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/Constants.h"
20 #include "llvm/DerivedTypes.h"
21 #include "llvm/Function.h"
22 #include "llvm/Instructions.h"
23 #include "llvm/IntrinsicInst.h"
24 #include "llvm/Intrinsics.h"
25 #include "llvm/LLVMContext.h"
26 #include "llvm/Metadata.h"
27 #include "llvm/Pass.h"
28 #include "llvm/Type.h"
29 #include "llvm/ADT/DenseMap.h"
30 #include "llvm/ADT/DenseSet.h"
31 #include "llvm/ADT/SmallVector.h"
32 #include "llvm/ADT/Statistic.h"
33 #include "llvm/ADT/STLExtras.h"
34 #include "llvm/ADT/StringExtras.h"
35 #include "llvm/Analysis/AliasAnalysis.h"
36 #include "llvm/Analysis/AliasSetTracker.h"
37 #include "llvm/Analysis/Dominators.h"
38 #include "llvm/Analysis/ScalarEvolution.h"
39 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
40 #include "llvm/Analysis/ValueTracking.h"
41 #include "llvm/Support/CommandLine.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/raw_ostream.h"
44 #include "llvm/Support/ValueHandle.h"
45 #include "llvm/DataLayout.h"
46 #include "llvm/TargetTransformInfo.h"
47 #include "llvm/Transforms/Utils/Local.h"
48 #include "llvm/Transforms/Vectorize.h"
54 IgnoreTargetInfo("bb-vectorize-ignore-target-info", cl::init(false),
55 cl::Hidden, cl::desc("Ignore target information"));
57 static cl::opt<unsigned>
58 ReqChainDepth("bb-vectorize-req-chain-depth", cl::init(6), cl::Hidden,
59 cl::desc("The required chain depth for vectorization"));
62 UseChainDepthWithTI("bb-vectorize-use-chain-depth", cl::init(false),
63 cl::Hidden, cl::desc("Use the chain depth requirement with"
64 " target information"));
66 static cl::opt<unsigned>
67 SearchLimit("bb-vectorize-search-limit", cl::init(400), cl::Hidden,
68 cl::desc("The maximum search distance for instruction pairs"));
71 SplatBreaksChain("bb-vectorize-splat-breaks-chain", cl::init(false), cl::Hidden,
72 cl::desc("Replicating one element to a pair breaks the chain"));
74 static cl::opt<unsigned>
75 VectorBits("bb-vectorize-vector-bits", cl::init(128), cl::Hidden,
76 cl::desc("The size of the native vector registers"));
78 static cl::opt<unsigned>
79 MaxIter("bb-vectorize-max-iter", cl::init(0), cl::Hidden,
80 cl::desc("The maximum number of pairing iterations"));
83 Pow2LenOnly("bb-vectorize-pow2-len-only", cl::init(false), cl::Hidden,
84 cl::desc("Don't try to form non-2^n-length vectors"));
86 static cl::opt<unsigned>
87 MaxInsts("bb-vectorize-max-instr-per-group", cl::init(500), cl::Hidden,
88 cl::desc("The maximum number of pairable instructions per group"));
90 static cl::opt<unsigned>
91 MaxCandPairsForCycleCheck("bb-vectorize-max-cycle-check-pairs", cl::init(200),
92 cl::Hidden, cl::desc("The maximum number of candidate pairs with which to use"
93 " a full cycle check"));
96 NoBools("bb-vectorize-no-bools", cl::init(false), cl::Hidden,
97 cl::desc("Don't try to vectorize boolean (i1) values"));
100 NoInts("bb-vectorize-no-ints", cl::init(false), cl::Hidden,
101 cl::desc("Don't try to vectorize integer values"));
104 NoFloats("bb-vectorize-no-floats", cl::init(false), cl::Hidden,
105 cl::desc("Don't try to vectorize floating-point values"));
107 // FIXME: This should default to false once pointer vector support works.
109 NoPointers("bb-vectorize-no-pointers", cl::init(/*false*/ true), cl::Hidden,
110 cl::desc("Don't try to vectorize pointer values"));
113 NoCasts("bb-vectorize-no-casts", cl::init(false), cl::Hidden,
114 cl::desc("Don't try to vectorize casting (conversion) operations"));
117 NoMath("bb-vectorize-no-math", cl::init(false), cl::Hidden,
118 cl::desc("Don't try to vectorize floating-point math intrinsics"));
121 NoFMA("bb-vectorize-no-fma", cl::init(false), cl::Hidden,
122 cl::desc("Don't try to vectorize the fused-multiply-add intrinsic"));
125 NoSelect("bb-vectorize-no-select", cl::init(false), cl::Hidden,
126 cl::desc("Don't try to vectorize select instructions"));
129 NoCmp("bb-vectorize-no-cmp", cl::init(false), cl::Hidden,
130 cl::desc("Don't try to vectorize comparison instructions"));
133 NoGEP("bb-vectorize-no-gep", cl::init(false), cl::Hidden,
134 cl::desc("Don't try to vectorize getelementptr instructions"));
137 NoMemOps("bb-vectorize-no-mem-ops", cl::init(false), cl::Hidden,
138 cl::desc("Don't try to vectorize loads and stores"));
141 AlignedOnly("bb-vectorize-aligned-only", cl::init(false), cl::Hidden,
142 cl::desc("Only generate aligned loads and stores"));
145 NoMemOpBoost("bb-vectorize-no-mem-op-boost",
146 cl::init(false), cl::Hidden,
147 cl::desc("Don't boost the chain-depth contribution of loads and stores"));
150 FastDep("bb-vectorize-fast-dep", cl::init(false), cl::Hidden,
151 cl::desc("Use a fast instruction dependency analysis"));
155 DebugInstructionExamination("bb-vectorize-debug-instruction-examination",
156 cl::init(false), cl::Hidden,
157 cl::desc("When debugging is enabled, output information on the"
158 " instruction-examination process"));
160 DebugCandidateSelection("bb-vectorize-debug-candidate-selection",
161 cl::init(false), cl::Hidden,
162 cl::desc("When debugging is enabled, output information on the"
163 " candidate-selection process"));
165 DebugPairSelection("bb-vectorize-debug-pair-selection",
166 cl::init(false), cl::Hidden,
167 cl::desc("When debugging is enabled, output information on the"
168 " pair-selection process"));
170 DebugCycleCheck("bb-vectorize-debug-cycle-check",
171 cl::init(false), cl::Hidden,
172 cl::desc("When debugging is enabled, output information on the"
173 " cycle-checking process"));
176 PrintAfterEveryPair("bb-vectorize-debug-print-after-every-pair",
177 cl::init(false), cl::Hidden,
178 cl::desc("When debugging is enabled, dump the basic block after"
179 " every pair is fused"));
182 STATISTIC(NumFusedOps, "Number of operations fused by bb-vectorize");
185 struct BBVectorize : public BasicBlockPass {
186 static char ID; // Pass identification, replacement for typeid
188 const VectorizeConfig Config;
190 BBVectorize(const VectorizeConfig &C = VectorizeConfig())
191 : BasicBlockPass(ID), Config(C) {
192 initializeBBVectorizePass(*PassRegistry::getPassRegistry());
195 BBVectorize(Pass *P, const VectorizeConfig &C)
196 : BasicBlockPass(ID), Config(C) {
197 AA = &P->getAnalysis<AliasAnalysis>();
198 DT = &P->getAnalysis<DominatorTree>();
199 SE = &P->getAnalysis<ScalarEvolution>();
200 TD = P->getAnalysisIfAvailable<DataLayout>();
201 TTI = IgnoreTargetInfo ? 0 :
202 P->getAnalysisIfAvailable<TargetTransformInfo>();
203 VTTI = TTI ? TTI->getVectorTargetTransformInfo() : 0;
206 typedef std::pair<Value *, Value *> ValuePair;
207 typedef std::pair<ValuePair, int> ValuePairWithCost;
208 typedef std::pair<ValuePair, size_t> ValuePairWithDepth;
209 typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair
210 typedef std::pair<VPPair, unsigned> VPPairWithType;
211 typedef std::pair<std::multimap<Value *, Value *>::iterator,
212 std::multimap<Value *, Value *>::iterator> VPIteratorPair;
213 typedef std::pair<std::multimap<ValuePair, ValuePair>::iterator,
214 std::multimap<ValuePair, ValuePair>::iterator>
221 TargetTransformInfo *TTI;
222 const VectorTargetTransformInfo *VTTI;
224 // FIXME: const correct?
226 bool vectorizePairs(BasicBlock &BB, bool NonPow2Len = false);
228 bool getCandidatePairs(BasicBlock &BB,
229 BasicBlock::iterator &Start,
230 std::multimap<Value *, Value *> &CandidatePairs,
231 DenseSet<ValuePair> &FixedOrderPairs,
232 DenseMap<ValuePair, int> &CandidatePairCostSavings,
233 std::vector<Value *> &PairableInsts, bool NonPow2Len);
235 // FIXME: The current implementation does not account for pairs that
236 // are connected in multiple ways. For example:
237 // C1 = A1 / A2; C2 = A2 / A1 (which may be both direct and a swap)
238 enum PairConnectionType {
239 PairConnectionDirect,
244 void computeConnectedPairs(std::multimap<Value *, Value *> &CandidatePairs,
245 std::vector<Value *> &PairableInsts,
246 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
247 DenseMap<VPPair, unsigned> &PairConnectionTypes);
249 void buildDepMap(BasicBlock &BB,
250 std::multimap<Value *, Value *> &CandidatePairs,
251 std::vector<Value *> &PairableInsts,
252 DenseSet<ValuePair> &PairableInstUsers);
254 void choosePairs(std::multimap<Value *, Value *> &CandidatePairs,
255 DenseMap<ValuePair, int> &CandidatePairCostSavings,
256 std::vector<Value *> &PairableInsts,
257 DenseSet<ValuePair> &FixedOrderPairs,
258 DenseMap<VPPair, unsigned> &PairConnectionTypes,
259 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
260 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps,
261 DenseSet<ValuePair> &PairableInstUsers,
262 DenseMap<Value *, Value *>& ChosenPairs);
264 void fuseChosenPairs(BasicBlock &BB,
265 std::vector<Value *> &PairableInsts,
266 DenseMap<Value *, Value *>& ChosenPairs,
267 DenseSet<ValuePair> &FixedOrderPairs,
268 DenseMap<VPPair, unsigned> &PairConnectionTypes,
269 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
270 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps);
273 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore);
275 bool areInstsCompatible(Instruction *I, Instruction *J,
276 bool IsSimpleLoadStore, bool NonPow2Len,
277 int &CostSavings, int &FixedOrder);
279 bool trackUsesOfI(DenseSet<Value *> &Users,
280 AliasSetTracker &WriteSet, Instruction *I,
281 Instruction *J, bool UpdateUsers = true,
282 std::multimap<Value *, Value *> *LoadMoveSet = 0);
284 void computePairsConnectedTo(
285 std::multimap<Value *, Value *> &CandidatePairs,
286 std::vector<Value *> &PairableInsts,
287 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
288 DenseMap<VPPair, unsigned> &PairConnectionTypes,
291 bool pairsConflict(ValuePair P, ValuePair Q,
292 DenseSet<ValuePair> &PairableInstUsers,
293 std::multimap<ValuePair, ValuePair> *PairableInstUserMap = 0);
295 bool pairWillFormCycle(ValuePair P,
296 std::multimap<ValuePair, ValuePair> &PairableInstUsers,
297 DenseSet<ValuePair> &CurrentPairs);
300 std::multimap<Value *, Value *> &CandidatePairs,
301 std::vector<Value *> &PairableInsts,
302 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
303 DenseSet<ValuePair> &PairableInstUsers,
304 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
305 DenseMap<Value *, Value *> &ChosenPairs,
306 DenseMap<ValuePair, size_t> &Tree,
307 DenseSet<ValuePair> &PrunedTree, ValuePair J,
310 void buildInitialTreeFor(
311 std::multimap<Value *, Value *> &CandidatePairs,
312 std::vector<Value *> &PairableInsts,
313 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
314 DenseSet<ValuePair> &PairableInstUsers,
315 DenseMap<Value *, Value *> &ChosenPairs,
316 DenseMap<ValuePair, size_t> &Tree, ValuePair J);
318 void findBestTreeFor(
319 std::multimap<Value *, Value *> &CandidatePairs,
320 DenseMap<ValuePair, int> &CandidatePairCostSavings,
321 std::vector<Value *> &PairableInsts,
322 DenseSet<ValuePair> &FixedOrderPairs,
323 DenseMap<VPPair, unsigned> &PairConnectionTypes,
324 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
325 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps,
326 DenseSet<ValuePair> &PairableInstUsers,
327 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
328 DenseMap<Value *, Value *> &ChosenPairs,
329 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth,
330 int &BestEffSize, VPIteratorPair ChoiceRange,
333 Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I,
334 Instruction *J, unsigned o);
336 void fillNewShuffleMask(LLVMContext& Context, Instruction *J,
337 unsigned MaskOffset, unsigned NumInElem,
338 unsigned NumInElem1, unsigned IdxOffset,
339 std::vector<Constant*> &Mask);
341 Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I,
344 bool expandIEChain(LLVMContext& Context, Instruction *I, Instruction *J,
345 unsigned o, Value *&LOp, unsigned numElemL,
346 Type *ArgTypeL, Type *ArgTypeR, bool IBeforeJ,
347 unsigned IdxOff = 0);
349 Value *getReplacementInput(LLVMContext& Context, Instruction *I,
350 Instruction *J, unsigned o, bool IBeforeJ);
352 void getReplacementInputsForPair(LLVMContext& Context, Instruction *I,
353 Instruction *J, SmallVector<Value *, 3> &ReplacedOperands,
356 void replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
357 Instruction *J, Instruction *K,
358 Instruction *&InsertionPt, Instruction *&K1,
361 void collectPairLoadMoveSet(BasicBlock &BB,
362 DenseMap<Value *, Value *> &ChosenPairs,
363 std::multimap<Value *, Value *> &LoadMoveSet,
366 void collectLoadMoveSet(BasicBlock &BB,
367 std::vector<Value *> &PairableInsts,
368 DenseMap<Value *, Value *> &ChosenPairs,
369 std::multimap<Value *, Value *> &LoadMoveSet);
371 bool canMoveUsesOfIAfterJ(BasicBlock &BB,
372 std::multimap<Value *, Value *> &LoadMoveSet,
373 Instruction *I, Instruction *J);
375 void moveUsesOfIAfterJ(BasicBlock &BB,
376 std::multimap<Value *, Value *> &LoadMoveSet,
377 Instruction *&InsertionPt,
378 Instruction *I, Instruction *J);
380 void combineMetadata(Instruction *K, const Instruction *J);
382 bool vectorizeBB(BasicBlock &BB) {
383 if (!DT->isReachableFromEntry(&BB)) {
384 DEBUG(dbgs() << "BBV: skipping unreachable " << BB.getName() <<
385 " in " << BB.getParent()->getName() << "\n");
389 DEBUG(if (VTTI) dbgs() << "BBV: using target information\n");
391 bool changed = false;
392 // Iterate a sufficient number of times to merge types of size 1 bit,
393 // then 2 bits, then 4, etc. up to half of the target vector width of the
394 // target vector register.
