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 std::multimap<Value *, 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(std::multimap<Value *, Value *> &CandidatePairs,
243 std::vector<Value *> &PairableInsts,
244 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
245 DenseMap<VPPair, unsigned> &PairConnectionTypes);
247 void buildDepMap(BasicBlock &BB,
248 std::multimap<Value *, Value *> &CandidatePairs,
249 std::vector<Value *> &PairableInsts,
250 DenseSet<ValuePair> &PairableInstUsers);
252 void choosePairs(std::multimap<Value *, Value *> &CandidatePairs,
253 DenseMap<ValuePair, int> &CandidatePairCostSavings,
254 std::vector<Value *> &PairableInsts,
255 DenseSet<ValuePair> &FixedOrderPairs,
256 DenseMap<VPPair, unsigned> &PairConnectionTypes,
257 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
258 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps,
259 DenseSet<ValuePair> &PairableInstUsers,
260 DenseMap<Value *, Value *>& ChosenPairs);
262 void fuseChosenPairs(BasicBlock &BB,
263 std::vector<Value *> &PairableInsts,
264 DenseMap<Value *, Value *>& ChosenPairs,
265 DenseSet<ValuePair> &FixedOrderPairs,
266 DenseMap<VPPair, unsigned> &PairConnectionTypes,
267 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
268 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps);
271 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore);
273 bool areInstsCompatible(Instruction *I, Instruction *J,
274 bool IsSimpleLoadStore, bool NonPow2Len,
275 int &CostSavings, int &FixedOrder);
277 bool trackUsesOfI(DenseSet<Value *> &Users,
278 AliasSetTracker &WriteSet, Instruction *I,
279 Instruction *J, bool UpdateUsers = true,
280 std::multimap<Value *, Value *> *LoadMoveSet = 0);
282 void computePairsConnectedTo(
283 std::multimap<Value *, Value *> &CandidatePairs,
284 std::vector<Value *> &PairableInsts,
285 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
286 DenseMap<VPPair, unsigned> &PairConnectionTypes,
289 bool pairsConflict(ValuePair P, ValuePair Q,
290 DenseSet<ValuePair> &PairableInstUsers,
291 std::multimap<ValuePair, ValuePair> *PairableInstUserMap = 0,
292 DenseSet<VPPair> *PairableInstUserPairSet = 0);
294 bool pairWillFormCycle(ValuePair P,
295 std::multimap<ValuePair, ValuePair> &PairableInstUsers,
296 DenseSet<ValuePair> &CurrentPairs);
299 std::multimap<Value *, Value *> &CandidatePairs,
300 std::vector<Value *> &PairableInsts,
301 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
302 DenseSet<ValuePair> &PairableInstUsers,
303 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
304 DenseSet<VPPair> &PairableInstUserPairSet,
305 DenseMap<Value *, Value *> &ChosenPairs,
306 DenseMap<ValuePair, size_t> &Tree,
307 DenseSet<ValuePair> &PrunedTree, ValuePair J,
310 void buildInitialTreeFor(
311 std::multimap<Value *, Value *> &CandidatePairs,
312 std::vector<Value *> &PairableInsts,
313 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
314 DenseSet<ValuePair> &PairableInstUsers,
315 DenseMap<Value *, Value *> &ChosenPairs,
316 DenseMap<ValuePair, size_t> &Tree, ValuePair J);
318 void findBestTreeFor(
319 std::multimap<Value *, Value *> &CandidatePairs,
320 DenseMap<ValuePair, int> &CandidatePairCostSavings,
321 std::vector<Value *> &PairableInsts,
322 DenseSet<ValuePair> &FixedOrderPairs,
323 DenseMap<VPPair, unsigned> &PairConnectionTypes,
324 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
325 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps,
326 DenseSet<ValuePair> &PairableInstUsers,
327 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
328 DenseSet<VPPair> &PairableInstUserPairSet,
329 DenseMap<Value *, Value *> &ChosenPairs,
330 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth,
331 int &BestEffSize, VPIteratorPair ChoiceRange,
334 Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I,
335 Instruction *J, unsigned o);
337 void fillNewShuffleMask(LLVMContext& Context, Instruction *J,
338 unsigned MaskOffset, unsigned NumInElem,
339 unsigned NumInElem1, unsigned IdxOffset,
340 std::vector<Constant*> &Mask);
342 Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I,
345 bool expandIEChain(LLVMContext& Context, Instruction *I, Instruction *J,
346 unsigned o, Value *&LOp, unsigned numElemL,
347 Type *ArgTypeL, Type *ArgTypeR, bool IBeforeJ,
348 unsigned IdxOff = 0);
350 Value *getReplacementInput(LLVMContext& Context, Instruction *I,
351 Instruction *J, unsigned o, bool IBeforeJ);
353 void getReplacementInputsForPair(LLVMContext& Context, Instruction *I,
354 Instruction *J, SmallVector<Value *, 3> &ReplacedOperands,
357 void replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
358 Instruction *J, Instruction *K,
359 Instruction *&InsertionPt, Instruction *&K1,
362 void collectPairLoadMoveSet(BasicBlock &BB,
363 DenseMap<Value *, Value *> &ChosenPairs,
364 std::multimap<Value *, Value *> &LoadMoveSet,
367 void collectLoadMoveSet(BasicBlock &BB,
368 std::vector<Value *> &PairableInsts,
369 DenseMap<Value *, Value *> &ChosenPairs,
370 std::multimap<Value *, Value *> &LoadMoveSet);
372 bool canMoveUsesOfIAfterJ(BasicBlock &BB,
373 std::multimap<Value *, Value *> &LoadMoveSet,
374 Instruction *I, Instruction *J);
376 void moveUsesOfIAfterJ(BasicBlock &BB,
377 std::multimap<Value *, Value *> &LoadMoveSet,
378 Instruction *&InsertionPt,
379 Instruction *I, Instruction *J);
381 void combineMetadata(Instruction *K, const Instruction *J);
383 bool vectorizeBB(BasicBlock &BB) {
384 if (!DT->isReachableFromEntry(&BB)) {
385 DEBUG(dbgs() << "BBV: skipping unreachable " << BB.getName() <<
386 " in " << BB.getParent()->getName() << "\n");
390 DEBUG(if (TTI) dbgs() << "BBV: using target information\n");
392 bool changed = false;
393 // Iterate a sufficient number of times to merge types of size 1 bit,
394 // then 2 bits, then 4, etc. up to half of the target vector width of the
395 // target vector register.
398 (TTI || v <= Config.VectorBits) &&
399 (!Config.MaxIter || n <= Config.MaxIter);
401 DEBUG(dbgs() << "BBV: fusing loop #" << n <<
402 " for " << BB.getName() << " in " <<
403 BB.getParent()->getName() << "...\n");
404 if (vectorizePairs(BB))
410 if (changed && !Pow2LenOnly) {
412 for (; !Config.MaxIter || n <= Config.MaxIter; ++n) {
413 DEBUG(dbgs() << "BBV: fusing for non-2^n-length vectors loop #: " <<
414 n << " for " << BB.getName() << " in " <<
415 BB.getParent()->getName() << "...\n");
416 if (!vectorizePairs(BB, true)) break;
420 DEBUG(dbgs() << "BBV: done!\n");
424 virtual bool runOnBasicBlock(BasicBlock &BB) {
425 AA = &getAnalysis<AliasAnalysis>();
426 DT = &getAnalysis<DominatorTree>();
427 SE = &getAnalysis<ScalarEvolution>();
428 TD = getAnalysisIfAvailable<DataLayout>();
429 TTI = IgnoreTargetInfo ? 0 : &getAnalysis<TargetTransformInfo>();
431 return vectorizeBB(BB);
434 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
435 BasicBlockPass::getAnalysisUsage(AU);
436 AU.addRequired<AliasAnalysis>();
437 AU.addRequired<DominatorTree>();
438 AU.addRequired<ScalarEvolution>();
439 AU.addRequired<TargetTransformInfo>();
440 AU.addPreserved<AliasAnalysis>();
441 AU.addPreserved<DominatorTree>();
442 AU.addPreserved<ScalarEvolution>();
443 AU.setPreservesCFG();
446 static inline VectorType *getVecTypeForPair(Type *ElemTy, Type *Elem2Ty) {
447 assert(ElemTy->getScalarType() == Elem2Ty->getScalarType() &&
448 "Cannot form vector from incompatible scalar types");
449 Type *STy = ElemTy->getScalarType();
452 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) {
453 numElem = VTy->getNumElements();
458 if (VectorType *VTy = dyn_cast<VectorType>(Elem2Ty)) {
459 numElem += VTy->getNumElements();
464 return VectorType::get(STy, numElem);
467 static inline void getInstructionTypes(Instruction *I,
468 Type *&T1, Type *&T2) {
469 if (isa<StoreInst>(I)) {
470 // For stores, it is the value type, not the pointer type that matters
471 // because the value is what will come from a vector register.
473 Value *IVal = cast<StoreInst>(I)->getValueOperand();
474 T1 = IVal->getType();
480 T2 = cast<CastInst>(I)->getSrcTy();
484 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
485 T2 = SI->getCondition()->getType();
486 } else if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(I)) {
487 T2 = SI->getOperand(0)->getType();
488 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) {
489 T2 = CI->getOperand(0)->getType();
493 // Returns the weight associated with the provided value. A chain of
494 // candidate pairs has a length given by the sum of the weights of its
495 // members (one weight per pair; the weight of each member of the pair
496 // is assumed to be the same). This length is then compared to the
497 // chain-length threshold to determine if a given chain is significant
498 // enough to be vectorized. The length is also used in comparing
499 // candidate chains where longer chains are considered to be better.
500 // Note: when this function returns 0, the resulting instructions are
501 // not actually fused.
502 inline size_t getDepthFactor(Value *V) {
503 // InsertElement and ExtractElement have a depth factor of zero. This is
504 // for two reasons: First, they cannot be usefully fused. Second, because
505 // the pass generates a lot of these, they can confuse the simple metric
506 // used to compare the trees in the next iteration. Thus, giving them a
507 // weight of zero allows the pass to essentially ignore them in
508 // subsequent iterations when looking for vectorization opportunities
509 // while still tracking dependency chains that flow through those
511 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V))
514 // Give a load or store half of the required depth so that load/store
515 // pairs will vectorize.
516 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V)))
517 return Config.ReqChainDepth/2;
522 // Returns the cost of the provided instruction using TTI.
523 // This does not handle loads and stores.
524 unsigned getInstrCost(unsigned Opcode, Type *T1, Type *T2) {
527 case Instruction::GetElementPtr:
528 // We mark this instruction as zero-cost because scalar GEPs are usually
529 // lowered to the intruction addressing mode. At the moment we don't
530 // generate vector GEPs.
532 case Instruction::Br:
533 return TTI->getCFInstrCost(Opcode);
534 case Instruction::PHI:
536 case Instruction::Add:
537 case Instruction::FAdd:
538 case Instruction::Sub:
539 case Instruction::FSub:
540 case Instruction::Mul:
541 case Instruction::FMul:
542 case Instruction::UDiv:
543 case Instruction::SDiv:
544 case Instruction::FDiv:
545 case Instruction::URem:
546 case Instruction::SRem:
547 case Instruction::FRem:
548 case Instruction::Shl:
549 case Instruction::LShr:
550 case Instruction::AShr:
551 case Instruction::And:
552 case Instruction::Or:
553 case Instruction::Xor:
554 return TTI->getArithmeticInstrCost(Opcode, T1);
555 case Instruction::Select:
556 case Instruction::ICmp:
557 case Instruction::FCmp:
558 return TTI->getCmpSelInstrCost(Opcode, T1, T2);
559 case Instruction::ZExt:
560 case Instruction::SExt:
561 case Instruction::FPToUI:
562 case Instruction::FPToSI:
563 case Instruction::FPExt:
564 case Instruction::PtrToInt:
565 case Instruction::IntToPtr:
566 case Instruction::SIToFP:
567 case Instruction::UIToFP:
568 case Instruction::Trunc:
569 case Instruction::FPTrunc:
570 case Instruction::BitCast:
571 case Instruction::ShuffleVector:
572 return TTI->getCastInstrCost(Opcode, T1, T2);
578 // This determines the relative offset of two loads or stores, returning
579 // true if the offset could be determined to be some constant value.
