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/ValueTracking.h"
33 #include "llvm/Constants.h"
34 #include "llvm/DataLayout.h"
35 #include "llvm/DerivedTypes.h"
36 #include "llvm/Function.h"
37 #include "llvm/Instructions.h"
38 #include "llvm/IntrinsicInst.h"
39 #include "llvm/Intrinsics.h"
40 #include "llvm/LLVMContext.h"
41 #include "llvm/Metadata.h"
42 #include "llvm/Pass.h"
43 #include "llvm/Support/CommandLine.h"
44 #include "llvm/Support/Debug.h"
45 #include "llvm/Support/ValueHandle.h"
46 #include "llvm/Support/raw_ostream.h"
47 #include "llvm/TargetTransformInfo.h"
48 #include "llvm/Transforms/Utils/Local.h"
49 #include "llvm/Type.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 :
203 P->getAnalysisIfAvailable<TargetTransformInfo>();
204 VTTI = TTI ? TTI->getVectorTargetTransformInfo() : 0;
207 typedef std::pair<Value *, Value *> ValuePair;
208 typedef std::pair<ValuePair, int> ValuePairWithCost;
209 typedef std::pair<ValuePair, size_t> ValuePairWithDepth;
210 typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair
211 typedef std::pair<VPPair, unsigned> VPPairWithType;
212 typedef std::pair<std::multimap<Value *, Value *>::iterator,
213 std::multimap<Value *, Value *>::iterator> VPIteratorPair;
214 typedef std::pair<std::multimap<ValuePair, ValuePair>::iterator,
215 std::multimap<ValuePair, ValuePair>::iterator>
222 TargetTransformInfo *TTI;
223 const VectorTargetTransformInfo *VTTI;
225 // FIXME: const correct?
227 bool vectorizePairs(BasicBlock &BB, bool NonPow2Len = false);
229 bool getCandidatePairs(BasicBlock &BB,
230 BasicBlock::iterator &Start,
231 std::multimap<Value *, Value *> &CandidatePairs,
232 DenseSet<ValuePair> &FixedOrderPairs,
233 DenseMap<ValuePair, int> &CandidatePairCostSavings,
234 std::vector<Value *> &PairableInsts, bool NonPow2Len);
236 // FIXME: The current implementation does not account for pairs that
237 // are connected in multiple ways. For example:
238 // C1 = A1 / A2; C2 = A2 / A1 (which may be both direct and a swap)
239 enum PairConnectionType {
240 PairConnectionDirect,
245 void computeConnectedPairs(std::multimap<Value *, Value *> &CandidatePairs,
246 std::vector<Value *> &PairableInsts,
247 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
248 DenseMap<VPPair, unsigned> &PairConnectionTypes);
250 void buildDepMap(BasicBlock &BB,
251 std::multimap<Value *, Value *> &CandidatePairs,
252 std::vector<Value *> &PairableInsts,
253 DenseSet<ValuePair> &PairableInstUsers);
255 void choosePairs(std::multimap<Value *, Value *> &CandidatePairs,
256 DenseMap<ValuePair, int> &CandidatePairCostSavings,
257 std::vector<Value *> &PairableInsts,
258 DenseSet<ValuePair> &FixedOrderPairs,
259 DenseMap<VPPair, unsigned> &PairConnectionTypes,
260 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
261 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps,
262 DenseSet<ValuePair> &PairableInstUsers,
263 DenseMap<Value *, Value *>& ChosenPairs);
265 void fuseChosenPairs(BasicBlock &BB,
266 std::vector<Value *> &PairableInsts,
267 DenseMap<Value *, Value *>& ChosenPairs,
268 DenseSet<ValuePair> &FixedOrderPairs,
269 DenseMap<VPPair, unsigned> &PairConnectionTypes,
270 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
271 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps);
274 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore);
276 bool areInstsCompatible(Instruction *I, Instruction *J,
277 bool IsSimpleLoadStore, bool NonPow2Len,
278 int &CostSavings, int &FixedOrder);
280 bool trackUsesOfI(DenseSet<Value *> &Users,
281 AliasSetTracker &WriteSet, Instruction *I,
282 Instruction *J, bool UpdateUsers = true,
283 std::multimap<Value *, Value *> *LoadMoveSet = 0);
285 void computePairsConnectedTo(
286 std::multimap<Value *, Value *> &CandidatePairs,
287 std::vector<Value *> &PairableInsts,
288 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
289 DenseMap<VPPair, unsigned> &PairConnectionTypes,
292 bool pairsConflict(ValuePair P, ValuePair Q,
293 DenseSet<ValuePair> &PairableInstUsers,
294 std::multimap<ValuePair, ValuePair> *PairableInstUserMap = 0);
296 bool pairWillFormCycle(ValuePair P,
297 std::multimap<ValuePair, ValuePair> &PairableInstUsers,
298 DenseSet<ValuePair> &CurrentPairs);
301 std::multimap<Value *, Value *> &CandidatePairs,
302 std::vector<Value *> &PairableInsts,
303 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
304 DenseSet<ValuePair> &PairableInstUsers,
305 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
306 DenseMap<Value *, Value *> &ChosenPairs,
307 DenseMap<ValuePair, size_t> &Tree,
308 DenseSet<ValuePair> &PrunedTree, ValuePair J,
311 void buildInitialTreeFor(
312 std::multimap<Value *, Value *> &CandidatePairs,
313 std::vector<Value *> &PairableInsts,
314 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
315 DenseSet<ValuePair> &PairableInstUsers,
316 DenseMap<Value *, Value *> &ChosenPairs,
317 DenseMap<ValuePair, size_t> &Tree, ValuePair J);
319 void findBestTreeFor(
320 std::multimap<Value *, Value *> &CandidatePairs,
321 DenseMap<ValuePair, int> &CandidatePairCostSavings,
322 std::vector<Value *> &PairableInsts,
323 DenseSet<ValuePair> &FixedOrderPairs,
324 DenseMap<VPPair, unsigned> &PairConnectionTypes,
325 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
326 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps,
327 DenseSet<ValuePair> &PairableInstUsers,
328 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
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 (VTTI) 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 (VTTI || 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 :
430 getAnalysisIfAvailable<TargetTransformInfo>();
431 VTTI = TTI ? TTI->getVectorTargetTransformInfo() : 0;
433 return vectorizeBB(BB);
436 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
437 BasicBlockPass::getAnalysisUsage(AU);
438 AU.addRequired<AliasAnalysis>();
439 AU.addRequired<DominatorTree>();
440 AU.addRequired<ScalarEvolution>();
441 AU.addPreserved<AliasAnalysis>();
442 AU.addPreserved<DominatorTree>();
443 AU.addPreserved<ScalarEvolution>();
444 AU.setPreservesCFG();
447 static inline VectorType *getVecTypeForPair(Type *ElemTy, Type *Elem2Ty) {
448 assert(ElemTy->getScalarType() == Elem2Ty->getScalarType() &&
449 "Cannot form vector from incompatible scalar types");
450 Type *STy = ElemTy->getScalarType();
453 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) {
454 numElem = VTy->getNumElements();
459 if (VectorType *VTy = dyn_cast<VectorType>(Elem2Ty)) {
460 numElem += VTy->getNumElements();
465 return VectorType::get(STy, numElem);
468 static inline void getInstructionTypes(Instruction *I,
469 Type *&T1, Type *&T2) {
470 if (isa<StoreInst>(I)) {
471 // For stores, it is the value type, not the pointer type that matters
472 // because the value is what will come from a vector register.
474 Value *IVal = cast<StoreInst>(I)->getValueOperand();
475 T1 = IVal->getType();
481 T2 = cast<CastInst>(I)->getSrcTy();
485 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
486 T2 = SI->getCondition()->getType();
487 } else if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(I)) {
488 T2 = SI->getOperand(0)->getType();
489 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) {
490 T2 = CI->getOperand(0)->getType();
494 // Returns the weight associated with the provided value. A chain of
495 // candidate pairs has a length given by the sum of the weights of its
496 // members (one weight per pair; the weight of each member of the pair
497 // is assumed to be the same). This length is then compared to the
498 // chain-length threshold to determine if a given chain is significant
499 // enough to be vectorized. The length is also used in comparing
500 // candidate chains where longer chains are considered to be better.
501 // Note: when this function returns 0, the resulting instructions are
502 // not actually fused.
503 inline size_t getDepthFactor(Value *V) {
504 // InsertElement and ExtractElement have a depth factor of zero. This is
505 // for two reasons: First, they cannot be usefully fused. Second, because
506 // the pass generates a lot of these, they can confuse the simple metric
507 // used to compare the trees in the next iteration. Thus, giving them a
508 // weight of zero allows the pass to essentially ignore them in
509 // subsequent iterations when looking for vectorization opportunities
510 // while still tracking dependency chains that flow through those
512 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V))
515 // Give a load or store half of the required depth so that load/store
516 // pairs will vectorize.
517 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V)))
518 return Config.ReqChainDepth/2;
523 // Returns the cost of the provided instruction using VTTI.
524 // This does not handle loads and stores.
525 unsigned getInstrCost(unsigned Opcode, Type *T1, Type *T2) {
528 case Instruction::GetElementPtr:
529 // We mark this instruction as zero-cost because scalar GEPs are usually
530 // lowered to the intruction addressing mode. At the moment we don't
531 // generate vector GEPs.
533 case Instruction::Br:
534 return VTTI->getCFInstrCost(Opcode);
535 case Instruction::PHI:
537 case Instruction::Add:
538 case Instruction::FAdd:
539 case Instruction::Sub:
540 case Instruction::FSub:
541 case Instruction::Mul:
542 case Instruction::FMul:
543 case Instruction::UDiv:
544 case Instruction::SDiv:
545 case Instruction::FDiv:
546 case Instruction::URem:
547 case Instruction::SRem:
548 case Instruction::FRem:
549 case Instruction::Shl:
550 case Instruction::LShr:
551 case Instruction::AShr:
552 case Instruction::And:
553 case Instruction::Or:
554 case Instruction::Xor:
555 return VTTI->getArithmeticInstrCost(Opcode, T1);
556 case Instruction::Select:
557 case Instruction::ICmp:
558 case Instruction::FCmp:
559 return VTTI->getCmpSelInstrCost(Opcode, T1, T2);
560 case Instruction::ZExt:
561 case Instruction::SExt:
562 case Instruction::FPToUI:
563 case Instruction::FPToSI:
564 case Instruction::FPExt:
565 case Instruction::PtrToInt:
566 case Instruction::IntToPtr:
567 case Instruction::SIToFP:
568 case Instruction::UIToFP:
569 case Instruction::Trunc:
570 case Instruction::FPTrunc:
571 case Instruction::BitCast:
572 case Instruction::ShuffleVector:
573 return VTTI->getCastInstrCost(Opcode, T1, T2);
579 // This determines the relative offset of two loads or stores, returning
580 // true if the offset could be determined to be some constant value.
581 // For example, if OffsetInElmts == 1, then J accesses the memory directly
582 // after I; if OffsetInElmts == -1 then I accesses the memory
584 bool getPairPtrInfo(Instruction *I, Instruction *J,
585 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment,
586 unsigned &IAddressSpace, unsigned &JAddressSpace,
587 int64_t &OffsetInElmts, bool ComputeOffset = true) {
589 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
590 LoadInst *LJ = cast<LoadInst>(J);
591 IPtr = LI->getPointerOperand();
592 JPtr = LJ->getPointerOperand();
593 IAlignment = LI->getAlignment();
594 JAlignment = LJ->getAlignment();
595 IAddressSpace = LI->getPointerAddressSpace();
596 JAddressSpace = LJ->getPointerAddressSpace();
598 StoreInst *SI = cast<StoreInst>(I), *SJ = cast<StoreInst>(J);
599 IPtr = SI->getPointerOperand();
600 JPtr = SJ->getPointerOperand();
601 IAlignment = SI->getAlignment();
602 JAlignment = SJ->getAlignment();
603 IAddressSpace = SI->getPointerAddressSpace();
604 JAddressSpace = SJ->getPointerAddressSpace();
610 const SCEV *IPtrSCEV = SE->getSCEV(IPtr);
611 const SCEV *JPtrSCEV = SE->getSCEV(JPtr);
613 // If this is a trivial offset, then we'll get something like
614 // 1*sizeof(type). With target data, which we need anyway, this will get
615 // constant folded into a number.