397 (VTTI || v <= Config.VectorBits) &&
398 (!Config.MaxIter || n <= Config.MaxIter);
400 DEBUG(dbgs() << "BBV: fusing loop #" << n <<
401 " for " << BB.getName() << " in " <<
402 BB.getParent()->getName() << "...\n");
403 if (vectorizePairs(BB))
409 if (changed && !Pow2LenOnly) {
411 for (; !Config.MaxIter || n <= Config.MaxIter; ++n) {
412 DEBUG(dbgs() << "BBV: fusing for non-2^n-length vectors loop #: " <<
413 n << " for " << BB.getName() << " in " <<
414 BB.getParent()->getName() << "...\n");
415 if (!vectorizePairs(BB, true)) break;
419 DEBUG(dbgs() << "BBV: done!\n");
423 virtual bool runOnBasicBlock(BasicBlock &BB) {
424 AA = &getAnalysis<AliasAnalysis>();
425 DT = &getAnalysis<DominatorTree>();
426 SE = &getAnalysis<ScalarEvolution>();
427 TD = getAnalysisIfAvailable<DataLayout>();
428 TTI = IgnoreTargetInfo ? 0 :
429 getAnalysisIfAvailable<TargetTransformInfo>();
430 VTTI = TTI ? TTI->getVectorTargetTransformInfo() : 0;
432 return vectorizeBB(BB);
435 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
436 BasicBlockPass::getAnalysisUsage(AU);
437 AU.addRequired<AliasAnalysis>();
438 AU.addRequired<DominatorTree>();
439 AU.addRequired<ScalarEvolution>();
440 AU.addPreserved<AliasAnalysis>();
441 AU.addPreserved<DominatorTree>();
442 AU.addPreserved<ScalarEvolution>();
443 AU.setPreservesCFG();
446 static inline VectorType *getVecTypeForPair(Type *ElemTy, Type *Elem2Ty) {
447 assert(ElemTy->getScalarType() == Elem2Ty->getScalarType() &&
448 "Cannot form vector from incompatible scalar types");
449 Type *STy = ElemTy->getScalarType();
452 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) {
453 numElem = VTy->getNumElements();
458 if (VectorType *VTy = dyn_cast<VectorType>(Elem2Ty)) {
459 numElem += VTy->getNumElements();
464 return VectorType::get(STy, numElem);
467 static inline void getInstructionTypes(Instruction *I,
468 Type *&T1, Type *&T2) {
469 if (isa<StoreInst>(I)) {
470 // For stores, it is the value type, not the pointer type that matters
471 // because the value is what will come from a vector register.
473 Value *IVal = cast<StoreInst>(I)->getValueOperand();
474 T1 = IVal->getType();
480 T2 = cast<CastInst>(I)->getSrcTy();
484 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
485 T2 = SI->getCondition()->getType();
486 } else if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(I)) {
487 T2 = SI->getOperand(0)->getType();
488 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) {
489 T2 = CI->getOperand(0)->getType();
493 // Returns the weight associated with the provided value. A chain of
494 // candidate pairs has a length given by the sum of the weights of its
495 // members (one weight per pair; the weight of each member of the pair
496 // is assumed to be the same). This length is then compared to the
497 // chain-length threshold to determine if a given chain is significant
498 // enough to be vectorized. The length is also used in comparing
499 // candidate chains where longer chains are considered to be better.
500 // Note: when this function returns 0, the resulting instructions are
501 // not actually fused.
502 inline size_t getDepthFactor(Value *V) {
503 // InsertElement and ExtractElement have a depth factor of zero. This is
504 // for two reasons: First, they cannot be usefully fused. Second, because
505 // the pass generates a lot of these, they can confuse the simple metric
506 // used to compare the trees in the next iteration. Thus, giving them a
507 // weight of zero allows the pass to essentially ignore them in
508 // subsequent iterations when looking for vectorization opportunities
509 // while still tracking dependency chains that flow through those
511 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V))
514 // Give a load or store half of the required depth so that load/store
515 // pairs will vectorize.
516 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V)))
517 return Config.ReqChainDepth/2;
522 // Returns the cost of the provided instruction using VTTI.
523 // This does not handle loads and stores.
524 unsigned getInstrCost(unsigned Opcode, Type *T1, Type *T2) {
527 case Instruction::GetElementPtr:
528 // We mark this instruction as zero-cost because scalar GEPs are usually
529 // lowered to the intruction addressing mode. At the moment we don't
530 // generate vector GEPs.
532 case Instruction::Br:
533 return VTTI->getCFInstrCost(Opcode);
534 case Instruction::PHI:
536 case Instruction::Add:
537 case Instruction::FAdd:
538 case Instruction::Sub:
539 case Instruction::FSub:
540 case Instruction::Mul:
541 case Instruction::FMul:
542 case Instruction::UDiv:
543 case Instruction::SDiv:
544 case Instruction::FDiv:
545 case Instruction::URem:
546 case Instruction::SRem:
547 case Instruction::FRem:
548 case Instruction::Shl:
549 case Instruction::LShr:
550 case Instruction::AShr:
551 case Instruction::And:
552 case Instruction::Or:
553 case Instruction::Xor:
554 return VTTI->getArithmeticInstrCost(Opcode, T1);
555 case Instruction::Select:
556 case Instruction::ICmp:
557 case Instruction::FCmp:
558 return VTTI->getCmpSelInstrCost(Opcode, T1, T2);
559 case Instruction::ZExt:
560 case Instruction::SExt:
561 case Instruction::FPToUI:
562 case Instruction::FPToSI:
563 case Instruction::FPExt:
564 case Instruction::PtrToInt:
565 case Instruction::IntToPtr:
566 case Instruction::SIToFP:
567 case Instruction::UIToFP:
568 case Instruction::Trunc:
569 case Instruction::FPTrunc:
570 case Instruction::BitCast:
571 case Instruction::ShuffleVector:
572 return VTTI->getCastInstrCost(Opcode, T1, T2);
578 // This determines the relative offset of two loads or stores, returning
579 // true if the offset could be determined to be some constant value.
580 // For example, if OffsetInElmts == 1, then J accesses the memory directly
581 // after I; if OffsetInElmts == -1 then I accesses the memory
583 bool getPairPtrInfo(Instruction *I, Instruction *J,
584 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment,
585 unsigned &IAddressSpace, unsigned &JAddressSpace,
586 int64_t &OffsetInElmts, bool ComputeOffset = true) {
588 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
589 LoadInst *LJ = cast<LoadInst>(J);
590 IPtr = LI->getPointerOperand();
591 JPtr = LJ->getPointerOperand();
592 IAlignment = LI->getAlignment();
593 JAlignment = LJ->getAlignment();
594 IAddressSpace = LI->getPointerAddressSpace();
595 JAddressSpace = LJ->getPointerAddressSpace();
597 StoreInst *SI = cast<StoreInst>(I), *SJ = cast<StoreInst>(J);
598 IPtr = SI->getPointerOperand();
599 JPtr = SJ->getPointerOperand();
600 IAlignment = SI->getAlignment();
601 JAlignment = SJ->getAlignment();
602 IAddressSpace = SI->getPointerAddressSpace();
603 JAddressSpace = SJ->getPointerAddressSpace();
609 const SCEV *IPtrSCEV = SE->getSCEV(IPtr);
610 const SCEV *JPtrSCEV = SE->getSCEV(JPtr);
612 // If this is a trivial offset, then we'll get something like
613 // 1*sizeof(type). With target data, which we need anyway, this will get
614 // constant folded into a number.
615 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV);
616 if (const SCEVConstant *ConstOffSCEV =
617 dyn_cast<SCEVConstant>(OffsetSCEV)) {
618 ConstantInt *IntOff = ConstOffSCEV->getValue();
619 int64_t Offset = IntOff->getSExtValue();
621 Type *VTy = cast<PointerType>(IPtr->getType())->getElementType();
622 int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy);
624 Type *VTy2 = cast<PointerType>(JPtr->getType())->getElementType();
625 if (VTy != VTy2 && Offset < 0) {
626 int64_t VTy2TSS = (int64_t) TD->getTypeStoreSize(VTy2);
627 OffsetInElmts = Offset/VTy2TSS;
628 return (abs64(Offset) % VTy2TSS) == 0;
631 OffsetInElmts = Offset/VTyTSS;
632 return (abs64(Offset) % VTyTSS) == 0;
638 // Returns true if the provided CallInst represents an intrinsic that can
640 bool isVectorizableIntrinsic(CallInst* I) {
641 Function *F = I->getCalledFunction();
642 if (!F) return false;
644 unsigned IID = F->getIntrinsicID();
645 if (!IID) return false;
650 case Intrinsic::sqrt:
651 case Intrinsic::powi:
655 case Intrinsic::log2:
656 case Intrinsic::log10:
658 case Intrinsic::exp2:
660 return Config.VectorizeMath;
662 return Config.VectorizeFMA;
666 // Returns true if J is the second element in some pair referenced by
667 // some multimap pair iterator pair.
668 template <typename V>
669 bool isSecondInIteratorPair(V J, std::pair<
670 typename std::multimap<V, V>::iterator,
671 typename std::multimap<V, V>::iterator> PairRange) {
672 for (typename std::multimap<V, V>::iterator K = PairRange.first;
673 K != PairRange.second; ++K)
674 if (K->second == J) return true;
680 // This function implements one vectorization iteration on the provided
681 // basic block. It returns true if the block is changed.
682 bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) {
684 BasicBlock::iterator Start = BB.getFirstInsertionPt();
686 std::vector<Value *> AllPairableInsts;
687 DenseMap<Value *, Value *> AllChosenPairs;
688 DenseSet<ValuePair> AllFixedOrderPairs;
689 DenseMap<VPPair, unsigned> AllPairConnectionTypes;
690 std::multimap<ValuePair, ValuePair> AllConnectedPairs, AllConnectedPairDeps;
693 std::vector<Value *> PairableInsts;
694 std::multimap<Value *, Value *> CandidatePairs;
695 DenseSet<ValuePair> FixedOrderPairs;
696 DenseMap<ValuePair, int> CandidatePairCostSavings;
697 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs,
699 CandidatePairCostSavings,
700 PairableInsts, NonPow2Len);
701 if (PairableInsts.empty()) continue;
703 // Now we have a map of all of the pairable instructions and we need to
704 // select the best possible pairing. A good pairing is one such that the
705 // users of the pair are also paired. This defines a (directed) forest
706 // over the pairs such that two pairs are connected iff the second pair
709 // Note that it only matters that both members of the second pair use some
710 // element of the first pair (to allow for splatting).