580 // For example, if OffsetInElmts == 1, then J accesses the memory directly
581 // after I; if OffsetInElmts == -1 then I accesses the memory
583 bool getPairPtrInfo(Instruction *I, Instruction *J,
584 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment,
585 unsigned &IAddressSpace, unsigned &JAddressSpace,
586 int64_t &OffsetInElmts, bool ComputeOffset = true) {
588 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
589 LoadInst *LJ = cast<LoadInst>(J);
590 IPtr = LI->getPointerOperand();
591 JPtr = LJ->getPointerOperand();
592 IAlignment = LI->getAlignment();
593 JAlignment = LJ->getAlignment();
594 IAddressSpace = LI->getPointerAddressSpace();
595 JAddressSpace = LJ->getPointerAddressSpace();
597 StoreInst *SI = cast<StoreInst>(I), *SJ = cast<StoreInst>(J);
598 IPtr = SI->getPointerOperand();
599 JPtr = SJ->getPointerOperand();
600 IAlignment = SI->getAlignment();
601 JAlignment = SJ->getAlignment();
602 IAddressSpace = SI->getPointerAddressSpace();
603 JAddressSpace = SJ->getPointerAddressSpace();
609 const SCEV *IPtrSCEV = SE->getSCEV(IPtr);
610 const SCEV *JPtrSCEV = SE->getSCEV(JPtr);
612 // If this is a trivial offset, then we'll get something like
613 // 1*sizeof(type). With target data, which we need anyway, this will get
614 // constant folded into a number.
615 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV);
616 if (const SCEVConstant *ConstOffSCEV =
617 dyn_cast<SCEVConstant>(OffsetSCEV)) {
618 ConstantInt *IntOff = ConstOffSCEV->getValue();
619 int64_t Offset = IntOff->getSExtValue();
621 Type *VTy = cast<PointerType>(IPtr->getType())->getElementType();
622 int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy);
624 Type *VTy2 = cast<PointerType>(JPtr->getType())->getElementType();
625 if (VTy != VTy2 && Offset < 0) {
626 int64_t VTy2TSS = (int64_t) TD->getTypeStoreSize(VTy2);
627 OffsetInElmts = Offset/VTy2TSS;
628 return (abs64(Offset) % VTy2TSS) == 0;
631 OffsetInElmts = Offset/VTyTSS;
632 return (abs64(Offset) % VTyTSS) == 0;
638 // Returns true if the provided CallInst represents an intrinsic that can
640 bool isVectorizableIntrinsic(CallInst* I) {
641 Function *F = I->getCalledFunction();
642 if (!F) return false;
644 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID();
645 if (!IID) return false;
650 case Intrinsic::sqrt:
651 case Intrinsic::powi:
655 case Intrinsic::log2:
656 case Intrinsic::log10:
658 case Intrinsic::exp2:
660 return Config.VectorizeMath;
662 case Intrinsic::fmuladd:
663 return Config.VectorizeFMA;
667 // Returns true if J is the second element in some pair referenced by
668 // some multimap pair iterator pair.
669 template <typename V>
670 bool isSecondInIteratorPair(V J, std::pair<
671 typename std::multimap<V, V>::iterator,
672 typename std::multimap<V, V>::iterator> PairRange) {
673 for (typename std::multimap<V, V>::iterator K = PairRange.first;
674 K != PairRange.second; ++K)
675 if (K->second == J) return true;
680 bool isPureIEChain(InsertElementInst *IE) {
681 InsertElementInst *IENext = IE;
683 if (!isa<UndefValue>(IENext->getOperand(0)) &&
684 !isa<InsertElementInst>(IENext->getOperand(0))) {
688 dyn_cast<InsertElementInst>(IENext->getOperand(0))));
694 // This function implements one vectorization iteration on the provided
695 // basic block. It returns true if the block is changed.
696 bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) {
698 BasicBlock::iterator Start = BB.getFirstInsertionPt();
700 std::vector<Value *> AllPairableInsts;
701 DenseMap<Value *, Value *> AllChosenPairs;
702 DenseSet<ValuePair> AllFixedOrderPairs;
703 DenseMap<VPPair, unsigned> AllPairConnectionTypes;
704 std::multimap<ValuePair, ValuePair> AllConnectedPairs, AllConnectedPairDeps;
707 std::vector<Value *> PairableInsts;
708 std::multimap<Value *, Value *> CandidatePairs;
709 DenseSet<ValuePair> FixedOrderPairs;
710 DenseMap<ValuePair, int> CandidatePairCostSavings;
711 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs,
713 CandidatePairCostSavings,
714 PairableInsts, NonPow2Len);
715 if (PairableInsts.empty()) continue;
717 // Now we have a map of all of the pairable instructions and we need to
718 // select the best possible pairing. A good pairing is one such that the
719 // users of the pair are also paired. This defines a (directed) forest
720 // over the pairs such that two pairs are connected iff the second pair
723 // Note that it only matters that both members of the second pair use some
724 // element of the first pair (to allow for splatting).
726 std::multimap<ValuePair, ValuePair> ConnectedPairs, ConnectedPairDeps;
727 DenseMap<VPPair, unsigned> PairConnectionTypes;
728 computeConnectedPairs(CandidatePairs, PairableInsts, ConnectedPairs,
729 PairConnectionTypes);
730 if (ConnectedPairs.empty()) continue;
732 for (std::multimap<ValuePair, ValuePair>::iterator
733 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
735 ConnectedPairDeps.insert(VPPair(I->second, I->first));
738 // Build the pairable-instruction dependency map
739 DenseSet<ValuePair> PairableInstUsers;
740 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers);
742 // There is now a graph of the connected pairs. For each variable, pick
743 // the pairing with the largest tree meeting the depth requirement on at
744 // least one branch. Then select all pairings that are part of that tree
745 // and remove them from the list of available pairings and pairable
748 DenseMap<Value *, Value *> ChosenPairs;
749 choosePairs(CandidatePairs, CandidatePairCostSavings,
750 PairableInsts, FixedOrderPairs, PairConnectionTypes,
751 ConnectedPairs, ConnectedPairDeps,
752 PairableInstUsers, ChosenPairs);
754 if (ChosenPairs.empty()) continue;
755 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(),
756 PairableInsts.end());
757 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end());
759 // Only for the chosen pairs, propagate information on fixed-order pairs,
760 // pair connections, and their types to the data structures used by the
761 // pair fusion procedures.
762 for (DenseMap<Value *, Value *>::iterator I = ChosenPairs.begin(),
763 IE = ChosenPairs.end(); I != IE; ++I) {
764 if (FixedOrderPairs.count(*I))
765 AllFixedOrderPairs.insert(*I);
766 else if (FixedOrderPairs.count(ValuePair(I->second, I->first)))
767 AllFixedOrderPairs.insert(ValuePair(I->second, I->first));
769 for (DenseMap<Value *, Value *>::iterator J = ChosenPairs.begin();
771 DenseMap<VPPair, unsigned>::iterator K =
772 PairConnectionTypes.find(VPPair(*I, *J));
773 if (K != PairConnectionTypes.end()) {
774 AllPairConnectionTypes.insert(*K);
776 K = PairConnectionTypes.find(VPPair(*J, *I));
777 if (K != PairConnectionTypes.end())
778 AllPairConnectionTypes.insert(*K);
783 for (std::multimap<ValuePair, ValuePair>::iterator
784 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
786 if (AllPairConnectionTypes.count(*I)) {
787 AllConnectedPairs.insert(*I);
788 AllConnectedPairDeps.insert(VPPair(I->second, I->first));
791 } while (ShouldContinue);
793 if (AllChosenPairs.empty()) return false;
794 NumFusedOps += AllChosenPairs.size();
796 // A set of pairs has now been selected. It is now necessary to replace the
797 // paired instructions with vector instructions. For this procedure each
798 // operand must be replaced with a vector operand. This vector is formed
799 // by using build_vector on the old operands. The replaced values are then
800 // replaced with a vector_extract on the result. Subsequent optimization
801 // passes should coalesce the build/extract combinations.
803 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs, AllFixedOrderPairs,
804 AllPairConnectionTypes,
805 AllConnectedPairs, AllConnectedPairDeps);
807 // It is important to cleanup here so that future iterations of this
808 // function have less work to do.
809 (void) SimplifyInstructionsInBlock(&BB, TD, AA->getTargetLibraryInfo());
813 // This function returns true if the provided instruction is capable of being
814 // fused into a vector instruction. This determination is based only on the
815 // type and other attributes of the instruction.
816 bool BBVectorize::isInstVectorizable(Instruction *I,
817 bool &IsSimpleLoadStore) {
818 IsSimpleLoadStore = false;
820 if (CallInst *C = dyn_cast<CallInst>(I)) {
821 if (!isVectorizableIntrinsic(C))
823 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) {
824 // Vectorize simple loads if possbile:
825 IsSimpleLoadStore = L->isSimple();
826 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
828 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) {
829 // Vectorize simple stores if possbile:
830 IsSimpleLoadStore = S->isSimple();
831 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
833 } else if (CastInst *C = dyn_cast<CastInst>(I)) {
834 // We can vectorize casts, but not casts of pointer types, etc.
835 if (!Config.VectorizeCasts)
838 Type *SrcTy = C->getSrcTy();
839 if (!SrcTy->isSingleValueType())
842 Type *DestTy = C->getDestTy();
843 if (!DestTy->isSingleValueType())
845 } else if (isa<SelectInst>(I)) {
846 if (!Config.VectorizeSelect)
848 } else if (isa<CmpInst>(I)) {
849 if (!Config.VectorizeCmp)
851 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) {
852 if (!Config.VectorizeGEP)
855 // Currently, vector GEPs exist only with one index.
856 if (G->getNumIndices() != 1)
858 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) ||
859 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) {
863 // We can't vectorize memory operations without target data
864 if (TD == 0 && IsSimpleLoadStore)
868 getInstructionTypes(I, T1, T2);
870 // Not every type can be vectorized...
871 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) ||
872 !(VectorType::isValidElementType(T2) || T2->isVectorTy()))
875 if (T1->getScalarSizeInBits() == 1) {
876 if (!Config.VectorizeBools)
879 if (!Config.VectorizeInts && T1->isIntOrIntVectorTy())
883 if (T2->getScalarSizeInBits() == 1) {
884 if (!Config.VectorizeBools)
887 if (!Config.VectorizeInts && T2->isIntOrIntVectorTy())
891 if (!Config.VectorizeFloats
892 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy()))
895 // Don't vectorize target-specific types.
896 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy())
898 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy())
901 if ((!Config.VectorizePointers || TD == 0) &&
902 (T1->getScalarType()->isPointerTy() ||
903 T2->getScalarType()->isPointerTy()))
906 if (!TTI && (T1->getPrimitiveSizeInBits() >= Config.VectorBits ||
907 T2->getPrimitiveSizeInBits() >= Config.VectorBits))
913 // This function returns true if the two provided instructions are compatible
914 // (meaning that they can be fused into a vector instruction). This assumes
915 // that I has already been determined to be vectorizable and that J is not
916 // in the use tree of I.
917 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J,
918 bool IsSimpleLoadStore, bool NonPow2Len,
919 int &CostSavings, int &FixedOrder) {
920 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I <<
921 " <-> " << *J << "\n");
926 // Loads and stores can be merged if they have different alignments,
927 // but are otherwise the same.
928 if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment |
929 (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0)))
932 Type *IT1, *IT2, *JT1, *JT2;
933 getInstructionTypes(I, IT1, IT2);
934 getInstructionTypes(J, JT1, JT2);
935 unsigned MaxTypeBits = std::max(
936 IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(),
937 IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits());
938 if (!TTI && MaxTypeBits > Config.VectorBits)
941 // FIXME: handle addsub-type operations!