616 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV);
617 if (const SCEVConstant *ConstOffSCEV =
618 dyn_cast<SCEVConstant>(OffsetSCEV)) {
619 ConstantInt *IntOff = ConstOffSCEV->getValue();
620 int64_t Offset = IntOff->getSExtValue();
622 Type *VTy = cast<PointerType>(IPtr->getType())->getElementType();
623 int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy);
625 Type *VTy2 = cast<PointerType>(JPtr->getType())->getElementType();
626 if (VTy != VTy2 && Offset < 0) {
627 int64_t VTy2TSS = (int64_t) TD->getTypeStoreSize(VTy2);
628 OffsetInElmts = Offset/VTy2TSS;
629 return (abs64(Offset) % VTy2TSS) == 0;
632 OffsetInElmts = Offset/VTyTSS;
633 return (abs64(Offset) % VTyTSS) == 0;
639 // Returns true if the provided CallInst represents an intrinsic that can
641 bool isVectorizableIntrinsic(CallInst* I) {
642 Function *F = I->getCalledFunction();
643 if (!F) return false;
645 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID();
646 if (!IID) return false;
651 case Intrinsic::sqrt:
652 case Intrinsic::powi:
656 case Intrinsic::log2:
657 case Intrinsic::log10:
659 case Intrinsic::exp2:
661 return Config.VectorizeMath;
663 case Intrinsic::fmuladd:
664 return Config.VectorizeFMA;
668 // Returns true if J is the second element in some pair referenced by
669 // some multimap pair iterator pair.
670 template <typename V>
671 bool isSecondInIteratorPair(V J, std::pair<
672 typename std::multimap<V, V>::iterator,
673 typename std::multimap<V, V>::iterator> PairRange) {
674 for (typename std::multimap<V, V>::iterator K = PairRange.first;
675 K != PairRange.second; ++K)
676 if (K->second == J) return true;
681 bool isPureIEChain(InsertElementInst *IE) {
682 InsertElementInst *IENext = IE;
684 if (!isa<UndefValue>(IENext->getOperand(0)) &&
685 !isa<InsertElementInst>(IENext->getOperand(0))) {
689 dyn_cast<InsertElementInst>(IENext->getOperand(0))));
695 // This function implements one vectorization iteration on the provided
696 // basic block. It returns true if the block is changed.
697 bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) {
699 BasicBlock::iterator Start = BB.getFirstInsertionPt();
701 std::vector<Value *> AllPairableInsts;
702 DenseMap<Value *, Value *> AllChosenPairs;
703 DenseSet<ValuePair> AllFixedOrderPairs;
704 DenseMap<VPPair, unsigned> AllPairConnectionTypes;
705 std::multimap<ValuePair, ValuePair> AllConnectedPairs, AllConnectedPairDeps;
708 std::vector<Value *> PairableInsts;
709 std::multimap<Value *, Value *> CandidatePairs;
710 DenseSet<ValuePair> FixedOrderPairs;
711 DenseMap<ValuePair, int> CandidatePairCostSavings;
712 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs,
714 CandidatePairCostSavings,
715 PairableInsts, NonPow2Len);
716 if (PairableInsts.empty()) continue;
718 // Now we have a map of all of the pairable instructions and we need to
719 // select the best possible pairing. A good pairing is one such that the
720 // users of the pair are also paired. This defines a (directed) forest
721 // over the pairs such that two pairs are connected iff the second pair
724 // Note that it only matters that both members of the second pair use some
725 // element of the first pair (to allow for splatting).
727 std::multimap<ValuePair, ValuePair> ConnectedPairs, ConnectedPairDeps;
728 DenseMap<VPPair, unsigned> PairConnectionTypes;
729 computeConnectedPairs(CandidatePairs, PairableInsts, ConnectedPairs,
730 PairConnectionTypes);
731 if (ConnectedPairs.empty()) continue;
733 for (std::multimap<ValuePair, ValuePair>::iterator
734 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
736 ConnectedPairDeps.insert(VPPair(I->second, I->first));
739 // Build the pairable-instruction dependency map
740 DenseSet<ValuePair> PairableInstUsers;
741 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers);
743 // There is now a graph of the connected pairs. For each variable, pick
744 // the pairing with the largest tree meeting the depth requirement on at
745 // least one branch. Then select all pairings that are part of that tree
746 // and remove them from the list of available pairings and pairable
749 DenseMap<Value *, Value *> ChosenPairs;
750 choosePairs(CandidatePairs, CandidatePairCostSavings,
751 PairableInsts, FixedOrderPairs, PairConnectionTypes,
752 ConnectedPairs, ConnectedPairDeps,
753 PairableInstUsers, ChosenPairs);
755 if (ChosenPairs.empty()) continue;
756 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(),
757 PairableInsts.end());
758 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end());
760 // Only for the chosen pairs, propagate information on fixed-order pairs,
761 // pair connections, and their types to the data structures used by the
762 // pair fusion procedures.
763 for (DenseMap<Value *, Value *>::iterator I = ChosenPairs.begin(),
764 IE = ChosenPairs.end(); I != IE; ++I) {
765 if (FixedOrderPairs.count(*I))
766 AllFixedOrderPairs.insert(*I);
767 else if (FixedOrderPairs.count(ValuePair(I->second, I->first)))
768 AllFixedOrderPairs.insert(ValuePair(I->second, I->first));
770 for (DenseMap<Value *, Value *>::iterator J = ChosenPairs.begin();
772 DenseMap<VPPair, unsigned>::iterator K =
773 PairConnectionTypes.find(VPPair(*I, *J));
774 if (K != PairConnectionTypes.end()) {
775 AllPairConnectionTypes.insert(*K);
777 K = PairConnectionTypes.find(VPPair(*J, *I));
778 if (K != PairConnectionTypes.end())
779 AllPairConnectionTypes.insert(*K);
784 for (std::multimap<ValuePair, ValuePair>::iterator
785 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
787 if (AllPairConnectionTypes.count(*I)) {
788 AllConnectedPairs.insert(*I);
789 AllConnectedPairDeps.insert(VPPair(I->second, I->first));
792 } while (ShouldContinue);
794 if (AllChosenPairs.empty()) return false;
795 NumFusedOps += AllChosenPairs.size();
797 // A set of pairs has now been selected. It is now necessary to replace the
798 // paired instructions with vector instructions. For this procedure each
799 // operand must be replaced with a vector operand. This vector is formed
800 // by using build_vector on the old operands. The replaced values are then
801 // replaced with a vector_extract on the result. Subsequent optimization
802 // passes should coalesce the build/extract combinations.
804 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs, AllFixedOrderPairs,
805 AllPairConnectionTypes,
806 AllConnectedPairs, AllConnectedPairDeps);
808 // It is important to cleanup here so that future iterations of this
809 // function have less work to do.
810 (void) SimplifyInstructionsInBlock(&BB, TD, AA->getTargetLibraryInfo());
814 // This function returns true if the provided instruction is capable of being
815 // fused into a vector instruction. This determination is based only on the
816 // type and other attributes of the instruction.
817 bool BBVectorize::isInstVectorizable(Instruction *I,
818 bool &IsSimpleLoadStore) {
819 IsSimpleLoadStore = false;
821 if (CallInst *C = dyn_cast<CallInst>(I)) {
822 if (!isVectorizableIntrinsic(C))
824 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) {
825 // Vectorize simple loads if possbile:
826 IsSimpleLoadStore = L->isSimple();
827 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
829 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) {
830 // Vectorize simple stores if possbile:
831 IsSimpleLoadStore = S->isSimple();
832 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
834 } else if (CastInst *C = dyn_cast<CastInst>(I)) {
835 // We can vectorize casts, but not casts of pointer types, etc.
836 if (!Config.VectorizeCasts)
839 Type *SrcTy = C->getSrcTy();
840 if (!SrcTy->isSingleValueType())
843 Type *DestTy = C->getDestTy();
844 if (!DestTy->isSingleValueType())
846 } else if (isa<SelectInst>(I)) {
847 if (!Config.VectorizeSelect)
849 } else if (isa<CmpInst>(I)) {
850 if (!Config.VectorizeCmp)
852 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) {
853 if (!Config.VectorizeGEP)
856 // Currently, vector GEPs exist only with one index.
857 if (G->getNumIndices() != 1)
859 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) ||
860 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) {
864 // We can't vectorize memory operations without target data
865 if (TD == 0 && IsSimpleLoadStore)
869 getInstructionTypes(I, T1, T2);
871 // Not every type can be vectorized...
872 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) ||
873 !(VectorType::isValidElementType(T2) || T2->isVectorTy()))
876 if (T1->getScalarSizeInBits() == 1) {
877 if (!Config.VectorizeBools)
880 if (!Config.VectorizeInts && T1->isIntOrIntVectorTy())
884 if (T2->getScalarSizeInBits() == 1) {
885 if (!Config.VectorizeBools)
888 if (!Config.VectorizeInts && T2->isIntOrIntVectorTy())
892 if (!Config.VectorizeFloats
893 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy()))
896 // Don't vectorize target-specific types.
897 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy())
899 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy())
902 if ((!Config.VectorizePointers || TD == 0) &&
903 (T1->getScalarType()->isPointerTy() ||
904 T2->getScalarType()->isPointerTy()))
907 if (!VTTI && (T1->getPrimitiveSizeInBits() >= Config.VectorBits ||
908 T2->getPrimitiveSizeInBits() >= Config.VectorBits))
914 // This function returns true if the two provided instructions are compatible
915 // (meaning that they can be fused into a vector instruction). This assumes
916 // that I has already been determined to be vectorizable and that J is not
917 // in the use tree of I.
918 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J,
919 bool IsSimpleLoadStore, bool NonPow2Len,
920 int &CostSavings, int &FixedOrder) {
921 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I <<
922 " <-> " << *J << "\n");
927 // Loads and stores can be merged if they have different alignments,
928 // but are otherwise the same.
929 if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment |
930 (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0)))
933 Type *IT1, *IT2, *JT1, *JT2;
934 getInstructionTypes(I, IT1, IT2);
935 getInstructionTypes(J, JT1, JT2);
936 unsigned MaxTypeBits = std::max(
937 IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(),
938 IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits());
939 if (!VTTI && MaxTypeBits > Config.VectorBits)
942 // FIXME: handle addsub-type operations!
