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();
491 // Returns the weight associated with the provided value. A chain of
492 // candidate pairs has a length given by the sum of the weights of its
493 // members (one weight per pair; the weight of each member of the pair
494 // is assumed to be the same). This length is then compared to the
495 // chain-length threshold to determine if a given chain is significant
496 // enough to be vectorized. The length is also used in comparing
497 // candidate chains where longer chains are considered to be better.
498 // Note: when this function returns 0, the resulting instructions are
499 // not actually fused.
500 inline size_t getDepthFactor(Value *V) {
501 // InsertElement and ExtractElement have a depth factor of zero. This is
502 // for two reasons: First, they cannot be usefully fused. Second, because
503 // the pass generates a lot of these, they can confuse the simple metric
504 // used to compare the trees in the next iteration. Thus, giving them a
505 // weight of zero allows the pass to essentially ignore them in
506 // subsequent iterations when looking for vectorization opportunities
507 // while still tracking dependency chains that flow through those
509 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V))
512 // Give a load or store half of the required depth so that load/store
513 // pairs will vectorize.
514 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V)))
515 return Config.ReqChainDepth/2;
520 // Returns the cost of the provided instruction using VTTI.
521 // This does not handle loads and stores.
522 unsigned getInstrCost(unsigned Opcode, Type *T1, Type *T2) {
525 case Instruction::GetElementPtr:
526 // We mark this instruction as zero-cost because scalar GEPs are usually
527 // lowered to the intruction addressing mode. At the moment we don't
528 // generate vector GEPs.
530 case Instruction::Br:
531 return VTTI->getCFInstrCost(Opcode);
532 case Instruction::PHI:
534 case Instruction::Add:
535 case Instruction::FAdd:
536 case Instruction::Sub:
537 case Instruction::FSub:
538 case Instruction::Mul:
539 case Instruction::FMul:
540 case Instruction::UDiv:
541 case Instruction::SDiv:
542 case Instruction::FDiv:
543 case Instruction::URem:
544 case Instruction::SRem:
545 case Instruction::FRem:
546 case Instruction::Shl:
547 case Instruction::LShr:
548 case Instruction::AShr:
549 case Instruction::And:
550 case Instruction::Or:
551 case Instruction::Xor:
552 return VTTI->getArithmeticInstrCost(Opcode, T1);
553 case Instruction::Select:
554 case Instruction::ICmp:
555 case Instruction::FCmp:
556 return VTTI->getCmpSelInstrCost(Opcode, T1, T2);
557 case Instruction::ZExt:
558 case Instruction::SExt:
559 case Instruction::FPToUI:
560 case Instruction::FPToSI:
561 case Instruction::FPExt:
562 case Instruction::PtrToInt:
563 case Instruction::IntToPtr:
564 case Instruction::SIToFP:
565 case Instruction::UIToFP:
566 case Instruction::Trunc:
567 case Instruction::FPTrunc:
568 case Instruction::BitCast:
569 case Instruction::ShuffleVector:
570 return VTTI->getCastInstrCost(Opcode, T1, T2);
576 // This determines the relative offset of two loads or stores, returning
577 // true if the offset could be determined to be some constant value.
578 // For example, if OffsetInElmts == 1, then J accesses the memory directly
579 // after I; if OffsetInElmts == -1 then I accesses the memory
581 bool getPairPtrInfo(Instruction *I, Instruction *J,
582 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment,
583 unsigned &IAddressSpace, unsigned &JAddressSpace,
584 int64_t &OffsetInElmts, bool ComputeOffset = true) {
586 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
587 LoadInst *LJ = cast<LoadInst>(J);
588 IPtr = LI->getPointerOperand();
589 JPtr = LJ->getPointerOperand();
590 IAlignment = LI->getAlignment();
591 JAlignment = LJ->getAlignment();
592 IAddressSpace = LI->getPointerAddressSpace();
593 JAddressSpace = LJ->getPointerAddressSpace();
595 StoreInst *SI = cast<StoreInst>(I), *SJ = cast<StoreInst>(J);
596 IPtr = SI->getPointerOperand();
597 JPtr = SJ->getPointerOperand();
598 IAlignment = SI->getAlignment();
599 JAlignment = SJ->getAlignment();
600 IAddressSpace = SI->getPointerAddressSpace();
601 JAddressSpace = SJ->getPointerAddressSpace();
607 const SCEV *IPtrSCEV = SE->getSCEV(IPtr);
608 const SCEV *JPtrSCEV = SE->getSCEV(JPtr);
610 // If this is a trivial offset, then we'll get something like
611 // 1*sizeof(type). With target data, which we need anyway, this will get
612 // constant folded into a number.
613 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV);
614 if (const SCEVConstant *ConstOffSCEV =
615 dyn_cast<SCEVConstant>(OffsetSCEV)) {
616 ConstantInt *IntOff = ConstOffSCEV->getValue();
617 int64_t Offset = IntOff->getSExtValue();
619 Type *VTy = cast<PointerType>(IPtr->getType())->getElementType();
620 int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy);
622 Type *VTy2 = cast<PointerType>(JPtr->getType())->getElementType();
623 if (VTy != VTy2 && Offset < 0) {
624 int64_t VTy2TSS = (int64_t) TD->getTypeStoreSize(VTy2);
625 OffsetInElmts = Offset/VTy2TSS;
626 return (abs64(Offset) % VTy2TSS) == 0;
629 OffsetInElmts = Offset/VTyTSS;
630 return (abs64(Offset) % VTyTSS) == 0;
636 // Returns true if the provided CallInst represents an intrinsic that can
638 bool isVectorizableIntrinsic(CallInst* I) {
639 Function *F = I->getCalledFunction();
640 if (!F) return false;
642 unsigned IID = F->getIntrinsicID();
643 if (!IID) return false;
648 case Intrinsic::sqrt:
649 case Intrinsic::powi:
653 case Intrinsic::log2:
654 case Intrinsic::log10:
656 case Intrinsic::exp2:
658 return Config.VectorizeMath;
660 return Config.VectorizeFMA;
664 // Returns true if J is the second element in some pair referenced by
665 // some multimap pair iterator pair.
666 template <typename V>
667 bool isSecondInIteratorPair(V J, std::pair<
668 typename std::multimap<V, V>::iterator,
669 typename std::multimap<V, V>::iterator> PairRange) {
670 for (typename std::multimap<V, V>::iterator K = PairRange.first;
671 K != PairRange.second; ++K)
672 if (K->second == J) return true;
678 // This function implements one vectorization iteration on the provided
679 // basic block. It returns true if the block is changed.
680 bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) {
682 BasicBlock::iterator Start = BB.getFirstInsertionPt();
684 std::vector<Value *> AllPairableInsts;
685 DenseMap<Value *, Value *> AllChosenPairs;
686 DenseSet<ValuePair> AllFixedOrderPairs;
687 DenseMap<VPPair, unsigned> AllPairConnectionTypes;
688 std::multimap<ValuePair, ValuePair> AllConnectedPairs, AllConnectedPairDeps;
691 std::vector<Value *> PairableInsts;
692 std::multimap<Value *, Value *> CandidatePairs;
693 DenseSet<ValuePair> FixedOrderPairs;
694 DenseMap<ValuePair, int> CandidatePairCostSavings;
695 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs,
697 CandidatePairCostSavings,
698 PairableInsts, NonPow2Len);
699 if (PairableInsts.empty()) continue;
701 // Now we have a map of all of the pairable instructions and we need to
702 // select the best possible pairing. A good pairing is one such that the
703 // users of the pair are also paired. This defines a (directed) forest
704 // over the pairs such that two pairs are connected iff the second pair
707 // Note that it only matters that both members of the second pair use some
708 // element of the first pair (to allow for splatting).
710 std::multimap<ValuePair, ValuePair> ConnectedPairs, ConnectedPairDeps;
711 DenseMap<VPPair, unsigned> PairConnectionTypes;
712 computeConnectedPairs(CandidatePairs, PairableInsts, ConnectedPairs,
713 PairConnectionTypes);
714 if (ConnectedPairs.empty()) continue;
716 for (std::multimap<ValuePair, ValuePair>::iterator
717 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
719 ConnectedPairDeps.insert(VPPair(I->second, I->first));
722 // Build the pairable-instruction dependency map
723 DenseSet<ValuePair> PairableInstUsers;
724 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers);
726 // There is now a graph of the connected pairs. For each variable, pick
727 // the pairing with the largest tree meeting the depth requirement on at
728 // least one branch. Then select all pairings that are part of that tree
729 // and remove them from the list of available pairings and pairable
732 DenseMap<Value *, Value *> ChosenPairs;
733 choosePairs(CandidatePairs, CandidatePairCostSavings,
734 PairableInsts, FixedOrderPairs, PairConnectionTypes,
735 ConnectedPairs, ConnectedPairDeps,
736 PairableInstUsers, ChosenPairs);
738 if (ChosenPairs.empty()) continue;
739 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(),
740 PairableInsts.end());
741 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end());
743 // Only for the chosen pairs, propagate information on fixed-order pairs,
744 // pair connections, and their types to the data structures used by the
745 // pair fusion procedures.
746 for (DenseMap<Value *, Value *>::iterator I = ChosenPairs.begin(),
747 IE = ChosenPairs.end(); I != IE; ++I) {
748 if (FixedOrderPairs.count(*I))
749 AllFixedOrderPairs.insert(*I);
750 else if (FixedOrderPairs.count(ValuePair(I->second, I->first)))
751 AllFixedOrderPairs.insert(ValuePair(I->second, I->first));
753 for (DenseMap<Value *, Value *>::iterator J = ChosenPairs.begin();
755 DenseMap<VPPair, unsigned>::iterator K =
756 PairConnectionTypes.find(VPPair(*I, *J));
757 if (K != PairConnectionTypes.end()) {
758 AllPairConnectionTypes.insert(*K);
760 K = PairConnectionTypes.find(VPPair(*J, *I));
761 if (K != PairConnectionTypes.end())
762 AllPairConnectionTypes.insert(*K);
767 for (std::multimap<ValuePair, ValuePair>::iterator
768 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
770 if (AllPairConnectionTypes.count(*I)) {
771 AllConnectedPairs.insert(*I);
772 AllConnectedPairDeps.insert(VPPair(I->second, I->first));
775 } while (ShouldContinue);
777 if (AllChosenPairs.empty()) return false;
778 NumFusedOps += AllChosenPairs.size();
780 // A set of pairs has now been selected. It is now necessary to replace the
781 // paired instructions with vector instructions. For this procedure each
782 // operand must be replaced with a vector operand. This vector is formed
783 // by using build_vector on the old operands. The replaced values are then
784 // replaced with a vector_extract on the result. Subsequent optimization
785 // passes should coalesce the build/extract combinations.