943 if (IsSimpleLoadStore) {
945 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
946 int64_t OffsetInElmts = 0;
947 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
948 IAddressSpace, JAddressSpace,
949 OffsetInElmts) && abs64(OffsetInElmts) == 1) {
950 FixedOrder = (int) OffsetInElmts;
951 unsigned BottomAlignment = IAlignment;
952 if (OffsetInElmts < 0) BottomAlignment = JAlignment;
954 Type *aTypeI = isa<StoreInst>(I) ?
955 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType();
956 Type *aTypeJ = isa<StoreInst>(J) ?
957 cast<StoreInst>(J)->getValueOperand()->getType() : J->getType();
958 Type *VType = getVecTypeForPair(aTypeI, aTypeJ);
960 if (Config.AlignedOnly) {
961 // An aligned load or store is possible only if the instruction
962 // with the lower offset has an alignment suitable for the
965 unsigned VecAlignment = TD->getPrefTypeAlignment(VType);
966 if (BottomAlignment < VecAlignment)
971 unsigned ICost = TTI->getMemoryOpCost(I->getOpcode(), aTypeI,
972 IAlignment, IAddressSpace);
973 unsigned JCost = TTI->getMemoryOpCost(J->getOpcode(), aTypeJ,
974 JAlignment, JAddressSpace);
975 unsigned VCost = TTI->getMemoryOpCost(I->getOpcode(), VType,
979 ICost += TTI->getAddressComputationCost(aTypeI);
980 JCost += TTI->getAddressComputationCost(aTypeJ);
981 VCost += TTI->getAddressComputationCost(VType);
983 if (VCost > ICost + JCost)
986 // We don't want to fuse to a type that will be split, even
987 // if the two input types will also be split and there is no other
989 unsigned VParts = TTI->getNumberOfParts(VType);
992 else if (!VParts && VCost == ICost + JCost)
995 CostSavings = ICost + JCost - VCost;
1001 unsigned ICost = getInstrCost(I->getOpcode(), IT1, IT2);
1002 unsigned JCost = getInstrCost(J->getOpcode(), JT1, JT2);
1003 Type *VT1 = getVecTypeForPair(IT1, JT1),
1004 *VT2 = getVecTypeForPair(IT2, JT2);
1006 // Note that this procedure is incorrect for insert and extract element
1007 // instructions (because combining these often results in a shuffle),
1008 // but this cost is ignored (because insert and extract element
1009 // instructions are assigned a zero depth factor and are not really
1010 // fused in general).
1011 unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2);
1013 if (VCost > ICost + JCost)
1016 // We don't want to fuse to a type that will be split, even
1017 // if the two input types will also be split and there is no other
1019 unsigned VParts1 = TTI->getNumberOfParts(VT1),
1020 VParts2 = TTI->getNumberOfParts(VT2);
1021 if (VParts1 > 1 || VParts2 > 1)
1023 else if ((!VParts1 || !VParts2) && VCost == ICost + JCost)
1026 CostSavings = ICost + JCost - VCost;
1029 // The powi intrinsic is special because only the first argument is
1030 // vectorized, the second arguments must be equal.
1031 CallInst *CI = dyn_cast<CallInst>(I);
1033 if (CI && (FI = CI->getCalledFunction())) {
1034 Intrinsic::ID IID = (Intrinsic::ID) FI->getIntrinsicID();
1035 if (IID == Intrinsic::powi) {
1036 Value *A1I = CI->getArgOperand(1),
1037 *A1J = cast<CallInst>(J)->getArgOperand(1);
1038 const SCEV *A1ISCEV = SE->getSCEV(A1I),
1039 *A1JSCEV = SE->getSCEV(A1J);
1040 return (A1ISCEV == A1JSCEV);
1044 SmallVector<Type*, 4> Tys;
1045 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
1046 Tys.push_back(CI->getArgOperand(i)->getType());
1047 unsigned ICost = TTI->getIntrinsicInstrCost(IID, IT1, Tys);
1050 CallInst *CJ = cast<CallInst>(J);
1051 for (unsigned i = 0, ie = CJ->getNumArgOperands(); i != ie; ++i)
1052 Tys.push_back(CJ->getArgOperand(i)->getType());
1053 unsigned JCost = TTI->getIntrinsicInstrCost(IID, JT1, Tys);
1056 assert(CI->getNumArgOperands() == CJ->getNumArgOperands() &&
1057 "Intrinsic argument counts differ");
1058 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
1059 if (IID == Intrinsic::powi && i == 1)
1060 Tys.push_back(CI->getArgOperand(i)->getType());
1062 Tys.push_back(getVecTypeForPair(CI->getArgOperand(i)->getType(),
1063 CJ->getArgOperand(i)->getType()));
1066 Type *RetTy = getVecTypeForPair(IT1, JT1);
1067 unsigned VCost = TTI->getIntrinsicInstrCost(IID, RetTy, Tys);
1069 if (VCost > ICost + JCost)
1072 // We don't want to fuse to a type that will be split, even
1073 // if the two input types will also be split and there is no other
1075 unsigned RetParts = TTI->getNumberOfParts(RetTy);
1078 else if (!RetParts && VCost == ICost + JCost)
1081 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
1082 if (!Tys[i]->isVectorTy())
1085 unsigned NumParts = TTI->getNumberOfParts(Tys[i]);
1088 else if (!NumParts && VCost == ICost + JCost)
1092 CostSavings = ICost + JCost - VCost;
1099 // Figure out whether or not J uses I and update the users and write-set
1100 // structures associated with I. Specifically, Users represents the set of
1101 // instructions that depend on I. WriteSet represents the set
1102 // of memory locations that are dependent on I. If UpdateUsers is true,
1103 // and J uses I, then Users is updated to contain J and WriteSet is updated
1104 // to contain any memory locations to which J writes. The function returns
1105 // true if J uses I. By default, alias analysis is used to determine
1106 // whether J reads from memory that overlaps with a location in WriteSet.
1107 // If LoadMoveSet is not null, then it is a previously-computed multimap
1108 // where the key is the memory-based user instruction and the value is
1109 // the instruction to be compared with I. So, if LoadMoveSet is provided,
1110 // then the alias analysis is not used. This is necessary because this
1111 // function is called during the process of moving instructions during
1112 // vectorization and the results of the alias analysis are not stable during
1114 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users,
1115 AliasSetTracker &WriteSet, Instruction *I,
1116 Instruction *J, bool UpdateUsers,
1117 std::multimap<Value *, Value *> *LoadMoveSet) {
1120 // This instruction may already be marked as a user due, for example, to
1121 // being a member of a selected pair.
1126 for (User::op_iterator JU = J->op_begin(), JE = J->op_end();
1129 if (I == V || Users.count(V)) {
1134 if (!UsesI && J->mayReadFromMemory()) {
1136 VPIteratorPair JPairRange = LoadMoveSet->equal_range(J);
1137 UsesI = isSecondInIteratorPair<Value*>(I, JPairRange);
1139 for (AliasSetTracker::iterator W = WriteSet.begin(),
1140 WE = WriteSet.end(); W != WE; ++W) {
1141 if (W->aliasesUnknownInst(J, *AA)) {
1149 if (UsesI && UpdateUsers) {
1150 if (J->mayWriteToMemory()) WriteSet.add(J);
1157 // This function iterates over all instruction pairs in the provided
1158 // basic block and collects all candidate pairs for vectorization.
1159 bool BBVectorize::getCandidatePairs(BasicBlock &BB,
1160 BasicBlock::iterator &Start,
1161 std::multimap<Value *, Value *> &CandidatePairs,
1162 DenseSet<ValuePair> &FixedOrderPairs,
1163 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1164 std::vector<Value *> &PairableInsts, bool NonPow2Len) {
1165 BasicBlock::iterator E = BB.end();
1166 if (Start == E) return false;
1168 bool ShouldContinue = false, IAfterStart = false;
1169 for (BasicBlock::iterator I = Start++; I != E; ++I) {
1170 if (I == Start) IAfterStart = true;
1172 bool IsSimpleLoadStore;
1173 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue;
1175 // Look for an instruction with which to pair instruction *I...
1176 DenseSet<Value *> Users;
1177 AliasSetTracker WriteSet(*AA);
1178 bool JAfterStart = IAfterStart;
1179 BasicBlock::iterator J = llvm::next(I);
1180 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) {
1181 if (J == Start) JAfterStart = true;
1183 // Determine if J uses I, if so, exit the loop.
1184 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep);
1185 if (Config.FastDep) {
1186 // Note: For this heuristic to be effective, independent operations
1187 // must tend to be intermixed. This is likely to be true from some
1188 // kinds of grouped loop unrolling (but not the generic LLVM pass),
1189 // but otherwise may require some kind of reordering pass.
1191 // When using fast dependency analysis,
1192 // stop searching after first use:
1195 if (UsesI) continue;
1198 // J does not use I, and comes before the first use of I, so it can be
1199 // merged with I if the instructions are compatible.
1200 int CostSavings, FixedOrder;
1201 if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len,
1202 CostSavings, FixedOrder)) continue;
1204 // J is a candidate for merging with I.
1205 if (!PairableInsts.size() ||
1206 PairableInsts[PairableInsts.size()-1] != I) {
1207 PairableInsts.push_back(I);
1210 CandidatePairs.insert(ValuePair(I, J));
1212 CandidatePairCostSavings.insert(ValuePairWithCost(ValuePair(I, J),
1215 if (FixedOrder == 1)
1216 FixedOrderPairs.insert(ValuePair(I, J));
1217 else if (FixedOrder == -1)
1218 FixedOrderPairs.insert(ValuePair(J, I));
1220 // The next call to this function must start after the last instruction
1221 // selected during this invocation.
1223 Start = llvm::next(J);
1224 IAfterStart = JAfterStart = false;
1227 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair "
1228 << *I << " <-> " << *J << " (cost savings: " <<
1229 CostSavings << ")\n");
1231 // If we have already found too many pairs, break here and this function
1232 // will be called again starting after the last instruction selected
1233 // during this invocation.
1234 if (PairableInsts.size() >= Config.MaxInsts) {
1235 ShouldContinue = true;
1244 DEBUG(dbgs() << "BBV: found " << PairableInsts.size()
1245 << " instructions with candidate pairs\n");
1247 return ShouldContinue;
1250 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that
1251 // it looks for pairs such that both members have an input which is an
1252 // output of PI or PJ.
1253 void BBVectorize::computePairsConnectedTo(
1254 std::multimap<Value *, Value *> &CandidatePairs,
1255 std::vector<Value *> &PairableInsts,
1256 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1257 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1261 // For each possible pairing for this variable, look at the uses of
1262 // the first value...
1263 for (Value::use_iterator I = P.first->use_begin(),
1264 E = P.first->use_end(); I != E; ++I) {
1265 if (isa<LoadInst>(*I)) {
1266 // A pair cannot be connected to a load because the load only takes one
1267 // operand (the address) and it is a scalar even after vectorization.
1269 } else if ((SI = dyn_cast<StoreInst>(*I)) &&
1270 P.first == SI->getPointerOperand()) {
1271 // Similarly, a pair cannot be connected to a store through its
1276 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I);
1278 // For each use of the first variable, look for uses of the second
1280 for (Value::use_iterator J = P.second->use_begin(),
1281 E2 = P.second->use_end(); J != E2; ++J) {
1282 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1283 P.second == SJ->getPointerOperand())
1286 VPIteratorPair JPairRange = CandidatePairs.equal_range(*J);
1289 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) {
1290 VPPair VP(P, ValuePair(*I, *J));
1291 ConnectedPairs.insert(VP);
1292 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionDirect));
1296 if (isSecondInIteratorPair<Value*>(*I, JPairRange)) {
1297 VPPair VP(P, ValuePair(*J, *I));
1298 ConnectedPairs.insert(VP);
1299 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSwap));
1303 if (Config.SplatBreaksChain) continue;
1304 // Look for cases where just the first value in the pair is used by
1305 // both members of another pair (splatting).
1306 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) {
1307 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1308 P.first == SJ->getPointerOperand())
1311 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) {
1312 VPPair VP(P, ValuePair(*I, *J));
1313 ConnectedPairs.insert(VP);
1314 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1319 if (Config.SplatBreaksChain) return;
1320 // Look for cases where just the second value in the pair is used by
1321 // both members of another pair (splatting).