944 if (IsSimpleLoadStore) {
946 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
947 int64_t OffsetInElmts = 0;
948 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
949 IAddressSpace, JAddressSpace,
950 OffsetInElmts) && abs64(OffsetInElmts) == 1) {
951 FixedOrder = (int) OffsetInElmts;
952 unsigned BottomAlignment = IAlignment;
953 if (OffsetInElmts < 0) BottomAlignment = JAlignment;
955 Type *aTypeI = isa<StoreInst>(I) ?
956 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType();
957 Type *aTypeJ = isa<StoreInst>(J) ?
958 cast<StoreInst>(J)->getValueOperand()->getType() : J->getType();
959 Type *VType = getVecTypeForPair(aTypeI, aTypeJ);
961 if (Config.AlignedOnly) {
962 // An aligned load or store is possible only if the instruction
963 // with the lower offset has an alignment suitable for the
966 unsigned VecAlignment = TD->getPrefTypeAlignment(VType);
967 if (BottomAlignment < VecAlignment)
972 unsigned ICost = VTTI->getMemoryOpCost(I->getOpcode(), aTypeI,
973 IAlignment, IAddressSpace);
974 unsigned JCost = VTTI->getMemoryOpCost(J->getOpcode(), aTypeJ,
975 JAlignment, JAddressSpace);
976 unsigned VCost = VTTI->getMemoryOpCost(I->getOpcode(), VType,
979 if (VCost > ICost + JCost)
982 // We don't want to fuse to a type that will be split, even
983 // if the two input types will also be split and there is no other
985 unsigned VParts = VTTI->getNumberOfParts(VType);
988 else if (!VParts && VCost == ICost + JCost)
991 CostSavings = ICost + JCost - VCost;
997 unsigned ICost = getInstrCost(I->getOpcode(), IT1, IT2);
998 unsigned JCost = getInstrCost(J->getOpcode(), JT1, JT2);
999 Type *VT1 = getVecTypeForPair(IT1, JT1),
1000 *VT2 = getVecTypeForPair(IT2, JT2);
1001 unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2);
1003 if (VCost > ICost + JCost)
1006 // We don't want to fuse to a type that will be split, even
1007 // if the two input types will also be split and there is no other
1009 unsigned VParts1 = VTTI->getNumberOfParts(VT1),
1010 VParts2 = VTTI->getNumberOfParts(VT2);
1011 if (VParts1 > 1 || VParts2 > 1)
1013 else if ((!VParts1 || !VParts2) && VCost == ICost + JCost)
1016 CostSavings = ICost + JCost - VCost;
1019 // The powi intrinsic is special because only the first argument is
1020 // vectorized, the second arguments must be equal.
1021 CallInst *CI = dyn_cast<CallInst>(I);
1023 if (CI && (FI = CI->getCalledFunction())) {
1024 Intrinsic::ID IID = (Intrinsic::ID) FI->getIntrinsicID();
1025 if (IID == Intrinsic::powi) {
1026 Value *A1I = CI->getArgOperand(1),
1027 *A1J = cast<CallInst>(J)->getArgOperand(1);
1028 const SCEV *A1ISCEV = SE->getSCEV(A1I),
1029 *A1JSCEV = SE->getSCEV(A1J);
1030 return (A1ISCEV == A1JSCEV);
1034 SmallVector<Type*, 4> Tys;
1035 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
1036 Tys.push_back(CI->getArgOperand(i)->getType());
1037 unsigned ICost = VTTI->getIntrinsicInstrCost(IID, IT1, Tys);
1040 CallInst *CJ = cast<CallInst>(J);
1041 for (unsigned i = 0, ie = CJ->getNumArgOperands(); i != ie; ++i)
1042 Tys.push_back(CJ->getArgOperand(i)->getType());
1043 unsigned JCost = VTTI->getIntrinsicInstrCost(IID, JT1, Tys);
1046 assert(CI->getNumArgOperands() == CJ->getNumArgOperands() &&
1047 "Intrinsic argument counts differ");
1048 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
1049 if (IID == Intrinsic::powi && i == 1)
1050 Tys.push_back(CI->getArgOperand(i)->getType());
1052 Tys.push_back(getVecTypeForPair(CI->getArgOperand(i)->getType(),
1053 CJ->getArgOperand(i)->getType()));
1056 Type *RetTy = getVecTypeForPair(IT1, JT1);
1057 unsigned VCost = VTTI->getIntrinsicInstrCost(IID, RetTy, Tys);
1059 if (VCost > ICost + JCost)
1062 // We don't want to fuse to a type that will be split, even
1063 // if the two input types will also be split and there is no other
1065 unsigned RetParts = VTTI->getNumberOfParts(RetTy);
1068 else if (!RetParts && VCost == ICost + JCost)
1071 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
1072 if (!Tys[i]->isVectorTy())
1075 unsigned NumParts = VTTI->getNumberOfParts(Tys[i]);
1078 else if (!NumParts && VCost == ICost + JCost)
1082 CostSavings = ICost + JCost - VCost;
1089 // Figure out whether or not J uses I and update the users and write-set
1090 // structures associated with I. Specifically, Users represents the set of
1091 // instructions that depend on I. WriteSet represents the set
1092 // of memory locations that are dependent on I. If UpdateUsers is true,
1093 // and J uses I, then Users is updated to contain J and WriteSet is updated
1094 // to contain any memory locations to which J writes. The function returns
1095 // true if J uses I. By default, alias analysis is used to determine
1096 // whether J reads from memory that overlaps with a location in WriteSet.
1097 // If LoadMoveSet is not null, then it is a previously-computed multimap
1098 // where the key is the memory-based user instruction and the value is
1099 // the instruction to be compared with I. So, if LoadMoveSet is provided,
1100 // then the alias analysis is not used. This is necessary because this
1101 // function is called during the process of moving instructions during
1102 // vectorization and the results of the alias analysis are not stable during
1104 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users,
1105 AliasSetTracker &WriteSet, Instruction *I,
1106 Instruction *J, bool UpdateUsers,
1107 std::multimap<Value *, Value *> *LoadMoveSet) {
1110 // This instruction may already be marked as a user due, for example, to
1111 // being a member of a selected pair.
1116 for (User::op_iterator JU = J->op_begin(), JE = J->op_end();
1119 if (I == V || Users.count(V)) {
1124 if (!UsesI && J->mayReadFromMemory()) {
1126 VPIteratorPair JPairRange = LoadMoveSet->equal_range(J);
1127 UsesI = isSecondInIteratorPair<Value*>(I, JPairRange);
1129 for (AliasSetTracker::iterator W = WriteSet.begin(),
1130 WE = WriteSet.end(); W != WE; ++W) {
1131 if (W->aliasesUnknownInst(J, *AA)) {
1139 if (UsesI && UpdateUsers) {
1140 if (J->mayWriteToMemory()) WriteSet.add(J);
1147 // This function iterates over all instruction pairs in the provided
1148 // basic block and collects all candidate pairs for vectorization.
1149 bool BBVectorize::getCandidatePairs(BasicBlock &BB,
1150 BasicBlock::iterator &Start,
1151 std::multimap<Value *, Value *> &CandidatePairs,
1152 DenseSet<ValuePair> &FixedOrderPairs,
1153 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1154 std::vector<Value *> &PairableInsts, bool NonPow2Len) {
1155 BasicBlock::iterator E = BB.end();
1156 if (Start == E) return false;
1158 bool ShouldContinue = false, IAfterStart = false;
1159 for (BasicBlock::iterator I = Start++; I != E; ++I) {
1160 if (I == Start) IAfterStart = true;
1162 bool IsSimpleLoadStore;
1163 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue;
1165 // Look for an instruction with which to pair instruction *I...
1166 DenseSet<Value *> Users;
1167 AliasSetTracker WriteSet(*AA);
1168 bool JAfterStart = IAfterStart;
1169 BasicBlock::iterator J = llvm::next(I);
1170 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) {
1171 if (J == Start) JAfterStart = true;
1173 // Determine if J uses I, if so, exit the loop.
1174 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep);
1175 if (Config.FastDep) {
1176 // Note: For this heuristic to be effective, independent operations
1177 // must tend to be intermixed. This is likely to be true from some
1178 // kinds of grouped loop unrolling (but not the generic LLVM pass),
1179 // but otherwise may require some kind of reordering pass.
1181 // When using fast dependency analysis,
1182 // stop searching after first use:
1185 if (UsesI) continue;
1188 // J does not use I, and comes before the first use of I, so it can be
1189 // merged with I if the instructions are compatible.
1190 int CostSavings, FixedOrder;
1191 if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len,
1192 CostSavings, FixedOrder)) continue;
1194 // J is a candidate for merging with I.
1195 if (!PairableInsts.size() ||
1196 PairableInsts[PairableInsts.size()-1] != I) {
1197 PairableInsts.push_back(I);
1200 CandidatePairs.insert(ValuePair(I, J));
1202 CandidatePairCostSavings.insert(ValuePairWithCost(ValuePair(I, J),
1205 if (FixedOrder == 1)
1206 FixedOrderPairs.insert(ValuePair(I, J));
1207 else if (FixedOrder == -1)
1208 FixedOrderPairs.insert(ValuePair(J, I));
1210 // The next call to this function must start after the last instruction
1211 // selected during this invocation.
1213 Start = llvm::next(J);
1214 IAfterStart = JAfterStart = false;
1217 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair "
1218 << *I << " <-> " << *J << " (cost savings: " <<
1219 CostSavings << ")\n");
1221 // If we have already found too many pairs, break here and this function
1222 // will be called again starting after the last instruction selected
1223 // during this invocation.
1224 if (PairableInsts.size() >= Config.MaxInsts) {
1225 ShouldContinue = true;
1234 DEBUG(dbgs() << "BBV: found " << PairableInsts.size()
1235 << " instructions with candidate pairs\n");
1237 return ShouldContinue;
1240 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that
1241 // it looks for pairs such that both members have an input which is an
1242 // output of PI or PJ.
1243 void BBVectorize::computePairsConnectedTo(
1244 std::multimap<Value *, Value *> &CandidatePairs,
1245 std::vector<Value *> &PairableInsts,
1246 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1247 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1251 // For each possible pairing for this variable, look at the uses of
1252 // the first value...
1253 for (Value::use_iterator I = P.first->use_begin(),
1254 E = P.first->use_end(); I != E; ++I) {
1255 if (isa<LoadInst>(*I)) {
1256 // A pair cannot be connected to a load because the load only takes one
1257 // operand (the address) and it is a scalar even after vectorization.
1259 } else if ((SI = dyn_cast<StoreInst>(*I)) &&
1260 P.first == SI->getPointerOperand()) {
1261 // Similarly, a pair cannot be connected to a store through its
1266 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I);
1268 // For each use of the first variable, look for uses of the second
1270 for (Value::use_iterator J = P.second->use_begin(),
1271 E2 = P.second->use_end(); J != E2; ++J) {
1272 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1273 P.second == SJ->getPointerOperand())
1276 VPIteratorPair JPairRange = CandidatePairs.equal_range(*J);
1279 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) {
1280 VPPair VP(P, ValuePair(*I, *J));
1281 ConnectedPairs.insert(VP);
1282 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionDirect));
1286 if (isSecondInIteratorPair<Value*>(*I, JPairRange)) {
1287 VPPair VP(P, ValuePair(*J, *I));
1288 ConnectedPairs.insert(VP);
1289 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSwap));
1293 if (Config.SplatBreaksChain) continue;
1294 // Look for cases where just the first value in the pair is used by
1295 // both members of another pair (splatting).
1296 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) {
1297 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1298 P.first == SJ->getPointerOperand())
1301 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) {
1302 VPPair VP(P, ValuePair(*I, *J));
1303 ConnectedPairs.insert(VP);
1304 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1309 if (Config.SplatBreaksChain) return;
1310 // Look for cases where just the second value in the pair is used by
1311 // both members of another pair (splatting).
1312 for (Value::use_iterator I = P.second->use_begin(),
1313 E = P.second->use_end(); I != E; ++I) {
1314 if (isa<LoadInst>(*I))
1316 else if ((SI = dyn_cast<StoreInst>(*I)) &&
1317 P.second == SI->getPointerOperand())
1320 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I);
1322 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) {
1323 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1324 P.second == SJ->getPointerOperand())
1327 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) {
1328 VPPair VP(P, ValuePair(*I, *J));
1329 ConnectedPairs.insert(VP);
1330 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1336 // This function figures out which pairs are connected. Two pairs are
1337 // connected if some output of the first pair forms an input to both members
1338 // of the second pair.