712 std::multimap<ValuePair, ValuePair> ConnectedPairs, ConnectedPairDeps;
713 DenseMap<VPPair, unsigned> PairConnectionTypes;
714 computeConnectedPairs(CandidatePairs, PairableInsts, ConnectedPairs,
715 PairConnectionTypes);
716 if (ConnectedPairs.empty()) continue;
718 for (std::multimap<ValuePair, ValuePair>::iterator
719 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
721 ConnectedPairDeps.insert(VPPair(I->second, I->first));
724 // Build the pairable-instruction dependency map
725 DenseSet<ValuePair> PairableInstUsers;
726 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers);
728 // There is now a graph of the connected pairs. For each variable, pick
729 // the pairing with the largest tree meeting the depth requirement on at
730 // least one branch. Then select all pairings that are part of that tree
731 // and remove them from the list of available pairings and pairable
734 DenseMap<Value *, Value *> ChosenPairs;
735 choosePairs(CandidatePairs, CandidatePairCostSavings,
736 PairableInsts, FixedOrderPairs, PairConnectionTypes,
737 ConnectedPairs, ConnectedPairDeps,
738 PairableInstUsers, ChosenPairs);
740 if (ChosenPairs.empty()) continue;
741 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(),
742 PairableInsts.end());
743 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end());
745 // Only for the chosen pairs, propagate information on fixed-order pairs,
746 // pair connections, and their types to the data structures used by the
747 // pair fusion procedures.
748 for (DenseMap<Value *, Value *>::iterator I = ChosenPairs.begin(),
749 IE = ChosenPairs.end(); I != IE; ++I) {
750 if (FixedOrderPairs.count(*I))
751 AllFixedOrderPairs.insert(*I);
752 else if (FixedOrderPairs.count(ValuePair(I->second, I->first)))
753 AllFixedOrderPairs.insert(ValuePair(I->second, I->first));
755 for (DenseMap<Value *, Value *>::iterator J = ChosenPairs.begin();
757 DenseMap<VPPair, unsigned>::iterator K =
758 PairConnectionTypes.find(VPPair(*I, *J));
759 if (K != PairConnectionTypes.end()) {
760 AllPairConnectionTypes.insert(*K);
762 K = PairConnectionTypes.find(VPPair(*J, *I));
763 if (K != PairConnectionTypes.end())
764 AllPairConnectionTypes.insert(*K);
769 for (std::multimap<ValuePair, ValuePair>::iterator
770 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
772 if (AllPairConnectionTypes.count(*I)) {
773 AllConnectedPairs.insert(*I);
774 AllConnectedPairDeps.insert(VPPair(I->second, I->first));
777 } while (ShouldContinue);
779 if (AllChosenPairs.empty()) return false;
780 NumFusedOps += AllChosenPairs.size();
782 // A set of pairs has now been selected. It is now necessary to replace the
783 // paired instructions with vector instructions. For this procedure each
784 // operand must be replaced with a vector operand. This vector is formed
785 // by using build_vector on the old operands. The replaced values are then
786 // replaced with a vector_extract on the result. Subsequent optimization
787 // passes should coalesce the build/extract combinations.
789 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs, AllFixedOrderPairs,
790 AllPairConnectionTypes,
791 AllConnectedPairs, AllConnectedPairDeps);
793 // It is important to cleanup here so that future iterations of this
794 // function have less work to do.
795 (void) SimplifyInstructionsInBlock(&BB, TD, AA->getTargetLibraryInfo());
799 // This function returns true if the provided instruction is capable of being
800 // fused into a vector instruction. This determination is based only on the
801 // type and other attributes of the instruction.
802 bool BBVectorize::isInstVectorizable(Instruction *I,
803 bool &IsSimpleLoadStore) {
804 IsSimpleLoadStore = false;
806 if (CallInst *C = dyn_cast<CallInst>(I)) {
807 if (!isVectorizableIntrinsic(C))
809 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) {
810 // Vectorize simple loads if possbile:
811 IsSimpleLoadStore = L->isSimple();
812 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
814 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) {
815 // Vectorize simple stores if possbile:
816 IsSimpleLoadStore = S->isSimple();
817 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
819 } else if (CastInst *C = dyn_cast<CastInst>(I)) {
820 // We can vectorize casts, but not casts of pointer types, etc.
821 if (!Config.VectorizeCasts)
824 Type *SrcTy = C->getSrcTy();
825 if (!SrcTy->isSingleValueType())
828 Type *DestTy = C->getDestTy();
829 if (!DestTy->isSingleValueType())
831 } else if (isa<SelectInst>(I)) {
832 if (!Config.VectorizeSelect)
834 } else if (isa<CmpInst>(I)) {
835 if (!Config.VectorizeCmp)
837 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) {
838 if (!Config.VectorizeGEP)
841 // Currently, vector GEPs exist only with one index.
842 if (G->getNumIndices() != 1)
844 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) ||
845 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) {
849 // We can't vectorize memory operations without target data
850 if (TD == 0 && IsSimpleLoadStore)
854 getInstructionTypes(I, T1, T2);
856 // Not every type can be vectorized...
857 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) ||
858 !(VectorType::isValidElementType(T2) || T2->isVectorTy()))
861 if (T1->getScalarSizeInBits() == 1) {
862 if (!Config.VectorizeBools)
865 if (!Config.VectorizeInts && T1->isIntOrIntVectorTy())
869 if (T2->getScalarSizeInBits() == 1) {
870 if (!Config.VectorizeBools)
873 if (!Config.VectorizeInts && T2->isIntOrIntVectorTy())
877 if (!Config.VectorizeFloats
878 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy()))
881 // Don't vectorize target-specific types.
882 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy())
884 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy())
887 if ((!Config.VectorizePointers || TD == 0) &&
888 (T1->getScalarType()->isPointerTy() ||
889 T2->getScalarType()->isPointerTy()))
892 if (!VTTI && (T1->getPrimitiveSizeInBits() >= Config.VectorBits ||
893 T2->getPrimitiveSizeInBits() >= Config.VectorBits))
899 // This function returns true if the two provided instructions are compatible
900 // (meaning that they can be fused into a vector instruction). This assumes
901 // that I has already been determined to be vectorizable and that J is not
902 // in the use tree of I.
903 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J,
904 bool IsSimpleLoadStore, bool NonPow2Len,
905 int &CostSavings, int &FixedOrder) {
906 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I <<
907 " <-> " << *J << "\n");
912 // Loads and stores can be merged if they have different alignments,
913 // but are otherwise the same.
914 if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment |
915 (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0)))
918 Type *IT1, *IT2, *JT1, *JT2;
919 getInstructionTypes(I, IT1, IT2);
920 getInstructionTypes(J, JT1, JT2);
921 unsigned MaxTypeBits = std::max(
922 IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(),
923 IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits());
924 if (!VTTI && MaxTypeBits > Config.VectorBits)
927 // FIXME: handle addsub-type operations!
929 if (IsSimpleLoadStore) {
931 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
932 int64_t OffsetInElmts = 0;
933 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
934 IAddressSpace, JAddressSpace,
935 OffsetInElmts) && abs64(OffsetInElmts) == 1) {
936 FixedOrder = (int) OffsetInElmts;
937 unsigned BottomAlignment = IAlignment;
938 if (OffsetInElmts < 0) BottomAlignment = JAlignment;
940 Type *aTypeI = isa<StoreInst>(I) ?
941 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType();
942 Type *aTypeJ = isa<StoreInst>(J) ?
943 cast<StoreInst>(J)->getValueOperand()->getType() : J->getType();
944 Type *VType = getVecTypeForPair(aTypeI, aTypeJ);
946 if (Config.AlignedOnly) {
947 // An aligned load or store is possible only if the instruction
948 // with the lower offset has an alignment suitable for the
951 unsigned VecAlignment = TD->getPrefTypeAlignment(VType);
952 if (BottomAlignment < VecAlignment)
957 unsigned ICost = VTTI->getMemoryOpCost(I->getOpcode(), I->getType(),
958 IAlignment, IAddressSpace);
959 unsigned JCost = VTTI->getMemoryOpCost(J->getOpcode(), J->getType(),
960 JAlignment, JAddressSpace);
961 unsigned VCost = VTTI->getMemoryOpCost(I->getOpcode(), VType,
964 if (VCost > ICost + JCost)
967 // We don't want to fuse to a type that will be split, even
968 // if the two input types will also be split and there is no other
970 unsigned VParts = VTTI->getNumberOfParts(VType);
973 else if (!VParts && VCost == ICost + JCost)
976 CostSavings = ICost + JCost - VCost;
982 unsigned ICost = getInstrCost(I->getOpcode(), IT1, IT2);
983 unsigned JCost = getInstrCost(J->getOpcode(), JT1, JT2);
984 Type *VT1 = getVecTypeForPair(IT1, JT1),
985 *VT2 = getVecTypeForPair(IT2, JT2);
986 unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2);
988 if (VCost > ICost + JCost)
991 // We don't want to fuse to a type that will be split, even
992 // if the two input types will also be split and there is no other
994 unsigned VParts1 = VTTI->getNumberOfParts(VT1),
995 VParts2 = VTTI->getNumberOfParts(VT2);
996 if (VParts1 > 1 || VParts2 > 1)
998 else if ((!VParts1 || !VParts2) && VCost == ICost + JCost)
1001 CostSavings = ICost + JCost - VCost;
1004 // The powi intrinsic is special because only the first argument is
1005 // vectorized, the second arguments must be equal.
1006 CallInst *CI = dyn_cast<CallInst>(I);
1008 if (CI && (FI = CI->getCalledFunction()) &&
1009 FI->getIntrinsicID() == Intrinsic::powi) {
1011 Value *A1I = CI->getArgOperand(1),
1012 *A1J = cast<CallInst>(J)->getArgOperand(1);
1013 const SCEV *A1ISCEV = SE->getSCEV(A1I),
1014 *A1JSCEV = SE->getSCEV(A1J);
1015 return (A1ISCEV == A1JSCEV);
1021 // Figure out whether or not J uses I and update the users and write-set
1022 // structures associated with I. Specifically, Users represents the set of
1023 // instructions that depend on I. WriteSet represents the set
1024 // of memory locations that are dependent on I. If UpdateUsers is true,
1025 // and J uses I, then Users is updated to contain J and WriteSet is updated
1026 // to contain any memory locations to which J writes. The function returns
1027 // true if J uses I. By default, alias analysis is used to determine
1028 // whether J reads from memory that overlaps with a location in WriteSet.
1029 // If LoadMoveSet is not null, then it is a previously-computed multimap
1030 // where the key is the memory-based user instruction and the value is
1031 // the instruction to be compared with I. So, if LoadMoveSet is provided,
1032 // then the alias analysis is not used. This is necessary because this
1033 // function is called during the process of moving instructions during
1034 // vectorization and the results of the alias analysis are not stable during
1036 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users,
1037 AliasSetTracker &WriteSet, Instruction *I,
1038 Instruction *J, bool UpdateUsers,
1039 std::multimap<Value *, Value *> *LoadMoveSet) {
1042 // This instruction may already be marked as a user due, for example, to
1043 // being a member of a selected pair.
1048 for (User::op_iterator JU = J->op_begin(), JE = J->op_end();
1051 if (I == V || Users.count(V)) {
1056 if (!UsesI && J->mayReadFromMemory()) {
1058 VPIteratorPair JPairRange = LoadMoveSet->equal_range(J);
1059 UsesI = isSecondInIteratorPair<Value*>(I, JPairRange);
1061 for (AliasSetTracker::iterator W = WriteSet.begin(),
1062 WE = WriteSet.end(); W != WE; ++W) {
1063 if (W->aliasesUnknownInst(J, *AA)) {
1071 if (UsesI && UpdateUsers) {
1072 if (J->mayWriteToMemory()) WriteSet.add(J);
1079 // This function iterates over all instruction pairs in the provided
1080 // basic block and collects all candidate pairs for vectorization.