1322 for (Value::use_iterator I = P.second->use_begin(),
1323 E = P.second->use_end(); I != E; ++I) {
1324 if (isa<LoadInst>(*I))
1326 else if ((SI = dyn_cast<StoreInst>(*I)) &&
1327 P.second == SI->getPointerOperand())
1330 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I);
1332 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) {
1333 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1334 P.second == SJ->getPointerOperand())
1337 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) {
1338 VPPair VP(P, ValuePair(*I, *J));
1339 ConnectedPairs.insert(VP);
1340 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1346 // This function figures out which pairs are connected. Two pairs are
1347 // connected if some output of the first pair forms an input to both members
1348 // of the second pair.
1349 void BBVectorize::computeConnectedPairs(
1350 std::multimap<Value *, Value *> &CandidatePairs,
1351 std::vector<Value *> &PairableInsts,
1352 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1353 DenseMap<VPPair, unsigned> &PairConnectionTypes) {
1355 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
1356 PE = PairableInsts.end(); PI != PE; ++PI) {
1357 VPIteratorPair choiceRange = CandidatePairs.equal_range(*PI);
1359 for (std::multimap<Value *, Value *>::iterator P = choiceRange.first;
1360 P != choiceRange.second; ++P)
1361 computePairsConnectedTo(CandidatePairs, PairableInsts,
1362 ConnectedPairs, PairConnectionTypes, *P);
1365 DEBUG(dbgs() << "BBV: found " << ConnectedPairs.size()
1366 << " pair connections.\n");
1369 // This function builds a set of use tuples such that <A, B> is in the set
1370 // if B is in the use tree of A. If B is in the use tree of A, then B
1371 // depends on the output of A.
1372 void BBVectorize::buildDepMap(
1374 std::multimap<Value *, Value *> &CandidatePairs,
1375 std::vector<Value *> &PairableInsts,
1376 DenseSet<ValuePair> &PairableInstUsers) {
1377 DenseSet<Value *> IsInPair;
1378 for (std::multimap<Value *, Value *>::iterator C = CandidatePairs.begin(),
1379 E = CandidatePairs.end(); C != E; ++C) {
1380 IsInPair.insert(C->first);
1381 IsInPair.insert(C->second);
1384 // Iterate through the basic block, recording all users of each
1385 // pairable instruction.
1387 BasicBlock::iterator E = BB.end();
1388 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) {
1389 if (IsInPair.find(I) == IsInPair.end()) continue;
1391 DenseSet<Value *> Users;
1392 AliasSetTracker WriteSet(*AA);
1393 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J)
1394 (void) trackUsesOfI(Users, WriteSet, I, J);
1396 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end();
1398 PairableInstUsers.insert(ValuePair(I, *U));
1402 // Returns true if an input to pair P is an output of pair Q and also an
1403 // input of pair Q is an output of pair P. If this is the case, then these
1404 // two pairs cannot be simultaneously fused.
1405 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q,
1406 DenseSet<ValuePair> &PairableInstUsers,
1407 std::multimap<ValuePair, ValuePair> *PairableInstUserMap,
1408 DenseSet<VPPair> *PairableInstUserPairSet) {
1409 // Two pairs are in conflict if they are mutual Users of eachother.
1410 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) ||
1411 PairableInstUsers.count(ValuePair(P.first, Q.second)) ||
1412 PairableInstUsers.count(ValuePair(P.second, Q.first)) ||
1413 PairableInstUsers.count(ValuePair(P.second, Q.second));
1414 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) ||
1415 PairableInstUsers.count(ValuePair(Q.first, P.second)) ||
1416 PairableInstUsers.count(ValuePair(Q.second, P.first)) ||
1417 PairableInstUsers.count(ValuePair(Q.second, P.second));
1418 if (PairableInstUserMap) {
1419 // FIXME: The expensive part of the cycle check is not so much the cycle
1420 // check itself but this edge insertion procedure. This needs some
1421 // profiling and probably a different data structure (same is true of
1422 // most uses of std::multimap).
1424 if (PairableInstUserPairSet->insert(VPPair(Q, P)).second)
1425 PairableInstUserMap->insert(VPPair(Q, P));
1428 if (PairableInstUserPairSet->insert(VPPair(P, Q)).second)
1429 PairableInstUserMap->insert(VPPair(P, Q));
1433 return (QUsesP && PUsesQ);
1436 // This function walks the use graph of current pairs to see if, starting
1437 // from P, the walk returns to P.
1438 bool BBVectorize::pairWillFormCycle(ValuePair P,
1439 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1440 DenseSet<ValuePair> &CurrentPairs) {
1441 DEBUG(if (DebugCycleCheck)
1442 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> "
1443 << *P.second << "\n");
1444 // A lookup table of visisted pairs is kept because the PairableInstUserMap
1445 // contains non-direct associations.
1446 DenseSet<ValuePair> Visited;
1447 SmallVector<ValuePair, 32> Q;
1448 // General depth-first post-order traversal:
1451 ValuePair QTop = Q.pop_back_val();
1452 Visited.insert(QTop);
1454 DEBUG(if (DebugCycleCheck)
1455 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> "
1456 << *QTop.second << "\n");
1457 VPPIteratorPair QPairRange = PairableInstUserMap.equal_range(QTop);
1458 for (std::multimap<ValuePair, ValuePair>::iterator C = QPairRange.first;
1459 C != QPairRange.second; ++C) {
1460 if (C->second == P) {
1462 << "BBV: rejected to prevent non-trivial cycle formation: "
1463 << *C->first.first << " <-> " << *C->first.second << "\n");
1467 if (CurrentPairs.count(C->second) && !Visited.count(C->second))
1468 Q.push_back(C->second);
1470 } while (!Q.empty());
1475 // This function builds the initial tree of connected pairs with the
1476 // pair J at the root.
1477 void BBVectorize::buildInitialTreeFor(
1478 std::multimap<Value *, Value *> &CandidatePairs,
1479 std::vector<Value *> &PairableInsts,
1480 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1481 DenseSet<ValuePair> &PairableInstUsers,
1482 DenseMap<Value *, Value *> &ChosenPairs,
1483 DenseMap<ValuePair, size_t> &Tree, ValuePair J) {
1484 // Each of these pairs is viewed as the root node of a Tree. The Tree
1485 // is then walked (depth-first). As this happens, we keep track of
1486 // the pairs that compose the Tree and the maximum depth of the Tree.
1487 SmallVector<ValuePairWithDepth, 32> Q;
1488 // General depth-first post-order traversal:
1489 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1491 ValuePairWithDepth QTop = Q.back();
1493 // Push each child onto the queue:
1494 bool MoreChildren = false;
1495 size_t MaxChildDepth = QTop.second;
1496 VPPIteratorPair qtRange = ConnectedPairs.equal_range(QTop.first);
1497 for (std::multimap<ValuePair, ValuePair>::iterator k = qtRange.first;
1498 k != qtRange.second; ++k) {
1499 // Make sure that this child pair is still a candidate:
1500 bool IsStillCand = false;
1501 VPIteratorPair checkRange =
1502 CandidatePairs.equal_range(k->second.first);
1503 for (std::multimap<Value *, Value *>::iterator m = checkRange.first;
1504 m != checkRange.second; ++m) {
1505 if (m->second == k->second.second) {
1512 DenseMap<ValuePair, size_t>::iterator C = Tree.find(k->second);
1513 if (C == Tree.end()) {
1514 size_t d = getDepthFactor(k->second.first);
1515 Q.push_back(ValuePairWithDepth(k->second, QTop.second+d));
1516 MoreChildren = true;
1518 MaxChildDepth = std::max(MaxChildDepth, C->second);
1523 if (!MoreChildren) {
1524 // Record the current pair as part of the Tree:
1525 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth));
1528 } while (!Q.empty());
1531 // Given some initial tree, prune it by removing conflicting pairs (pairs
1532 // that cannot be simultaneously chosen for vectorization).
1533 void BBVectorize::pruneTreeFor(
1534 std::multimap<Value *, Value *> &CandidatePairs,
1535 std::vector<Value *> &PairableInsts,
1536 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1537 DenseSet<ValuePair> &PairableInstUsers,
1538 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1539 DenseSet<VPPair> &PairableInstUserPairSet,
1540 DenseMap<Value *, Value *> &ChosenPairs,
1541 DenseMap<ValuePair, size_t> &Tree,
1542 DenseSet<ValuePair> &PrunedTree, ValuePair J,
1543 bool UseCycleCheck) {
1544 SmallVector<ValuePairWithDepth, 32> Q;
1545 // General depth-first post-order traversal:
1546 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1548 ValuePairWithDepth QTop = Q.pop_back_val();
1549 PrunedTree.insert(QTop.first);
1551 // Visit each child, pruning as necessary...
1552 SmallVector<ValuePairWithDepth, 8> BestChildren;
1553 VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first);
1554 for (std::multimap<ValuePair, ValuePair>::iterator K = QTopRange.first;
1555 K != QTopRange.second; ++K) {
1556 DenseMap<ValuePair, size_t>::iterator C = Tree.find(K->second);
1557 if (C == Tree.end()) continue;
1559 // This child is in the Tree, now we need to make sure it is the
1560 // best of any conflicting children. There could be multiple
1561 // conflicting children, so first, determine if we're keeping
1562 // this child, then delete conflicting children as necessary.
1564 // It is also necessary to guard against pairing-induced
1565 // dependencies. Consider instructions a .. x .. y .. b
1566 // such that (a,b) are to be fused and (x,y) are to be fused
1567 // but a is an input to x and b is an output from y. This
1568 // means that y cannot be moved after b but x must be moved
1569 // after b for (a,b) to be fused. In other words, after
1570 // fusing (a,b) we have y .. a/b .. x where y is an input
1571 // to a/b and x is an output to a/b: x and y can no longer
1572 // be legally fused. To prevent this condition, we must
1573 // make sure that a child pair added to the Tree is not
1574 // both an input and output of an already-selected pair.
1576 // Pairing-induced dependencies can also form from more complicated
1577 // cycles. The pair vs. pair conflicts are easy to check, and so
1578 // that is done explicitly for "fast rejection", and because for
1579 // child vs. child conflicts, we may prefer to keep the current
1580 // pair in preference to the already-selected child.
1581 DenseSet<ValuePair> CurrentPairs;
1584 for (SmallVector<ValuePairWithDepth, 8>::iterator C2
1585 = BestChildren.begin(), E2 = BestChildren.end();
1587 if (C2->first.first == C->first.first ||
1588 C2->first.first == C->first.second ||
1589 C2->first.second == C->first.first ||
1590 C2->first.second == C->first.second ||
1591 pairsConflict(C2->first, C->first, PairableInstUsers,
1592 UseCycleCheck ? &PairableInstUserMap : 0,
1593 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1594 if (C2->second >= C->second) {
1599 CurrentPairs.insert(C2->first);
1602 if (!CanAdd) continue;
1604 // Even worse, this child could conflict with another node already
1605 // selected for the Tree. If that is the case, ignore this child.
1606 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(),
1607 E2 = PrunedTree.end(); T != E2; ++T) {
1608 if (T->first == C->first.first ||
1609 T->first == C->first.second ||
1610 T->second == C->first.first ||
1611 T->second == C->first.second ||
1612 pairsConflict(*T, C->first, PairableInstUsers,
1613 UseCycleCheck ? &PairableInstUserMap : 0,
1614 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1619 CurrentPairs.insert(*T);
1621 if (!CanAdd) continue;
1623 // And check the queue too...
1624 for (SmallVector<ValuePairWithDepth, 32>::iterator C2 = Q.begin(),
1625 E2 = Q.end(); C2 != E2; ++C2) {
1626 if (C2->first.first == C->first.first ||
1627 C2->first.first == C->first.second ||
1628 C2->first.second == C->first.first ||
1629 C2->first.second == C->first.second ||
1630 pairsConflict(C2->first, C->first, PairableInstUsers,
1631 UseCycleCheck ? &PairableInstUserMap : 0,
1632 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1637 CurrentPairs.insert(C2->first);
1639 if (!CanAdd) continue;
1641 // Last but not least, check for a conflict with any of the
1642 // already-chosen pairs.