1339 void BBVectorize::computeConnectedPairs(
1340 std::multimap<Value *, Value *> &CandidatePairs,
1341 std::vector<Value *> &PairableInsts,
1342 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1343 DenseMap<VPPair, unsigned> &PairConnectionTypes) {
1345 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
1346 PE = PairableInsts.end(); PI != PE; ++PI) {
1347 VPIteratorPair choiceRange = CandidatePairs.equal_range(*PI);
1349 for (std::multimap<Value *, Value *>::iterator P = choiceRange.first;
1350 P != choiceRange.second; ++P)
1351 computePairsConnectedTo(CandidatePairs, PairableInsts,
1352 ConnectedPairs, PairConnectionTypes, *P);
1355 DEBUG(dbgs() << "BBV: found " << ConnectedPairs.size()
1356 << " pair connections.\n");
1359 // This function builds a set of use tuples such that <A, B> is in the set
1360 // if B is in the use tree of A. If B is in the use tree of A, then B
1361 // depends on the output of A.
1362 void BBVectorize::buildDepMap(
1364 std::multimap<Value *, Value *> &CandidatePairs,
1365 std::vector<Value *> &PairableInsts,
1366 DenseSet<ValuePair> &PairableInstUsers) {
1367 DenseSet<Value *> IsInPair;
1368 for (std::multimap<Value *, Value *>::iterator C = CandidatePairs.begin(),
1369 E = CandidatePairs.end(); C != E; ++C) {
1370 IsInPair.insert(C->first);
1371 IsInPair.insert(C->second);
1374 // Iterate through the basic block, recording all Users of each
1375 // pairable instruction.
1377 BasicBlock::iterator E = BB.end();
1378 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) {
1379 if (IsInPair.find(I) == IsInPair.end()) continue;
1381 DenseSet<Value *> Users;
1382 AliasSetTracker WriteSet(*AA);
1383 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J)
1384 (void) trackUsesOfI(Users, WriteSet, I, J);
1386 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end();
1388 PairableInstUsers.insert(ValuePair(I, *U));
1392 // Returns true if an input to pair P is an output of pair Q and also an
1393 // input of pair Q is an output of pair P. If this is the case, then these
1394 // two pairs cannot be simultaneously fused.
1395 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q,
1396 DenseSet<ValuePair> &PairableInstUsers,
1397 std::multimap<ValuePair, ValuePair> *PairableInstUserMap) {
1398 // Two pairs are in conflict if they are mutual Users of eachother.
1399 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) ||
1400 PairableInstUsers.count(ValuePair(P.first, Q.second)) ||
1401 PairableInstUsers.count(ValuePair(P.second, Q.first)) ||
1402 PairableInstUsers.count(ValuePair(P.second, Q.second));
1403 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) ||
1404 PairableInstUsers.count(ValuePair(Q.first, P.second)) ||
1405 PairableInstUsers.count(ValuePair(Q.second, P.first)) ||
1406 PairableInstUsers.count(ValuePair(Q.second, P.second));
1407 if (PairableInstUserMap) {
1408 // FIXME: The expensive part of the cycle check is not so much the cycle
1409 // check itself but this edge insertion procedure. This needs some
1410 // profiling and probably a different data structure (same is true of
1411 // most uses of std::multimap).
1413 VPPIteratorPair QPairRange = PairableInstUserMap->equal_range(Q);
1414 if (!isSecondInIteratorPair(P, QPairRange))
1415 PairableInstUserMap->insert(VPPair(Q, P));
1418 VPPIteratorPair PPairRange = PairableInstUserMap->equal_range(P);
1419 if (!isSecondInIteratorPair(Q, PPairRange))
1420 PairableInstUserMap->insert(VPPair(P, Q));
1424 return (QUsesP && PUsesQ);
1427 // This function walks the use graph of current pairs to see if, starting
1428 // from P, the walk returns to P.
1429 bool BBVectorize::pairWillFormCycle(ValuePair P,
1430 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1431 DenseSet<ValuePair> &CurrentPairs) {
1432 DEBUG(if (DebugCycleCheck)
1433 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> "
1434 << *P.second << "\n");
1435 // A lookup table of visisted pairs is kept because the PairableInstUserMap
1436 // contains non-direct associations.
1437 DenseSet<ValuePair> Visited;
1438 SmallVector<ValuePair, 32> Q;
1439 // General depth-first post-order traversal:
1442 ValuePair QTop = Q.pop_back_val();
1443 Visited.insert(QTop);
1445 DEBUG(if (DebugCycleCheck)
1446 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> "
1447 << *QTop.second << "\n");
1448 VPPIteratorPair QPairRange = PairableInstUserMap.equal_range(QTop);
1449 for (std::multimap<ValuePair, ValuePair>::iterator C = QPairRange.first;
1450 C != QPairRange.second; ++C) {
1451 if (C->second == P) {
1453 << "BBV: rejected to prevent non-trivial cycle formation: "
1454 << *C->first.first << " <-> " << *C->first.second << "\n");
1458 if (CurrentPairs.count(C->second) && !Visited.count(C->second))
1459 Q.push_back(C->second);
1461 } while (!Q.empty());
1466 // This function builds the initial tree of connected pairs with the
1467 // pair J at the root.
1468 void BBVectorize::buildInitialTreeFor(
1469 std::multimap<Value *, Value *> &CandidatePairs,
1470 std::vector<Value *> &PairableInsts,
1471 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1472 DenseSet<ValuePair> &PairableInstUsers,
1473 DenseMap<Value *, Value *> &ChosenPairs,
1474 DenseMap<ValuePair, size_t> &Tree, ValuePair J) {
1475 // Each of these pairs is viewed as the root node of a Tree. The Tree
1476 // is then walked (depth-first). As this happens, we keep track of
1477 // the pairs that compose the Tree and the maximum depth of the Tree.
1478 SmallVector<ValuePairWithDepth, 32> Q;
1479 // General depth-first post-order traversal:
1480 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1482 ValuePairWithDepth QTop = Q.back();
1484 // Push each child onto the queue:
1485 bool MoreChildren = false;
1486 size_t MaxChildDepth = QTop.second;
1487 VPPIteratorPair qtRange = ConnectedPairs.equal_range(QTop.first);
1488 for (std::multimap<ValuePair, ValuePair>::iterator k = qtRange.first;
1489 k != qtRange.second; ++k) {
1490 // Make sure that this child pair is still a candidate:
1491 bool IsStillCand = false;
1492 VPIteratorPair checkRange =
1493 CandidatePairs.equal_range(k->second.first);
1494 for (std::multimap<Value *, Value *>::iterator m = checkRange.first;
1495 m != checkRange.second; ++m) {
1496 if (m->second == k->second.second) {
1503 DenseMap<ValuePair, size_t>::iterator C = Tree.find(k->second);
1504 if (C == Tree.end()) {
1505 size_t d = getDepthFactor(k->second.first);
1506 Q.push_back(ValuePairWithDepth(k->second, QTop.second+d));
1507 MoreChildren = true;
1509 MaxChildDepth = std::max(MaxChildDepth, C->second);
1514 if (!MoreChildren) {
1515 // Record the current pair as part of the Tree:
1516 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth));
1519 } while (!Q.empty());
1522 // Given some initial tree, prune it by removing conflicting pairs (pairs
1523 // that cannot be simultaneously chosen for vectorization).
1524 void BBVectorize::pruneTreeFor(
1525 std::multimap<Value *, Value *> &CandidatePairs,
1526 std::vector<Value *> &PairableInsts,
1527 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1528 DenseSet<ValuePair> &PairableInstUsers,
1529 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1530 DenseMap<Value *, Value *> &ChosenPairs,
1531 DenseMap<ValuePair, size_t> &Tree,
1532 DenseSet<ValuePair> &PrunedTree, ValuePair J,
1533 bool UseCycleCheck) {
1534 SmallVector<ValuePairWithDepth, 32> Q;
1535 // General depth-first post-order traversal:
1536 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1538 ValuePairWithDepth QTop = Q.pop_back_val();
1539 PrunedTree.insert(QTop.first);
1541 // Visit each child, pruning as necessary...
1542 SmallVector<ValuePairWithDepth, 8> BestChildren;
1543 VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first);
1544 for (std::multimap<ValuePair, ValuePair>::iterator K = QTopRange.first;
1545 K != QTopRange.second; ++K) {
1546 DenseMap<ValuePair, size_t>::iterator C = Tree.find(K->second);
1547 if (C == Tree.end()) continue;
1549 // This child is in the Tree, now we need to make sure it is the
1550 // best of any conflicting children. There could be multiple
1551 // conflicting children, so first, determine if we're keeping
1552 // this child, then delete conflicting children as necessary.
1554 // It is also necessary to guard against pairing-induced
1555 // dependencies. Consider instructions a .. x .. y .. b
1556 // such that (a,b) are to be fused and (x,y) are to be fused
1557 // but a is an input to x and b is an output from y. This
1558 // means that y cannot be moved after b but x must be moved
1559 // after b for (a,b) to be fused. In other words, after
1560 // fusing (a,b) we have y .. a/b .. x where y is an input
1561 // to a/b and x is an output to a/b: x and y can no longer
1562 // be legally fused. To prevent this condition, we must
1563 // make sure that a child pair added to the Tree is not
1564 // both an input and output of an already-selected pair.
1566 // Pairing-induced dependencies can also form from more complicated
1567 // cycles. The pair vs. pair conflicts are easy to check, and so
1568 // that is done explicitly for "fast rejection", and because for
1569 // child vs. child conflicts, we may prefer to keep the current
1570 // pair in preference to the already-selected child.
1571 DenseSet<ValuePair> CurrentPairs;
1574 for (SmallVector<ValuePairWithDepth, 8>::iterator C2
1575 = BestChildren.begin(), E2 = BestChildren.end();
1577 if (C2->first.first == C->first.first ||
1578 C2->first.first == C->first.second ||
1579 C2->first.second == C->first.first ||
1580 C2->first.second == C->first.second ||
1581 pairsConflict(C2->first, C->first, PairableInstUsers,
1582 UseCycleCheck ? &PairableInstUserMap : 0)) {
1583 if (C2->second >= C->second) {
1588 CurrentPairs.insert(C2->first);
1591 if (!CanAdd) continue;
1593 // Even worse, this child could conflict with another node already
1594 // selected for the Tree. If that is the case, ignore this child.
1595 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(),
1596 E2 = PrunedTree.end(); T != E2; ++T) {
1597 if (T->first == C->first.first ||
1598 T->first == C->first.second ||
1599 T->second == C->first.first ||
1600 T->second == C->first.second ||
1601 pairsConflict(*T, C->first, PairableInstUsers,
1602 UseCycleCheck ? &PairableInstUserMap : 0)) {
1607 CurrentPairs.insert(*T);
1609 if (!CanAdd) continue;
1611 // And check the queue too...
1612 for (SmallVector<ValuePairWithDepth, 32>::iterator C2 = Q.begin(),
1613 E2 = Q.end(); C2 != E2; ++C2) {
1614 if (C2->first.first == C->first.first ||
1615 C2->first.first == C->first.second ||
1616 C2->first.second == C->first.first ||
1617 C2->first.second == C->first.second ||
1618 pairsConflict(C2->first, C->first, PairableInstUsers,
1619 UseCycleCheck ? &PairableInstUserMap : 0)) {
1624 CurrentPairs.insert(C2->first);
1626 if (!CanAdd) continue;
1628 // Last but not least, check for a conflict with any of the
1629 // already-chosen pairs.
1630 for (DenseMap<Value *, Value *>::iterator C2 =
1631 ChosenPairs.begin(), E2 = ChosenPairs.end();
1633 if (pairsConflict(*C2, C->first, PairableInstUsers,
1634 UseCycleCheck ? &PairableInstUserMap : 0)) {
1639 CurrentPairs.insert(*C2);
1641 if (!CanAdd) continue;
1643 // To check for non-trivial cycles formed by the addition of the
1644 // current pair we've formed a list of all relevant pairs, now use a
1645 // graph walk to check for a cycle. We start from the current pair and
1646 // walk the use tree to see if we again reach the current pair. If we
1647 // do, then the current pair is rejected.
