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();
489 // Returns the weight associated with the provided value. A chain of
490 // candidate pairs has a length given by the sum of the weights of its
491 // members (one weight per pair; the weight of each member of the pair
492 // is assumed to be the same). This length is then compared to the
493 // chain-length threshold to determine if a given chain is significant
494 // enough to be vectorized. The length is also used in comparing
495 // candidate chains where longer chains are considered to be better.
496 // Note: when this function returns 0, the resulting instructions are
497 // not actually fused.
498 inline size_t getDepthFactor(Value *V) {
499 // InsertElement and ExtractElement have a depth factor of zero. This is
500 // for two reasons: First, they cannot be usefully fused. Second, because
501 // the pass generates a lot of these, they can confuse the simple metric
502 // used to compare the trees in the next iteration. Thus, giving them a
503 // weight of zero allows the pass to essentially ignore them in
504 // subsequent iterations when looking for vectorization opportunities
505 // while still tracking dependency chains that flow through those
507 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V))
510 // Give a load or store half of the required depth so that load/store
511 // pairs will vectorize.
512 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V)))
513 return Config.ReqChainDepth/2;
518 // Returns the cost of the provided instruction using VTTI.
519 // This does not handle loads and stores.
520 unsigned getInstrCost(unsigned Opcode, Type *T1, Type *T2) {
523 case Instruction::GetElementPtr:
524 // We mark this instruction as zero-cost because scalar GEPs are usually
525 // lowered to the intruction addressing mode. At the moment we don't
526 // generate vector GEPs.
528 case Instruction::Br:
529 return VTTI->getCFInstrCost(Opcode);
530 case Instruction::PHI:
532 case Instruction::Add:
533 case Instruction::FAdd:
534 case Instruction::Sub:
535 case Instruction::FSub:
536 case Instruction::Mul:
537 case Instruction::FMul:
538 case Instruction::UDiv:
539 case Instruction::SDiv:
540 case Instruction::FDiv:
541 case Instruction::URem:
542 case Instruction::SRem:
543 case Instruction::FRem:
544 case Instruction::Shl:
545 case Instruction::LShr:
546 case Instruction::AShr:
547 case Instruction::And:
548 case Instruction::Or:
549 case Instruction::Xor:
550 return VTTI->getArithmeticInstrCost(Opcode, T1);
551 case Instruction::Select:
552 case Instruction::ICmp:
553 case Instruction::FCmp:
554 return VTTI->getCmpSelInstrCost(Opcode, T1, T2);
555 case Instruction::ZExt:
556 case Instruction::SExt:
557 case Instruction::FPToUI:
558 case Instruction::FPToSI:
559 case Instruction::FPExt:
560 case Instruction::PtrToInt:
561 case Instruction::IntToPtr:
562 case Instruction::SIToFP:
563 case Instruction::UIToFP:
564 case Instruction::Trunc:
565 case Instruction::FPTrunc:
566 case Instruction::BitCast:
567 case Instruction::ShuffleVector:
568 return VTTI->getCastInstrCost(Opcode, T1, T2);
574 // This determines the relative offset of two loads or stores, returning
575 // true if the offset could be determined to be some constant value.
576 // For example, if OffsetInElmts == 1, then J accesses the memory directly
577 // after I; if OffsetInElmts == -1 then I accesses the memory
579 bool getPairPtrInfo(Instruction *I, Instruction *J,
580 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment,
581 unsigned &IAddressSpace, unsigned &JAddressSpace,
582 int64_t &OffsetInElmts, bool ComputeOffset = true) {
584 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
585 LoadInst *LJ = cast<LoadInst>(J);
586 IPtr = LI->getPointerOperand();
587 JPtr = LJ->getPointerOperand();
588 IAlignment = LI->getAlignment();
589 JAlignment = LJ->getAlignment();
590 IAddressSpace = LI->getPointerAddressSpace();
591 JAddressSpace = LJ->getPointerAddressSpace();
593 StoreInst *SI = cast<StoreInst>(I), *SJ = cast<StoreInst>(J);
594 IPtr = SI->getPointerOperand();
595 JPtr = SJ->getPointerOperand();
596 IAlignment = SI->getAlignment();
597 JAlignment = SJ->getAlignment();
598 IAddressSpace = SI->getPointerAddressSpace();
599 JAddressSpace = SJ->getPointerAddressSpace();
605 const SCEV *IPtrSCEV = SE->getSCEV(IPtr);
606 const SCEV *JPtrSCEV = SE->getSCEV(JPtr);
608 // If this is a trivial offset, then we'll get something like
609 // 1*sizeof(type). With target data, which we need anyway, this will get
610 // constant folded into a number.