1643 for (DenseMap<Value *, Value *>::iterator C2 =
1644 ChosenPairs.begin(), E2 = ChosenPairs.end();
1646 if (pairsConflict(*C2, C->first, PairableInstUsers,
1647 UseCycleCheck ? &PairableInstUserMap : 0,
1648 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1653 CurrentPairs.insert(*C2);
1655 if (!CanAdd) continue;
1657 // To check for non-trivial cycles formed by the addition of the
1658 // current pair we've formed a list of all relevant pairs, now use a
1659 // graph walk to check for a cycle. We start from the current pair and
1660 // walk the use tree to see if we again reach the current pair. If we
1661 // do, then the current pair is rejected.
1663 // FIXME: It may be more efficient to use a topological-ordering
1664 // algorithm to improve the cycle check. This should be investigated.
1665 if (UseCycleCheck &&
1666 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs))
1669 // This child can be added, but we may have chosen it in preference
1670 // to an already-selected child. Check for this here, and if a
1671 // conflict is found, then remove the previously-selected child
1672 // before adding this one in its place.
1673 for (SmallVector<ValuePairWithDepth, 8>::iterator C2
1674 = BestChildren.begin(); C2 != BestChildren.end();) {
1675 if (C2->first.first == C->first.first ||
1676 C2->first.first == C->first.second ||
1677 C2->first.second == C->first.first ||
1678 C2->first.second == C->first.second ||
1679 pairsConflict(C2->first, C->first, PairableInstUsers))
1680 C2 = BestChildren.erase(C2);
1685 BestChildren.push_back(ValuePairWithDepth(C->first, C->second));
1688 for (SmallVector<ValuePairWithDepth, 8>::iterator C
1689 = BestChildren.begin(), E2 = BestChildren.end();
1691 size_t DepthF = getDepthFactor(C->first.first);
1692 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF));
1694 } while (!Q.empty());
1697 // This function finds the best tree of mututally-compatible connected
1698 // pairs, given the choice of root pairs as an iterator range.
1699 void BBVectorize::findBestTreeFor(
1700 std::multimap<Value *, Value *> &CandidatePairs,
1701 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1702 std::vector<Value *> &PairableInsts,
1703 DenseSet<ValuePair> &FixedOrderPairs,
1704 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1705 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1706 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps,
1707 DenseSet<ValuePair> &PairableInstUsers,
1708 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1709 DenseSet<VPPair> &PairableInstUserPairSet,
1710 DenseMap<Value *, Value *> &ChosenPairs,
1711 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth,
1712 int &BestEffSize, VPIteratorPair ChoiceRange,
1713 bool UseCycleCheck) {
1714 for (std::multimap<Value *, Value *>::iterator J = ChoiceRange.first;
1715 J != ChoiceRange.second; ++J) {
1717 // Before going any further, make sure that this pair does not
1718 // conflict with any already-selected pairs (see comment below
1719 // near the Tree pruning for more details).
1720 DenseSet<ValuePair> ChosenPairSet;
1721 bool DoesConflict = false;
1722 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(),
1723 E = ChosenPairs.end(); C != E; ++C) {
1724 if (pairsConflict(*C, *J, PairableInstUsers,
1725 UseCycleCheck ? &PairableInstUserMap : 0,
1726 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1727 DoesConflict = true;
1731 ChosenPairSet.insert(*C);
1733 if (DoesConflict) continue;
1735 if (UseCycleCheck &&
1736 pairWillFormCycle(*J, PairableInstUserMap, ChosenPairSet))
1739 DenseMap<ValuePair, size_t> Tree;
1740 buildInitialTreeFor(CandidatePairs, PairableInsts, ConnectedPairs,
1741 PairableInstUsers, ChosenPairs, Tree, *J);
1743 // Because we'll keep the child with the largest depth, the largest
1744 // depth is still the same in the unpruned Tree.
1745 size_t MaxDepth = Tree.lookup(*J);
1747 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found Tree for pair {"
1748 << *J->first << " <-> " << *J->second << "} of depth " <<
1749 MaxDepth << " and size " << Tree.size() << "\n");
1751 // At this point the Tree has been constructed, but, may contain
1752 // contradictory children (meaning that different children of
1753 // some tree node may be attempting to fuse the same instruction).
1754 // So now we walk the tree again, in the case of a conflict,
1755 // keep only the child with the largest depth. To break a tie,
1756 // favor the first child.
1758 DenseSet<ValuePair> PrunedTree;
1759 pruneTreeFor(CandidatePairs, PairableInsts, ConnectedPairs,
1760 PairableInstUsers, PairableInstUserMap, PairableInstUserPairSet,
1761 ChosenPairs, Tree, PrunedTree, *J, UseCycleCheck);
1765 DenseSet<Value *> PrunedTreeInstrs;
1766 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
1767 E = PrunedTree.end(); S != E; ++S) {
1768 PrunedTreeInstrs.insert(S->first);
1769 PrunedTreeInstrs.insert(S->second);
1772 // The set of pairs that have already contributed to the total cost.
1773 DenseSet<ValuePair> IncomingPairs;
1775 // If the cost model were perfect, this might not be necessary; but we
1776 // need to make sure that we don't get stuck vectorizing our own
1778 bool HasNontrivialInsts = false;
1780 // The node weights represent the cost savings associated with
1781 // fusing the pair of instructions.
1782 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
1783 E = PrunedTree.end(); S != E; ++S) {
1784 if (!isa<ShuffleVectorInst>(S->first) &&
1785 !isa<InsertElementInst>(S->first) &&
1786 !isa<ExtractElementInst>(S->first))
1787 HasNontrivialInsts = true;
1789 bool FlipOrder = false;
1791 if (getDepthFactor(S->first)) {
1792 int ESContrib = CandidatePairCostSavings.find(*S)->second;
1793 DEBUG(if (DebugPairSelection) dbgs() << "\tweight {"
1794 << *S->first << " <-> " << *S->second << "} = " <<
1796 EffSize += ESContrib;
1799 // The edge weights contribute in a negative sense: they represent
1800 // the cost of shuffles.
1801 VPPIteratorPair IP = ConnectedPairDeps.equal_range(*S);
1802 if (IP.first != ConnectedPairDeps.end()) {
1803 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
1804 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first;
1805 Q != IP.second; ++Q) {
1806 if (!PrunedTree.count(Q->second))
1808 DenseMap<VPPair, unsigned>::iterator R =
1809 PairConnectionTypes.find(VPPair(Q->second, Q->first));
1810 assert(R != PairConnectionTypes.end() &&
1811 "Cannot find pair connection type");
1812 if (R->second == PairConnectionDirect)
1814 else if (R->second == PairConnectionSwap)
1818 // If there are more swaps than direct connections, then
1819 // the pair order will be flipped during fusion. So the real
1820 // number of swaps is the minimum number.
1821 FlipOrder = !FixedOrderPairs.count(*S) &&
1822 ((NumDepsSwap > NumDepsDirect) ||
1823 FixedOrderPairs.count(ValuePair(S->second, S->first)));
1825 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first;
1826 Q != IP.second; ++Q) {
1827 if (!PrunedTree.count(Q->second))
1829 DenseMap<VPPair, unsigned>::iterator R =
1830 PairConnectionTypes.find(VPPair(Q->second, Q->first));
1831 assert(R != PairConnectionTypes.end() &&
1832 "Cannot find pair connection type");
1833 Type *Ty1 = Q->second.first->getType(),
1834 *Ty2 = Q->second.second->getType();
1835 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1836 if ((R->second == PairConnectionDirect && FlipOrder) ||
1837 (R->second == PairConnectionSwap && !FlipOrder) ||
1838 R->second == PairConnectionSplat) {
1839 int ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1842 if (VTy->getVectorNumElements() == 2) {
1843 if (R->second == PairConnectionSplat)
1844 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1845 TargetTransformInfo::SK_Broadcast, VTy));
1847 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1848 TargetTransformInfo::SK_Reverse, VTy));
1851 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1852 *Q->second.first << " <-> " << *Q->second.second <<
1854 *S->first << " <-> " << *S->second << "} = " <<
1856 EffSize -= ESContrib;
1861 // Compute the cost of outgoing edges. We assume that edges outgoing
1862 // to shuffles, inserts or extracts can be merged, and so contribute
1863 // no additional cost.
1864 if (!S->first->getType()->isVoidTy()) {
1865 Type *Ty1 = S->first->getType(),
1866 *Ty2 = S->second->getType();
1867 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1869 bool NeedsExtraction = false;
1870 for (Value::use_iterator I = S->first->use_begin(),
1871 IE = S->first->use_end(); I != IE; ++I) {
1872 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) {
1873 // Shuffle can be folded if it has no other input
1874 if (isa<UndefValue>(SI->getOperand(1)))
1877 if (isa<ExtractElementInst>(*I))
1879 if (PrunedTreeInstrs.count(*I))
1881 NeedsExtraction = true;
1885 if (NeedsExtraction) {
1887 if (Ty1->isVectorTy()) {
1888 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1890 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1891 TargetTransformInfo::SK_ExtractSubvector, VTy, 0, Ty1));
1893 ESContrib = (int) TTI->getVectorInstrCost(
1894 Instruction::ExtractElement, VTy, 0);
1896 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1897 *S->first << "} = " << ESContrib << "\n");
1898 EffSize -= ESContrib;
1901 NeedsExtraction = false;
1902 for (Value::use_iterator I = S->second->use_begin(),
1903 IE = S->second->use_end(); I != IE; ++I) {
1904 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) {
1905 // Shuffle can be folded if it has no other input
1906 if (isa<UndefValue>(SI->getOperand(1)))
1909 if (isa<ExtractElementInst>(*I))
1911 if (PrunedTreeInstrs.count(*I))
1913 NeedsExtraction = true;
1917 if (NeedsExtraction) {
1919 if (Ty2->isVectorTy()) {
1920 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1922 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1923 TargetTransformInfo::SK_ExtractSubvector, VTy,
1924 Ty1->isVectorTy() ? Ty1->getVectorNumElements() : 1, Ty2));
1926 ESContrib = (int) TTI->getVectorInstrCost(
1927 Instruction::ExtractElement, VTy, 1);
1928 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1929 *S->second << "} = " << ESContrib << "\n");
1930 EffSize -= ESContrib;
1934 // Compute the cost of incoming edges.
1935 if (!isa<LoadInst>(S->first) && !isa<StoreInst>(S->first)) {
1936 Instruction *S1 = cast<Instruction>(S->first),
1937 *S2 = cast<Instruction>(S->second);
1938 for (unsigned o = 0; o < S1->getNumOperands(); ++o) {
1939 Value *O1 = S1->getOperand(o), *O2 = S2->getOperand(o);
1941 // Combining constants into vector constants (or small vector
1942 // constants into larger ones are assumed free).
1943 if (isa<Constant>(O1) && isa<Constant>(O2))
1949 ValuePair VP = ValuePair(O1, O2);
1950 ValuePair VPR = ValuePair(O2, O1);
1952 // Internal edges are not handled here.
1953 if (PrunedTree.count(VP) || PrunedTree.count(VPR))
1956 Type *Ty1 = O1->getType(),
1957 *Ty2 = O2->getType();
1958 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1960 // Combining vector operations of the same type is also assumed
1961 // folded with other operations.
1963 // If both are insert elements, then both can be widened.
1964 InsertElementInst *IEO1 = dyn_cast<InsertElementInst>(O1),
1965 *IEO2 = dyn_cast<InsertElementInst>(O2);
1966 if (IEO1 && IEO2 && isPureIEChain(IEO1) && isPureIEChain(IEO2))
1968 // If both are extract elements, and both have the same input
1969 // type, then they can be replaced with a shuffle
1970 ExtractElementInst *EIO1 = dyn_cast<ExtractElementInst>(O1),
1971 *EIO2 = dyn_cast<ExtractElementInst>(O2);
1973 EIO1->getOperand(0)->getType() ==
1974 EIO2->getOperand(0)->getType())
1976 // If both are a shuffle with equal operand types and only two
1977 // unqiue operands, then they can be replaced with a single
1979 ShuffleVectorInst *SIO1 = dyn_cast<ShuffleVectorInst>(O1),
1980 *SIO2 = dyn_cast<ShuffleVectorInst>(O2);
1982 SIO1->getOperand(0)->getType() ==
1983 SIO2->getOperand(0)->getType()) {
1984 SmallSet<Value *, 4> SIOps;
1985 SIOps.insert(SIO1->getOperand(0));
1986 SIOps.insert(SIO1->getOperand(1));
1987 SIOps.insert(SIO2->getOperand(0));
1988 SIOps.insert(SIO2->getOperand(1));
1989 if (SIOps.size() <= 2)
1995 // This pair has already been formed.