1649 // FIXME: It may be more efficient to use a topological-ordering
1650 // algorithm to improve the cycle check. This should be investigated.
1651 if (UseCycleCheck &&
1652 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs))
1655 // This child can be added, but we may have chosen it in preference
1656 // to an already-selected child. Check for this here, and if a
1657 // conflict is found, then remove the previously-selected child
1658 // before adding this one in its place.
1659 for (SmallVector<ValuePairWithDepth, 8>::iterator C2
1660 = BestChildren.begin(); C2 != BestChildren.end();) {
1661 if (C2->first.first == C->first.first ||
1662 C2->first.first == C->first.second ||
1663 C2->first.second == C->first.first ||
1664 C2->first.second == C->first.second ||
1665 pairsConflict(C2->first, C->first, PairableInstUsers))
1666 C2 = BestChildren.erase(C2);
1671 BestChildren.push_back(ValuePairWithDepth(C->first, C->second));
1674 for (SmallVector<ValuePairWithDepth, 8>::iterator C
1675 = BestChildren.begin(), E2 = BestChildren.end();
1677 size_t DepthF = getDepthFactor(C->first.first);
1678 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF));
1680 } while (!Q.empty());
1683 // This function finds the best tree of mututally-compatible connected
1684 // pairs, given the choice of root pairs as an iterator range.
1685 void BBVectorize::findBestTreeFor(
1686 std::multimap<Value *, Value *> &CandidatePairs,
1687 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1688 std::vector<Value *> &PairableInsts,
1689 DenseSet<ValuePair> &FixedOrderPairs,
1690 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1691 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1692 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps,
1693 DenseSet<ValuePair> &PairableInstUsers,
1694 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1695 DenseMap<Value *, Value *> &ChosenPairs,
1696 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth,
1697 int &BestEffSize, VPIteratorPair ChoiceRange,
1698 bool UseCycleCheck) {
1699 for (std::multimap<Value *, Value *>::iterator J = ChoiceRange.first;
1700 J != ChoiceRange.second; ++J) {
1702 // Before going any further, make sure that this pair does not
1703 // conflict with any already-selected pairs (see comment below
1704 // near the Tree pruning for more details).
1705 DenseSet<ValuePair> ChosenPairSet;
1706 bool DoesConflict = false;
1707 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(),
1708 E = ChosenPairs.end(); C != E; ++C) {
1709 if (pairsConflict(*C, *J, PairableInstUsers,
1710 UseCycleCheck ? &PairableInstUserMap : 0)) {
1711 DoesConflict = true;
1715 ChosenPairSet.insert(*C);
1717 if (DoesConflict) continue;
1719 if (UseCycleCheck &&
1720 pairWillFormCycle(*J, PairableInstUserMap, ChosenPairSet))
1723 DenseMap<ValuePair, size_t> Tree;
1724 buildInitialTreeFor(CandidatePairs, PairableInsts, ConnectedPairs,
1725 PairableInstUsers, ChosenPairs, Tree, *J);
1727 // Because we'll keep the child with the largest depth, the largest
1728 // depth is still the same in the unpruned Tree.
1729 size_t MaxDepth = Tree.lookup(*J);
1731 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found Tree for pair {"
1732 << *J->first << " <-> " << *J->second << "} of depth " <<
1733 MaxDepth << " and size " << Tree.size() << "\n");
1735 // At this point the Tree has been constructed, but, may contain
1736 // contradictory children (meaning that different children of
1737 // some tree node may be attempting to fuse the same instruction).
1738 // So now we walk the tree again, in the case of a conflict,
1739 // keep only the child with the largest depth. To break a tie,
1740 // favor the first child.
1742 DenseSet<ValuePair> PrunedTree;
1743 pruneTreeFor(CandidatePairs, PairableInsts, ConnectedPairs,
1744 PairableInstUsers, PairableInstUserMap, ChosenPairs, Tree,
1745 PrunedTree, *J, UseCycleCheck);
1749 DenseSet<Value *> PrunedTreeInstrs;
1750 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
1751 E = PrunedTree.end(); S != E; ++S) {
1752 PrunedTreeInstrs.insert(S->first);
1753 PrunedTreeInstrs.insert(S->second);
1756 // The set of pairs that have already contributed to the total cost.
1757 DenseSet<ValuePair> IncomingPairs;
1759 // If the cost model were perfect, this might not be necessary; but we
1760 // need to make sure that we don't get stuck vectorizing our own
1762 bool HasNontrivialInsts = false;
1764 // The node weights represent the cost savings associated with
1765 // fusing the pair of instructions.
1766 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
1767 E = PrunedTree.end(); S != E; ++S) {
1768 if (!isa<ShuffleVectorInst>(S->first) &&
1769 !isa<InsertElementInst>(S->first) &&
1770 !isa<ExtractElementInst>(S->first))
1771 HasNontrivialInsts = true;
1773 bool FlipOrder = false;
1775 if (getDepthFactor(S->first)) {
1776 int ESContrib = CandidatePairCostSavings.find(*S)->second;
1777 DEBUG(if (DebugPairSelection) dbgs() << "\tweight {"
1778 << *S->first << " <-> " << *S->second << "} = " <<
1780 EffSize += ESContrib;
1783 // The edge weights contribute in a negative sense: they represent
1784 // the cost of shuffles.
1785 VPPIteratorPair IP = ConnectedPairDeps.equal_range(*S);
1786 if (IP.first != ConnectedPairDeps.end()) {
1787 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
1788 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first;
1789 Q != IP.second; ++Q) {
1790 if (!PrunedTree.count(Q->second))
1792 DenseMap<VPPair, unsigned>::iterator R =
1793 PairConnectionTypes.find(VPPair(Q->second, Q->first));
1794 assert(R != PairConnectionTypes.end() &&
1795 "Cannot find pair connection type");
1796 if (R->second == PairConnectionDirect)
1798 else if (R->second == PairConnectionSwap)
1802 // If there are more swaps than direct connections, then
1803 // the pair order will be flipped during fusion. So the real
1804 // number of swaps is the minimum number.
1805 FlipOrder = !FixedOrderPairs.count(*S) &&
1806 ((NumDepsSwap > NumDepsDirect) ||
1807 FixedOrderPairs.count(ValuePair(S->second, S->first)));
1809 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first;
1810 Q != IP.second; ++Q) {
1811 if (!PrunedTree.count(Q->second))
1813 DenseMap<VPPair, unsigned>::iterator R =
1814 PairConnectionTypes.find(VPPair(Q->second, Q->first));
1815 assert(R != PairConnectionTypes.end() &&
1816 "Cannot find pair connection type");
1817 Type *Ty1 = Q->second.first->getType(),
1818 *Ty2 = Q->second.second->getType();
1819 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1820 if ((R->second == PairConnectionDirect && FlipOrder) ||
1821 (R->second == PairConnectionSwap && !FlipOrder) ||
1822 R->second == PairConnectionSplat) {
1823 int ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1825 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1826 *Q->second.first << " <-> " << *Q->second.second <<
1828 *S->first << " <-> " << *S->second << "} = " <<
1830 EffSize -= ESContrib;
1835 // Compute the cost of outgoing edges. We assume that edges outgoing
1836 // to shuffles, inserts or extracts can be merged, and so contribute
1837 // no additional cost.
1838 if (!S->first->getType()->isVoidTy()) {
1839 Type *Ty1 = S->first->getType(),
1840 *Ty2 = S->second->getType();
1841 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1843 bool NeedsExtraction = false;
1844 for (Value::use_iterator I = S->first->use_begin(),
1845 IE = S->first->use_end(); I != IE; ++I) {
1846 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) {
1847 // Shuffle can be folded if it has no other input
1848 if (isa<UndefValue>(SI->getOperand(1)))
1851 if (isa<ExtractElementInst>(*I))
1853 if (PrunedTreeInstrs.count(*I))
1855 NeedsExtraction = true;
1859 if (NeedsExtraction) {
1861 if (Ty1->isVectorTy())
1862 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1865 ESContrib = (int) VTTI->getVectorInstrCost(
1866 Instruction::ExtractElement, VTy, 0);
1868 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1869 *S->first << "} = " << ESContrib << "\n");
1870 EffSize -= ESContrib;
1873 NeedsExtraction = false;
1874 for (Value::use_iterator I = S->second->use_begin(),
1875 IE = S->second->use_end(); I != IE; ++I) {
1876 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) {
1877 // Shuffle can be folded if it has no other input
1878 if (isa<UndefValue>(SI->getOperand(1)))
1881 if (isa<ExtractElementInst>(*I))
1883 if (PrunedTreeInstrs.count(*I))
1885 NeedsExtraction = true;
1889 if (NeedsExtraction) {
1891 if (Ty2->isVectorTy())
1892 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1895 ESContrib = (int) VTTI->getVectorInstrCost(
1896 Instruction::ExtractElement, VTy, 1);
1897 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1898 *S->second << "} = " << ESContrib << "\n");
1899 EffSize -= ESContrib;
1903 // Compute the cost of incoming edges.
1904 if (!isa<LoadInst>(S->first) && !isa<StoreInst>(S->first)) {
1905 Instruction *S1 = cast<Instruction>(S->first),
1906 *S2 = cast<Instruction>(S->second);
1907 for (unsigned o = 0; o < S1->getNumOperands(); ++o) {
1908 Value *O1 = S1->getOperand(o), *O2 = S2->getOperand(o);
1910 // Combining constants into vector constants (or small vector
1911 // constants into larger ones are assumed free).
1912 if (isa<Constant>(O1) && isa<Constant>(O2))
1918 ValuePair VP = ValuePair(O1, O2);
1919 ValuePair VPR = ValuePair(O2, O1);
1921 // Internal edges are not handled here.
1922 if (PrunedTree.count(VP) || PrunedTree.count(VPR))
1925 Type *Ty1 = O1->getType(),
1926 *Ty2 = O2->getType();
1927 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1929 // Combining vector operations of the same type is also assumed
1930 // folded with other operations.
1932 // If both are insert elements, then both can be widened.
1933 InsertElementInst *IEO1 = dyn_cast<InsertElementInst>(O1),
1934 *IEO2 = dyn_cast<InsertElementInst>(O2);
1935 if (IEO1 && IEO2 && isPureIEChain(IEO1) && isPureIEChain(IEO2))
1937 // If both are extract elements, and both have the same input
1938 // type, then they can be replaced with a shuffle
1939 ExtractElementInst *EIO1 = dyn_cast<ExtractElementInst>(O1),
1940 *EIO2 = dyn_cast<ExtractElementInst>(O2);
1942 EIO1->getOperand(0)->getType() ==
1943 EIO2->getOperand(0)->getType())
1945 // If both are a shuffle with equal operand types and only two
1946 // unqiue operands, then they can be replaced with a single
1948 ShuffleVectorInst *SIO1 = dyn_cast<ShuffleVectorInst>(O1),
1949 *SIO2 = dyn_cast<ShuffleVectorInst>(O2);
1951 SIO1->getOperand(0)->getType() ==
1952 SIO2->getOperand(0)->getType()) {
1953 SmallSet<Value *, 4> SIOps;
1954 SIOps.insert(SIO1->getOperand(0));
1955 SIOps.insert(SIO1->getOperand(1));
1956 SIOps.insert(SIO2->getOperand(0));
1957 SIOps.insert(SIO2->getOperand(1));
1958 if (SIOps.size() <= 2)
1964 // This pair has already been formed.
1965 if (IncomingPairs.count(VP)) {
1967 } else if (IncomingPairs.count(VPR)) {
1968 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1970 } else if (!Ty1->isVectorTy() && !Ty2->isVectorTy()) {
1971 ESContrib = (int) VTTI->getVectorInstrCost(
1972 Instruction::InsertElement, VTy, 0);
1973 ESContrib += (int) VTTI->getVectorInstrCost(
1974 Instruction::InsertElement, VTy, 1);
1975 } else if (!Ty1->isVectorTy()) {
1976 // O1 needs to be inserted into a vector of size O2, and then
1977 // both need to be shuffled together.