787 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs, AllFixedOrderPairs,
788 AllPairConnectionTypes,
789 AllConnectedPairs, AllConnectedPairDeps);
791 // It is important to cleanup here so that future iterations of this
792 // function have less work to do.
793 (void) SimplifyInstructionsInBlock(&BB, TD, AA->getTargetLibraryInfo());
797 // This function returns true if the provided instruction is capable of being
798 // fused into a vector instruction. This determination is based only on the
799 // type and other attributes of the instruction.
800 bool BBVectorize::isInstVectorizable(Instruction *I,
801 bool &IsSimpleLoadStore) {
802 IsSimpleLoadStore = false;
804 if (CallInst *C = dyn_cast<CallInst>(I)) {
805 if (!isVectorizableIntrinsic(C))
807 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) {
808 // Vectorize simple loads if possbile:
809 IsSimpleLoadStore = L->isSimple();
810 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
812 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) {
813 // Vectorize simple stores if possbile:
814 IsSimpleLoadStore = S->isSimple();
815 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
817 } else if (CastInst *C = dyn_cast<CastInst>(I)) {
818 // We can vectorize casts, but not casts of pointer types, etc.
819 if (!Config.VectorizeCasts)
822 Type *SrcTy = C->getSrcTy();
823 if (!SrcTy->isSingleValueType())
826 Type *DestTy = C->getDestTy();
827 if (!DestTy->isSingleValueType())
829 } else if (isa<SelectInst>(I)) {
830 if (!Config.VectorizeSelect)
832 } else if (isa<CmpInst>(I)) {
833 if (!Config.VectorizeCmp)
835 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) {
836 if (!Config.VectorizeGEP)
839 // Currently, vector GEPs exist only with one index.
840 if (G->getNumIndices() != 1)
842 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) ||
843 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) {
847 // We can't vectorize memory operations without target data
848 if (TD == 0 && IsSimpleLoadStore)
852 getInstructionTypes(I, T1, T2);
854 // Not every type can be vectorized...
855 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) ||
856 !(VectorType::isValidElementType(T2) || T2->isVectorTy()))
859 if (T1->getScalarSizeInBits() == 1) {
860 if (!Config.VectorizeBools)
863 if (!Config.VectorizeInts && T1->isIntOrIntVectorTy())
867 if (T2->getScalarSizeInBits() == 1) {
868 if (!Config.VectorizeBools)
871 if (!Config.VectorizeInts && T2->isIntOrIntVectorTy())
875 if (!Config.VectorizeFloats
876 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy()))
879 // Don't vectorize target-specific types.
880 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy())
882 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy())
885 if ((!Config.VectorizePointers || TD == 0) &&
886 (T1->getScalarType()->isPointerTy() ||
887 T2->getScalarType()->isPointerTy()))
890 if (!VTTI && (T1->getPrimitiveSizeInBits() >= Config.VectorBits ||
891 T2->getPrimitiveSizeInBits() >= Config.VectorBits))
897 // This function returns true if the two provided instructions are compatible
898 // (meaning that they can be fused into a vector instruction). This assumes
899 // that I has already been determined to be vectorizable and that J is not
900 // in the use tree of I.
901 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J,
902 bool IsSimpleLoadStore, bool NonPow2Len,
903 int &CostSavings, int &FixedOrder) {
904 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I <<
905 " <-> " << *J << "\n");
910 // Loads and stores can be merged if they have different alignments,
911 // but are otherwise the same.
912 if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment |
913 (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0)))
916 Type *IT1, *IT2, *JT1, *JT2;
917 getInstructionTypes(I, IT1, IT2);
918 getInstructionTypes(J, JT1, JT2);
919 unsigned MaxTypeBits = std::max(
920 IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(),
921 IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits());
922 if (!VTTI && MaxTypeBits > Config.VectorBits)
925 // FIXME: handle addsub-type operations!
927 if (IsSimpleLoadStore) {
929 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
930 int64_t OffsetInElmts = 0;
931 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
932 IAddressSpace, JAddressSpace,
933 OffsetInElmts) && abs64(OffsetInElmts) == 1) {
934 FixedOrder = (int) OffsetInElmts;
935 unsigned BottomAlignment = IAlignment;
936 if (OffsetInElmts < 0) BottomAlignment = JAlignment;
938 Type *aTypeI = isa<StoreInst>(I) ?
939 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType();
940 Type *aTypeJ = isa<StoreInst>(J) ?
941 cast<StoreInst>(J)->getValueOperand()->getType() : J->getType();
942 Type *VType = getVecTypeForPair(aTypeI, aTypeJ);
944 if (Config.AlignedOnly) {
945 // An aligned load or store is possible only if the instruction
946 // with the lower offset has an alignment suitable for the
949 unsigned VecAlignment = TD->getPrefTypeAlignment(VType);
950 if (BottomAlignment < VecAlignment)
955 unsigned ICost = VTTI->getMemoryOpCost(I->getOpcode(), I->getType(),
956 IAlignment, IAddressSpace);
957 unsigned JCost = VTTI->getMemoryOpCost(J->getOpcode(), J->getType(),
958 JAlignment, JAddressSpace);
959 unsigned VCost = VTTI->getMemoryOpCost(I->getOpcode(), VType,
962 if (VCost > ICost + JCost)
965 // We don't want to fuse to a type that will be split, even
966 // if the two input types will also be split and there is no other
968 unsigned VParts = VTTI->getNumberOfParts(VType);
971 else if (!VParts && VCost == ICost + JCost)
974 CostSavings = ICost + JCost - VCost;
980 unsigned ICost = getInstrCost(I->getOpcode(), IT1, IT2);
981 unsigned JCost = getInstrCost(J->getOpcode(), JT1, JT2);
982 Type *VT1 = getVecTypeForPair(IT1, JT1),
983 *VT2 = getVecTypeForPair(IT2, JT2);
984 unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2);
986 if (VCost > ICost + JCost)
989 // We don't want to fuse to a type that will be split, even
990 // if the two input types will also be split and there is no other
992 unsigned VParts1 = VTTI->getNumberOfParts(VT1),
993 VParts2 = VTTI->getNumberOfParts(VT2);
994 if (VParts1 > 1 || VParts2 > 1)
996 else if ((!VParts1 || !VParts2) && VCost == ICost + JCost)
999 CostSavings = ICost + JCost - VCost;
1002 // The powi intrinsic is special because only the first argument is
1003 // vectorized, the second arguments must be equal.
1004 CallInst *CI = dyn_cast<CallInst>(I);
1006 if (CI && (FI = CI->getCalledFunction()) &&
1007 FI->getIntrinsicID() == Intrinsic::powi) {
1009 Value *A1I = CI->getArgOperand(1),
1010 *A1J = cast<CallInst>(J)->getArgOperand(1);
1011 const SCEV *A1ISCEV = SE->getSCEV(A1I),
1012 *A1JSCEV = SE->getSCEV(A1J);
1013 return (A1ISCEV == A1JSCEV);
1019 // Figure out whether or not J uses I and update the users and write-set
1020 // structures associated with I. Specifically, Users represents the set of
1021 // instructions that depend on I. WriteSet represents the set
1022 // of memory locations that are dependent on I. If UpdateUsers is true,
1023 // and J uses I, then Users is updated to contain J and WriteSet is updated
1024 // to contain any memory locations to which J writes. The function returns
1025 // true if J uses I. By default, alias analysis is used to determine
1026 // whether J reads from memory that overlaps with a location in WriteSet.
1027 // If LoadMoveSet is not null, then it is a previously-computed multimap
1028 // where the key is the memory-based user instruction and the value is
1029 // the instruction to be compared with I. So, if LoadMoveSet is provided,
1030 // then the alias analysis is not used. This is necessary because this
1031 // function is called during the process of moving instructions during
1032 // vectorization and the results of the alias analysis are not stable during
1034 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users,
1035 AliasSetTracker &WriteSet, Instruction *I,
1036 Instruction *J, bool UpdateUsers,
1037 std::multimap<Value *, Value *> *LoadMoveSet) {
1040 // This instruction may already be marked as a user due, for example, to
1041 // being a member of a selected pair.
1046 for (User::op_iterator JU = J->op_begin(), JE = J->op_end();
1049 if (I == V || Users.count(V)) {
1054 if (!UsesI && J->mayReadFromMemory()) {
1056 VPIteratorPair JPairRange = LoadMoveSet->equal_range(J);
1057 UsesI = isSecondInIteratorPair<Value*>(I, JPairRange);
1059 for (AliasSetTracker::iterator W = WriteSet.begin(),
1060 WE = WriteSet.end(); W != WE; ++W) {
1061 if (W->aliasesUnknownInst(J, *AA)) {
1069 if (UsesI && UpdateUsers) {
1070 if (J->mayWriteToMemory()) WriteSet.add(J);
1077 // This function iterates over all instruction pairs in the provided
1078 // basic block and collects all candidate pairs for vectorization.
1079 bool BBVectorize::getCandidatePairs(BasicBlock &BB,
1080 BasicBlock::iterator &Start,
1081 std::multimap<Value *, Value *> &CandidatePairs,
1082 DenseSet<ValuePair> &FixedOrderPairs,
1083 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1084 std::vector<Value *> &PairableInsts, bool NonPow2Len) {
1085 BasicBlock::iterator E = BB.end();
1086 if (Start == E) return false;
1088 bool ShouldContinue = false, IAfterStart = false;
1089 for (BasicBlock::iterator I = Start++; I != E; ++I) {
1090 if (I == Start) IAfterStart = true;
1092 bool IsSimpleLoadStore;
1093 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue;
1095 // Look for an instruction with which to pair instruction *I...
1096 DenseSet<Value *> Users;
1097 AliasSetTracker WriteSet(*AA);
1098 bool JAfterStart = IAfterStart;
1099 BasicBlock::iterator J = llvm::next(I);
1100 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) {
1101 if (J == Start) JAfterStart = true;
1103 // Determine if J uses I, if so, exit the loop.