1996 if (IncomingPairs.count(VP)) {
1998 } else if (IncomingPairs.count(VPR)) {
1999 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2002 if (VTy->getVectorNumElements() == 2)
2003 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
2004 TargetTransformInfo::SK_Reverse, VTy));
2005 } else if (!Ty1->isVectorTy() && !Ty2->isVectorTy()) {
2006 ESContrib = (int) TTI->getVectorInstrCost(
2007 Instruction::InsertElement, VTy, 0);
2008 ESContrib += (int) TTI->getVectorInstrCost(
2009 Instruction::InsertElement, VTy, 1);
2010 } else if (!Ty1->isVectorTy()) {
2011 // O1 needs to be inserted into a vector of size O2, and then
2012 // both need to be shuffled together.
2013 ESContrib = (int) TTI->getVectorInstrCost(
2014 Instruction::InsertElement, Ty2, 0);
2015 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2017 } else if (!Ty2->isVectorTy()) {
2018 // O2 needs to be inserted into a vector of size O1, and then
2019 // both need to be shuffled together.
2020 ESContrib = (int) TTI->getVectorInstrCost(
2021 Instruction::InsertElement, Ty1, 0);
2022 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2025 Type *TyBig = Ty1, *TySmall = Ty2;
2026 if (Ty2->getVectorNumElements() > Ty1->getVectorNumElements())
2027 std::swap(TyBig, TySmall);
2029 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2031 if (TyBig != TySmall)
2032 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2036 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {"
2037 << *O1 << " <-> " << *O2 << "} = " <<
2039 EffSize -= ESContrib;
2040 IncomingPairs.insert(VP);
2045 if (!HasNontrivialInsts) {
2046 DEBUG(if (DebugPairSelection) dbgs() <<
2047 "\tNo non-trivial instructions in tree;"
2048 " override to zero effective size\n");
2052 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
2053 E = PrunedTree.end(); S != E; ++S)
2054 EffSize += (int) getDepthFactor(S->first);
2057 DEBUG(if (DebugPairSelection)
2058 dbgs() << "BBV: found pruned Tree for pair {"
2059 << *J->first << " <-> " << *J->second << "} of depth " <<
2060 MaxDepth << " and size " << PrunedTree.size() <<
2061 " (effective size: " << EffSize << ")\n");
2062 if (((TTI && !UseChainDepthWithTI) ||
2063 MaxDepth >= Config.ReqChainDepth) &&
2064 EffSize > 0 && EffSize > BestEffSize) {
2065 BestMaxDepth = MaxDepth;
2066 BestEffSize = EffSize;
2067 BestTree = PrunedTree;
2072 // Given the list of candidate pairs, this function selects those
2073 // that will be fused into vector instructions.
2074 void BBVectorize::choosePairs(
2075 std::multimap<Value *, Value *> &CandidatePairs,
2076 DenseMap<ValuePair, int> &CandidatePairCostSavings,
2077 std::vector<Value *> &PairableInsts,
2078 DenseSet<ValuePair> &FixedOrderPairs,
2079 DenseMap<VPPair, unsigned> &PairConnectionTypes,
2080 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
2081 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps,
2082 DenseSet<ValuePair> &PairableInstUsers,
2083 DenseMap<Value *, Value *>& ChosenPairs) {
2084 bool UseCycleCheck =
2085 CandidatePairs.size() <= Config.MaxCandPairsForCycleCheck;
2086 std::multimap<ValuePair, ValuePair> PairableInstUserMap;
2087 DenseSet<VPPair> PairableInstUserPairSet;
2088 for (std::vector<Value *>::iterator I = PairableInsts.begin(),
2089 E = PairableInsts.end(); I != E; ++I) {
2090 // The number of possible pairings for this variable:
2091 size_t NumChoices = CandidatePairs.count(*I);
2092 if (!NumChoices) continue;
2094 VPIteratorPair ChoiceRange = CandidatePairs.equal_range(*I);
2096 // The best pair to choose and its tree:
2097 size_t BestMaxDepth = 0;
2098 int BestEffSize = 0;
2099 DenseSet<ValuePair> BestTree;
2100 findBestTreeFor(CandidatePairs, CandidatePairCostSavings,
2101 PairableInsts, FixedOrderPairs, PairConnectionTypes,
2102 ConnectedPairs, ConnectedPairDeps,
2103 PairableInstUsers, PairableInstUserMap,
2104 PairableInstUserPairSet, ChosenPairs,
2105 BestTree, BestMaxDepth, BestEffSize, ChoiceRange,
2108 // A tree has been chosen (or not) at this point. If no tree was
2109 // chosen, then this instruction, I, cannot be paired (and is no longer
2112 DEBUG(if (BestTree.size() > 0)
2113 dbgs() << "BBV: selected pairs in the best tree for: "
2114 << *cast<Instruction>(*I) << "\n");
2116 for (DenseSet<ValuePair>::iterator S = BestTree.begin(),
2117 SE2 = BestTree.end(); S != SE2; ++S) {
2118 // Insert the members of this tree into the list of chosen pairs.
2119 ChosenPairs.insert(ValuePair(S->first, S->second));
2120 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " <<
2121 *S->second << "\n");
2123 // Remove all candidate pairs that have values in the chosen tree.
2124 for (std::multimap<Value *, Value *>::iterator K =
2125 CandidatePairs.begin(); K != CandidatePairs.end();) {
2126 if (K->first == S->first || K->second == S->first ||
2127 K->second == S->second || K->first == S->second) {
2128 // Don't remove the actual pair chosen so that it can be used
2129 // in subsequent tree selections.
2130 if (!(K->first == S->first && K->second == S->second))
2131 CandidatePairs.erase(K++);
2141 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n");
2144 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o,
2149 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) +
2150 (n > 0 ? "." + utostr(n) : "")).str();
2153 // Returns the value that is to be used as the pointer input to the vector
2154 // instruction that fuses I with J.
2155 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context,
2156 Instruction *I, Instruction *J, unsigned o) {
2158 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
2159 int64_t OffsetInElmts;
2161 // Note: the analysis might fail here, that is why the pair order has
2162 // been precomputed (OffsetInElmts must be unused here).
2163 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
2164 IAddressSpace, JAddressSpace,
2165 OffsetInElmts, false);
2167 // The pointer value is taken to be the one with the lowest offset.
2170 Type *ArgTypeI = cast<PointerType>(IPtr->getType())->getElementType();
2171 Type *ArgTypeJ = cast<PointerType>(JPtr->getType())->getElementType();
2172 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2173 Type *VArgPtrType = PointerType::get(VArgType,
2174 cast<PointerType>(IPtr->getType())->getAddressSpace());
2175 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o),
2176 /* insert before */ I);
2179 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J,
2180 unsigned MaskOffset, unsigned NumInElem,
2181 unsigned NumInElem1, unsigned IdxOffset,
2182 std::vector<Constant*> &Mask) {
2183 unsigned NumElem1 = cast<VectorType>(J->getType())->getNumElements();
2184 for (unsigned v = 0; v < NumElem1; ++v) {
2185 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v);
2187 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context));
2189 unsigned mm = m + (int) IdxOffset;
2190 if (m >= (int) NumInElem1)
2191 mm += (int) NumInElem;
2193 Mask[v+MaskOffset] =
2194 ConstantInt::get(Type::getInt32Ty(Context), mm);
2199 // Returns the value that is to be used as the vector-shuffle mask to the
2200 // vector instruction that fuses I with J.
2201 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context,
2202 Instruction *I, Instruction *J) {
2203 // This is the shuffle mask. We need to append the second
2204 // mask to the first, and the numbers need to be adjusted.
2206 Type *ArgTypeI = I->getType();
2207 Type *ArgTypeJ = J->getType();
2208 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2210 unsigned NumElemI = cast<VectorType>(ArgTypeI)->getNumElements();
2212 // Get the total number of elements in the fused vector type.
2213 // By definition, this must equal the number of elements in
2215 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements();
2216 std::vector<Constant*> Mask(NumElem);
2218 Type *OpTypeI = I->getOperand(0)->getType();
2219 unsigned NumInElemI = cast<VectorType>(OpTypeI)->getNumElements();
2220 Type *OpTypeJ = J->getOperand(0)->getType();
2221 unsigned NumInElemJ = cast<VectorType>(OpTypeJ)->getNumElements();
2223 // The fused vector will be:
2224 // -----------------------------------------------------
2225 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ |
2226 // -----------------------------------------------------
2227 // from which we'll extract NumElem total elements (where the first NumElemI
2228 // of them come from the mask in I and the remainder come from the mask
2231 // For the mask from the first pair...
2232 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI,
2235 // For the mask from the second pair...
2236 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ,
2239 return ConstantVector::get(Mask);
2242 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I,
2243 Instruction *J, unsigned o, Value *&LOp,
2245 Type *ArgTypeL, Type *ArgTypeH,
2246 bool IBeforeJ, unsigned IdxOff) {
2247 bool ExpandedIEChain = false;
2248 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) {
2249 // If we have a pure insertelement chain, then this can be rewritten
2250 // into a chain that directly builds the larger type.
2251 if (isPureIEChain(LIE)) {
2252 SmallVector<Value *, 8> VectElemts(numElemL,
2253 UndefValue::get(ArgTypeL->getScalarType()));
2254 InsertElementInst *LIENext = LIE;
2257 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue();
2258 VectElemts[Idx] = LIENext->getOperand(1);
2260 dyn_cast<InsertElementInst>(LIENext->getOperand(0))));
2263 Value *LIEPrev = UndefValue::get(ArgTypeH);
2264 for (unsigned i = 0; i < numElemL; ++i) {
2265 if (isa<UndefValue>(VectElemts[i])) continue;
2266 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i],
2267 ConstantInt::get(Type::getInt32Ty(Context),
2269 getReplacementName(IBeforeJ ? I : J,
2271 LIENext->insertBefore(IBeforeJ ? J : I);
2275 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH);
2276 ExpandedIEChain = true;
2280 return ExpandedIEChain;
2283 // Returns the value to be used as the specified operand of the vector
2284 // instruction that fuses I with J.
2285 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I,
2286 Instruction *J, unsigned o, bool IBeforeJ) {
2287 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2288 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1);
2290 // Compute the fused vector type for this operand
2291 Type *ArgTypeI = I->getOperand(o)->getType();
2292 Type *ArgTypeJ = J->getOperand(o)->getType();
2293 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2295 Instruction *L = I, *H = J;
2296 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ;
2299 if (ArgTypeL->isVectorTy())
2300 numElemL = cast<VectorType>(ArgTypeL)->getNumElements();
2305 if (ArgTypeH->isVectorTy())
2306 numElemH = cast<VectorType>(ArgTypeH)->getNumElements();
2310 Value *LOp = L->getOperand(o);
2311 Value *HOp = H->getOperand(o);
2312 unsigned numElem = VArgType->getNumElements();
2314 // First, we check if we can reuse the "original" vector outputs (if these
2315 // exist). We might need a shuffle.
2316 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp);
2317 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp);
2318 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp);
2319 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp);
2321 // FIXME: If we're fusing shuffle instructions, then we can't apply this
2322 // optimization. The input vectors to the shuffle might be a different
2323 // length from the shuffle outputs. Unfortunately, the replacement
2324 // shuffle mask has already been formed, and the mask entries are sensitive
2325 // to the sizes of the inputs.
2326 bool IsSizeChangeShuffle =
2327 isa<ShuffleVectorInst>(L) &&
2328 (LOp->getType() != L->getType() || HOp->getType() != H->getType());
2330 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) {
2331 // We can have at most two unique vector inputs.