1081 bool BBVectorize::getCandidatePairs(BasicBlock &BB,
1082 BasicBlock::iterator &Start,
1083 std::multimap<Value *, Value *> &CandidatePairs,
1084 DenseSet<ValuePair> &FixedOrderPairs,
1085 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1086 std::vector<Value *> &PairableInsts, bool NonPow2Len) {
1087 BasicBlock::iterator E = BB.end();
1088 if (Start == E) return false;
1090 bool ShouldContinue = false, IAfterStart = false;
1091 for (BasicBlock::iterator I = Start++; I != E; ++I) {
1092 if (I == Start) IAfterStart = true;
1094 bool IsSimpleLoadStore;
1095 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue;
1097 // Look for an instruction with which to pair instruction *I...
1098 DenseSet<Value *> Users;
1099 AliasSetTracker WriteSet(*AA);
1100 bool JAfterStart = IAfterStart;
1101 BasicBlock::iterator J = llvm::next(I);
1102 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) {
1103 if (J == Start) JAfterStart = true;
1105 // Determine if J uses I, if so, exit the loop.
1106 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep);
1107 if (Config.FastDep) {
1108 // Note: For this heuristic to be effective, independent operations
1109 // must tend to be intermixed. This is likely to be true from some
1110 // kinds of grouped loop unrolling (but not the generic LLVM pass),
1111 // but otherwise may require some kind of reordering pass.
1113 // When using fast dependency analysis,
1114 // stop searching after first use:
1117 if (UsesI) continue;
1120 // J does not use I, and comes before the first use of I, so it can be
1121 // merged with I if the instructions are compatible.
1122 int CostSavings, FixedOrder;
1123 if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len,
1124 CostSavings, FixedOrder)) continue;
1126 // J is a candidate for merging with I.
1127 if (!PairableInsts.size() ||
1128 PairableInsts[PairableInsts.size()-1] != I) {
1129 PairableInsts.push_back(I);
1132 CandidatePairs.insert(ValuePair(I, J));
1134 CandidatePairCostSavings.insert(ValuePairWithCost(ValuePair(I, J),
1137 if (FixedOrder == 1)
1138 FixedOrderPairs.insert(ValuePair(I, J));
1139 else if (FixedOrder == -1)
1140 FixedOrderPairs.insert(ValuePair(J, I));
1142 // The next call to this function must start after the last instruction
1143 // selected during this invocation.
1145 Start = llvm::next(J);
1146 IAfterStart = JAfterStart = false;
1149 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair "
1150 << *I << " <-> " << *J << " (cost savings: " <<
1151 CostSavings << ")\n");
1153 // If we have already found too many pairs, break here and this function
1154 // will be called again starting after the last instruction selected
1155 // during this invocation.
1156 if (PairableInsts.size() >= Config.MaxInsts) {
1157 ShouldContinue = true;
1166 DEBUG(dbgs() << "BBV: found " << PairableInsts.size()
1167 << " instructions with candidate pairs\n");
1169 return ShouldContinue;
1172 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that
1173 // it looks for pairs such that both members have an input which is an
1174 // output of PI or PJ.
1175 void BBVectorize::computePairsConnectedTo(
1176 std::multimap<Value *, Value *> &CandidatePairs,
1177 std::vector<Value *> &PairableInsts,
1178 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1179 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1183 // For each possible pairing for this variable, look at the uses of
1184 // the first value...
1185 for (Value::use_iterator I = P.first->use_begin(),
1186 E = P.first->use_end(); I != E; ++I) {
1187 if (isa<LoadInst>(*I)) {
1188 // A pair cannot be connected to a load because the load only takes one
1189 // operand (the address) and it is a scalar even after vectorization.
1191 } else if ((SI = dyn_cast<StoreInst>(*I)) &&
1192 P.first == SI->getPointerOperand()) {
1193 // Similarly, a pair cannot be connected to a store through its
1198 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I);
1200 // For each use of the first variable, look for uses of the second
1202 for (Value::use_iterator J = P.second->use_begin(),
1203 E2 = P.second->use_end(); J != E2; ++J) {
1204 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1205 P.second == SJ->getPointerOperand())
1208 VPIteratorPair JPairRange = CandidatePairs.equal_range(*J);
1211 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) {
1212 VPPair VP(P, ValuePair(*I, *J));
1213 ConnectedPairs.insert(VP);
1214 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionDirect));
1218 if (isSecondInIteratorPair<Value*>(*I, JPairRange)) {
1219 VPPair VP(P, ValuePair(*J, *I));
1220 ConnectedPairs.insert(VP);
1221 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSwap));
1225 if (Config.SplatBreaksChain) continue;
1226 // Look for cases where just the first value in the pair is used by
1227 // both members of another pair (splatting).
1228 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) {
1229 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1230 P.first == SJ->getPointerOperand())
1233 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) {
1234 VPPair VP(P, ValuePair(*I, *J));
1235 ConnectedPairs.insert(VP);
1236 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1241 if (Config.SplatBreaksChain) return;
1242 // Look for cases where just the second value in the pair is used by
1243 // both members of another pair (splatting).
1244 for (Value::use_iterator I = P.second->use_begin(),
1245 E = P.second->use_end(); I != E; ++I) {
1246 if (isa<LoadInst>(*I))
1248 else if ((SI = dyn_cast<StoreInst>(*I)) &&
1249 P.second == SI->getPointerOperand())
1252 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I);
1254 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) {
1255 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1256 P.second == SJ->getPointerOperand())
1259 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) {
1260 VPPair VP(P, ValuePair(*I, *J));
1261 ConnectedPairs.insert(VP);
1262 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1268 // This function figures out which pairs are connected. Two pairs are
1269 // connected if some output of the first pair forms an input to both members
1270 // of the second pair.
1271 void BBVectorize::computeConnectedPairs(
1272 std::multimap<Value *, Value *> &CandidatePairs,
1273 std::vector<Value *> &PairableInsts,
1274 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1275 DenseMap<VPPair, unsigned> &PairConnectionTypes) {
1277 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
1278 PE = PairableInsts.end(); PI != PE; ++PI) {
1279 VPIteratorPair choiceRange = CandidatePairs.equal_range(*PI);
1281 for (std::multimap<Value *, Value *>::iterator P = choiceRange.first;
1282 P != choiceRange.second; ++P)
1283 computePairsConnectedTo(CandidatePairs, PairableInsts,
1284 ConnectedPairs, PairConnectionTypes, *P);
1287 DEBUG(dbgs() << "BBV: found " << ConnectedPairs.size()
1288 << " pair connections.\n");
1291 // This function builds a set of use tuples such that <A, B> is in the set
1292 // if B is in the use tree of A. If B is in the use tree of A, then B
1293 // depends on the output of A.
1294 void BBVectorize::buildDepMap(
1296 std::multimap<Value *, Value *> &CandidatePairs,
1297 std::vector<Value *> &PairableInsts,
1298 DenseSet<ValuePair> &PairableInstUsers) {
1299 DenseSet<Value *> IsInPair;
1300 for (std::multimap<Value *, Value *>::iterator C = CandidatePairs.begin(),
1301 E = CandidatePairs.end(); C != E; ++C) {
1302 IsInPair.insert(C->first);
1303 IsInPair.insert(C->second);
1306 // Iterate through the basic block, recording all Users of each
1307 // pairable instruction.
1309 BasicBlock::iterator E = BB.end();
1310 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) {
1311 if (IsInPair.find(I) == IsInPair.end()) continue;
1313 DenseSet<Value *> Users;
1314 AliasSetTracker WriteSet(*AA);
1315 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J)
1316 (void) trackUsesOfI(Users, WriteSet, I, J);
1318 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end();
1320 PairableInstUsers.insert(ValuePair(I, *U));
1324 // Returns true if an input to pair P is an output of pair Q and also an
1325 // input of pair Q is an output of pair P. If this is the case, then these
1326 // two pairs cannot be simultaneously fused.
1327 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q,
1328 DenseSet<ValuePair> &PairableInstUsers,
1329 std::multimap<ValuePair, ValuePair> *PairableInstUserMap) {
1330 // Two pairs are in conflict if they are mutual Users of eachother.
1331 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) ||
1332 PairableInstUsers.count(ValuePair(P.first, Q.second)) ||
1333 PairableInstUsers.count(ValuePair(P.second, Q.first)) ||
1334 PairableInstUsers.count(ValuePair(P.second, Q.second));
1335 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) ||
1336 PairableInstUsers.count(ValuePair(Q.first, P.second)) ||
1337 PairableInstUsers.count(ValuePair(Q.second, P.first)) ||
1338 PairableInstUsers.count(ValuePair(Q.second, P.second));
1339 if (PairableInstUserMap) {
1340 // FIXME: The expensive part of the cycle check is not so much the cycle
1341 // check itself but this edge insertion procedure. This needs some
1342 // profiling and probably a different data structure (same is true of
1343 // most uses of std::multimap).
1345 VPPIteratorPair QPairRange = PairableInstUserMap->equal_range(Q);
1346 if (!isSecondInIteratorPair(P, QPairRange))
1347 PairableInstUserMap->insert(VPPair(Q, P));
1350 VPPIteratorPair PPairRange = PairableInstUserMap->equal_range(P);
1351 if (!isSecondInIteratorPair(Q, PPairRange))
1352 PairableInstUserMap->insert(VPPair(P, Q));
1356 return (QUsesP && PUsesQ);
1359 // This function walks the use graph of current pairs to see if, starting
1360 // from P, the walk returns to P.
1361 bool BBVectorize::pairWillFormCycle(ValuePair P,
1362 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1363 DenseSet<ValuePair> &CurrentPairs) {
1364 DEBUG(if (DebugCycleCheck)
1365 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> "
1366 << *P.second << "\n");
1367 // A lookup table of visisted pairs is kept because the PairableInstUserMap
1368 // contains non-direct associations.
1369 DenseSet<ValuePair> Visited;
1370 SmallVector<ValuePair, 32> Q;
1371 // General depth-first post-order traversal:
1374 ValuePair QTop = Q.pop_back_val();
1375 Visited.insert(QTop);
1377 DEBUG(if (DebugCycleCheck)
1378 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> "
1379 << *QTop.second << "\n");
1380 VPPIteratorPair QPairRange = PairableInstUserMap.equal_range(QTop);
1381 for (std::multimap<ValuePair, ValuePair>::iterator C = QPairRange.first;
1382 C != QPairRange.second; ++C) {
1383 if (C->second == P) {
1385 << "BBV: rejected to prevent non-trivial cycle formation: "
1386 << *C->first.first << " <-> " << *C->first.second << "\n");
1390 if (CurrentPairs.count(C->second) && !Visited.count(C->second))
1391 Q.push_back(C->second);
1393 } while (!Q.empty());
1398 // This function builds the initial tree of connected pairs with the
1399 // pair J at the root.
1400 void BBVectorize::buildInitialTreeFor(
1401 std::multimap<Value *, Value *> &CandidatePairs,
1402 std::vector<Value *> &PairableInsts,
1403 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1404 DenseSet<ValuePair> &PairableInstUsers,
1405 DenseMap<Value *, Value *> &ChosenPairs,
1406 DenseMap<ValuePair, size_t> &Tree, ValuePair J) {
1407 // Each of these pairs is viewed as the root node of a Tree. The Tree
1408 // is then walked (depth-first). As this happens, we keep track of
1409 // the pairs that compose the Tree and the maximum depth of the Tree.
1410 SmallVector<ValuePairWithDepth, 32> Q;
1411 // General depth-first post-order traversal:
1412 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1414 ValuePairWithDepth QTop = Q.back();
1416 // Push each child onto the queue:
1417 bool MoreChildren = false;
1418 size_t MaxChildDepth = QTop.second;
1419 VPPIteratorPair qtRange = ConnectedPairs.equal_range(QTop.first);
1420 for (std::multimap<ValuePair, ValuePair>::iterator k = qtRange.first;
1421 k != qtRange.second; ++k) {
1422 // Make sure that this child pair is still a candidate:
1423 bool IsStillCand = false;
1424 VPIteratorPair checkRange =
1425 CandidatePairs.equal_range(k->second.first);
1426 for (std::multimap<Value *, Value *>::iterator m = checkRange.first;
1427 m != checkRange.second; ++m) {
1428 if (m->second == k->second.second) {
1435 DenseMap<ValuePair, size_t>::iterator C = Tree.find(k->second);
1436 if (C == Tree.end()) {
1437 size_t d = getDepthFactor(k->second.first);
1438 Q.push_back(ValuePairWithDepth(k->second, QTop.second+d));
1439 MoreChildren = true;
1441 MaxChildDepth = std::max(MaxChildDepth, C->second);
1446 if (!MoreChildren) {
1447 // Record the current pair as part of the Tree:
1448 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth));
1451 } while (!Q.empty());
1454 // Given some initial tree, prune it by removing conflicting pairs (pairs
1455 // that cannot be simultaneously chosen for vectorization).