1104 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep);
1105 if (Config.FastDep) {
1106 // Note: For this heuristic to be effective, independent operations
1107 // must tend to be intermixed. This is likely to be true from some
1108 // kinds of grouped loop unrolling (but not the generic LLVM pass),
1109 // but otherwise may require some kind of reordering pass.
1111 // When using fast dependency analysis,
1112 // stop searching after first use:
1115 if (UsesI) continue;
1118 // J does not use I, and comes before the first use of I, so it can be
1119 // merged with I if the instructions are compatible.
1120 int CostSavings, FixedOrder;
1121 if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len,
1122 CostSavings, FixedOrder)) continue;
1124 // J is a candidate for merging with I.
1125 if (!PairableInsts.size() ||
1126 PairableInsts[PairableInsts.size()-1] != I) {
1127 PairableInsts.push_back(I);
1130 CandidatePairs.insert(ValuePair(I, J));
1132 CandidatePairCostSavings.insert(ValuePairWithCost(ValuePair(I, J),
1135 if (FixedOrder == 1)
1136 FixedOrderPairs.insert(ValuePair(I, J));
1137 else if (FixedOrder == -1)
1138 FixedOrderPairs.insert(ValuePair(J, I));
1140 // The next call to this function must start after the last instruction
1141 // selected during this invocation.
1143 Start = llvm::next(J);
1144 IAfterStart = JAfterStart = false;
1147 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair "
1148 << *I << " <-> " << *J << " (cost savings: " <<
1149 CostSavings << ")\n");
1151 // If we have already found too many pairs, break here and this function
1152 // will be called again starting after the last instruction selected
1153 // during this invocation.
1154 if (PairableInsts.size() >= Config.MaxInsts) {
1155 ShouldContinue = true;
1164 DEBUG(dbgs() << "BBV: found " << PairableInsts.size()
1165 << " instructions with candidate pairs\n");
1167 return ShouldContinue;
1170 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that
1171 // it looks for pairs such that both members have an input which is an
1172 // output of PI or PJ.
1173 void BBVectorize::computePairsConnectedTo(
1174 std::multimap<Value *, Value *> &CandidatePairs,
1175 std::vector<Value *> &PairableInsts,
1176 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1177 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1181 // For each possible pairing for this variable, look at the uses of
1182 // the first value...
1183 for (Value::use_iterator I = P.first->use_begin(),
1184 E = P.first->use_end(); I != E; ++I) {
1185 if (isa<LoadInst>(*I)) {
1186 // A pair cannot be connected to a load because the load only takes one
1187 // operand (the address) and it is a scalar even after vectorization.
1189 } else if ((SI = dyn_cast<StoreInst>(*I)) &&
1190 P.first == SI->getPointerOperand()) {
1191 // Similarly, a pair cannot be connected to a store through its
1196 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I);
1198 // For each use of the first variable, look for uses of the second
1200 for (Value::use_iterator J = P.second->use_begin(),
1201 E2 = P.second->use_end(); J != E2; ++J) {
1202 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1203 P.second == SJ->getPointerOperand())
1206 VPIteratorPair JPairRange = CandidatePairs.equal_range(*J);
1209 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) {
1210 VPPair VP(P, ValuePair(*I, *J));
1211 ConnectedPairs.insert(VP);
1212 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionDirect));
1216 if (isSecondInIteratorPair<Value*>(*I, JPairRange)) {
1217 VPPair VP(P, ValuePair(*J, *I));
1218 ConnectedPairs.insert(VP);
1219 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSwap));
1223 if (Config.SplatBreaksChain) continue;
1224 // Look for cases where just the first value in the pair is used by
1225 // both members of another pair (splatting).
1226 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) {
1227 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1228 P.first == SJ->getPointerOperand())
1231 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) {
1232 VPPair VP(P, ValuePair(*I, *J));
1233 ConnectedPairs.insert(VP);
1234 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1239 if (Config.SplatBreaksChain) return;
1240 // Look for cases where just the second value in the pair is used by
1241 // both members of another pair (splatting).
1242 for (Value::use_iterator I = P.second->use_begin(),
1243 E = P.second->use_end(); I != E; ++I) {
1244 if (isa<LoadInst>(*I))
1246 else if ((SI = dyn_cast<StoreInst>(*I)) &&
1247 P.second == SI->getPointerOperand())
1250 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I);
1252 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) {
1253 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1254 P.second == SJ->getPointerOperand())
1257 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) {
1258 VPPair VP(P, ValuePair(*I, *J));
1259 ConnectedPairs.insert(VP);
1260 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1266 // This function figures out which pairs are connected. Two pairs are
1267 // connected if some output of the first pair forms an input to both members
1268 // of the second pair.
1269 void BBVectorize::computeConnectedPairs(
1270 std::multimap<Value *, Value *> &CandidatePairs,
1271 std::vector<Value *> &PairableInsts,
1272 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1273 DenseMap<VPPair, unsigned> &PairConnectionTypes) {
1275 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
1276 PE = PairableInsts.end(); PI != PE; ++PI) {
1277 VPIteratorPair choiceRange = CandidatePairs.equal_range(*PI);
1279 for (std::multimap<Value *, Value *>::iterator P = choiceRange.first;
1280 P != choiceRange.second; ++P)
1281 computePairsConnectedTo(CandidatePairs, PairableInsts,
1282 ConnectedPairs, PairConnectionTypes, *P);
1285 DEBUG(dbgs() << "BBV: found " << ConnectedPairs.size()
1286 << " pair connections.\n");
1289 // This function builds a set of use tuples such that <A, B> is in the set
1290 // if B is in the use tree of A. If B is in the use tree of A, then B
1291 // depends on the output of A.
1292 void BBVectorize::buildDepMap(
1294 std::multimap<Value *, Value *> &CandidatePairs,
1295 std::vector<Value *> &PairableInsts,
1296 DenseSet<ValuePair> &PairableInstUsers) {
1297 DenseSet<Value *> IsInPair;
1298 for (std::multimap<Value *, Value *>::iterator C = CandidatePairs.begin(),
1299 E = CandidatePairs.end(); C != E; ++C) {
1300 IsInPair.insert(C->first);
1301 IsInPair.insert(C->second);
1304 // Iterate through the basic block, recording all Users of each
1305 // pairable instruction.
1307 BasicBlock::iterator E = BB.end();
1308 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) {
1309 if (IsInPair.find(I) == IsInPair.end()) continue;
1311 DenseSet<Value *> Users;
1312 AliasSetTracker WriteSet(*AA);
1313 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J)
1314 (void) trackUsesOfI(Users, WriteSet, I, J);
1316 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end();
1318 PairableInstUsers.insert(ValuePair(I, *U));
1322 // Returns true if an input to pair P is an output of pair Q and also an
1323 // input of pair Q is an output of pair P. If this is the case, then these
1324 // two pairs cannot be simultaneously fused.
1325 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q,
1326 DenseSet<ValuePair> &PairableInstUsers,
1327 std::multimap<ValuePair, ValuePair> *PairableInstUserMap) {
1328 // Two pairs are in conflict if they are mutual Users of eachother.
1329 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) ||
1330 PairableInstUsers.count(ValuePair(P.first, Q.second)) ||
1331 PairableInstUsers.count(ValuePair(P.second, Q.first)) ||
1332 PairableInstUsers.count(ValuePair(P.second, Q.second));
1333 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) ||
1334 PairableInstUsers.count(ValuePair(Q.first, P.second)) ||
1335 PairableInstUsers.count(ValuePair(Q.second, P.first)) ||
1336 PairableInstUsers.count(ValuePair(Q.second, P.second));
1337 if (PairableInstUserMap) {
1338 // FIXME: The expensive part of the cycle check is not so much the cycle
1339 // check itself but this edge insertion procedure. This needs some
1340 // profiling and probably a different data structure (same is true of
1341 // most uses of std::multimap).
1343 VPPIteratorPair QPairRange = PairableInstUserMap->equal_range(Q);
1344 if (!isSecondInIteratorPair(P, QPairRange))
1345 PairableInstUserMap->insert(VPPair(Q, P));
1348 VPPIteratorPair PPairRange = PairableInstUserMap->equal_range(P);
1349 if (!isSecondInIteratorPair(Q, PPairRange))
1350 PairableInstUserMap->insert(VPPair(P, Q));
1354 return (QUsesP && PUsesQ);
1357 // This function walks the use graph of current pairs to see if, starting
1358 // from P, the walk returns to P.
1359 bool BBVectorize::pairWillFormCycle(ValuePair P,
1360 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1361 DenseSet<ValuePair> &CurrentPairs) {
1362 DEBUG(if (DebugCycleCheck)
1363 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> "
1364 << *P.second << "\n");
1365 // A lookup table of visisted pairs is kept because the PairableInstUserMap
1366 // contains non-direct associations.
1367 DenseSet<ValuePair> Visited;
1368 SmallVector<ValuePair, 32> Q;
1369 // General depth-first post-order traversal:
1372 ValuePair QTop = Q.pop_back_val();
1373 Visited.insert(QTop);
1375 DEBUG(if (DebugCycleCheck)
1376 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> "
1377 << *QTop.second << "\n");
1378 VPPIteratorPair QPairRange = PairableInstUserMap.equal_range(QTop);
1379 for (std::multimap<ValuePair, ValuePair>::iterator C = QPairRange.first;
1380 C != QPairRange.second; ++C) {
1381 if (C->second == P) {
1383 << "BBV: rejected to prevent non-trivial cycle formation: "
1384 << *C->first.first << " <-> " << *C->first.second << "\n");
1388 if (CurrentPairs.count(C->second) && !Visited.count(C->second))
1389 Q.push_back(C->second);
1391 } while (!Q.empty());
1396 // This function builds the initial tree of connected pairs with the
1397 // pair J at the root.
1398 void BBVectorize::buildInitialTreeFor(
1399 std::multimap<Value *, Value *> &CandidatePairs,
1400 std::vector<Value *> &PairableInsts,
1401 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1402 DenseSet<ValuePair> &PairableInstUsers,
1403 DenseMap<Value *, Value *> &ChosenPairs,
1404 DenseMap<ValuePair, size_t> &Tree, ValuePair J) {
1405 // Each of these pairs is viewed as the root node of a Tree. The Tree
1406 // is then walked (depth-first). As this happens, we keep track of
1407 // the pairs that compose the Tree and the maximum depth of the Tree.