2332 bool CanUseInputs = true;
2335 I1 = LEE->getOperand(0);
2337 I1 = LSV->getOperand(0);
2338 I2 = LSV->getOperand(1);
2339 if (I2 == I1 || isa<UndefValue>(I2))
2344 Value *I3 = HEE->getOperand(0);
2345 if (!I2 && I3 != I1)
2347 else if (I3 != I1 && I3 != I2)
2348 CanUseInputs = false;
2350 Value *I3 = HSV->getOperand(0);
2351 if (!I2 && I3 != I1)
2353 else if (I3 != I1 && I3 != I2)
2354 CanUseInputs = false;
2357 Value *I4 = HSV->getOperand(1);
2358 if (!isa<UndefValue>(I4)) {
2359 if (!I2 && I4 != I1)
2361 else if (I4 != I1 && I4 != I2)
2362 CanUseInputs = false;
2369 cast<VectorType>(cast<Instruction>(LOp)->getOperand(0)->getType())
2372 cast<VectorType>(cast<Instruction>(HOp)->getOperand(0)->getType())
2375 // We have one or two input vectors. We need to map each index of the
2376 // operands to the index of the original vector.
2377 SmallVector<std::pair<int, int>, 8> II(numElem);
2378 for (unsigned i = 0; i < numElemL; ++i) {
2382 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue();
2383 INum = LEE->getOperand(0) == I1 ? 0 : 1;
2385 Idx = LSV->getMaskValue(i);
2386 if (Idx < (int) LOpElem) {
2387 INum = LSV->getOperand(0) == I1 ? 0 : 1;
2390 INum = LSV->getOperand(1) == I1 ? 0 : 1;
2394 II[i] = std::pair<int, int>(Idx, INum);
2396 for (unsigned i = 0; i < numElemH; ++i) {
2400 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue();
2401 INum = HEE->getOperand(0) == I1 ? 0 : 1;
2403 Idx = HSV->getMaskValue(i);
2404 if (Idx < (int) HOpElem) {
2405 INum = HSV->getOperand(0) == I1 ? 0 : 1;
2408 INum = HSV->getOperand(1) == I1 ? 0 : 1;
2412 II[i + numElemL] = std::pair<int, int>(Idx, INum);
2415 // We now have an array which tells us from which index of which
2416 // input vector each element of the operand comes.
2417 VectorType *I1T = cast<VectorType>(I1->getType());
2418 unsigned I1Elem = I1T->getNumElements();
2421 // In this case there is only one underlying vector input. Check for
2422 // the trivial case where we can use the input directly.
2423 if (I1Elem == numElem) {
2424 bool ElemInOrder = true;
2425 for (unsigned i = 0; i < numElem; ++i) {
2426 if (II[i].first != (int) i && II[i].first != -1) {
2427 ElemInOrder = false;
2436 // A shuffle is needed.
2437 std::vector<Constant *> Mask(numElem);
2438 for (unsigned i = 0; i < numElem; ++i) {
2439 int Idx = II[i].first;
2441 Mask[i] = UndefValue::get(Type::getInt32Ty(Context));
2443 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2447 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2448 ConstantVector::get(Mask),
2449 getReplacementName(IBeforeJ ? I : J,
2451 S->insertBefore(IBeforeJ ? J : I);
2455 VectorType *I2T = cast<VectorType>(I2->getType());
2456 unsigned I2Elem = I2T->getNumElements();
2458 // This input comes from two distinct vectors. The first step is to
2459 // make sure that both vectors are the same length. If not, the
2460 // smaller one will need to grow before they can be shuffled together.
2461 if (I1Elem < I2Elem) {
2462 std::vector<Constant *> Mask(I2Elem);
2464 for (; v < I1Elem; ++v)
2465 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2466 for (; v < I2Elem; ++v)
2467 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2469 Instruction *NewI1 =
2470 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2471 ConstantVector::get(Mask),
2472 getReplacementName(IBeforeJ ? I : J,
2474 NewI1->insertBefore(IBeforeJ ? J : I);
2478 } else if (I1Elem > I2Elem) {
2479 std::vector<Constant *> Mask(I1Elem);
2481 for (; v < I2Elem; ++v)
2482 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2483 for (; v < I1Elem; ++v)
2484 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2486 Instruction *NewI2 =
2487 new ShuffleVectorInst(I2, UndefValue::get(I2T),
2488 ConstantVector::get(Mask),
2489 getReplacementName(IBeforeJ ? I : J,
2491 NewI2->insertBefore(IBeforeJ ? J : I);
2497 // Now that both I1 and I2 are the same length we can shuffle them
2498 // together (and use the result).
2499 std::vector<Constant *> Mask(numElem);
2500 for (unsigned v = 0; v < numElem; ++v) {
2501 if (II[v].first == -1) {
2502 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2504 int Idx = II[v].first + II[v].second * I1Elem;
2505 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2509 Instruction *NewOp =
2510 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask),
2511 getReplacementName(IBeforeJ ? I : J, true, o));
2512 NewOp->insertBefore(IBeforeJ ? J : I);
2517 Type *ArgType = ArgTypeL;
2518 if (numElemL < numElemH) {
2519 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH,
2520 ArgTypeL, VArgType, IBeforeJ, 1)) {
2521 // This is another short-circuit case: we're combining a scalar into
2522 // a vector that is formed by an IE chain. We've just expanded the IE
2523 // chain, now insert the scalar and we're done.
2525 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0,
2526 getReplacementName(IBeforeJ ? I : J, true, o));
2527 S->insertBefore(IBeforeJ ? J : I);
2529 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL,
2530 ArgTypeH, IBeforeJ)) {
2531 // The two vector inputs to the shuffle must be the same length,
2532 // so extend the smaller vector to be the same length as the larger one.
2536 std::vector<Constant *> Mask(numElemH);
2538 for (; v < numElemL; ++v)
2539 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2540 for (; v < numElemH; ++v)
2541 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2543 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL),
2544 ConstantVector::get(Mask),
2545 getReplacementName(IBeforeJ ? I : J,
2548 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0,
2549 getReplacementName(IBeforeJ ? I : J,
2553 NLOp->insertBefore(IBeforeJ ? J : I);
2558 } else if (numElemL > numElemH) {
2559 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL,
2560 ArgTypeH, VArgType, IBeforeJ)) {
2562 InsertElementInst::Create(LOp, HOp,
2563 ConstantInt::get(Type::getInt32Ty(Context),
2565 getReplacementName(IBeforeJ ? I : J,
2567 S->insertBefore(IBeforeJ ? J : I);
2569 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH,
2570 ArgTypeL, IBeforeJ)) {
2573 std::vector<Constant *> Mask(numElemL);
2575 for (; v < numElemH; ++v)
2576 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2577 for (; v < numElemL; ++v)
2578 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2580 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH),
2581 ConstantVector::get(Mask),
2582 getReplacementName(IBeforeJ ? I : J,
2585 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0,
2586 getReplacementName(IBeforeJ ? I : J,
2590 NHOp->insertBefore(IBeforeJ ? J : I);
2595 if (ArgType->isVectorTy()) {
2596 unsigned numElem = cast<VectorType>(VArgType)->getNumElements();
2597 std::vector<Constant*> Mask(numElem);
2598 for (unsigned v = 0; v < numElem; ++v) {
2600 // If the low vector was expanded, we need to skip the extra
2601 // undefined entries.
2602 if (v >= numElemL && numElemH > numElemL)
2603 Idx += (numElemH - numElemL);
2604 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2607 Instruction *BV = new ShuffleVectorInst(LOp, HOp,
2608 ConstantVector::get(Mask),
2609 getReplacementName(IBeforeJ ? I : J, true, o));
2610 BV->insertBefore(IBeforeJ ? J : I);
2614 Instruction *BV1 = InsertElementInst::Create(
2615 UndefValue::get(VArgType), LOp, CV0,
2616 getReplacementName(IBeforeJ ? I : J,
2618 BV1->insertBefore(IBeforeJ ? J : I);
2619 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1,
2620 getReplacementName(IBeforeJ ? I : J,
2622 BV2->insertBefore(IBeforeJ ? J : I);
2626 // This function creates an array of values that will be used as the inputs
2627 // to the vector instruction that fuses I with J.
2628 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context,
2629 Instruction *I, Instruction *J,
2630 SmallVector<Value *, 3> &ReplacedOperands,
2632 unsigned NumOperands = I->getNumOperands();
2634 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) {
2635 // Iterate backward so that we look at the store pointer
2636 // first and know whether or not we need to flip the inputs.
2638 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) {
2639 // This is the pointer for a load/store instruction.
2640 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o);
2642 } else if (isa<CallInst>(I)) {
2643 Function *F = cast<CallInst>(I)->getCalledFunction();
2644 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID();
2645 if (o == NumOperands-1) {
2646 BasicBlock &BB = *I->getParent();
2648 Module *M = BB.getParent()->getParent();
2649 Type *ArgTypeI = I->getType();
2650 Type *ArgTypeJ = J->getType();
2651 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2653 ReplacedOperands[o] = Intrinsic::getDeclaration(M, IID, VArgType);
2655 } else if (IID == Intrinsic::powi && o == 1) {
2656 // The second argument of powi is a single integer and we've already
2657 // checked that both arguments are equal. As a result, we just keep
2658 // I's second argument.
2659 ReplacedOperands[o] = I->getOperand(o);
2662 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) {
2663 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J);
2667 ReplacedOperands[o] = getReplacementInput(Context, I, J, o, IBeforeJ);
2671 // This function creates two values that represent the outputs of the
2672 // original I and J instructions. These are generally vector shuffles
2673 // or extracts. In many cases, these will end up being unused and, thus,
2674 // eliminated by later passes.
2675 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
2676 Instruction *J, Instruction *K,
2677 Instruction *&InsertionPt,
2678 Instruction *&K1, Instruction *&K2) {
2679 if (isa<StoreInst>(I)) {
2680 AA->replaceWithNewValue(I, K);
2681 AA->replaceWithNewValue(J, K);
2683 Type *IType = I->getType();
2684 Type *JType = J->getType();
2686 VectorType *VType = getVecTypeForPair(IType, JType);
2687 unsigned numElem = VType->getNumElements();
2689 unsigned numElemI, numElemJ;
2690 if (IType->isVectorTy())
2691 numElemI = cast<VectorType>(IType)->getNumElements();
2695 if (JType->isVectorTy())
2696 numElemJ = cast<VectorType>(JType)->getNumElements();
2700 if (IType->isVectorTy()) {
2701 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI);
2702 for (unsigned v = 0; v < numElemI; ++v) {
2703 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2704 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v);
2707 K1 = new ShuffleVectorInst(K, UndefValue::get(VType),
2708 ConstantVector::get( Mask1),
2709 getReplacementName(K, false, 1));
2711 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2712 K1 = ExtractElementInst::Create(K, CV0,
2713 getReplacementName(K, false, 1));
2716 if (JType->isVectorTy()) {
2717 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ);
2718 for (unsigned v = 0; v < numElemJ; ++v) {
2719 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2720 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v);
2723 K2 = new ShuffleVectorInst(K, UndefValue::get(VType),
2724 ConstantVector::get( Mask2),
2725 getReplacementName(K, false, 2));
2727 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1);
2728 K2 = ExtractElementInst::Create(K, CV1,
2729 getReplacementName(K, false, 2));
2733 K2->insertAfter(K1);
2738 // Move all uses of the function I (including pairing-induced uses) after J.
2739 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB,
2740 std::multimap<Value *, Value *> &LoadMoveSet,
2741 Instruction *I, Instruction *J) {
2742 // Skip to the first instruction past I.
2743 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2745 DenseSet<Value *> Users;
2746 AliasSetTracker WriteSet(*AA);
2747 for (; cast<Instruction>(L) != J; ++L)
2748 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet);
2750 assert(cast<Instruction>(L) == J &&
2751 "Tracking has not proceeded far enough to check for dependencies");
2752 // If J is now in the use set of I, then trackUsesOfI will return true
2753 // and we have a dependency cycle (and the fusing operation must abort).
2754 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSet);
2757 // Move all uses of the function I (including pairing-induced uses) after J.
2758 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB,
2759 std::multimap<Value *, Value *> &LoadMoveSet,
2760 Instruction *&InsertionPt,
2761 Instruction *I, Instruction *J) {
2762 // Skip to the first instruction past I.