1456 void BBVectorize::pruneTreeFor(
1457 std::multimap<Value *, Value *> &CandidatePairs,
1458 std::vector<Value *> &PairableInsts,
1459 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1460 DenseSet<ValuePair> &PairableInstUsers,
1461 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1462 DenseMap<Value *, Value *> &ChosenPairs,
1463 DenseMap<ValuePair, size_t> &Tree,
1464 DenseSet<ValuePair> &PrunedTree, ValuePair J,
1465 bool UseCycleCheck) {
1466 SmallVector<ValuePairWithDepth, 32> Q;
1467 // General depth-first post-order traversal:
1468 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1470 ValuePairWithDepth QTop = Q.pop_back_val();
1471 PrunedTree.insert(QTop.first);
1473 // Visit each child, pruning as necessary...
1474 DenseMap<ValuePair, size_t> BestChildren;
1475 VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first);
1476 for (std::multimap<ValuePair, ValuePair>::iterator K = QTopRange.first;
1477 K != QTopRange.second; ++K) {
1478 DenseMap<ValuePair, size_t>::iterator C = Tree.find(K->second);
1479 if (C == Tree.end()) continue;
1481 // This child is in the Tree, now we need to make sure it is the
1482 // best of any conflicting children. There could be multiple
1483 // conflicting children, so first, determine if we're keeping
1484 // this child, then delete conflicting children as necessary.
1486 // It is also necessary to guard against pairing-induced
1487 // dependencies. Consider instructions a .. x .. y .. b
1488 // such that (a,b) are to be fused and (x,y) are to be fused
1489 // but a is an input to x and b is an output from y. This
1490 // means that y cannot be moved after b but x must be moved
1491 // after b for (a,b) to be fused. In other words, after
1492 // fusing (a,b) we have y .. a/b .. x where y is an input
1493 // to a/b and x is an output to a/b: x and y can no longer
1494 // be legally fused. To prevent this condition, we must
1495 // make sure that a child pair added to the Tree is not
1496 // both an input and output of an already-selected pair.
1498 // Pairing-induced dependencies can also form from more complicated
1499 // cycles. The pair vs. pair conflicts are easy to check, and so
1500 // that is done explicitly for "fast rejection", and because for
1501 // child vs. child conflicts, we may prefer to keep the current
1502 // pair in preference to the already-selected child.
1503 DenseSet<ValuePair> CurrentPairs;
1506 for (DenseMap<ValuePair, size_t>::iterator C2
1507 = BestChildren.begin(), E2 = BestChildren.end();
1509 if (C2->first.first == C->first.first ||
1510 C2->first.first == C->first.second ||
1511 C2->first.second == C->first.first ||
1512 C2->first.second == C->first.second ||
1513 pairsConflict(C2->first, C->first, PairableInstUsers,
1514 UseCycleCheck ? &PairableInstUserMap : 0)) {
1515 if (C2->second >= C->second) {
1520 CurrentPairs.insert(C2->first);
1523 if (!CanAdd) continue;
1525 // Even worse, this child could conflict with another node already
1526 // selected for the Tree. If that is the case, ignore this child.
1527 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(),
1528 E2 = PrunedTree.end(); T != E2; ++T) {
1529 if (T->first == C->first.first ||
1530 T->first == C->first.second ||
1531 T->second == C->first.first ||
1532 T->second == C->first.second ||
1533 pairsConflict(*T, C->first, PairableInstUsers,
1534 UseCycleCheck ? &PairableInstUserMap : 0)) {
1539 CurrentPairs.insert(*T);
1541 if (!CanAdd) continue;
1543 // And check the queue too...
1544 for (SmallVector<ValuePairWithDepth, 32>::iterator C2 = Q.begin(),
1545 E2 = Q.end(); C2 != E2; ++C2) {
1546 if (C2->first.first == C->first.first ||
1547 C2->first.first == C->first.second ||
1548 C2->first.second == C->first.first ||
1549 C2->first.second == C->first.second ||
1550 pairsConflict(C2->first, C->first, PairableInstUsers,
1551 UseCycleCheck ? &PairableInstUserMap : 0)) {
1556 CurrentPairs.insert(C2->first);
1558 if (!CanAdd) continue;
1560 // Last but not least, check for a conflict with any of the
1561 // already-chosen pairs.
1562 for (DenseMap<Value *, Value *>::iterator C2 =
1563 ChosenPairs.begin(), E2 = ChosenPairs.end();
1565 if (pairsConflict(*C2, C->first, PairableInstUsers,
1566 UseCycleCheck ? &PairableInstUserMap : 0)) {
1571 CurrentPairs.insert(*C2);
1573 if (!CanAdd) continue;
1575 // To check for non-trivial cycles formed by the addition of the
1576 // current pair we've formed a list of all relevant pairs, now use a
1577 // graph walk to check for a cycle. We start from the current pair and
1578 // walk the use tree to see if we again reach the current pair. If we
1579 // do, then the current pair is rejected.
1581 // FIXME: It may be more efficient to use a topological-ordering
1582 // algorithm to improve the cycle check. This should be investigated.
1583 if (UseCycleCheck &&
1584 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs))
1587 // This child can be added, but we may have chosen it in preference
1588 // to an already-selected child. Check for this here, and if a
1589 // conflict is found, then remove the previously-selected child
1590 // before adding this one in its place.
1591 for (DenseMap<ValuePair, size_t>::iterator C2
1592 = BestChildren.begin(); C2 != BestChildren.end();) {
1593 if (C2->first.first == C->first.first ||
1594 C2->first.first == C->first.second ||
1595 C2->first.second == C->first.first ||
1596 C2->first.second == C->first.second ||
1597 pairsConflict(C2->first, C->first, PairableInstUsers))
1598 BestChildren.erase(C2++);
1603 BestChildren.insert(ValuePairWithDepth(C->first, C->second));
1606 for (DenseMap<ValuePair, size_t>::iterator C
1607 = BestChildren.begin(), E2 = BestChildren.end();
1609 size_t DepthF = getDepthFactor(C->first.first);
1610 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF));
1612 } while (!Q.empty());
1615 // This function finds the best tree of mututally-compatible connected
1616 // pairs, given the choice of root pairs as an iterator range.
1617 void BBVectorize::findBestTreeFor(
1618 std::multimap<Value *, Value *> &CandidatePairs,
1619 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1620 std::vector<Value *> &PairableInsts,
1621 DenseSet<ValuePair> &FixedOrderPairs,
1622 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1623 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1624 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps,
1625 DenseSet<ValuePair> &PairableInstUsers,
1626 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1627 DenseMap<Value *, Value *> &ChosenPairs,
1628 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth,
1629 int &BestEffSize, VPIteratorPair ChoiceRange,
1630 bool UseCycleCheck) {
1631 for (std::multimap<Value *, Value *>::iterator J = ChoiceRange.first;
1632 J != ChoiceRange.second; ++J) {
1634 // Before going any further, make sure that this pair does not
1635 // conflict with any already-selected pairs (see comment below
1636 // near the Tree pruning for more details).
1637 DenseSet<ValuePair> ChosenPairSet;
1638 bool DoesConflict = false;
1639 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(),
1640 E = ChosenPairs.end(); C != E; ++C) {
1641 if (pairsConflict(*C, *J, PairableInstUsers,
1642 UseCycleCheck ? &PairableInstUserMap : 0)) {
1643 DoesConflict = true;
1647 ChosenPairSet.insert(*C);
1649 if (DoesConflict) continue;
1651 if (UseCycleCheck &&
1652 pairWillFormCycle(*J, PairableInstUserMap, ChosenPairSet))
1655 DenseMap<ValuePair, size_t> Tree;
1656 buildInitialTreeFor(CandidatePairs, PairableInsts, ConnectedPairs,
1657 PairableInstUsers, ChosenPairs, Tree, *J);
1659 // Because we'll keep the child with the largest depth, the largest
1660 // depth is still the same in the unpruned Tree.
1661 size_t MaxDepth = Tree.lookup(*J);
1663 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found Tree for pair {"
1664 << *J->first << " <-> " << *J->second << "} of depth " <<
1665 MaxDepth << " and size " << Tree.size() << "\n");
1667 // At this point the Tree has been constructed, but, may contain
1668 // contradictory children (meaning that different children of
1669 // some tree node may be attempting to fuse the same instruction).
1670 // So now we walk the tree again, in the case of a conflict,
1671 // keep only the child with the largest depth. To break a tie,
1672 // favor the first child.
1674 DenseSet<ValuePair> PrunedTree;
1675 pruneTreeFor(CandidatePairs, PairableInsts, ConnectedPairs,
1676 PairableInstUsers, PairableInstUserMap, ChosenPairs, Tree,
1677 PrunedTree, *J, UseCycleCheck);
1681 DenseSet<Value *> PrunedTreeInstrs;
1682 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
1683 E = PrunedTree.end(); S != E; ++S) {
1684 PrunedTreeInstrs.insert(S->first);
1685 PrunedTreeInstrs.insert(S->second);
1688 // The set of pairs that have already contributed to the total cost.
1689 DenseSet<ValuePair> IncomingPairs;
1691 // The node weights represent the cost savings associated with
1692 // fusing the pair of instructions.
1693 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
1694 E = PrunedTree.end(); S != E; ++S) {
1695 bool FlipOrder = false;
1697 if (getDepthFactor(S->first)) {
1698 int ESContrib = CandidatePairCostSavings.find(*S)->second;
1699 DEBUG(if (DebugPairSelection) dbgs() << "\tweight {"
1700 << *S->first << " <-> " << *S->second << "} = " <<
1702 EffSize += ESContrib;
1705 // The edge weights contribute in a negative sense: they represent
1706 // the cost of shuffles.
1707 VPPIteratorPair IP = ConnectedPairDeps.equal_range(*S);
1708 if (IP.first != ConnectedPairDeps.end()) {
1709 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
1710 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first;
1711 Q != IP.second; ++Q) {
1712 if (!PrunedTree.count(Q->second))
1714 DenseMap<VPPair, unsigned>::iterator R =
1715 PairConnectionTypes.find(VPPair(Q->second, Q->first));
1716 assert(R != PairConnectionTypes.end() &&
1717 "Cannot find pair connection type");
1718 if (R->second == PairConnectionDirect)
1720 else if (R->second == PairConnectionSwap)
1724 // If there are more swaps than direct connections, then
1725 // the pair order will be flipped during fusion. So the real
1726 // number of swaps is the minimum number.
1727 FlipOrder = !FixedOrderPairs.count(*S) &&
1728 ((NumDepsSwap > NumDepsDirect) ||
1729 FixedOrderPairs.count(ValuePair(S->second, S->first)));
1731 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first;
1732 Q != IP.second; ++Q) {
1733 if (!PrunedTree.count(Q->second))
1735 DenseMap<VPPair, unsigned>::iterator R =
1736 PairConnectionTypes.find(VPPair(Q->second, Q->first));
1737 assert(R != PairConnectionTypes.end() &&
1738 "Cannot find pair connection type");
1739 Type *Ty1 = Q->second.first->getType(),
1740 *Ty2 = Q->second.second->getType();
1741 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1742 if ((R->second == PairConnectionDirect && FlipOrder) ||
1743 (R->second == PairConnectionSwap && !FlipOrder) ||
1744 R->second == PairConnectionSplat) {
1745 int ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1747 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1748 *Q->second.first << " <-> " << *Q->second.second <<
1750 *S->first << " <-> " << *S->second << "} = " <<
1752 EffSize -= ESContrib;
1757 // Compute the cost of outgoing edges. We assume that edges outgoing
1758 // to shuffles, inserts or extracts can be merged, and so contribute
1759 // no additional cost.
1760 if (!S->first->getType()->isVoidTy()) {
1761 Type *Ty1 = S->first->getType(),
1762 *Ty2 = S->second->getType();
1763 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1765 bool NeedsExtraction = false;
1766 for (Value::use_iterator I = S->first->use_begin(),
1767 IE = S->first->use_end(); I != IE; ++I) {
1768 if (isa<ShuffleVectorInst>(*I) ||
1769 isa<InsertElementInst>(*I) ||
1770 isa<ExtractElementInst>(*I))
1772 if (PrunedTreeInstrs.count(*I))
1774 NeedsExtraction = true;
1778 if (NeedsExtraction) {
1780 if (Ty1->isVectorTy())
1781 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1784 ESContrib = (int) VTTI->getVectorInstrCost(
1785 Instruction::ExtractElement, VTy, 0);
1787 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1788 *S->first << "} = " << ESContrib << "\n");
1789 EffSize -= ESContrib;
1792 NeedsExtraction = false;
1793 for (Value::use_iterator I = S->second->use_begin(),
1794 IE = S->second->use_end(); I != IE; ++I) {
1795 if (isa<ShuffleVectorInst>(*I) ||
1796 isa<InsertElementInst>(*I) ||
1797 isa<ExtractElementInst>(*I))
1799 if (PrunedTreeInstrs.count(*I))
1801 NeedsExtraction = true;
1805 if (NeedsExtraction) {
1807 if (Ty2->isVectorTy())
1808 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1811 ESContrib = (int) VTTI->getVectorInstrCost(
1812 Instruction::ExtractElement, VTy, 1);
1813 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1814 *S->second << "} = " << ESContrib << "\n");
1815 EffSize -= ESContrib;
1819 // Compute the cost of incoming edges.