1408 SmallVector<ValuePairWithDepth, 32> Q;
1409 // General depth-first post-order traversal:
1410 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1412 ValuePairWithDepth QTop = Q.back();
1414 // Push each child onto the queue:
1415 bool MoreChildren = false;
1416 size_t MaxChildDepth = QTop.second;
1417 VPPIteratorPair qtRange = ConnectedPairs.equal_range(QTop.first);
1418 for (std::multimap<ValuePair, ValuePair>::iterator k = qtRange.first;
1419 k != qtRange.second; ++k) {
1420 // Make sure that this child pair is still a candidate:
1421 bool IsStillCand = false;
1422 VPIteratorPair checkRange =
1423 CandidatePairs.equal_range(k->second.first);
1424 for (std::multimap<Value *, Value *>::iterator m = checkRange.first;
1425 m != checkRange.second; ++m) {
1426 if (m->second == k->second.second) {
1433 DenseMap<ValuePair, size_t>::iterator C = Tree.find(k->second);
1434 if (C == Tree.end()) {
1435 size_t d = getDepthFactor(k->second.first);
1436 Q.push_back(ValuePairWithDepth(k->second, QTop.second+d));
1437 MoreChildren = true;
1439 MaxChildDepth = std::max(MaxChildDepth, C->second);
1444 if (!MoreChildren) {
1445 // Record the current pair as part of the Tree:
1446 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth));
1449 } while (!Q.empty());
1452 // Given some initial tree, prune it by removing conflicting pairs (pairs
1453 // that cannot be simultaneously chosen for vectorization).
1454 void BBVectorize::pruneTreeFor(
1455 std::multimap<Value *, Value *> &CandidatePairs,
1456 std::vector<Value *> &PairableInsts,
1457 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1458 DenseSet<ValuePair> &PairableInstUsers,
1459 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1460 DenseMap<Value *, Value *> &ChosenPairs,
1461 DenseMap<ValuePair, size_t> &Tree,
1462 DenseSet<ValuePair> &PrunedTree, ValuePair J,
1463 bool UseCycleCheck) {
1464 SmallVector<ValuePairWithDepth, 32> Q;
1465 // General depth-first post-order traversal:
1466 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1468 ValuePairWithDepth QTop = Q.pop_back_val();
1469 PrunedTree.insert(QTop.first);
1471 // Visit each child, pruning as necessary...
1472 DenseMap<ValuePair, size_t> BestChildren;
1473 VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first);
1474 for (std::multimap<ValuePair, ValuePair>::iterator K = QTopRange.first;
1475 K != QTopRange.second; ++K) {
1476 DenseMap<ValuePair, size_t>::iterator C = Tree.find(K->second);
1477 if (C == Tree.end()) continue;
1479 // This child is in the Tree, now we need to make sure it is the
1480 // best of any conflicting children. There could be multiple
1481 // conflicting children, so first, determine if we're keeping
1482 // this child, then delete conflicting children as necessary.
1484 // It is also necessary to guard against pairing-induced
1485 // dependencies. Consider instructions a .. x .. y .. b
1486 // such that (a,b) are to be fused and (x,y) are to be fused
1487 // but a is an input to x and b is an output from y. This
1488 // means that y cannot be moved after b but x must be moved
1489 // after b for (a,b) to be fused. In other words, after
1490 // fusing (a,b) we have y .. a/b .. x where y is an input
1491 // to a/b and x is an output to a/b: x and y can no longer
1492 // be legally fused. To prevent this condition, we must
1493 // make sure that a child pair added to the Tree is not
1494 // both an input and output of an already-selected pair.
1496 // Pairing-induced dependencies can also form from more complicated
1497 // cycles. The pair vs. pair conflicts are easy to check, and so
1498 // that is done explicitly for "fast rejection", and because for
1499 // child vs. child conflicts, we may prefer to keep the current
1500 // pair in preference to the already-selected child.
1501 DenseSet<ValuePair> CurrentPairs;
1504 for (DenseMap<ValuePair, size_t>::iterator C2
1505 = BestChildren.begin(), E2 = BestChildren.end();
1507 if (C2->first.first == C->first.first ||
1508 C2->first.first == C->first.second ||
1509 C2->first.second == C->first.first ||
1510 C2->first.second == C->first.second ||
1511 pairsConflict(C2->first, C->first, PairableInstUsers,
1512 UseCycleCheck ? &PairableInstUserMap : 0)) {
1513 if (C2->second >= C->second) {
1518 CurrentPairs.insert(C2->first);
1521 if (!CanAdd) continue;
1523 // Even worse, this child could conflict with another node already
1524 // selected for the Tree. If that is the case, ignore this child.
1525 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(),
1526 E2 = PrunedTree.end(); T != E2; ++T) {
1527 if (T->first == C->first.first ||
1528 T->first == C->first.second ||
1529 T->second == C->first.first ||
1530 T->second == C->first.second ||
1531 pairsConflict(*T, C->first, PairableInstUsers,
1532 UseCycleCheck ? &PairableInstUserMap : 0)) {
1537 CurrentPairs.insert(*T);
1539 if (!CanAdd) continue;
1541 // And check the queue too...
1542 for (SmallVector<ValuePairWithDepth, 32>::iterator C2 = Q.begin(),
1543 E2 = Q.end(); C2 != E2; ++C2) {
1544 if (C2->first.first == C->first.first ||
1545 C2->first.first == C->first.second ||
1546 C2->first.second == C->first.first ||
1547 C2->first.second == C->first.second ||
1548 pairsConflict(C2->first, C->first, PairableInstUsers,
1549 UseCycleCheck ? &PairableInstUserMap : 0)) {
1554 CurrentPairs.insert(C2->first);
1556 if (!CanAdd) continue;
1558 // Last but not least, check for a conflict with any of the
1559 // already-chosen pairs.
1560 for (DenseMap<Value *, Value *>::iterator C2 =
1561 ChosenPairs.begin(), E2 = ChosenPairs.end();
1563 if (pairsConflict(*C2, C->first, PairableInstUsers,
1564 UseCycleCheck ? &PairableInstUserMap : 0)) {
1569 CurrentPairs.insert(*C2);
1571 if (!CanAdd) continue;
1573 // To check for non-trivial cycles formed by the addition of the
1574 // current pair we've formed a list of all relevant pairs, now use a
1575 // graph walk to check for a cycle. We start from the current pair and
1576 // walk the use tree to see if we again reach the current pair. If we
1577 // do, then the current pair is rejected.
1579 // FIXME: It may be more efficient to use a topological-ordering
1580 // algorithm to improve the cycle check. This should be investigated.
1581 if (UseCycleCheck &&
1582 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs))
1585 // This child can be added, but we may have chosen it in preference
1586 // to an already-selected child. Check for this here, and if a
1587 // conflict is found, then remove the previously-selected child
1588 // before adding this one in its place.
1589 for (DenseMap<ValuePair, size_t>::iterator C2
1590 = BestChildren.begin(); C2 != BestChildren.end();) {
1591 if (C2->first.first == C->first.first ||
1592 C2->first.first == C->first.second ||
1593 C2->first.second == C->first.first ||
1594 C2->first.second == C->first.second ||
1595 pairsConflict(C2->first, C->first, PairableInstUsers))
1596 BestChildren.erase(C2++);
1601 BestChildren.insert(ValuePairWithDepth(C->first, C->second));
1604 for (DenseMap<ValuePair, size_t>::iterator C
1605 = BestChildren.begin(), E2 = BestChildren.end();
1607 size_t DepthF = getDepthFactor(C->first.first);
1608 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF));
1610 } while (!Q.empty());
1613 // This function finds the best tree of mututally-compatible connected
1614 // pairs, given the choice of root pairs as an iterator range.
1615 void BBVectorize::findBestTreeFor(
1616 std::multimap<Value *, Value *> &CandidatePairs,
1617 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1618 std::vector<Value *> &PairableInsts,
1619 DenseSet<ValuePair> &FixedOrderPairs,
1620 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1621 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1622 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps,
1623 DenseSet<ValuePair> &PairableInstUsers,
1624 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1625 DenseMap<Value *, Value *> &ChosenPairs,
1626 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth,
1627 int &BestEffSize, VPIteratorPair ChoiceRange,
1628 bool UseCycleCheck) {
1629 for (std::multimap<Value *, Value *>::iterator J = ChoiceRange.first;
1630 J != ChoiceRange.second; ++J) {
1632 // Before going any further, make sure that this pair does not
1633 // conflict with any already-selected pairs (see comment below
1634 // near the Tree pruning for more details).
1635 DenseSet<ValuePair> ChosenPairSet;
1636 bool DoesConflict = false;
1637 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(),
1638 E = ChosenPairs.end(); C != E; ++C) {
1639 if (pairsConflict(*C, *J, PairableInstUsers,
1640 UseCycleCheck ? &PairableInstUserMap : 0)) {
1641 DoesConflict = true;
1645 ChosenPairSet.insert(*C);
1647 if (DoesConflict) continue;
1649 if (UseCycleCheck &&
1650 pairWillFormCycle(*J, PairableInstUserMap, ChosenPairSet))
1653 DenseMap<ValuePair, size_t> Tree;
1654 buildInitialTreeFor(CandidatePairs, PairableInsts, ConnectedPairs,
1655 PairableInstUsers, ChosenPairs, Tree, *J);
1657 // Because we'll keep the child with the largest depth, the largest
1658 // depth is still the same in the unpruned Tree.
1659 size_t MaxDepth = Tree.lookup(*J);
1661 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found Tree for pair {"
1662 << *J->first << " <-> " << *J->second << "} of depth " <<
1663 MaxDepth << " and size " << Tree.size() << "\n");
1665 // At this point the Tree has been constructed, but, may contain
1666 // contradictory children (meaning that different children of
1667 // some tree node may be attempting to fuse the same instruction).
1668 // So now we walk the tree again, in the case of a conflict,
1669 // keep only the child with the largest depth. To break a tie,
1670 // favor the first child.
1672 DenseSet<ValuePair> PrunedTree;
1673 pruneTreeFor(CandidatePairs, PairableInsts, ConnectedPairs,
1674 PairableInstUsers, PairableInstUserMap, ChosenPairs, Tree,
1675 PrunedTree, *J, UseCycleCheck);
1679 DenseSet<Value *> PrunedTreeInstrs;
1680 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
1681 E = PrunedTree.end(); S != E; ++S) {
1682 PrunedTreeInstrs.insert(S->first);
1683 PrunedTreeInstrs.insert(S->second);
1686 // The set of pairs that have already contributed to the total cost.