2763 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2765 DenseSet<Value *> Users;
2766 AliasSetTracker WriteSet(*AA);
2767 for (; cast<Instruction>(L) != J;) {
2768 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet)) {
2769 // Move this instruction
2770 Instruction *InstToMove = L; ++L;
2772 DEBUG(dbgs() << "BBV: moving: " << *InstToMove <<
2773 " to after " << *InsertionPt << "\n");
2774 InstToMove->removeFromParent();
2775 InstToMove->insertAfter(InsertionPt);
2776 InsertionPt = InstToMove;
2783 // Collect all load instruction that are in the move set of a given first
2784 // pair member. These loads depend on the first instruction, I, and so need
2785 // to be moved after J (the second instruction) when the pair is fused.
2786 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB,
2787 DenseMap<Value *, Value *> &ChosenPairs,
2788 std::multimap<Value *, Value *> &LoadMoveSet,
2790 // Skip to the first instruction past I.
2791 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2793 DenseSet<Value *> Users;
2794 AliasSetTracker WriteSet(*AA);
2796 // Note: We cannot end the loop when we reach J because J could be moved
2797 // farther down the use chain by another instruction pairing. Also, J
2798 // could be before I if this is an inverted input.
2799 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) {
2800 if (trackUsesOfI(Users, WriteSet, I, L)) {
2801 if (L->mayReadFromMemory())
2802 LoadMoveSet.insert(ValuePair(L, I));
2807 // In cases where both load/stores and the computation of their pointers
2808 // are chosen for vectorization, we can end up in a situation where the
2809 // aliasing analysis starts returning different query results as the
2810 // process of fusing instruction pairs continues. Because the algorithm
2811 // relies on finding the same use trees here as were found earlier, we'll
2812 // need to precompute the necessary aliasing information here and then
2813 // manually update it during the fusion process.
2814 void BBVectorize::collectLoadMoveSet(BasicBlock &BB,
2815 std::vector<Value *> &PairableInsts,
2816 DenseMap<Value *, Value *> &ChosenPairs,
2817 std::multimap<Value *, Value *> &LoadMoveSet) {
2818 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
2819 PIE = PairableInsts.end(); PI != PIE; ++PI) {
2820 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI);
2821 if (P == ChosenPairs.end()) continue;
2823 Instruction *I = cast<Instruction>(P->first);
2824 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, I);
2828 // When the first instruction in each pair is cloned, it will inherit its
2829 // parent's metadata. This metadata must be combined with that of the other
2830 // instruction in a safe way.
2831 void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) {
2832 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata;
2833 K->getAllMetadataOtherThanDebugLoc(Metadata);
2834 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) {
2835 unsigned Kind = Metadata[i].first;
2836 MDNode *JMD = J->getMetadata(Kind);
2837 MDNode *KMD = Metadata[i].second;
2841 K->setMetadata(Kind, 0); // Remove unknown metadata
2843 case LLVMContext::MD_tbaa:
2844 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD));
2846 case LLVMContext::MD_fpmath:
2847 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD));
2853 // This function fuses the chosen instruction pairs into vector instructions,
2854 // taking care preserve any needed scalar outputs and, then, it reorders the
2855 // remaining instructions as needed (users of the first member of the pair
2856 // need to be moved to after the location of the second member of the pair
2857 // because the vector instruction is inserted in the location of the pair's
2859 void BBVectorize::fuseChosenPairs(BasicBlock &BB,
2860 std::vector<Value *> &PairableInsts,
2861 DenseMap<Value *, Value *> &ChosenPairs,
2862 DenseSet<ValuePair> &FixedOrderPairs,
2863 DenseMap<VPPair, unsigned> &PairConnectionTypes,
2864 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
2865 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps) {
2866 LLVMContext& Context = BB.getContext();
2868 // During the vectorization process, the order of the pairs to be fused
2869 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs
2870 // list. After a pair is fused, the flipped pair is removed from the list.
2871 DenseSet<ValuePair> FlippedPairs;
2872 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(),
2873 E = ChosenPairs.end(); P != E; ++P)
2874 FlippedPairs.insert(ValuePair(P->second, P->first));
2875 for (DenseSet<ValuePair>::iterator P = FlippedPairs.begin(),
2876 E = FlippedPairs.end(); P != E; ++P)
2877 ChosenPairs.insert(*P);
2879 std::multimap<Value *, Value *> LoadMoveSet;
2880 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, LoadMoveSet);
2882 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n");
2884 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) {
2885 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI);
2886 if (P == ChosenPairs.end()) {
2891 if (getDepthFactor(P->first) == 0) {
2892 // These instructions are not really fused, but are tracked as though
2893 // they are. Any case in which it would be interesting to fuse them
2894 // will be taken care of by InstCombine.
2900 Instruction *I = cast<Instruction>(P->first),
2901 *J = cast<Instruction>(P->second);
2903 DEBUG(dbgs() << "BBV: fusing: " << *I <<
2904 " <-> " << *J << "\n");
2906 // Remove the pair and flipped pair from the list.
2907 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second);
2908 assert(FP != ChosenPairs.end() && "Flipped pair not found in list");
2909 ChosenPairs.erase(FP);
2910 ChosenPairs.erase(P);
2912 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSet, I, J)) {
2913 DEBUG(dbgs() << "BBV: fusion of: " << *I <<
2915 " aborted because of non-trivial dependency cycle\n");
2921 // If the pair must have the other order, then flip it.
2922 bool FlipPairOrder = FixedOrderPairs.count(ValuePair(J, I));
2923 if (!FlipPairOrder && !FixedOrderPairs.count(ValuePair(I, J))) {
2924 // This pair does not have a fixed order, and so we might want to
2925 // flip it if that will yield fewer shuffles. We count the number
2926 // of dependencies connected via swaps, and those directly connected,
2927 // and flip the order if the number of swaps is greater.
2928 bool OrigOrder = true;
2929 VPPIteratorPair IP = ConnectedPairDeps.equal_range(ValuePair(I, J));
2930 if (IP.first == ConnectedPairDeps.end()) {
2931 IP = ConnectedPairDeps.equal_range(ValuePair(J, I));
2935 if (IP.first != ConnectedPairDeps.end()) {
2936 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
2937 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first;
2938 Q != IP.second; ++Q) {
2939 DenseMap<VPPair, unsigned>::iterator R =
2940 PairConnectionTypes.find(VPPair(Q->second, Q->first));
2941 assert(R != PairConnectionTypes.end() &&
2942 "Cannot find pair connection type");
2943 if (R->second == PairConnectionDirect)
2945 else if (R->second == PairConnectionSwap)
2950 std::swap(NumDepsDirect, NumDepsSwap);
2952 if (NumDepsSwap > NumDepsDirect) {
2953 FlipPairOrder = true;
2954 DEBUG(dbgs() << "BBV: reordering pair: " << *I <<
2955 " <-> " << *J << "\n");
2960 Instruction *L = I, *H = J;
2964 // If the pair being fused uses the opposite order from that in the pair
2965 // connection map, then we need to flip the types.
2966 VPPIteratorPair IP = ConnectedPairs.equal_range(ValuePair(H, L));
2967 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first;
2968 Q != IP.second; ++Q) {
2969 DenseMap<VPPair, unsigned>::iterator R = PairConnectionTypes.find(*Q);
2970 assert(R != PairConnectionTypes.end() &&
2971 "Cannot find pair connection type");
2972 if (R->second == PairConnectionDirect)
2973 R->second = PairConnectionSwap;
2974 else if (R->second == PairConnectionSwap)
2975 R->second = PairConnectionDirect;
2978 bool LBeforeH = !FlipPairOrder;
2979 unsigned NumOperands = I->getNumOperands();
2980 SmallVector<Value *, 3> ReplacedOperands(NumOperands);
2981 getReplacementInputsForPair(Context, L, H, ReplacedOperands,
2984 // Make a copy of the original operation, change its type to the vector
2985 // type and replace its operands with the vector operands.
2986 Instruction *K = L->clone();
2989 else if (H->hasName())
2992 if (!isa<StoreInst>(K))
2993 K->mutateType(getVecTypeForPair(L->getType(), H->getType()));
2995 combineMetadata(K, H);
2996 K->intersectOptionalDataWith(H);
2998 for (unsigned o = 0; o < NumOperands; ++o)
2999 K->setOperand(o, ReplacedOperands[o]);
3003 // Instruction insertion point:
3004 Instruction *InsertionPt = K;
3005 Instruction *K1 = 0, *K2 = 0;
3006 replaceOutputsOfPair(Context, L, H, K, InsertionPt, K1, K2);
3008 // The use tree of the first original instruction must be moved to after
3009 // the location of the second instruction. The entire use tree of the
3010 // first instruction is disjoint from the input tree of the second
3011 // (by definition), and so commutes with it.
3013 moveUsesOfIAfterJ(BB, LoadMoveSet, InsertionPt, I, J);
3015 if (!isa<StoreInst>(I)) {
3016 L->replaceAllUsesWith(K1);
3017 H->replaceAllUsesWith(K2);
3018 AA->replaceWithNewValue(L, K1);
3019 AA->replaceWithNewValue(H, K2);
3022 // Instructions that may read from memory may be in the load move set.
3023 // Once an instruction is fused, we no longer need its move set, and so
3024 // the values of the map never need to be updated. However, when a load
3025 // is fused, we need to merge the entries from both instructions in the
3026 // pair in case those instructions were in the move set of some other
3027 // yet-to-be-fused pair. The loads in question are the keys of the map.
3028 if (I->mayReadFromMemory()) {
3029 std::vector<ValuePair> NewSetMembers;
3030 VPIteratorPair IPairRange = LoadMoveSet.equal_range(I);
3031 VPIteratorPair JPairRange = LoadMoveSet.equal_range(J);
3032 for (std::multimap<Value *, Value *>::iterator N = IPairRange.first;
3033 N != IPairRange.second; ++N)
3034 NewSetMembers.push_back(ValuePair(K, N->second));
3035 for (std::multimap<Value *, Value *>::iterator N = JPairRange.first;
3036 N != JPairRange.second; ++N)
3037 NewSetMembers.push_back(ValuePair(K, N->second));
3038 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(),
3039 AE = NewSetMembers.end(); A != AE; ++A)
3040 LoadMoveSet.insert(*A);
3043 // Before removing I, set the iterator to the next instruction.
3044 PI = llvm::next(BasicBlock::iterator(I));
3045 if (cast<Instruction>(PI) == J)
3050 I->eraseFromParent();
3051 J->eraseFromParent();
3053 DEBUG(if (PrintAfterEveryPair) dbgs() << "BBV: block is now: \n" <<
3057 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n");
3061 char BBVectorize::ID = 0;
3062 static const char bb_vectorize_name[] = "Basic-Block Vectorization";
3063 INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
3064 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
3065 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
3066 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
3067 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3068 INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
3070 BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) {
3071 return new BBVectorize(C);
3075 llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) {
3076 BBVectorize BBVectorizer(P, C);
3077 return BBVectorizer.vectorizeBB(BB);
3080 //===----------------------------------------------------------------------===//
3081 VectorizeConfig::VectorizeConfig() {
3082 VectorBits = ::VectorBits;
3083 VectorizeBools = !::NoBools;
3084 VectorizeInts = !::NoInts;
3085 VectorizeFloats = !::NoFloats;
3086 VectorizePointers = !::NoPointers;
3087 VectorizeCasts = !::NoCasts;
3088 VectorizeMath = !::NoMath;
3089 VectorizeFMA = !::NoFMA;
3090 VectorizeSelect = !::NoSelect;
3091 VectorizeCmp = !::NoCmp;
3092 VectorizeGEP = !::NoGEP;
3093 VectorizeMemOps = !::NoMemOps;
3094 AlignedOnly = ::AlignedOnly;
3095 ReqChainDepth= ::ReqChainDepth;
3096 SearchLimit = ::SearchLimit;
3097 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck;
3098 SplatBreaksChain = ::SplatBreaksChain;
3099 MaxInsts = ::MaxInsts;
3100 MaxIter = ::MaxIter;
3101 Pow2LenOnly = ::Pow2LenOnly;
3102 NoMemOpBoost = ::NoMemOpBoost;
3103 FastDep = ::FastDep;