1820 if (!isa<LoadInst>(S->first) && !isa<StoreInst>(S->first)) {
1821 Instruction *S1 = cast<Instruction>(S->first),
1822 *S2 = cast<Instruction>(S->second);
1823 for (unsigned o = 0; o < S1->getNumOperands(); ++o) {
1824 Value *O1 = S1->getOperand(o), *O2 = S2->getOperand(o);
1826 // Combining constants into vector constants (or small vector
1827 // constants into larger ones are assumed free).
1828 if (isa<Constant>(O1) && isa<Constant>(O2))
1834 ValuePair VP = ValuePair(O1, O2);
1835 ValuePair VPR = ValuePair(O2, O1);
1837 // Internal edges are not handled here.
1838 if (PrunedTree.count(VP) || PrunedTree.count(VPR))
1841 Type *Ty1 = O1->getType(),
1842 *Ty2 = O2->getType();
1843 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1845 // Combining vector operations of the same type is also assumed
1846 // folded with other operations.
1848 (isa<ShuffleVectorInst>(O1) ||
1849 isa<InsertElementInst>(O1) ||
1850 isa<InsertElementInst>(O1)) &&
1851 (isa<ShuffleVectorInst>(O2) ||
1852 isa<InsertElementInst>(O2) ||
1853 isa<InsertElementInst>(O2)))
1857 // This pair has already been formed.
1858 if (IncomingPairs.count(VP)) {
1860 } else if (IncomingPairs.count(VPR)) {
1861 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1863 } else if (!Ty1->isVectorTy() && !Ty2->isVectorTy()) {
1864 ESContrib = (int) VTTI->getVectorInstrCost(
1865 Instruction::InsertElement, VTy, 0);
1866 ESContrib += (int) VTTI->getVectorInstrCost(
1867 Instruction::InsertElement, VTy, 1);
1868 } else if (!Ty1->isVectorTy()) {
1869 // O1 needs to be inserted into a vector of size O2, and then
1870 // both need to be shuffled together.
1871 ESContrib = (int) VTTI->getVectorInstrCost(
1872 Instruction::InsertElement, Ty2, 0);
1873 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
1875 } else if (!Ty2->isVectorTy()) {
1876 // O2 needs to be inserted into a vector of size O1, and then
1877 // both need to be shuffled together.
1878 ESContrib = (int) VTTI->getVectorInstrCost(
1879 Instruction::InsertElement, Ty1, 0);
1880 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
1883 Type *TyBig = Ty1, *TySmall = Ty2;
1884 if (Ty2->getVectorNumElements() > Ty1->getVectorNumElements())
1885 std::swap(TyBig, TySmall);
1887 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1889 if (TyBig != TySmall)
1890 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
1894 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {"
1895 << *O1 << " <-> " << *O2 << "} = " <<
1897 EffSize -= ESContrib;
1898 IncomingPairs.insert(VP);
1903 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
1904 E = PrunedTree.end(); S != E; ++S)
1905 EffSize += (int) getDepthFactor(S->first);
1908 DEBUG(if (DebugPairSelection)
1909 dbgs() << "BBV: found pruned Tree for pair {"
1910 << *J->first << " <-> " << *J->second << "} of depth " <<
1911 MaxDepth << " and size " << PrunedTree.size() <<
1912 " (effective size: " << EffSize << ")\n");
1913 if (((VTTI && !UseChainDepthWithTI) ||
1914 MaxDepth >= Config.ReqChainDepth) &&
1915 EffSize > 0 && EffSize > BestEffSize) {
1916 BestMaxDepth = MaxDepth;
1917 BestEffSize = EffSize;
1918 BestTree = PrunedTree;
1923 // Given the list of candidate pairs, this function selects those
1924 // that will be fused into vector instructions.
1925 void BBVectorize::choosePairs(
1926 std::multimap<Value *, Value *> &CandidatePairs,
1927 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1928 std::vector<Value *> &PairableInsts,
1929 DenseSet<ValuePair> &FixedOrderPairs,
1930 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1931 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1932 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps,
1933 DenseSet<ValuePair> &PairableInstUsers,
1934 DenseMap<Value *, Value *>& ChosenPairs) {
1935 bool UseCycleCheck =
1936 CandidatePairs.size() <= Config.MaxCandPairsForCycleCheck;
1937 std::multimap<ValuePair, ValuePair> PairableInstUserMap;
1938 for (std::vector<Value *>::iterator I = PairableInsts.begin(),
1939 E = PairableInsts.end(); I != E; ++I) {
1940 // The number of possible pairings for this variable:
1941 size_t NumChoices = CandidatePairs.count(*I);
1942 if (!NumChoices) continue;
1944 VPIteratorPair ChoiceRange = CandidatePairs.equal_range(*I);
1946 // The best pair to choose and its tree:
1947 size_t BestMaxDepth = 0;
1948 int BestEffSize = 0;
1949 DenseSet<ValuePair> BestTree;
1950 findBestTreeFor(CandidatePairs, CandidatePairCostSavings,
1951 PairableInsts, FixedOrderPairs, PairConnectionTypes,
1952 ConnectedPairs, ConnectedPairDeps,
1953 PairableInstUsers, PairableInstUserMap, ChosenPairs,
1954 BestTree, BestMaxDepth, BestEffSize, ChoiceRange,
1957 // A tree has been chosen (or not) at this point. If no tree was
1958 // chosen, then this instruction, I, cannot be paired (and is no longer
1961 DEBUG(if (BestTree.size() > 0)
1962 dbgs() << "BBV: selected pairs in the best tree for: "
1963 << *cast<Instruction>(*I) << "\n");
1965 for (DenseSet<ValuePair>::iterator S = BestTree.begin(),
1966 SE2 = BestTree.end(); S != SE2; ++S) {
1967 // Insert the members of this tree into the list of chosen pairs.
1968 ChosenPairs.insert(ValuePair(S->first, S->second));
1969 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " <<
1970 *S->second << "\n");
1972 // Remove all candidate pairs that have values in the chosen tree.
1973 for (std::multimap<Value *, Value *>::iterator K =
1974 CandidatePairs.begin(); K != CandidatePairs.end();) {
1975 if (K->first == S->first || K->second == S->first ||
1976 K->second == S->second || K->first == S->second) {
1977 // Don't remove the actual pair chosen so that it can be used
1978 // in subsequent tree selections.
1979 if (!(K->first == S->first && K->second == S->second))
1980 CandidatePairs.erase(K++);
1990 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n");
1993 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o,
1998 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) +
1999 (n > 0 ? "." + utostr(n) : "")).str();
2002 // Returns the value that is to be used as the pointer input to the vector
2003 // instruction that fuses I with J.
2004 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context,
2005 Instruction *I, Instruction *J, unsigned o) {
2007 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
2008 int64_t OffsetInElmts;
2010 // Note: the analysis might fail here, that is why the pair order has
2011 // been precomputed (OffsetInElmts must be unused here).
2012 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
2013 IAddressSpace, JAddressSpace,
2014 OffsetInElmts, false);
2016 // The pointer value is taken to be the one with the lowest offset.
2019 Type *ArgTypeI = cast<PointerType>(IPtr->getType())->getElementType();
2020 Type *ArgTypeJ = cast<PointerType>(JPtr->getType())->getElementType();
2021 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2022 Type *VArgPtrType = PointerType::get(VArgType,
2023 cast<PointerType>(IPtr->getType())->getAddressSpace());
2024 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o),
2025 /* insert before */ I);
2028 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J,
2029 unsigned MaskOffset, unsigned NumInElem,
2030 unsigned NumInElem1, unsigned IdxOffset,
2031 std::vector<Constant*> &Mask) {
2032 unsigned NumElem1 = cast<VectorType>(J->getType())->getNumElements();
2033 for (unsigned v = 0; v < NumElem1; ++v) {
2034 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v);
2036 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context));
2038 unsigned mm = m + (int) IdxOffset;
2039 if (m >= (int) NumInElem1)
2040 mm += (int) NumInElem;
2042 Mask[v+MaskOffset] =
2043 ConstantInt::get(Type::getInt32Ty(Context), mm);
2048 // Returns the value that is to be used as the vector-shuffle mask to the
2049 // vector instruction that fuses I with J.
2050 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context,
2051 Instruction *I, Instruction *J) {
2052 // This is the shuffle mask. We need to append the second
2053 // mask to the first, and the numbers need to be adjusted.
2055 Type *ArgTypeI = I->getType();
2056 Type *ArgTypeJ = J->getType();
2057 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2059 unsigned NumElemI = cast<VectorType>(ArgTypeI)->getNumElements();
2061 // Get the total number of elements in the fused vector type.
2062 // By definition, this must equal the number of elements in
2064 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements();
2065 std::vector<Constant*> Mask(NumElem);
2067 Type *OpTypeI = I->getOperand(0)->getType();
2068 unsigned NumInElemI = cast<VectorType>(OpTypeI)->getNumElements();
2069 Type *OpTypeJ = J->getOperand(0)->getType();
2070 unsigned NumInElemJ = cast<VectorType>(OpTypeJ)->getNumElements();
2072 // The fused vector will be:
2073 // -----------------------------------------------------
2074 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ |
2075 // -----------------------------------------------------
2076 // from which we'll extract NumElem total elements (where the first NumElemI
2077 // of them come from the mask in I and the remainder come from the mask
2080 // For the mask from the first pair...
2081 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI,
2084 // For the mask from the second pair...
2085 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ,
2088 return ConstantVector::get(Mask);
2091 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I,
2092 Instruction *J, unsigned o, Value *&LOp,
2094 Type *ArgTypeL, Type *ArgTypeH,
2095 bool IBeforeJ, unsigned IdxOff) {
2096 bool ExpandedIEChain = false;
2097 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) {
2098 // If we have a pure insertelement chain, then this can be rewritten
2099 // into a chain that directly builds the larger type.
2100 bool PureChain = true;
2101 InsertElementInst *LIENext = LIE;
2103 if (!isa<UndefValue>(LIENext->getOperand(0)) &&
2104 !isa<InsertElementInst>(LIENext->getOperand(0))) {
2109 dyn_cast<InsertElementInst>(LIENext->getOperand(0))));
2112 SmallVector<Value *, 8> VectElemts(numElemL,
2113 UndefValue::get(ArgTypeL->getScalarType()));
2114 InsertElementInst *LIENext = LIE;
2117 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue();
2118 VectElemts[Idx] = LIENext->getOperand(1);
2120 dyn_cast<InsertElementInst>(LIENext->getOperand(0))));
2123 Value *LIEPrev = UndefValue::get(ArgTypeH);
2124 for (unsigned i = 0; i < numElemL; ++i) {
2125 if (isa<UndefValue>(VectElemts[i])) continue;
2126 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i],
2127 ConstantInt::get(Type::getInt32Ty(Context),
2129 getReplacementName(IBeforeJ ? I : J,
2131 LIENext->insertBefore(IBeforeJ ? J : I);
2135 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH);
2136 ExpandedIEChain = true;
2140 return ExpandedIEChain;
2143 // Returns the value to be used as the specified operand of the vector
2144 // instruction that fuses I with J.
2145 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I,
2146 Instruction *J, unsigned o, bool IBeforeJ) {
2147 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2148 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1);
2150 // Compute the fused vector type for this operand
2151 Type *ArgTypeI = I->getOperand(o)->getType();
2152 Type *ArgTypeJ = J->getOperand(o)->getType();
2153 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2155 Instruction *L = I, *H = J;
2156 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ;
2159 if (ArgTypeL->isVectorTy())
2160 numElemL = cast<VectorType>(ArgTypeL)->getNumElements();
2165 if (ArgTypeH->isVectorTy())
2166 numElemH = cast<VectorType>(ArgTypeH)->getNumElements();
2170 Value *LOp = L->getOperand(o);
2171 Value *HOp = H->getOperand(o);
2172 unsigned numElem = VArgType->getNumElements();
2174 // First, we check if we can reuse the "original" vector outputs (if these
2175 // exist). We might need a shuffle.