1687 DenseSet<ValuePair> IncomingPairs;
1689 // The node weights represent the cost savings associated with
1690 // fusing the pair of instructions.
1691 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
1692 E = PrunedTree.end(); S != E; ++S) {
1693 bool FlipOrder = false;
1695 if (getDepthFactor(S->first)) {
1696 int ESContrib = CandidatePairCostSavings.find(*S)->second;
1697 DEBUG(if (DebugPairSelection) dbgs() << "\tweight {"
1698 << *S->first << " <-> " << *S->second << "} = " <<
1700 EffSize += ESContrib;
1703 // The edge weights contribute in a negative sense: they represent
1704 // the cost of shuffles.
1705 VPPIteratorPair IP = ConnectedPairDeps.equal_range(*S);
1706 if (IP.first != ConnectedPairDeps.end()) {
1707 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
1708 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first;
1709 Q != IP.second; ++Q) {
1710 if (!PrunedTree.count(Q->second))
1712 DenseMap<VPPair, unsigned>::iterator R =
1713 PairConnectionTypes.find(VPPair(Q->second, Q->first));
1714 assert(R != PairConnectionTypes.end() &&
1715 "Cannot find pair connection type");
1716 if (R->second == PairConnectionDirect)
1718 else if (R->second == PairConnectionSwap)
1722 // If there are more swaps than direct connections, then
1723 // the pair order will be flipped during fusion. So the real
1724 // number of swaps is the minimum number.
1725 FlipOrder = !FixedOrderPairs.count(*S) &&
1726 ((NumDepsSwap > NumDepsDirect) ||
1727 FixedOrderPairs.count(ValuePair(S->second, S->first)));
1729 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first;
1730 Q != IP.second; ++Q) {
1731 if (!PrunedTree.count(Q->second))
1733 DenseMap<VPPair, unsigned>::iterator R =
1734 PairConnectionTypes.find(VPPair(Q->second, Q->first));
1735 assert(R != PairConnectionTypes.end() &&
1736 "Cannot find pair connection type");
1737 Type *Ty1 = Q->second.first->getType(),
1738 *Ty2 = Q->second.second->getType();
1739 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1740 if ((R->second == PairConnectionDirect && FlipOrder) ||
1741 (R->second == PairConnectionSwap && !FlipOrder) ||
1742 R->second == PairConnectionSplat) {
1743 int ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1745 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1746 *Q->second.first << " <-> " << *Q->second.second <<
1748 *S->first << " <-> " << *S->second << "} = " <<
1750 EffSize -= ESContrib;
1755 // Compute the cost of outgoing edges. We assume that edges outgoing
1756 // to shuffles, inserts or extracts can be merged, and so contribute
1757 // no additional cost.
1758 if (!S->first->getType()->isVoidTy()) {
1759 Type *Ty1 = S->first->getType(),
1760 *Ty2 = S->second->getType();
1761 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1763 bool NeedsExtraction = false;
1764 for (Value::use_iterator I = S->first->use_begin(),
1765 IE = S->first->use_end(); I != IE; ++I) {
1766 if (isa<ShuffleVectorInst>(*I) ||
1767 isa<InsertElementInst>(*I) ||
1768 isa<ExtractElementInst>(*I))
1770 if (PrunedTreeInstrs.count(*I))
1772 NeedsExtraction = true;
1776 if (NeedsExtraction) {
1778 if (Ty1->isVectorTy())
1779 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1782 ESContrib = (int) VTTI->getVectorInstrCost(
1783 Instruction::ExtractElement, VTy, 0);
1785 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1786 *S->first << "} = " << ESContrib << "\n");
1787 EffSize -= ESContrib;
1790 NeedsExtraction = false;
1791 for (Value::use_iterator I = S->second->use_begin(),
1792 IE = S->second->use_end(); I != IE; ++I) {
1793 if (isa<ShuffleVectorInst>(*I) ||
1794 isa<InsertElementInst>(*I) ||
1795 isa<ExtractElementInst>(*I))
1797 if (PrunedTreeInstrs.count(*I))
1799 NeedsExtraction = true;
1803 if (NeedsExtraction) {
1805 if (Ty2->isVectorTy())
1806 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1809 ESContrib = (int) VTTI->getVectorInstrCost(
1810 Instruction::ExtractElement, VTy, 1);
1811 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1812 *S->second << "} = " << ESContrib << "\n");
1813 EffSize -= ESContrib;
1817 // Compute the cost of incoming edges.
1818 if (!isa<LoadInst>(S->first) && !isa<StoreInst>(S->first)) {
1819 Instruction *S1 = cast<Instruction>(S->first),
1820 *S2 = cast<Instruction>(S->second);
1821 for (unsigned o = 0; o < S1->getNumOperands(); ++o) {
1822 Value *O1 = S1->getOperand(o), *O2 = S2->getOperand(o);
1824 // Combining constants into vector constants (or small vector
1825 // constants into larger ones are assumed free).
1826 if (isa<Constant>(O1) && isa<Constant>(O2))
1832 ValuePair VP = ValuePair(O1, O2);
1833 ValuePair VPR = ValuePair(O2, O1);
1835 // Internal edges are not handled here.
1836 if (PrunedTree.count(VP) || PrunedTree.count(VPR))
1839 Type *Ty1 = O1->getType(),
1840 *Ty2 = O2->getType();
1841 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1843 // Combining vector operations of the same type is also assumed
1844 // folded with other operations.
1846 (isa<ShuffleVectorInst>(O1) ||
1847 isa<InsertElementInst>(O1) ||
1848 isa<InsertElementInst>(O1)) &&
1849 (isa<ShuffleVectorInst>(O2) ||
1850 isa<InsertElementInst>(O2) ||
1851 isa<InsertElementInst>(O2)))
1855 // This pair has already been formed.
1856 if (IncomingPairs.count(VP)) {
1858 } else if (IncomingPairs.count(VPR)) {
1859 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1861 } else if (!Ty1->isVectorTy() && !Ty2->isVectorTy()) {
1862 ESContrib = (int) VTTI->getVectorInstrCost(
1863 Instruction::InsertElement, VTy, 0);
1864 ESContrib += (int) VTTI->getVectorInstrCost(
1865 Instruction::InsertElement, VTy, 1);
1866 } else if (!Ty1->isVectorTy()) {
1867 // O1 needs to be inserted into a vector of size O2, and then
1868 // both need to be shuffled together.
1869 ESContrib = (int) VTTI->getVectorInstrCost(
1870 Instruction::InsertElement, Ty2, 0);
1871 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
1873 } else if (!Ty2->isVectorTy()) {
1874 // O2 needs to be inserted into a vector of size O1, and then
1875 // both need to be shuffled together.
1876 ESContrib = (int) VTTI->getVectorInstrCost(
1877 Instruction::InsertElement, Ty1, 0);
1878 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
1881 Type *TyBig = Ty1, *TySmall = Ty2;
1882 if (Ty2->getVectorNumElements() > Ty1->getVectorNumElements())
1883 std::swap(TyBig, TySmall);
1885 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1887 if (TyBig != TySmall)
1888 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
1892 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {"
1893 << *O1 << " <-> " << *O2 << "} = " <<
1895 EffSize -= ESContrib;
1896 IncomingPairs.insert(VP);
1901 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
1902 E = PrunedTree.end(); S != E; ++S)
1903 EffSize += (int) getDepthFactor(S->first);
1906 DEBUG(if (DebugPairSelection)
1907 dbgs() << "BBV: found pruned Tree for pair {"
1908 << *J->first << " <-> " << *J->second << "} of depth " <<
1909 MaxDepth << " and size " << PrunedTree.size() <<
1910 " (effective size: " << EffSize << ")\n");
1911 if (((VTTI && !UseChainDepthWithTI) ||
1912 MaxDepth >= Config.ReqChainDepth) &&
1913 EffSize > 0 && EffSize > BestEffSize) {
1914 BestMaxDepth = MaxDepth;
1915 BestEffSize = EffSize;
1916 BestTree = PrunedTree;
1921 // Given the list of candidate pairs, this function selects those
1922 // that will be fused into vector instructions.
1923 void BBVectorize::choosePairs(
1924 std::multimap<Value *, Value *> &CandidatePairs,
1925 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1926 std::vector<Value *> &PairableInsts,
1927 DenseSet<ValuePair> &FixedOrderPairs,
1928 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1929 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1930 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps,
1931 DenseSet<ValuePair> &PairableInstUsers,
1932 DenseMap<Value *, Value *>& ChosenPairs) {
1933 bool UseCycleCheck =
1934 CandidatePairs.size() <= Config.MaxCandPairsForCycleCheck;
1935 std::multimap<ValuePair, ValuePair> PairableInstUserMap;
1936 for (std::vector<Value *>::iterator I = PairableInsts.begin(),
1937 E = PairableInsts.end(); I != E; ++I) {
1938 // The number of possible pairings for this variable:
1939 size_t NumChoices = CandidatePairs.count(*I);
1940 if (!NumChoices) continue;
1942 VPIteratorPair ChoiceRange = CandidatePairs.equal_range(*I);
1944 // The best pair to choose and its tree:
1945 size_t BestMaxDepth = 0;
1946 int BestEffSize = 0;
1947 DenseSet<ValuePair> BestTree;
1948 findBestTreeFor(CandidatePairs, CandidatePairCostSavings,
1949 PairableInsts, FixedOrderPairs, PairConnectionTypes,
1950 ConnectedPairs, ConnectedPairDeps,
1951 PairableInstUsers, PairableInstUserMap, ChosenPairs,
1952 BestTree, BestMaxDepth, BestEffSize, ChoiceRange,
1955 // A tree has been chosen (or not) at this point. If no tree was
1956 // chosen, then this instruction, I, cannot be paired (and is no longer
1959 DEBUG(if (BestTree.size() > 0)
1960 dbgs() << "BBV: selected pairs in the best tree for: "
1961 << *cast<Instruction>(*I) << "\n");
1963 for (DenseSet<ValuePair>::iterator S = BestTree.begin(),
1964 SE2 = BestTree.end(); S != SE2; ++S) {
1965 // Insert the members of this tree into the list of chosen pairs.
1966 ChosenPairs.insert(ValuePair(S->first, S->second));
1967 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " <<
1968 *S->second << "\n");
1970 // Remove all candidate pairs that have values in the chosen tree.