1978 ESContrib = (int) VTTI->getVectorInstrCost(
1979 Instruction::InsertElement, Ty2, 0);
1980 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
1982 } else if (!Ty2->isVectorTy()) {
1983 // O2 needs to be inserted into a vector of size O1, and then
1984 // both need to be shuffled together.
1985 ESContrib = (int) VTTI->getVectorInstrCost(
1986 Instruction::InsertElement, Ty1, 0);
1987 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
1990 Type *TyBig = Ty1, *TySmall = Ty2;
1991 if (Ty2->getVectorNumElements() > Ty1->getVectorNumElements())
1992 std::swap(TyBig, TySmall);
1994 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1996 if (TyBig != TySmall)
1997 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2001 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {"
2002 << *O1 << " <-> " << *O2 << "} = " <<
2004 EffSize -= ESContrib;
2005 IncomingPairs.insert(VP);
2010 if (!HasNontrivialInsts) {
2011 DEBUG(if (DebugPairSelection) dbgs() <<
2012 "\tNo non-trivial instructions in tree;"
2013 " override to zero effective size\n");
2017 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
2018 E = PrunedTree.end(); S != E; ++S)
2019 EffSize += (int) getDepthFactor(S->first);
2022 DEBUG(if (DebugPairSelection)
2023 dbgs() << "BBV: found pruned Tree for pair {"
2024 << *J->first << " <-> " << *J->second << "} of depth " <<
2025 MaxDepth << " and size " << PrunedTree.size() <<
2026 " (effective size: " << EffSize << ")\n");
2027 if (((VTTI && !UseChainDepthWithTI) ||
2028 MaxDepth >= Config.ReqChainDepth) &&
2029 EffSize > 0 && EffSize > BestEffSize) {
2030 BestMaxDepth = MaxDepth;
2031 BestEffSize = EffSize;
2032 BestTree = PrunedTree;
2037 // Given the list of candidate pairs, this function selects those
2038 // that will be fused into vector instructions.
2039 void BBVectorize::choosePairs(
2040 std::multimap<Value *, Value *> &CandidatePairs,
2041 DenseMap<ValuePair, int> &CandidatePairCostSavings,
2042 std::vector<Value *> &PairableInsts,
2043 DenseSet<ValuePair> &FixedOrderPairs,
2044 DenseMap<VPPair, unsigned> &PairConnectionTypes,
2045 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
2046 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps,
2047 DenseSet<ValuePair> &PairableInstUsers,
2048 DenseMap<Value *, Value *>& ChosenPairs) {
2049 bool UseCycleCheck =
2050 CandidatePairs.size() <= Config.MaxCandPairsForCycleCheck;
2051 std::multimap<ValuePair, ValuePair> PairableInstUserMap;
2052 for (std::vector<Value *>::iterator I = PairableInsts.begin(),
2053 E = PairableInsts.end(); I != E; ++I) {
2054 // The number of possible pairings for this variable:
2055 size_t NumChoices = CandidatePairs.count(*I);
2056 if (!NumChoices) continue;
2058 VPIteratorPair ChoiceRange = CandidatePairs.equal_range(*I);
2060 // The best pair to choose and its tree:
2061 size_t BestMaxDepth = 0;
2062 int BestEffSize = 0;
2063 DenseSet<ValuePair> BestTree;
2064 findBestTreeFor(CandidatePairs, CandidatePairCostSavings,
2065 PairableInsts, FixedOrderPairs, PairConnectionTypes,
2066 ConnectedPairs, ConnectedPairDeps,
2067 PairableInstUsers, PairableInstUserMap, ChosenPairs,
2068 BestTree, BestMaxDepth, BestEffSize, ChoiceRange,
2071 // A tree has been chosen (or not) at this point. If no tree was
2072 // chosen, then this instruction, I, cannot be paired (and is no longer
2075 DEBUG(if (BestTree.size() > 0)
2076 dbgs() << "BBV: selected pairs in the best tree for: "
2077 << *cast<Instruction>(*I) << "\n");
2079 for (DenseSet<ValuePair>::iterator S = BestTree.begin(),
2080 SE2 = BestTree.end(); S != SE2; ++S) {
2081 // Insert the members of this tree into the list of chosen pairs.
2082 ChosenPairs.insert(ValuePair(S->first, S->second));
2083 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " <<
2084 *S->second << "\n");
2086 // Remove all candidate pairs that have values in the chosen tree.
2087 for (std::multimap<Value *, Value *>::iterator K =
2088 CandidatePairs.begin(); K != CandidatePairs.end();) {
2089 if (K->first == S->first || K->second == S->first ||
2090 K->second == S->second || K->first == S->second) {
2091 // Don't remove the actual pair chosen so that it can be used
2092 // in subsequent tree selections.
2093 if (!(K->first == S->first && K->second == S->second))
2094 CandidatePairs.erase(K++);
2104 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n");
2107 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o,
2112 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) +
2113 (n > 0 ? "." + utostr(n) : "")).str();
2116 // Returns the value that is to be used as the pointer input to the vector
2117 // instruction that fuses I with J.
2118 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context,
2119 Instruction *I, Instruction *J, unsigned o) {
2121 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
2122 int64_t OffsetInElmts;
2124 // Note: the analysis might fail here, that is why the pair order has
2125 // been precomputed (OffsetInElmts must be unused here).
2126 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
2127 IAddressSpace, JAddressSpace,
2128 OffsetInElmts, false);
2130 // The pointer value is taken to be the one with the lowest offset.
2133 Type *ArgTypeI = cast<PointerType>(IPtr->getType())->getElementType();
2134 Type *ArgTypeJ = cast<PointerType>(JPtr->getType())->getElementType();
2135 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2136 Type *VArgPtrType = PointerType::get(VArgType,
2137 cast<PointerType>(IPtr->getType())->getAddressSpace());
2138 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o),
2139 /* insert before */ I);
2142 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J,
2143 unsigned MaskOffset, unsigned NumInElem,
2144 unsigned NumInElem1, unsigned IdxOffset,
2145 std::vector<Constant*> &Mask) {
2146 unsigned NumElem1 = cast<VectorType>(J->getType())->getNumElements();
2147 for (unsigned v = 0; v < NumElem1; ++v) {
2148 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v);
2150 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context));
2152 unsigned mm = m + (int) IdxOffset;
2153 if (m >= (int) NumInElem1)
2154 mm += (int) NumInElem;
2156 Mask[v+MaskOffset] =
2157 ConstantInt::get(Type::getInt32Ty(Context), mm);
2162 // Returns the value that is to be used as the vector-shuffle mask to the
2163 // vector instruction that fuses I with J.
2164 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context,
2165 Instruction *I, Instruction *J) {
2166 // This is the shuffle mask. We need to append the second
2167 // mask to the first, and the numbers need to be adjusted.
2169 Type *ArgTypeI = I->getType();
2170 Type *ArgTypeJ = J->getType();
2171 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2173 unsigned NumElemI = cast<VectorType>(ArgTypeI)->getNumElements();
2175 // Get the total number of elements in the fused vector type.
2176 // By definition, this must equal the number of elements in
2178 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements();
2179 std::vector<Constant*> Mask(NumElem);
2181 Type *OpTypeI = I->getOperand(0)->getType();
2182 unsigned NumInElemI = cast<VectorType>(OpTypeI)->getNumElements();
2183 Type *OpTypeJ = J->getOperand(0)->getType();
2184 unsigned NumInElemJ = cast<VectorType>(OpTypeJ)->getNumElements();
2186 // The fused vector will be:
2187 // -----------------------------------------------------
2188 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ |
2189 // -----------------------------------------------------
2190 // from which we'll extract NumElem total elements (where the first NumElemI
2191 // of them come from the mask in I and the remainder come from the mask
2194 // For the mask from the first pair...
2195 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI,
2198 // For the mask from the second pair...
2199 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ,
2202 return ConstantVector::get(Mask);
2205 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I,
2206 Instruction *J, unsigned o, Value *&LOp,
2208 Type *ArgTypeL, Type *ArgTypeH,
2209 bool IBeforeJ, unsigned IdxOff) {
2210 bool ExpandedIEChain = false;
2211 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) {
2212 // If we have a pure insertelement chain, then this can be rewritten
2213 // into a chain that directly builds the larger type.
2214 if (isPureIEChain(LIE)) {
2215 SmallVector<Value *, 8> VectElemts(numElemL,
2216 UndefValue::get(ArgTypeL->getScalarType()));
2217 InsertElementInst *LIENext = LIE;
2220 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue();
2221 VectElemts[Idx] = LIENext->getOperand(1);
2223 dyn_cast<InsertElementInst>(LIENext->getOperand(0))));
2226 Value *LIEPrev = UndefValue::get(ArgTypeH);
2227 for (unsigned i = 0; i < numElemL; ++i) {
2228 if (isa<UndefValue>(VectElemts[i])) continue;
2229 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i],
2230 ConstantInt::get(Type::getInt32Ty(Context),
2232 getReplacementName(IBeforeJ ? I : J,
2234 LIENext->insertBefore(IBeforeJ ? J : I);
2238 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH);
2239 ExpandedIEChain = true;
2243 return ExpandedIEChain;
2246 // Returns the value to be used as the specified operand of the vector
2247 // instruction that fuses I with J.
2248 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I,
2249 Instruction *J, unsigned o, bool IBeforeJ) {
2250 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2251 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1);
2253 // Compute the fused vector type for this operand
2254 Type *ArgTypeI = I->getOperand(o)->getType();
2255 Type *ArgTypeJ = J->getOperand(o)->getType();
2256 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2258 Instruction *L = I, *H = J;
2259 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ;
2262 if (ArgTypeL->isVectorTy())
2263 numElemL = cast<VectorType>(ArgTypeL)->getNumElements();
2268 if (ArgTypeH->isVectorTy())
2269 numElemH = cast<VectorType>(ArgTypeH)->getNumElements();
2273 Value *LOp = L->getOperand(o);
2274 Value *HOp = H->getOperand(o);
2275 unsigned numElem = VArgType->getNumElements();
2277 // First, we check if we can reuse the "original" vector outputs (if these
2278 // exist). We might need a shuffle.
2279 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp);
2280 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp);
2281 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp);
2282 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp);
2284 // FIXME: If we're fusing shuffle instructions, then we can't apply this
2285 // optimization. The input vectors to the shuffle might be a different
2286 // length from the shuffle outputs. Unfortunately, the replacement
2287 // shuffle mask has already been formed, and the mask entries are sensitive
2288 // to the sizes of the inputs.
2289 bool IsSizeChangeShuffle =
2290 isa<ShuffleVectorInst>(L) &&
2291 (LOp->getType() != L->getType() || HOp->getType() != H->getType());
2293 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) {
2294 // We can have at most two unique vector inputs.
2295 bool CanUseInputs = true;
2298 I1 = LEE->getOperand(0);
2300 I1 = LSV->getOperand(0);
2301 I2 = LSV->getOperand(1);
2302 if (I2 == I1 || isa<UndefValue>(I2))
2307 Value *I3 = HEE->getOperand(0);
2308 if (!I2 && I3 != I1)
2310 else if (I3 != I1 && I3 != I2)
2311 CanUseInputs = false;
2313 Value *I3 = HSV->getOperand(0);
2314 if (!I2 && I3 != I1)
2316 else if (I3 != I1 && I3 != I2)
2317 CanUseInputs = false;
2320 Value *I4 = HSV->getOperand(1);
2321 if (!isa<UndefValue>(I4)) {
2322 if (!I2 && I4 != I1)
2324 else if (I4 != I1 && I4 != I2)
2325 CanUseInputs = false;
2332 cast<VectorType>(cast<Instruction>(LOp)->getOperand(0)->getType())
2335 cast<VectorType>(cast<Instruction>(HOp)->getOperand(0)->getType())
2338 // We have one or two input vectors. We need to map each index of the
2339 // operands to the index of the original vector.