2176 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp);
2177 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp);
2178 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp);
2179 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp);
2181 // FIXME: If we're fusing shuffle instructions, then we can't apply this
2182 // optimization. The input vectors to the shuffle might be a different
2183 // length from the shuffle outputs. Unfortunately, the replacement
2184 // shuffle mask has already been formed, and the mask entries are sensitive
2185 // to the sizes of the inputs.
2186 bool IsSizeChangeShuffle =
2187 isa<ShuffleVectorInst>(L) &&
2188 (LOp->getType() != L->getType() || HOp->getType() != H->getType());
2190 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) {
2191 // We can have at most two unique vector inputs.
2192 bool CanUseInputs = true;
2195 I1 = LEE->getOperand(0);
2197 I1 = LSV->getOperand(0);
2198 I2 = LSV->getOperand(1);
2199 if (I2 == I1 || isa<UndefValue>(I2))
2204 Value *I3 = HEE->getOperand(0);
2205 if (!I2 && I3 != I1)
2207 else if (I3 != I1 && I3 != I2)
2208 CanUseInputs = false;
2210 Value *I3 = HSV->getOperand(0);
2211 if (!I2 && I3 != I1)
2213 else if (I3 != I1 && I3 != I2)
2214 CanUseInputs = false;
2217 Value *I4 = HSV->getOperand(1);
2218 if (!isa<UndefValue>(I4)) {
2219 if (!I2 && I4 != I1)
2221 else if (I4 != I1 && I4 != I2)
2222 CanUseInputs = false;
2229 cast<VectorType>(cast<Instruction>(LOp)->getOperand(0)->getType())
2232 cast<VectorType>(cast<Instruction>(HOp)->getOperand(0)->getType())
2235 // We have one or two input vectors. We need to map each index of the
2236 // operands to the index of the original vector.
2237 SmallVector<std::pair<int, int>, 8> II(numElem);
2238 for (unsigned i = 0; i < numElemL; ++i) {
2242 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue();
2243 INum = LEE->getOperand(0) == I1 ? 0 : 1;
2245 Idx = LSV->getMaskValue(i);
2246 if (Idx < (int) LOpElem) {
2247 INum = LSV->getOperand(0) == I1 ? 0 : 1;
2250 INum = LSV->getOperand(1) == I1 ? 0 : 1;
2254 II[i] = std::pair<int, int>(Idx, INum);
2256 for (unsigned i = 0; i < numElemH; ++i) {
2260 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue();
2261 INum = HEE->getOperand(0) == I1 ? 0 : 1;
2263 Idx = HSV->getMaskValue(i);
2264 if (Idx < (int) HOpElem) {
2265 INum = HSV->getOperand(0) == I1 ? 0 : 1;
2268 INum = HSV->getOperand(1) == I1 ? 0 : 1;
2272 II[i + numElemL] = std::pair<int, int>(Idx, INum);
2275 // We now have an array which tells us from which index of which
2276 // input vector each element of the operand comes.
2277 VectorType *I1T = cast<VectorType>(I1->getType());
2278 unsigned I1Elem = I1T->getNumElements();
2281 // In this case there is only one underlying vector input. Check for
2282 // the trivial case where we can use the input directly.
2283 if (I1Elem == numElem) {
2284 bool ElemInOrder = true;
2285 for (unsigned i = 0; i < numElem; ++i) {
2286 if (II[i].first != (int) i && II[i].first != -1) {
2287 ElemInOrder = false;
2296 // A shuffle is needed.
2297 std::vector<Constant *> Mask(numElem);
2298 for (unsigned i = 0; i < numElem; ++i) {
2299 int Idx = II[i].first;
2301 Mask[i] = UndefValue::get(Type::getInt32Ty(Context));
2303 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2307 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2308 ConstantVector::get(Mask),
2309 getReplacementName(IBeforeJ ? I : J,
2311 S->insertBefore(IBeforeJ ? J : I);
2315 VectorType *I2T = cast<VectorType>(I2->getType());
2316 unsigned I2Elem = I2T->getNumElements();
2318 // This input comes from two distinct vectors. The first step is to
2319 // make sure that both vectors are the same length. If not, the
2320 // smaller one will need to grow before they can be shuffled together.
2321 if (I1Elem < I2Elem) {
2322 std::vector<Constant *> Mask(I2Elem);
2324 for (; v < I1Elem; ++v)
2325 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2326 for (; v < I2Elem; ++v)
2327 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2329 Instruction *NewI1 =
2330 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2331 ConstantVector::get(Mask),
2332 getReplacementName(IBeforeJ ? I : J,
2334 NewI1->insertBefore(IBeforeJ ? J : I);
2338 } else if (I1Elem > I2Elem) {
2339 std::vector<Constant *> Mask(I1Elem);
2341 for (; v < I2Elem; ++v)
2342 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2343 for (; v < I1Elem; ++v)
2344 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2346 Instruction *NewI2 =
2347 new ShuffleVectorInst(I2, UndefValue::get(I2T),
2348 ConstantVector::get(Mask),
2349 getReplacementName(IBeforeJ ? I : J,
2351 NewI2->insertBefore(IBeforeJ ? J : I);
2357 // Now that both I1 and I2 are the same length we can shuffle them
2358 // together (and use the result).
2359 std::vector<Constant *> Mask(numElem);
2360 for (unsigned v = 0; v < numElem; ++v) {
2361 if (II[v].first == -1) {
2362 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2364 int Idx = II[v].first + II[v].second * I1Elem;
2365 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2369 Instruction *NewOp =
2370 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask),
2371 getReplacementName(IBeforeJ ? I : J, true, o));
2372 NewOp->insertBefore(IBeforeJ ? J : I);
2377 Type *ArgType = ArgTypeL;
2378 if (numElemL < numElemH) {
2379 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH,
2380 ArgTypeL, VArgType, IBeforeJ, 1)) {
2381 // This is another short-circuit case: we're combining a scalar into
2382 // a vector that is formed by an IE chain. We've just expanded the IE
2383 // chain, now insert the scalar and we're done.
2385 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0,
2386 getReplacementName(IBeforeJ ? I : J, true, o));
2387 S->insertBefore(IBeforeJ ? J : I);
2389 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL,
2390 ArgTypeH, IBeforeJ)) {
2391 // The two vector inputs to the shuffle must be the same length,
2392 // so extend the smaller vector to be the same length as the larger one.
2396 std::vector<Constant *> Mask(numElemH);
2398 for (; v < numElemL; ++v)
2399 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2400 for (; v < numElemH; ++v)
2401 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2403 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL),
2404 ConstantVector::get(Mask),
2405 getReplacementName(IBeforeJ ? I : J,
2408 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0,
2409 getReplacementName(IBeforeJ ? I : J,
2413 NLOp->insertBefore(IBeforeJ ? J : I);
2418 } else if (numElemL > numElemH) {
2419 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL,
2420 ArgTypeH, VArgType, IBeforeJ)) {
2422 InsertElementInst::Create(LOp, HOp,
2423 ConstantInt::get(Type::getInt32Ty(Context),
2425 getReplacementName(IBeforeJ ? I : J,
2427 S->insertBefore(IBeforeJ ? J : I);
2429 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH,
2430 ArgTypeL, IBeforeJ)) {
2433 std::vector<Constant *> Mask(numElemL);
2435 for (; v < numElemH; ++v)
2436 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2437 for (; v < numElemL; ++v)
2438 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2440 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH),
2441 ConstantVector::get(Mask),
2442 getReplacementName(IBeforeJ ? I : J,
2445 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0,
2446 getReplacementName(IBeforeJ ? I : J,
2450 NHOp->insertBefore(IBeforeJ ? J : I);
2455 if (ArgType->isVectorTy()) {
2456 unsigned numElem = cast<VectorType>(VArgType)->getNumElements();
2457 std::vector<Constant*> Mask(numElem);
2458 for (unsigned v = 0; v < numElem; ++v) {
2460 // If the low vector was expanded, we need to skip the extra
2461 // undefined entries.
2462 if (v >= numElemL && numElemH > numElemL)
2463 Idx += (numElemH - numElemL);
2464 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2467 Instruction *BV = new ShuffleVectorInst(LOp, HOp,
2468 ConstantVector::get(Mask),
2469 getReplacementName(IBeforeJ ? I : J, true, o));
2470 BV->insertBefore(IBeforeJ ? J : I);
2474 Instruction *BV1 = InsertElementInst::Create(
2475 UndefValue::get(VArgType), LOp, CV0,
2476 getReplacementName(IBeforeJ ? I : J,
2478 BV1->insertBefore(IBeforeJ ? J : I);
2479 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1,
2480 getReplacementName(IBeforeJ ? I : J,
2482 BV2->insertBefore(IBeforeJ ? J : I);
2486 // This function creates an array of values that will be used as the inputs
2487 // to the vector instruction that fuses I with J.
2488 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context,
2489 Instruction *I, Instruction *J,
2490 SmallVector<Value *, 3> &ReplacedOperands,
2492 unsigned NumOperands = I->getNumOperands();
2494 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) {
2495 // Iterate backward so that we look at the store pointer
2496 // first and know whether or not we need to flip the inputs.
2498 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) {
2499 // This is the pointer for a load/store instruction.
2500 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o);
2502 } else if (isa<CallInst>(I)) {
2503 Function *F = cast<CallInst>(I)->getCalledFunction();
2504 unsigned IID = F->getIntrinsicID();
2505 if (o == NumOperands-1) {
2506 BasicBlock &BB = *I->getParent();
2508 Module *M = BB.getParent()->getParent();
2509 Type *ArgTypeI = I->getType();
2510 Type *ArgTypeJ = J->getType();
2511 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2513 ReplacedOperands[o] = Intrinsic::getDeclaration(M,
2514 (Intrinsic::ID) IID, VArgType);
2516 } else if (IID == Intrinsic::powi && o == 1) {
2517 // The second argument of powi is a single integer and we've already
2518 // checked that both arguments are equal. As a result, we just keep
2519 // I's second argument.
2520 ReplacedOperands[o] = I->getOperand(o);
2523 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) {
2524 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J);
2528 ReplacedOperands[o] = getReplacementInput(Context, I, J, o, IBeforeJ);
2532 // This function creates two values that represent the outputs of the
2533 // original I and J instructions. These are generally vector shuffles
2534 // or extracts. In many cases, these will end up being unused and, thus,
2535 // eliminated by later passes.
2536 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
2537 Instruction *J, Instruction *K,
2538 Instruction *&InsertionPt,
2539 Instruction *&K1, Instruction *&K2) {
2540 if (isa<StoreInst>(I)) {
2541 AA->replaceWithNewValue(I, K);
2542 AA->replaceWithNewValue(J, K);
2544 Type *IType = I->getType();
2545 Type *JType = J->getType();
2547 VectorType *VType = getVecTypeForPair(IType, JType);
2548 unsigned numElem = VType->getNumElements();
2550 unsigned numElemI, numElemJ;
2551 if (IType->isVectorTy())
2552 numElemI = cast<VectorType>(IType)->getNumElements();
2556 if (JType->isVectorTy())
2557 numElemJ = cast<VectorType>(JType)->getNumElements();
2561 if (IType->isVectorTy()) {
2562 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI);
2563 for (unsigned v = 0; v < numElemI; ++v) {
2564 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2565 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v);
2568 K1 = new ShuffleVectorInst(K, UndefValue::get(VType),
2569 ConstantVector::get( Mask1),
2570 getReplacementName(K, false, 1));
2572 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2573 K1 = ExtractElementInst::Create(K, CV0,
2574 getReplacementName(K, false, 1));
2577 if (JType->isVectorTy()) {
2578 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ);
2579 for (unsigned v = 0; v < numElemJ; ++v) {
2580 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2581 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v);
2584 K2 = new ShuffleVectorInst(K, UndefValue::get(VType),
2585 ConstantVector::get( Mask2),
2586 getReplacementName(K, false, 2));
2588 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1);
2589 K2 = ExtractElementInst::Create(K, CV1,
2590 getReplacementName(K, false, 2));
2594 K2->insertAfter(K1);
2599 // Move all uses of the function I (including pairing-induced uses) after J.
2600 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB,
2601 std::multimap<Value *, Value *> &LoadMoveSet,
2602 Instruction *I, Instruction *J) {
2603 // Skip to the first instruction past I.
2604 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2606 DenseSet<Value *> Users;
2607 AliasSetTracker WriteSet(*AA);
2608 for (; cast<Instruction>(L) != J; ++L)
2609 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet);
2611 assert(cast<Instruction>(L) == J &&
2612 "Tracking has not proceeded far enough to check for dependencies");
2613 // If J is now in the use set of I, then trackUsesOfI will return true
2614 // and we have a dependency cycle (and the fusing operation must abort).