1971 for (std::multimap<Value *, Value *>::iterator K =
1972 CandidatePairs.begin(); K != CandidatePairs.end();) {
1973 if (K->first == S->first || K->second == S->first ||
1974 K->second == S->second || K->first == S->second) {
1975 // Don't remove the actual pair chosen so that it can be used
1976 // in subsequent tree selections.
1977 if (!(K->first == S->first && K->second == S->second))
1978 CandidatePairs.erase(K++);
1988 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n");
1991 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o,
1996 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) +
1997 (n > 0 ? "." + utostr(n) : "")).str();
2000 // Returns the value that is to be used as the pointer input to the vector
2001 // instruction that fuses I with J.
2002 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context,
2003 Instruction *I, Instruction *J, unsigned o) {
2005 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
2006 int64_t OffsetInElmts;
2008 // Note: the analysis might fail here, that is why the pair order has
2009 // been precomputed (OffsetInElmts must be unused here).
2010 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
2011 IAddressSpace, JAddressSpace,
2012 OffsetInElmts, false);
2014 // The pointer value is taken to be the one with the lowest offset.
2017 Type *ArgTypeI = cast<PointerType>(IPtr->getType())->getElementType();
2018 Type *ArgTypeJ = cast<PointerType>(JPtr->getType())->getElementType();
2019 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2020 Type *VArgPtrType = PointerType::get(VArgType,
2021 cast<PointerType>(IPtr->getType())->getAddressSpace());
2022 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o),
2023 /* insert before */ I);
2026 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J,
2027 unsigned MaskOffset, unsigned NumInElem,
2028 unsigned NumInElem1, unsigned IdxOffset,
2029 std::vector<Constant*> &Mask) {
2030 unsigned NumElem1 = cast<VectorType>(J->getType())->getNumElements();
2031 for (unsigned v = 0; v < NumElem1; ++v) {
2032 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v);
2034 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context));
2036 unsigned mm = m + (int) IdxOffset;
2037 if (m >= (int) NumInElem1)
2038 mm += (int) NumInElem;
2040 Mask[v+MaskOffset] =
2041 ConstantInt::get(Type::getInt32Ty(Context), mm);
2046 // Returns the value that is to be used as the vector-shuffle mask to the
2047 // vector instruction that fuses I with J.
2048 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context,
2049 Instruction *I, Instruction *J) {
2050 // This is the shuffle mask. We need to append the second
2051 // mask to the first, and the numbers need to be adjusted.
2053 Type *ArgTypeI = I->getType();
2054 Type *ArgTypeJ = J->getType();
2055 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2057 unsigned NumElemI = cast<VectorType>(ArgTypeI)->getNumElements();
2059 // Get the total number of elements in the fused vector type.
2060 // By definition, this must equal the number of elements in
2062 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements();
2063 std::vector<Constant*> Mask(NumElem);
2065 Type *OpTypeI = I->getOperand(0)->getType();
2066 unsigned NumInElemI = cast<VectorType>(OpTypeI)->getNumElements();
2067 Type *OpTypeJ = J->getOperand(0)->getType();
2068 unsigned NumInElemJ = cast<VectorType>(OpTypeJ)->getNumElements();
2070 // The fused vector will be:
2071 // -----------------------------------------------------
2072 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ |
2073 // -----------------------------------------------------
2074 // from which we'll extract NumElem total elements (where the first NumElemI
2075 // of them come from the mask in I and the remainder come from the mask
2078 // For the mask from the first pair...
2079 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI,
2082 // For the mask from the second pair...
2083 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ,
2086 return ConstantVector::get(Mask);
2089 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I,
2090 Instruction *J, unsigned o, Value *&LOp,
2092 Type *ArgTypeL, Type *ArgTypeH,
2093 bool IBeforeJ, unsigned IdxOff) {
2094 bool ExpandedIEChain = false;
2095 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) {
2096 // If we have a pure insertelement chain, then this can be rewritten
2097 // into a chain that directly builds the larger type.
2098 bool PureChain = true;
2099 InsertElementInst *LIENext = LIE;
2101 if (!isa<UndefValue>(LIENext->getOperand(0)) &&
2102 !isa<InsertElementInst>(LIENext->getOperand(0))) {
2107 dyn_cast<InsertElementInst>(LIENext->getOperand(0))));
2110 SmallVector<Value *, 8> VectElemts(numElemL,
2111 UndefValue::get(ArgTypeL->getScalarType()));
2112 InsertElementInst *LIENext = LIE;
2115 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue();
2116 VectElemts[Idx] = LIENext->getOperand(1);
2118 dyn_cast<InsertElementInst>(LIENext->getOperand(0))));
2121 Value *LIEPrev = UndefValue::get(ArgTypeH);
2122 for (unsigned i = 0; i < numElemL; ++i) {
2123 if (isa<UndefValue>(VectElemts[i])) continue;
2124 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i],
2125 ConstantInt::get(Type::getInt32Ty(Context),
2127 getReplacementName(IBeforeJ ? I : J,
2129 LIENext->insertBefore(IBeforeJ ? J : I);
2133 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH);
2134 ExpandedIEChain = true;
2138 return ExpandedIEChain;
2141 // Returns the value to be used as the specified operand of the vector
2142 // instruction that fuses I with J.
2143 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I,
2144 Instruction *J, unsigned o, bool IBeforeJ) {
2145 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2146 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1);
2148 // Compute the fused vector type for this operand
2149 Type *ArgTypeI = I->getOperand(o)->getType();
2150 Type *ArgTypeJ = J->getOperand(o)->getType();
2151 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2153 Instruction *L = I, *H = J;
2154 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ;
2157 if (ArgTypeL->isVectorTy())
2158 numElemL = cast<VectorType>(ArgTypeL)->getNumElements();
2163 if (ArgTypeH->isVectorTy())
2164 numElemH = cast<VectorType>(ArgTypeH)->getNumElements();
2168 Value *LOp = L->getOperand(o);
2169 Value *HOp = H->getOperand(o);
2170 unsigned numElem = VArgType->getNumElements();
2172 // First, we check if we can reuse the "original" vector outputs (if these
2173 // exist). We might need a shuffle.
2174 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp);
2175 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp);
2176 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp);
2177 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp);
2179 // FIXME: If we're fusing shuffle instructions, then we can't apply this
2180 // optimization. The input vectors to the shuffle might be a different
2181 // length from the shuffle outputs. Unfortunately, the replacement
2182 // shuffle mask has already been formed, and the mask entries are sensitive
2183 // to the sizes of the inputs.
2184 bool IsSizeChangeShuffle =
2185 isa<ShuffleVectorInst>(L) &&
2186 (LOp->getType() != L->getType() || HOp->getType() != H->getType());
2188 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) {
2189 // We can have at most two unique vector inputs.
2190 bool CanUseInputs = true;
2193 I1 = LEE->getOperand(0);
2195 I1 = LSV->getOperand(0);
2196 I2 = LSV->getOperand(1);
2197 if (I2 == I1 || isa<UndefValue>(I2))
2202 Value *I3 = HEE->getOperand(0);
2203 if (!I2 && I3 != I1)
2205 else if (I3 != I1 && I3 != I2)
2206 CanUseInputs = false;
2208 Value *I3 = HSV->getOperand(0);
2209 if (!I2 && I3 != I1)
2211 else if (I3 != I1 && I3 != I2)
2212 CanUseInputs = false;
2215 Value *I4 = HSV->getOperand(1);
2216 if (!isa<UndefValue>(I4)) {
2217 if (!I2 && I4 != I1)
2219 else if (I4 != I1 && I4 != I2)
2220 CanUseInputs = false;
2227 cast<VectorType>(cast<Instruction>(LOp)->getOperand(0)->getType())
2230 cast<VectorType>(cast<Instruction>(HOp)->getOperand(0)->getType())
2233 // We have one or two input vectors. We need to map each index of the
2234 // operands to the index of the original vector.
2235 SmallVector<std::pair<int, int>, 8> II(numElem);
2236 for (unsigned i = 0; i < numElemL; ++i) {
2240 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue();
2241 INum = LEE->getOperand(0) == I1 ? 0 : 1;
2243 Idx = LSV->getMaskValue(i);
2244 if (Idx < (int) LOpElem) {
2245 INum = LSV->getOperand(0) == I1 ? 0 : 1;
2248 INum = LSV->getOperand(1) == I1 ? 0 : 1;
2252 II[i] = std::pair<int, int>(Idx, INum);
2254 for (unsigned i = 0; i < numElemH; ++i) {
2258 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue();
2259 INum = HEE->getOperand(0) == I1 ? 0 : 1;
2261 Idx = HSV->getMaskValue(i);
2262 if (Idx < (int) HOpElem) {
2263 INum = HSV->getOperand(0) == I1 ? 0 : 1;
2266 INum = HSV->getOperand(1) == I1 ? 0 : 1;
2270 II[i + numElemL] = std::pair<int, int>(Idx, INum);
2273 // We now have an array which tells us from which index of which
2274 // input vector each element of the operand comes.
2275 VectorType *I1T = cast<VectorType>(I1->getType());
2276 unsigned I1Elem = I1T->getNumElements();
2279 // In this case there is only one underlying vector input. Check for
2280 // the trivial case where we can use the input directly.
2281 if (I1Elem == numElem) {
2282 bool ElemInOrder = true;
2283 for (unsigned i = 0; i < numElem; ++i) {
2284 if (II[i].first != (int) i && II[i].first != -1) {
2285 ElemInOrder = false;
2294 // A shuffle is needed.
2295 std::vector<Constant *> Mask(numElem);
2296 for (unsigned i = 0; i < numElem; ++i) {
2297 int Idx = II[i].first;
2299 Mask[i] = UndefValue::get(Type::getInt32Ty(Context));
2301 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2305 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2306 ConstantVector::get(Mask),
2307 getReplacementName(IBeforeJ ? I : J,
2309 S->insertBefore(IBeforeJ ? J : I);
2313 VectorType *I2T = cast<VectorType>(I2->getType());
2314 unsigned I2Elem = I2T->getNumElements();
2316 // This input comes from two distinct vectors. The first step is to
2317 // make sure that both vectors are the same length. If not, the
2318 // smaller one will need to grow before they can be shuffled together.