2340 SmallVector<std::pair<int, int>, 8> II(numElem);
2341 for (unsigned i = 0; i < numElemL; ++i) {
2345 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue();
2346 INum = LEE->getOperand(0) == I1 ? 0 : 1;
2348 Idx = LSV->getMaskValue(i);
2349 if (Idx < (int) LOpElem) {
2350 INum = LSV->getOperand(0) == I1 ? 0 : 1;
2353 INum = LSV->getOperand(1) == I1 ? 0 : 1;
2357 II[i] = std::pair<int, int>(Idx, INum);
2359 for (unsigned i = 0; i < numElemH; ++i) {
2363 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue();
2364 INum = HEE->getOperand(0) == I1 ? 0 : 1;
2366 Idx = HSV->getMaskValue(i);
2367 if (Idx < (int) HOpElem) {
2368 INum = HSV->getOperand(0) == I1 ? 0 : 1;
2371 INum = HSV->getOperand(1) == I1 ? 0 : 1;
2375 II[i + numElemL] = std::pair<int, int>(Idx, INum);
2378 // We now have an array which tells us from which index of which
2379 // input vector each element of the operand comes.
2380 VectorType *I1T = cast<VectorType>(I1->getType());
2381 unsigned I1Elem = I1T->getNumElements();
2384 // In this case there is only one underlying vector input. Check for
2385 // the trivial case where we can use the input directly.
2386 if (I1Elem == numElem) {
2387 bool ElemInOrder = true;
2388 for (unsigned i = 0; i < numElem; ++i) {
2389 if (II[i].first != (int) i && II[i].first != -1) {
2390 ElemInOrder = false;
2399 // A shuffle is needed.
2400 std::vector<Constant *> Mask(numElem);
2401 for (unsigned i = 0; i < numElem; ++i) {
2402 int Idx = II[i].first;
2404 Mask[i] = UndefValue::get(Type::getInt32Ty(Context));
2406 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2410 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2411 ConstantVector::get(Mask),
2412 getReplacementName(IBeforeJ ? I : J,
2414 S->insertBefore(IBeforeJ ? J : I);
2418 VectorType *I2T = cast<VectorType>(I2->getType());
2419 unsigned I2Elem = I2T->getNumElements();
2421 // This input comes from two distinct vectors. The first step is to
2422 // make sure that both vectors are the same length. If not, the
2423 // smaller one will need to grow before they can be shuffled together.
2424 if (I1Elem < I2Elem) {
2425 std::vector<Constant *> Mask(I2Elem);
2427 for (; v < I1Elem; ++v)
2428 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2429 for (; v < I2Elem; ++v)
2430 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2432 Instruction *NewI1 =
2433 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2434 ConstantVector::get(Mask),
2435 getReplacementName(IBeforeJ ? I : J,
2437 NewI1->insertBefore(IBeforeJ ? J : I);
2441 } else if (I1Elem > I2Elem) {
2442 std::vector<Constant *> Mask(I1Elem);
2444 for (; v < I2Elem; ++v)
2445 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2446 for (; v < I1Elem; ++v)
2447 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2449 Instruction *NewI2 =
2450 new ShuffleVectorInst(I2, UndefValue::get(I2T),
2451 ConstantVector::get(Mask),
2452 getReplacementName(IBeforeJ ? I : J,
2454 NewI2->insertBefore(IBeforeJ ? J : I);
2460 // Now that both I1 and I2 are the same length we can shuffle them
2461 // together (and use the result).
2462 std::vector<Constant *> Mask(numElem);
2463 for (unsigned v = 0; v < numElem; ++v) {
2464 if (II[v].first == -1) {
2465 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2467 int Idx = II[v].first + II[v].second * I1Elem;
2468 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2472 Instruction *NewOp =
2473 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask),
2474 getReplacementName(IBeforeJ ? I : J, true, o));
2475 NewOp->insertBefore(IBeforeJ ? J : I);
2480 Type *ArgType = ArgTypeL;
2481 if (numElemL < numElemH) {
2482 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH,
2483 ArgTypeL, VArgType, IBeforeJ, 1)) {
2484 // This is another short-circuit case: we're combining a scalar into
2485 // a vector that is formed by an IE chain. We've just expanded the IE
2486 // chain, now insert the scalar and we're done.
2488 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0,
2489 getReplacementName(IBeforeJ ? I : J, true, o));
2490 S->insertBefore(IBeforeJ ? J : I);
2492 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL,
2493 ArgTypeH, IBeforeJ)) {
2494 // The two vector inputs to the shuffle must be the same length,
2495 // so extend the smaller vector to be the same length as the larger one.
2499 std::vector<Constant *> Mask(numElemH);
2501 for (; v < numElemL; ++v)
2502 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2503 for (; v < numElemH; ++v)
2504 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2506 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL),
2507 ConstantVector::get(Mask),
2508 getReplacementName(IBeforeJ ? I : J,
2511 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0,
2512 getReplacementName(IBeforeJ ? I : J,
2516 NLOp->insertBefore(IBeforeJ ? J : I);
2521 } else if (numElemL > numElemH) {
2522 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL,
2523 ArgTypeH, VArgType, IBeforeJ)) {
2525 InsertElementInst::Create(LOp, HOp,
2526 ConstantInt::get(Type::getInt32Ty(Context),
2528 getReplacementName(IBeforeJ ? I : J,
2530 S->insertBefore(IBeforeJ ? J : I);
2532 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH,
2533 ArgTypeL, IBeforeJ)) {
2536 std::vector<Constant *> Mask(numElemL);
2538 for (; v < numElemH; ++v)
2539 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2540 for (; v < numElemL; ++v)
2541 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2543 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH),
2544 ConstantVector::get(Mask),
2545 getReplacementName(IBeforeJ ? I : J,
2548 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0,
2549 getReplacementName(IBeforeJ ? I : J,
2553 NHOp->insertBefore(IBeforeJ ? J : I);
2558 if (ArgType->isVectorTy()) {
2559 unsigned numElem = cast<VectorType>(VArgType)->getNumElements();
2560 std::vector<Constant*> Mask(numElem);
2561 for (unsigned v = 0; v < numElem; ++v) {
2563 // If the low vector was expanded, we need to skip the extra
2564 // undefined entries.
2565 if (v >= numElemL && numElemH > numElemL)
2566 Idx += (numElemH - numElemL);
2567 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2570 Instruction *BV = new ShuffleVectorInst(LOp, HOp,
2571 ConstantVector::get(Mask),
2572 getReplacementName(IBeforeJ ? I : J, true, o));
2573 BV->insertBefore(IBeforeJ ? J : I);
2577 Instruction *BV1 = InsertElementInst::Create(
2578 UndefValue::get(VArgType), LOp, CV0,
2579 getReplacementName(IBeforeJ ? I : J,
2581 BV1->insertBefore(IBeforeJ ? J : I);
2582 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1,
2583 getReplacementName(IBeforeJ ? I : J,
2585 BV2->insertBefore(IBeforeJ ? J : I);
2589 // This function creates an array of values that will be used as the inputs
2590 // to the vector instruction that fuses I with J.
2591 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context,
2592 Instruction *I, Instruction *J,
2593 SmallVector<Value *, 3> &ReplacedOperands,
2595 unsigned NumOperands = I->getNumOperands();
2597 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) {
2598 // Iterate backward so that we look at the store pointer
2599 // first and know whether or not we need to flip the inputs.
2601 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) {
2602 // This is the pointer for a load/store instruction.
2603 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o);
2605 } else if (isa<CallInst>(I)) {
2606 Function *F = cast<CallInst>(I)->getCalledFunction();
2607 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID();
2608 if (o == NumOperands-1) {
2609 BasicBlock &BB = *I->getParent();
2611 Module *M = BB.getParent()->getParent();
2612 Type *ArgTypeI = I->getType();
2613 Type *ArgTypeJ = J->getType();
2614 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2616 ReplacedOperands[o] = Intrinsic::getDeclaration(M, IID, VArgType);
2618 } else if (IID == Intrinsic::powi && o == 1) {
2619 // The second argument of powi is a single integer and we've already
2620 // checked that both arguments are equal. As a result, we just keep
2621 // I's second argument.
2622 ReplacedOperands[o] = I->getOperand(o);
2625 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) {
2626 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J);
2630 ReplacedOperands[o] = getReplacementInput(Context, I, J, o, IBeforeJ);
2634 // This function creates two values that represent the outputs of the
2635 // original I and J instructions. These are generally vector shuffles
2636 // or extracts. In many cases, these will end up being unused and, thus,
2637 // eliminated by later passes.
2638 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
2639 Instruction *J, Instruction *K,
2640 Instruction *&InsertionPt,
2641 Instruction *&K1, Instruction *&K2) {
2642 if (isa<StoreInst>(I)) {
2643 AA->replaceWithNewValue(I, K);
2644 AA->replaceWithNewValue(J, K);
2646 Type *IType = I->getType();
2647 Type *JType = J->getType();
2649 VectorType *VType = getVecTypeForPair(IType, JType);
2650 unsigned numElem = VType->getNumElements();
2652 unsigned numElemI, numElemJ;
2653 if (IType->isVectorTy())
2654 numElemI = cast<VectorType>(IType)->getNumElements();
2658 if (JType->isVectorTy())
2659 numElemJ = cast<VectorType>(JType)->getNumElements();
2663 if (IType->isVectorTy()) {
2664 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI);
2665 for (unsigned v = 0; v < numElemI; ++v) {
2666 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2667 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v);
2670 K1 = new ShuffleVectorInst(K, UndefValue::get(VType),
2671 ConstantVector::get( Mask1),
2672 getReplacementName(K, false, 1));
2674 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2675 K1 = ExtractElementInst::Create(K, CV0,
2676 getReplacementName(K, false, 1));
2679 if (JType->isVectorTy()) {
2680 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ);
2681 for (unsigned v = 0; v < numElemJ; ++v) {
2682 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2683 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v);
2686 K2 = new ShuffleVectorInst(K, UndefValue::get(VType),
2687 ConstantVector::get( Mask2),
2688 getReplacementName(K, false, 2));
2690 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1);
2691 K2 = ExtractElementInst::Create(K, CV1,
2692 getReplacementName(K, false, 2));
2696 K2->insertAfter(K1);
2701 // Move all uses of the function I (including pairing-induced uses) after J.
2702 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB,
2703 std::multimap<Value *, Value *> &LoadMoveSet,
2704 Instruction *I, Instruction *J) {
2705 // Skip to the first instruction past I.
2706 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2708 DenseSet<Value *> Users;
2709 AliasSetTracker WriteSet(*AA);
2710 for (; cast<Instruction>(L) != J; ++L)
2711 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet);
2713 assert(cast<Instruction>(L) == J &&
2714 "Tracking has not proceeded far enough to check for dependencies");
2715 // If J is now in the use set of I, then trackUsesOfI will return true
2716 // and we have a dependency cycle (and the fusing operation must abort).
2717 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSet);
2720 // Move all uses of the function I (including pairing-induced uses) after J.
2721 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB,
2722 std::multimap<Value *, Value *> &LoadMoveSet,
2723 Instruction *&InsertionPt,
2724 Instruction *I, Instruction *J) {
2725 // Skip to the first instruction past I.
2726 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2728 DenseSet<Value *> Users;
2729 AliasSetTracker WriteSet(*AA);
2730 for (; cast<Instruction>(L) != J;) {
2731 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet)) {
2732 // Move this instruction
2733 Instruction *InstToMove = L; ++L;
2735 DEBUG(dbgs() << "BBV: moving: " << *InstToMove <<
2736 " to after " << *InsertionPt << "\n");
2737 InstToMove->removeFromParent();
2738 InstToMove->insertAfter(InsertionPt);
2739 InsertionPt = InstToMove;
2746 // Collect all load instruction that are in the move set of a given first
2747 // pair member. These loads depend on the first instruction, I, and so need
2748 // to be moved after J (the second instruction) when the pair is fused.