2615 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSet);
2618 // Move all uses of the function I (including pairing-induced uses) after J.
2619 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB,
2620 std::multimap<Value *, Value *> &LoadMoveSet,
2621 Instruction *&InsertionPt,
2622 Instruction *I, Instruction *J) {
2623 // Skip to the first instruction past I.
2624 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2626 DenseSet<Value *> Users;
2627 AliasSetTracker WriteSet(*AA);
2628 for (; cast<Instruction>(L) != J;) {
2629 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet)) {
2630 // Move this instruction
2631 Instruction *InstToMove = L; ++L;
2633 DEBUG(dbgs() << "BBV: moving: " << *InstToMove <<
2634 " to after " << *InsertionPt << "\n");
2635 InstToMove->removeFromParent();
2636 InstToMove->insertAfter(InsertionPt);
2637 InsertionPt = InstToMove;
2644 // Collect all load instruction that are in the move set of a given first
2645 // pair member. These loads depend on the first instruction, I, and so need
2646 // to be moved after J (the second instruction) when the pair is fused.
2647 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB,
2648 DenseMap<Value *, Value *> &ChosenPairs,
2649 std::multimap<Value *, Value *> &LoadMoveSet,
2651 // Skip to the first instruction past I.
2652 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2654 DenseSet<Value *> Users;
2655 AliasSetTracker WriteSet(*AA);
2657 // Note: We cannot end the loop when we reach J because J could be moved
2658 // farther down the use chain by another instruction pairing. Also, J
2659 // could be before I if this is an inverted input.
2660 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) {
2661 if (trackUsesOfI(Users, WriteSet, I, L)) {
2662 if (L->mayReadFromMemory())
2663 LoadMoveSet.insert(ValuePair(L, I));
2668 // In cases where both load/stores and the computation of their pointers
2669 // are chosen for vectorization, we can end up in a situation where the
2670 // aliasing analysis starts returning different query results as the
2671 // process of fusing instruction pairs continues. Because the algorithm
2672 // relies on finding the same use trees here as were found earlier, we'll
2673 // need to precompute the necessary aliasing information here and then
2674 // manually update it during the fusion process.
2675 void BBVectorize::collectLoadMoveSet(BasicBlock &BB,
2676 std::vector<Value *> &PairableInsts,
2677 DenseMap<Value *, Value *> &ChosenPairs,
2678 std::multimap<Value *, Value *> &LoadMoveSet) {
2679 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
2680 PIE = PairableInsts.end(); PI != PIE; ++PI) {
2681 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI);
2682 if (P == ChosenPairs.end()) continue;
2684 Instruction *I = cast<Instruction>(P->first);
2685 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, I);
2689 // When the first instruction in each pair is cloned, it will inherit its
2690 // parent's metadata. This metadata must be combined with that of the other
2691 // instruction in a safe way.
2692 void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) {
2693 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata;
2694 K->getAllMetadataOtherThanDebugLoc(Metadata);
2695 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) {
2696 unsigned Kind = Metadata[i].first;
2697 MDNode *JMD = J->getMetadata(Kind);
2698 MDNode *KMD = Metadata[i].second;
2702 K->setMetadata(Kind, 0); // Remove unknown metadata
2704 case LLVMContext::MD_tbaa:
2705 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD));
2707 case LLVMContext::MD_fpmath:
2708 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD));
2714 // This function fuses the chosen instruction pairs into vector instructions,
2715 // taking care preserve any needed scalar outputs and, then, it reorders the
2716 // remaining instructions as needed (users of the first member of the pair
2717 // need to be moved to after the location of the second member of the pair
2718 // because the vector instruction is inserted in the location of the pair's
2720 void BBVectorize::fuseChosenPairs(BasicBlock &BB,
2721 std::vector<Value *> &PairableInsts,
2722 DenseMap<Value *, Value *> &ChosenPairs,
2723 DenseSet<ValuePair> &FixedOrderPairs,
2724 DenseMap<VPPair, unsigned> &PairConnectionTypes,
2725 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
2726 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps) {
2727 LLVMContext& Context = BB.getContext();
2729 // During the vectorization process, the order of the pairs to be fused
2730 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs
2731 // list. After a pair is fused, the flipped pair is removed from the list.
2732 DenseSet<ValuePair> FlippedPairs;
2733 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(),
2734 E = ChosenPairs.end(); P != E; ++P)
2735 FlippedPairs.insert(ValuePair(P->second, P->first));
2736 for (DenseSet<ValuePair>::iterator P = FlippedPairs.begin(),
2737 E = FlippedPairs.end(); P != E; ++P)
2738 ChosenPairs.insert(*P);
2740 std::multimap<Value *, Value *> LoadMoveSet;
2741 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, LoadMoveSet);
2743 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n");
2745 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) {
2746 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI);
2747 if (P == ChosenPairs.end()) {
2752 if (getDepthFactor(P->first) == 0) {
2753 // These instructions are not really fused, but are tracked as though
2754 // they are. Any case in which it would be interesting to fuse them
2755 // will be taken care of by InstCombine.
2761 Instruction *I = cast<Instruction>(P->first),
2762 *J = cast<Instruction>(P->second);
2764 DEBUG(dbgs() << "BBV: fusing: " << *I <<
2765 " <-> " << *J << "\n");
2767 // Remove the pair and flipped pair from the list.
2768 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second);
2769 assert(FP != ChosenPairs.end() && "Flipped pair not found in list");
2770 ChosenPairs.erase(FP);
2771 ChosenPairs.erase(P);
2773 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSet, I, J)) {
2774 DEBUG(dbgs() << "BBV: fusion of: " << *I <<
2776 " aborted because of non-trivial dependency cycle\n");
2782 // If the pair must have the other order, then flip it.
2783 bool FlipPairOrder = FixedOrderPairs.count(ValuePair(J, I));
2784 if (!FlipPairOrder && !FixedOrderPairs.count(ValuePair(I, J))) {
2785 // This pair does not have a fixed order, and so we might want to
2786 // flip it if that will yield fewer shuffles. We count the number
2787 // of dependencies connected via swaps, and those directly connected,
2788 // and flip the order if the number of swaps is greater.
2789 bool OrigOrder = true;
2790 VPPIteratorPair IP = ConnectedPairDeps.equal_range(ValuePair(I, J));
2791 if (IP.first == ConnectedPairDeps.end()) {
2792 IP = ConnectedPairDeps.equal_range(ValuePair(J, I));
2796 if (IP.first != ConnectedPairDeps.end()) {
2797 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
2798 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first;
2799 Q != IP.second; ++Q) {
2800 DenseMap<VPPair, unsigned>::iterator R =
2801 PairConnectionTypes.find(VPPair(Q->second, Q->first));
2802 assert(R != PairConnectionTypes.end() &&
2803 "Cannot find pair connection type");
2804 if (R->second == PairConnectionDirect)
2806 else if (R->second == PairConnectionSwap)
2811 std::swap(NumDepsDirect, NumDepsSwap);
2813 if (NumDepsSwap > NumDepsDirect) {
2814 FlipPairOrder = true;
2815 DEBUG(dbgs() << "BBV: reordering pair: " << *I <<
2816 " <-> " << *J << "\n");
2821 Instruction *L = I, *H = J;
2825 // If the pair being fused uses the opposite order from that in the pair
2826 // connection map, then we need to flip the types.
2827 VPPIteratorPair IP = ConnectedPairs.equal_range(ValuePair(H, L));
2828 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first;
2829 Q != IP.second; ++Q) {
2830 DenseMap<VPPair, unsigned>::iterator R = PairConnectionTypes.find(*Q);
2831 assert(R != PairConnectionTypes.end() &&
2832 "Cannot find pair connection type");
2833 if (R->second == PairConnectionDirect)
2834 R->second = PairConnectionSwap;
2835 else if (R->second == PairConnectionSwap)
2836 R->second = PairConnectionDirect;
2839 bool LBeforeH = !FlipPairOrder;
2840 unsigned NumOperands = I->getNumOperands();
2841 SmallVector<Value *, 3> ReplacedOperands(NumOperands);
2842 getReplacementInputsForPair(Context, L, H, ReplacedOperands,
2845 // Make a copy of the original operation, change its type to the vector
2846 // type and replace its operands with the vector operands.
2847 Instruction *K = L->clone();
2850 else if (H->hasName())
2853 if (!isa<StoreInst>(K))
2854 K->mutateType(getVecTypeForPair(L->getType(), H->getType()));
2856 combineMetadata(K, H);
2858 for (unsigned o = 0; o < NumOperands; ++o)
2859 K->setOperand(o, ReplacedOperands[o]);
2863 // Instruction insertion point:
2864 Instruction *InsertionPt = K;
2865 Instruction *K1 = 0, *K2 = 0;
2866 replaceOutputsOfPair(Context, L, H, K, InsertionPt, K1, K2);
2868 // The use tree of the first original instruction must be moved to after
2869 // the location of the second instruction. The entire use tree of the
2870 // first instruction is disjoint from the input tree of the second
2871 // (by definition), and so commutes with it.
2873 moveUsesOfIAfterJ(BB, LoadMoveSet, InsertionPt, I, J);
2875 if (!isa<StoreInst>(I)) {
2876 L->replaceAllUsesWith(K1);
2877 H->replaceAllUsesWith(K2);
2878 AA->replaceWithNewValue(L, K1);
2879 AA->replaceWithNewValue(H, K2);
2882 // Instructions that may read from memory may be in the load move set.
2883 // Once an instruction is fused, we no longer need its move set, and so
2884 // the values of the map never need to be updated. However, when a load
2885 // is fused, we need to merge the entries from both instructions in the
2886 // pair in case those instructions were in the move set of some other
2887 // yet-to-be-fused pair. The loads in question are the keys of the map.
2888 if (I->mayReadFromMemory()) {
2889 std::vector<ValuePair> NewSetMembers;
2890 VPIteratorPair IPairRange = LoadMoveSet.equal_range(I);
2891 VPIteratorPair JPairRange = LoadMoveSet.equal_range(J);
2892 for (std::multimap<Value *, Value *>::iterator N = IPairRange.first;
2893 N != IPairRange.second; ++N)
2894 NewSetMembers.push_back(ValuePair(K, N->second));
2895 for (std::multimap<Value *, Value *>::iterator N = JPairRange.first;
2896 N != JPairRange.second; ++N)
2897 NewSetMembers.push_back(ValuePair(K, N->second));
2898 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(),
2899 AE = NewSetMembers.end(); A != AE; ++A)
2900 LoadMoveSet.insert(*A);
2903 // Before removing I, set the iterator to the next instruction.
2904 PI = llvm::next(BasicBlock::iterator(I));
2905 if (cast<Instruction>(PI) == J)
2910 I->eraseFromParent();
2911 J->eraseFromParent();
2913 DEBUG(if (PrintAfterEveryPair) dbgs() << "BBV: block is now: \n" <<
2917 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n");
2921 char BBVectorize::ID = 0;
2922 static const char bb_vectorize_name[] = "Basic-Block Vectorization";
2923 INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
2924 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
2925 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
2926 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
2927 INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
2929 BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) {
2930 return new BBVectorize(C);
2934 llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) {
2935 BBVectorize BBVectorizer(P, C);
2936 return BBVectorizer.vectorizeBB(BB);
2939 //===----------------------------------------------------------------------===//
2940 VectorizeConfig::VectorizeConfig() {
2941 VectorBits = ::VectorBits;
2942 VectorizeBools = !::NoBools;
2943 VectorizeInts = !::NoInts;
2944 VectorizeFloats = !::NoFloats;
2945 VectorizePointers = !::NoPointers;
2946 VectorizeCasts = !::NoCasts;
2947 VectorizeMath = !::NoMath;
2948 VectorizeFMA = !::NoFMA;
2949 VectorizeSelect = !::NoSelect;
2950 VectorizeCmp = !::NoCmp;
2951 VectorizeGEP = !::NoGEP;
2952 VectorizeMemOps = !::NoMemOps;
2953 AlignedOnly = ::AlignedOnly;
2954 ReqChainDepth= ::ReqChainDepth;
2955 SearchLimit = ::SearchLimit;
2956 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck;
2957 SplatBreaksChain = ::SplatBreaksChain;
2958 MaxInsts = ::MaxInsts;
2959 MaxIter = ::MaxIter;
2960 Pow2LenOnly = ::Pow2LenOnly;
2961 NoMemOpBoost = ::NoMemOpBoost;
2962 FastDep = ::FastDep;