2319 if (I1Elem < I2Elem) {
2320 std::vector<Constant *> Mask(I2Elem);
2322 for (; v < I1Elem; ++v)
2323 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2324 for (; v < I2Elem; ++v)
2325 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2327 Instruction *NewI1 =
2328 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2329 ConstantVector::get(Mask),
2330 getReplacementName(IBeforeJ ? I : J,
2332 NewI1->insertBefore(IBeforeJ ? J : I);
2336 } else if (I1Elem > I2Elem) {
2337 std::vector<Constant *> Mask(I1Elem);
2339 for (; v < I2Elem; ++v)
2340 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2341 for (; v < I1Elem; ++v)
2342 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2344 Instruction *NewI2 =
2345 new ShuffleVectorInst(I2, UndefValue::get(I2T),
2346 ConstantVector::get(Mask),
2347 getReplacementName(IBeforeJ ? I : J,
2349 NewI2->insertBefore(IBeforeJ ? J : I);
2355 // Now that both I1 and I2 are the same length we can shuffle them
2356 // together (and use the result).
2357 std::vector<Constant *> Mask(numElem);
2358 for (unsigned v = 0; v < numElem; ++v) {
2359 if (II[v].first == -1) {
2360 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2362 int Idx = II[v].first + II[v].second * I1Elem;
2363 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2367 Instruction *NewOp =
2368 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask),
2369 getReplacementName(IBeforeJ ? I : J, true, o));
2370 NewOp->insertBefore(IBeforeJ ? J : I);
2375 Type *ArgType = ArgTypeL;
2376 if (numElemL < numElemH) {
2377 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH,
2378 ArgTypeL, VArgType, IBeforeJ, 1)) {
2379 // This is another short-circuit case: we're combining a scalar into
2380 // a vector that is formed by an IE chain. We've just expanded the IE
2381 // chain, now insert the scalar and we're done.
2383 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0,
2384 getReplacementName(IBeforeJ ? I : J, true, o));
2385 S->insertBefore(IBeforeJ ? J : I);
2387 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL,
2388 ArgTypeH, IBeforeJ)) {
2389 // The two vector inputs to the shuffle must be the same length,
2390 // so extend the smaller vector to be the same length as the larger one.
2394 std::vector<Constant *> Mask(numElemH);
2396 for (; v < numElemL; ++v)
2397 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2398 for (; v < numElemH; ++v)
2399 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2401 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL),
2402 ConstantVector::get(Mask),
2403 getReplacementName(IBeforeJ ? I : J,
2406 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0,
2407 getReplacementName(IBeforeJ ? I : J,
2411 NLOp->insertBefore(IBeforeJ ? J : I);
2416 } else if (numElemL > numElemH) {
2417 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL,
2418 ArgTypeH, VArgType, IBeforeJ)) {
2420 InsertElementInst::Create(LOp, HOp,
2421 ConstantInt::get(Type::getInt32Ty(Context),
2423 getReplacementName(IBeforeJ ? I : J,
2425 S->insertBefore(IBeforeJ ? J : I);
2427 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH,
2428 ArgTypeL, IBeforeJ)) {
2431 std::vector<Constant *> Mask(numElemL);
2433 for (; v < numElemH; ++v)
2434 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2435 for (; v < numElemL; ++v)
2436 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2438 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH),
2439 ConstantVector::get(Mask),
2440 getReplacementName(IBeforeJ ? I : J,
2443 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0,
2444 getReplacementName(IBeforeJ ? I : J,
2448 NHOp->insertBefore(IBeforeJ ? J : I);
2453 if (ArgType->isVectorTy()) {
2454 unsigned numElem = cast<VectorType>(VArgType)->getNumElements();
2455 std::vector<Constant*> Mask(numElem);
2456 for (unsigned v = 0; v < numElem; ++v) {
2458 // If the low vector was expanded, we need to skip the extra
2459 // undefined entries.
2460 if (v >= numElemL && numElemH > numElemL)
2461 Idx += (numElemH - numElemL);
2462 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2465 Instruction *BV = new ShuffleVectorInst(LOp, HOp,
2466 ConstantVector::get(Mask),
2467 getReplacementName(IBeforeJ ? I : J, true, o));
2468 BV->insertBefore(IBeforeJ ? J : I);
2472 Instruction *BV1 = InsertElementInst::Create(
2473 UndefValue::get(VArgType), LOp, CV0,
2474 getReplacementName(IBeforeJ ? I : J,
2476 BV1->insertBefore(IBeforeJ ? J : I);
2477 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1,
2478 getReplacementName(IBeforeJ ? I : J,
2480 BV2->insertBefore(IBeforeJ ? J : I);
2484 // This function creates an array of values that will be used as the inputs
2485 // to the vector instruction that fuses I with J.
2486 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context,
2487 Instruction *I, Instruction *J,
2488 SmallVector<Value *, 3> &ReplacedOperands,
2490 unsigned NumOperands = I->getNumOperands();
2492 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) {
2493 // Iterate backward so that we look at the store pointer
2494 // first and know whether or not we need to flip the inputs.
2496 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) {
2497 // This is the pointer for a load/store instruction.
2498 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o);
2500 } else if (isa<CallInst>(I)) {
2501 Function *F = cast<CallInst>(I)->getCalledFunction();
2502 unsigned IID = F->getIntrinsicID();
2503 if (o == NumOperands-1) {
2504 BasicBlock &BB = *I->getParent();
2506 Module *M = BB.getParent()->getParent();
2507 Type *ArgTypeI = I->getType();
2508 Type *ArgTypeJ = J->getType();
2509 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2511 ReplacedOperands[o] = Intrinsic::getDeclaration(M,
2512 (Intrinsic::ID) IID, VArgType);
2514 } else if (IID == Intrinsic::powi && o == 1) {
2515 // The second argument of powi is a single integer and we've already
2516 // checked that both arguments are equal. As a result, we just keep
2517 // I's second argument.
2518 ReplacedOperands[o] = I->getOperand(o);
2521 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) {
2522 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J);
2526 ReplacedOperands[o] = getReplacementInput(Context, I, J, o, IBeforeJ);
2530 // This function creates two values that represent the outputs of the
2531 // original I and J instructions. These are generally vector shuffles
2532 // or extracts. In many cases, these will end up being unused and, thus,
2533 // eliminated by later passes.
2534 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
2535 Instruction *J, Instruction *K,
2536 Instruction *&InsertionPt,
2537 Instruction *&K1, Instruction *&K2) {
2538 if (isa<StoreInst>(I)) {
2539 AA->replaceWithNewValue(I, K);
2540 AA->replaceWithNewValue(J, K);
2542 Type *IType = I->getType();
2543 Type *JType = J->getType();
2545 VectorType *VType = getVecTypeForPair(IType, JType);
2546 unsigned numElem = VType->getNumElements();
2548 unsigned numElemI, numElemJ;
2549 if (IType->isVectorTy())
2550 numElemI = cast<VectorType>(IType)->getNumElements();
2554 if (JType->isVectorTy())
2555 numElemJ = cast<VectorType>(JType)->getNumElements();
2559 if (IType->isVectorTy()) {
2560 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI);
2561 for (unsigned v = 0; v < numElemI; ++v) {
2562 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2563 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v);
2566 K1 = new ShuffleVectorInst(K, UndefValue::get(VType),
2567 ConstantVector::get( Mask1),
2568 getReplacementName(K, false, 1));
2570 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2571 K1 = ExtractElementInst::Create(K, CV0,
2572 getReplacementName(K, false, 1));
2575 if (JType->isVectorTy()) {
2576 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ);
2577 for (unsigned v = 0; v < numElemJ; ++v) {
2578 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2579 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v);
2582 K2 = new ShuffleVectorInst(K, UndefValue::get(VType),
2583 ConstantVector::get( Mask2),
2584 getReplacementName(K, false, 2));
2586 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1);
2587 K2 = ExtractElementInst::Create(K, CV1,
2588 getReplacementName(K, false, 2));
2592 K2->insertAfter(K1);
2597 // Move all uses of the function I (including pairing-induced uses) after J.
2598 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB,
2599 std::multimap<Value *, Value *> &LoadMoveSet,
2600 Instruction *I, Instruction *J) {
2601 // Skip to the first instruction past I.
2602 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2604 DenseSet<Value *> Users;
2605 AliasSetTracker WriteSet(*AA);
2606 for (; cast<Instruction>(L) != J; ++L)
2607 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet);
2609 assert(cast<Instruction>(L) == J &&
2610 "Tracking has not proceeded far enough to check for dependencies");
2611 // If J is now in the use set of I, then trackUsesOfI will return true
2612 // and we have a dependency cycle (and the fusing operation must abort).
2613 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSet);
2616 // Move all uses of the function I (including pairing-induced uses) after J.
2617 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB,
2618 std::multimap<Value *, Value *> &LoadMoveSet,
2619 Instruction *&InsertionPt,
2620 Instruction *I, Instruction *J) {
2621 // Skip to the first instruction past I.
2622 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2624 DenseSet<Value *> Users;
2625 AliasSetTracker WriteSet(*AA);
2626 for (; cast<Instruction>(L) != J;) {
2627 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet)) {
2628 // Move this instruction
2629 Instruction *InstToMove = L; ++L;
2631 DEBUG(dbgs() << "BBV: moving: " << *InstToMove <<
2632 " to after " << *InsertionPt << "\n");
2633 InstToMove->removeFromParent();
2634 InstToMove->insertAfter(InsertionPt);
2635 InsertionPt = InstToMove;
2642 // Collect all load instruction that are in the move set of a given first
2643 // pair member. These loads depend on the first instruction, I, and so need
2644 // to be moved after J (the second instruction) when the pair is fused.
2645 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB,
2646 DenseMap<Value *, Value *> &ChosenPairs,
2647 std::multimap<Value *, Value *> &LoadMoveSet,
2649 // Skip to the first instruction past I.
2650 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2652 DenseSet<Value *> Users;
2653 AliasSetTracker WriteSet(*AA);
2655 // Note: We cannot end the loop when we reach J because J could be moved
2656 // farther down the use chain by another instruction pairing. Also, J
2657 // could be before I if this is an inverted input.
2658 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) {
2659 if (trackUsesOfI(Users, WriteSet, I, L)) {
2660 if (L->mayReadFromMemory())
2661 LoadMoveSet.insert(ValuePair(L, I));
2666 // In cases where both load/stores and the computation of their pointers
2667 // are chosen for vectorization, we can end up in a situation where the
2668 // aliasing analysis starts returning different query results as the
2669 // process of fusing instruction pairs continues. Because the algorithm
2670 // relies on finding the same use trees here as were found earlier, we'll
2671 // need to precompute the necessary aliasing information here and then
2672 // manually update it during the fusion process.