2749 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB,
2750 DenseMap<Value *, Value *> &ChosenPairs,
2751 std::multimap<Value *, Value *> &LoadMoveSet,
2753 // Skip to the first instruction past I.
2754 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2756 DenseSet<Value *> Users;
2757 AliasSetTracker WriteSet(*AA);
2759 // Note: We cannot end the loop when we reach J because J could be moved
2760 // farther down the use chain by another instruction pairing. Also, J
2761 // could be before I if this is an inverted input.
2762 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) {
2763 if (trackUsesOfI(Users, WriteSet, I, L)) {
2764 if (L->mayReadFromMemory())
2765 LoadMoveSet.insert(ValuePair(L, I));
2770 // In cases where both load/stores and the computation of their pointers
2771 // are chosen for vectorization, we can end up in a situation where the
2772 // aliasing analysis starts returning different query results as the
2773 // process of fusing instruction pairs continues. Because the algorithm
2774 // relies on finding the same use trees here as were found earlier, we'll
2775 // need to precompute the necessary aliasing information here and then
2776 // manually update it during the fusion process.
2777 void BBVectorize::collectLoadMoveSet(BasicBlock &BB,
2778 std::vector<Value *> &PairableInsts,
2779 DenseMap<Value *, Value *> &ChosenPairs,
2780 std::multimap<Value *, Value *> &LoadMoveSet) {
2781 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
2782 PIE = PairableInsts.end(); PI != PIE; ++PI) {
2783 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI);
2784 if (P == ChosenPairs.end()) continue;
2786 Instruction *I = cast<Instruction>(P->first);
2787 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, I);
2791 // When the first instruction in each pair is cloned, it will inherit its
2792 // parent's metadata. This metadata must be combined with that of the other
2793 // instruction in a safe way.
2794 void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) {
2795 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata;
2796 K->getAllMetadataOtherThanDebugLoc(Metadata);
2797 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) {
2798 unsigned Kind = Metadata[i].first;
2799 MDNode *JMD = J->getMetadata(Kind);
2800 MDNode *KMD = Metadata[i].second;
2804 K->setMetadata(Kind, 0); // Remove unknown metadata
2806 case LLVMContext::MD_tbaa:
2807 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD));
2809 case LLVMContext::MD_fpmath:
2810 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD));
2816 // This function fuses the chosen instruction pairs into vector instructions,
2817 // taking care preserve any needed scalar outputs and, then, it reorders the
2818 // remaining instructions as needed (users of the first member of the pair
2819 // need to be moved to after the location of the second member of the pair
2820 // because the vector instruction is inserted in the location of the pair's
2822 void BBVectorize::fuseChosenPairs(BasicBlock &BB,
2823 std::vector<Value *> &PairableInsts,
2824 DenseMap<Value *, Value *> &ChosenPairs,
2825 DenseSet<ValuePair> &FixedOrderPairs,
2826 DenseMap<VPPair, unsigned> &PairConnectionTypes,
2827 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
2828 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps) {
2829 LLVMContext& Context = BB.getContext();
2831 // During the vectorization process, the order of the pairs to be fused
2832 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs
2833 // list. After a pair is fused, the flipped pair is removed from the list.
2834 DenseSet<ValuePair> FlippedPairs;
2835 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(),
2836 E = ChosenPairs.end(); P != E; ++P)
2837 FlippedPairs.insert(ValuePair(P->second, P->first));
2838 for (DenseSet<ValuePair>::iterator P = FlippedPairs.begin(),
2839 E = FlippedPairs.end(); P != E; ++P)
2840 ChosenPairs.insert(*P);
2842 std::multimap<Value *, Value *> LoadMoveSet;
2843 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, LoadMoveSet);
2845 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n");
2847 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) {
2848 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI);
2849 if (P == ChosenPairs.end()) {
2854 if (getDepthFactor(P->first) == 0) {
2855 // These instructions are not really fused, but are tracked as though
2856 // they are. Any case in which it would be interesting to fuse them
2857 // will be taken care of by InstCombine.
2863 Instruction *I = cast<Instruction>(P->first),
2864 *J = cast<Instruction>(P->second);
2866 DEBUG(dbgs() << "BBV: fusing: " << *I <<
2867 " <-> " << *J << "\n");
2869 // Remove the pair and flipped pair from the list.
2870 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second);
2871 assert(FP != ChosenPairs.end() && "Flipped pair not found in list");
2872 ChosenPairs.erase(FP);
2873 ChosenPairs.erase(P);
2875 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSet, I, J)) {
2876 DEBUG(dbgs() << "BBV: fusion of: " << *I <<
2878 " aborted because of non-trivial dependency cycle\n");
2884 // If the pair must have the other order, then flip it.
2885 bool FlipPairOrder = FixedOrderPairs.count(ValuePair(J, I));
2886 if (!FlipPairOrder && !FixedOrderPairs.count(ValuePair(I, J))) {
2887 // This pair does not have a fixed order, and so we might want to
2888 // flip it if that will yield fewer shuffles. We count the number
2889 // of dependencies connected via swaps, and those directly connected,
2890 // and flip the order if the number of swaps is greater.
2891 bool OrigOrder = true;
2892 VPPIteratorPair IP = ConnectedPairDeps.equal_range(ValuePair(I, J));
2893 if (IP.first == ConnectedPairDeps.end()) {
2894 IP = ConnectedPairDeps.equal_range(ValuePair(J, I));
2898 if (IP.first != ConnectedPairDeps.end()) {
2899 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
2900 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first;
2901 Q != IP.second; ++Q) {
2902 DenseMap<VPPair, unsigned>::iterator R =
2903 PairConnectionTypes.find(VPPair(Q->second, Q->first));
2904 assert(R != PairConnectionTypes.end() &&
2905 "Cannot find pair connection type");
2906 if (R->second == PairConnectionDirect)
2908 else if (R->second == PairConnectionSwap)
2913 std::swap(NumDepsDirect, NumDepsSwap);
2915 if (NumDepsSwap > NumDepsDirect) {
2916 FlipPairOrder = true;
2917 DEBUG(dbgs() << "BBV: reordering pair: " << *I <<
2918 " <-> " << *J << "\n");
2923 Instruction *L = I, *H = J;
2927 // If the pair being fused uses the opposite order from that in the pair
2928 // connection map, then we need to flip the types.
2929 VPPIteratorPair IP = ConnectedPairs.equal_range(ValuePair(H, L));
2930 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first;
2931 Q != IP.second; ++Q) {
2932 DenseMap<VPPair, unsigned>::iterator R = PairConnectionTypes.find(*Q);
2933 assert(R != PairConnectionTypes.end() &&
2934 "Cannot find pair connection type");
2935 if (R->second == PairConnectionDirect)
2936 R->second = PairConnectionSwap;
2937 else if (R->second == PairConnectionSwap)
2938 R->second = PairConnectionDirect;
2941 bool LBeforeH = !FlipPairOrder;
2942 unsigned NumOperands = I->getNumOperands();
2943 SmallVector<Value *, 3> ReplacedOperands(NumOperands);
2944 getReplacementInputsForPair(Context, L, H, ReplacedOperands,
2947 // Make a copy of the original operation, change its type to the vector
2948 // type and replace its operands with the vector operands.
2949 Instruction *K = L->clone();
2952 else if (H->hasName())
2955 if (!isa<StoreInst>(K))
2956 K->mutateType(getVecTypeForPair(L->getType(), H->getType()));
2958 combineMetadata(K, H);
2959 K->intersectOptionalDataWith(H);
2961 for (unsigned o = 0; o < NumOperands; ++o)
2962 K->setOperand(o, ReplacedOperands[o]);
2966 // Instruction insertion point:
2967 Instruction *InsertionPt = K;
2968 Instruction *K1 = 0, *K2 = 0;
2969 replaceOutputsOfPair(Context, L, H, K, InsertionPt, K1, K2);
2971 // The use tree of the first original instruction must be moved to after
2972 // the location of the second instruction. The entire use tree of the
2973 // first instruction is disjoint from the input tree of the second
2974 // (by definition), and so commutes with it.
2976 moveUsesOfIAfterJ(BB, LoadMoveSet, InsertionPt, I, J);
2978 if (!isa<StoreInst>(I)) {
2979 L->replaceAllUsesWith(K1);
2980 H->replaceAllUsesWith(K2);
2981 AA->replaceWithNewValue(L, K1);
2982 AA->replaceWithNewValue(H, K2);
2985 // Instructions that may read from memory may be in the load move set.
2986 // Once an instruction is fused, we no longer need its move set, and so
2987 // the values of the map never need to be updated. However, when a load
2988 // is fused, we need to merge the entries from both instructions in the
2989 // pair in case those instructions were in the move set of some other
2990 // yet-to-be-fused pair. The loads in question are the keys of the map.
2991 if (I->mayReadFromMemory()) {
2992 std::vector<ValuePair> NewSetMembers;
2993 VPIteratorPair IPairRange = LoadMoveSet.equal_range(I);
2994 VPIteratorPair JPairRange = LoadMoveSet.equal_range(J);
2995 for (std::multimap<Value *, Value *>::iterator N = IPairRange.first;
2996 N != IPairRange.second; ++N)
2997 NewSetMembers.push_back(ValuePair(K, N->second));
2998 for (std::multimap<Value *, Value *>::iterator N = JPairRange.first;
2999 N != JPairRange.second; ++N)
3000 NewSetMembers.push_back(ValuePair(K, N->second));
3001 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(),
3002 AE = NewSetMembers.end(); A != AE; ++A)
3003 LoadMoveSet.insert(*A);
3006 // Before removing I, set the iterator to the next instruction.
3007 PI = llvm::next(BasicBlock::iterator(I));
3008 if (cast<Instruction>(PI) == J)
3013 I->eraseFromParent();
3014 J->eraseFromParent();
3016 DEBUG(if (PrintAfterEveryPair) dbgs() << "BBV: block is now: \n" <<
3020 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n");
3024 char BBVectorize::ID = 0;
3025 static const char bb_vectorize_name[] = "Basic-Block Vectorization";
3026 INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
3027 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
3028 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
3029 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3030 INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
3032 BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) {
3033 return new BBVectorize(C);
3037 llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) {
3038 BBVectorize BBVectorizer(P, C);
3039 return BBVectorizer.vectorizeBB(BB);
3042 //===----------------------------------------------------------------------===//
3043 VectorizeConfig::VectorizeConfig() {
3044 VectorBits = ::VectorBits;
3045 VectorizeBools = !::NoBools;
3046 VectorizeInts = !::NoInts;
3047 VectorizeFloats = !::NoFloats;
3048 VectorizePointers = !::NoPointers;
3049 VectorizeCasts = !::NoCasts;
3050 VectorizeMath = !::NoMath;
3051 VectorizeFMA = !::NoFMA;
3052 VectorizeSelect = !::NoSelect;
3053 VectorizeCmp = !::NoCmp;
3054 VectorizeGEP = !::NoGEP;
3055 VectorizeMemOps = !::NoMemOps;
3056 AlignedOnly = ::AlignedOnly;
3057 ReqChainDepth= ::ReqChainDepth;
3058 SearchLimit = ::SearchLimit;
3059 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck;
3060 SplatBreaksChain = ::SplatBreaksChain;
3061 MaxInsts = ::MaxInsts;
3062 MaxIter = ::MaxIter;
3063 Pow2LenOnly = ::Pow2LenOnly;
3064 NoMemOpBoost = ::NoMemOpBoost;
3065 FastDep = ::FastDep;