 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
3108 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
3109 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
3110 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3111 INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
3113 BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) {
3114 return new BBVectorize(C);
3118 llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) {
3119 BBVectorize BBVectorizer(P, C);
3120 return BBVectorizer.vectorizeBB(BB);
3123 //===----------------------------------------------------------------------===//
3124 VectorizeConfig::VectorizeConfig() {
3125 VectorBits = ::VectorBits;
3126 VectorizeBools = !::NoBools;
3127 VectorizeInts = !::NoInts;
3128 VectorizeFloats = !::NoFloats;
3129 VectorizePointers = !::NoPointers;
3130 VectorizeCasts = !::NoCasts;
3131 VectorizeMath = !::NoMath;
3132 VectorizeFMA = !::NoFMA;
3133 VectorizeSelect = !::NoSelect;
3134 VectorizeCmp = !::NoCmp;
3135 VectorizeGEP = !::NoGEP;
3136 VectorizeMemOps = !::NoMemOps;
3137 AlignedOnly = ::AlignedOnly;
3138 ReqChainDepth= ::ReqChainDepth;
3139 SearchLimit = ::SearchLimit;
3140 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck;
3141 SplatBreaksChain = ::SplatBreaksChain;
3142 MaxInsts = ::MaxInsts;
3143 MaxIter = ::MaxIter;
3144 Pow2LenOnly = ::Pow2LenOnly;
3145 NoMemOpBoost = ::NoMemOpBoost;
3146 FastDep = ::FastDep;