2673 void BBVectorize::collectLoadMoveSet(BasicBlock &BB,
2674 std::vector<Value *> &PairableInsts,
2675 DenseMap<Value *, Value *> &ChosenPairs,
2676 std::multimap<Value *, Value *> &LoadMoveSet) {
2677 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
2678 PIE = PairableInsts.end(); PI != PIE; ++PI) {
2679 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI);
2680 if (P == ChosenPairs.end()) continue;
2682 Instruction *I = cast<Instruction>(P->first);
2683 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, I);
2687 // When the first instruction in each pair is cloned, it will inherit its
2688 // parent's metadata. This metadata must be combined with that of the other
2689 // instruction in a safe way.
2690 void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) {
2691 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata;
2692 K->getAllMetadataOtherThanDebugLoc(Metadata);
2693 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) {
2694 unsigned Kind = Metadata[i].first;
2695 MDNode *JMD = J->getMetadata(Kind);
2696 MDNode *KMD = Metadata[i].second;
2700 K->setMetadata(Kind, 0); // Remove unknown metadata
2702 case LLVMContext::MD_tbaa:
2703 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD));
2705 case LLVMContext::MD_fpmath:
2706 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD));
2712 // This function fuses the chosen instruction pairs into vector instructions,
2713 // taking care preserve any needed scalar outputs and, then, it reorders the
2714 // remaining instructions as needed (users of the first member of the pair
2715 // need to be moved to after the location of the second member of the pair
2716 // because the vector instruction is inserted in the location of the pair's
2718 void BBVectorize::fuseChosenPairs(BasicBlock &BB,
2719 std::vector<Value *> &PairableInsts,
2720 DenseMap<Value *, Value *> &ChosenPairs,
2721 DenseSet<ValuePair> &FixedOrderPairs,
2722 DenseMap<VPPair, unsigned> &PairConnectionTypes,
2723 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
2724 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps) {
2725 LLVMContext& Context = BB.getContext();
2727 // During the vectorization process, the order of the pairs to be fused
2728 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs
2729 // list. After a pair is fused, the flipped pair is removed from the list.
2730 DenseSet<ValuePair> FlippedPairs;
2731 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(),
2732 E = ChosenPairs.end(); P != E; ++P)
2733 FlippedPairs.insert(ValuePair(P->second, P->first));
2734 for (DenseSet<ValuePair>::iterator P = FlippedPairs.begin(),
2735 E = FlippedPairs.end(); P != E; ++P)
2736 ChosenPairs.insert(*P);
2738 std::multimap<Value *, Value *> LoadMoveSet;
2739 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, LoadMoveSet);
2741 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n");
2743 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) {
2744 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI);
2745 if (P == ChosenPairs.end()) {
2750 if (getDepthFactor(P->first) == 0) {
2751 // These instructions are not really fused, but are tracked as though
2752 // they are. Any case in which it would be interesting to fuse them
2753 // will be taken care of by InstCombine.
2759 Instruction *I = cast<Instruction>(P->first),
2760 *J = cast<Instruction>(P->second);
2762 DEBUG(dbgs() << "BBV: fusing: " << *I <<
2763 " <-> " << *J << "\n");
2765 // Remove the pair and flipped pair from the list.
2766 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second);
2767 assert(FP != ChosenPairs.end() && "Flipped pair not found in list");
2768 ChosenPairs.erase(FP);
2769 ChosenPairs.erase(P);
2771 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSet, I, J)) {
2772 DEBUG(dbgs() << "BBV: fusion of: " << *I <<
2774 " aborted because of non-trivial dependency cycle\n");
2780 // If the pair must have the other order, then flip it.
2781 bool FlipPairOrder = FixedOrderPairs.count(ValuePair(J, I));
2782 if (!FlipPairOrder && !FixedOrderPairs.count(ValuePair(I, J))) {
2783 // This pair does not have a fixed order, and so we might want to
2784 // flip it if that will yield fewer shuffles. We count the number
2785 // of dependencies connected via swaps, and those directly connected,
2786 // and flip the order if the number of swaps is greater.
2787 bool OrigOrder = true;
2788 VPPIteratorPair IP = ConnectedPairDeps.equal_range(ValuePair(I, J));
2789 if (IP.first == ConnectedPairDeps.end()) {
2790 IP = ConnectedPairDeps.equal_range(ValuePair(J, I));
2794 if (IP.first != ConnectedPairDeps.end()) {
2795 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
2796 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first;
2797 Q != IP.second; ++Q) {
2798 DenseMap<VPPair, unsigned>::iterator R =
2799 PairConnectionTypes.find(VPPair(Q->second, Q->first));
2800 assert(R != PairConnectionTypes.end() &&
2801 "Cannot find pair connection type");
2802 if (R->second == PairConnectionDirect)
2804 else if (R->second == PairConnectionSwap)
2809 std::swap(NumDepsDirect, NumDepsSwap);
2811 if (NumDepsSwap > NumDepsDirect) {
2812 FlipPairOrder = true;
2813 DEBUG(dbgs() << "BBV: reordering pair: " << *I <<
2814 " <-> " << *J << "\n");
2819 Instruction *L = I, *H = J;
2823 // If the pair being fused uses the opposite order from that in the pair
2824 // connection map, then we need to flip the types.
2825 VPPIteratorPair IP = ConnectedPairs.equal_range(ValuePair(H, L));
2826 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first;
2827 Q != IP.second; ++Q) {
2828 DenseMap<VPPair, unsigned>::iterator R = PairConnectionTypes.find(*Q);
2829 assert(R != PairConnectionTypes.end() &&
2830 "Cannot find pair connection type");
2831 if (R->second == PairConnectionDirect)
2832 R->second = PairConnectionSwap;
2833 else if (R->second == PairConnectionSwap)
2834 R->second = PairConnectionDirect;
2837 bool LBeforeH = !FlipPairOrder;
2838 unsigned NumOperands = I->getNumOperands();
2839 SmallVector<Value *, 3> ReplacedOperands(NumOperands);
2840 getReplacementInputsForPair(Context, L, H, ReplacedOperands,
2843 // Make a copy of the original operation, change its type to the vector
2844 // type and replace its operands with the vector operands.
2845 Instruction *K = L->clone();
2848 else if (H->hasName())
2851 if (!isa<StoreInst>(K))
2852 K->mutateType(getVecTypeForPair(L->getType(), H->getType()));
2854 combineMetadata(K, H);
2856 for (unsigned o = 0; o < NumOperands; ++o)
2857 K->setOperand(o, ReplacedOperands[o]);
2861 // Instruction insertion point:
2862 Instruction *InsertionPt = K;
2863 Instruction *K1 = 0, *K2 = 0;
2864 replaceOutputsOfPair(Context, L, H, K, InsertionPt, K1, K2);
2866 // The use tree of the first original instruction must be moved to after
2867 // the location of the second instruction. The entire use tree of the
2868 // first instruction is disjoint from the input tree of the second
2869 // (by definition), and so commutes with it.
2871 moveUsesOfIAfterJ(BB, LoadMoveSet, InsertionPt, I, J);
2873 if (!isa<StoreInst>(I)) {
2874 L->replaceAllUsesWith(K1);
2875 H->replaceAllUsesWith(K2);
2876 AA->replaceWithNewValue(L, K1);
2877 AA->replaceWithNewValue(H, K2);
2880 // Instructions that may read from memory may be in the load move set.
2881 // Once an instruction is fused, we no longer need its move set, and so
2882 // the values of the map never need to be updated. However, when a load
2883 // is fused, we need to merge the entries from both instructions in the
2884 // pair in case those instructions were in the move set of some other
2885 // yet-to-be-fused pair. The loads in question are the keys of the map.
2886 if (I->mayReadFromMemory()) {
2887 std::vector<ValuePair> NewSetMembers;
2888 VPIteratorPair IPairRange = LoadMoveSet.equal_range(I);
2889 VPIteratorPair JPairRange = LoadMoveSet.equal_range(J);
2890 for (std::multimap<Value *, Value *>::iterator N = IPairRange.first;
2891 N != IPairRange.second; ++N)
2892 NewSetMembers.push_back(ValuePair(K, N->second));
2893 for (std::multimap<Value *, Value *>::iterator N = JPairRange.first;
2894 N != JPairRange.second; ++N)
2895 NewSetMembers.push_back(ValuePair(K, N->second));
2896 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(),
2897 AE = NewSetMembers.end(); A != AE; ++A)
2898 LoadMoveSet.insert(*A);
2901 // Before removing I, set the iterator to the next instruction.
2902 PI = llvm::next(BasicBlock::iterator(I));
2903 if (cast<Instruction>(PI) == J)
2908 I->eraseFromParent();
2909 J->eraseFromParent();
2911 DEBUG(if (PrintAfterEveryPair) dbgs() << "BBV: block is now: \n" <<
2915 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n");
2919 char BBVectorize::ID = 0;
2920 static const char bb_vectorize_name[] = "Basic-Block Vectorization";
2921 INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
2922 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
2923 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
2924 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
2925 INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
2927 BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) {
2928 return new BBVectorize(C);
2932 llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) {
2933 BBVectorize BBVectorizer(P, C);
2934 return BBVectorizer.vectorizeBB(BB);
2937 //===----------------------------------------------------------------------===//
2938 VectorizeConfig::VectorizeConfig() {
2939 VectorBits = ::VectorBits;
2940 VectorizeBools = !::NoBools;
2941 VectorizeInts = !::NoInts;
2942 VectorizeFloats = !::NoFloats;
2943 VectorizePointers = !::NoPointers;
2944 VectorizeCasts = !::NoCasts;
2945 VectorizeMath = !::NoMath;
2946 VectorizeFMA = !::NoFMA;
2947 VectorizeSelect = !::NoSelect;
2948 VectorizeCmp = !::NoCmp;
2949 VectorizeGEP = !::NoGEP;
2950 VectorizeMemOps = !::NoMemOps;
2951 AlignedOnly = ::AlignedOnly;
2952 ReqChainDepth= ::ReqChainDepth;
2953 SearchLimit = ::SearchLimit;
2954 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck;
2955 SplatBreaksChain = ::SplatBreaksChain;
2956 MaxInsts = ::MaxInsts;
2957 MaxIter = ::MaxIter;
2958 Pow2LenOnly = ::Pow2LenOnly;
2959 NoMemOpBoost = ::NoMemOpBoost;
2960 FastDep = ::FastDep;