611 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV);
612 if (const SCEVConstant *ConstOffSCEV =
613 dyn_cast<SCEVConstant>(OffsetSCEV)) {
614 ConstantInt *IntOff = ConstOffSCEV->getValue();
615 int64_t Offset = IntOff->getSExtValue();
617 Type *VTy = cast<PointerType>(IPtr->getType())->getElementType();
618 int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy);
620 Type *VTy2 = cast<PointerType>(JPtr->getType())->getElementType();
621 if (VTy != VTy2 && Offset < 0) {
622 int64_t VTy2TSS = (int64_t) TD->getTypeStoreSize(VTy2);
623 OffsetInElmts = Offset/VTy2TSS;
624 return (abs64(Offset) % VTy2TSS) == 0;
627 OffsetInElmts = Offset/VTyTSS;
628 return (abs64(Offset) % VTyTSS) == 0;
634 // Returns true if the provided CallInst represents an intrinsic that can
636 bool isVectorizableIntrinsic(CallInst* I) {
637 Function *F = I->getCalledFunction();
638 if (!F) return false;
640 unsigned IID = F->getIntrinsicID();
641 if (!IID) return false;
646 case Intrinsic::sqrt:
647 case Intrinsic::powi:
651 case Intrinsic::log2:
652 case Intrinsic::log10:
654 case Intrinsic::exp2:
656 return Config.VectorizeMath;
658 return Config.VectorizeFMA;
662 // Returns true if J is the second element in some pair referenced by
663 // some multimap pair iterator pair.
664 template <typename V>
665 bool isSecondInIteratorPair(V J, std::pair<
666 typename std::multimap<V, V>::iterator,
667 typename std::multimap<V, V>::iterator> PairRange) {
668 for (typename std::multimap<V, V>::iterator K = PairRange.first;
669 K != PairRange.second; ++K)
670 if (K->second == J) return true;
676 // This function implements one vectorization iteration on the provided
677 // basic block. It returns true if the block is changed.
678 bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) {
680 BasicBlock::iterator Start = BB.getFirstInsertionPt();
682 std::vector<Value *> AllPairableInsts;
683 DenseMap<Value *, Value *> AllChosenPairs;
684 DenseSet<ValuePair> AllFixedOrderPairs;
685 DenseMap<VPPair, unsigned> AllPairConnectionTypes;
686 std::multimap<ValuePair, ValuePair> AllConnectedPairs, AllConnectedPairDeps;
689 std::vector<Value *> PairableInsts;
690 std::multimap<Value *, Value *> CandidatePairs;
691 DenseSet<ValuePair> FixedOrderPairs;
692 DenseMap<ValuePair, int> CandidatePairCostSavings;
693 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs,
695 CandidatePairCostSavings,
696 PairableInsts, NonPow2Len);
697 if (PairableInsts.empty()) continue;
699 // Now we have a map of all of the pairable instructions and we need to
700 // select the best possible pairing. A good pairing is one such that the
701 // users of the pair are also paired. This defines a (directed) forest
702 // over the pairs such that two pairs are connected iff the second pair
705 // Note that it only matters that both members of the second pair use some
706 // element of the first pair (to allow for splatting).
708 std::multimap<ValuePair, ValuePair> ConnectedPairs, ConnectedPairDeps;
709 DenseMap<VPPair, unsigned> PairConnectionTypes;
710 computeConnectedPairs(CandidatePairs, PairableInsts, ConnectedPairs,
711 PairConnectionTypes);
712 if (ConnectedPairs.empty()) continue;
714 for (std::multimap<ValuePair, ValuePair>::iterator
715 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
717 ConnectedPairDeps.insert(VPPair(I->second, I->first));
720 // Build the pairable-instruction dependency map
721 DenseSet<ValuePair> PairableInstUsers;
722 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers);
724 // There is now a graph of the connected pairs. For each variable, pick
725 // the pairing with the largest tree meeting the depth requirement on at
726 // least one branch. Then select all pairings that are part of that tree
727 // and remove them from the list of available pairings and pairable
730 DenseMap<Value *, Value *> ChosenPairs;
731 choosePairs(CandidatePairs, CandidatePairCostSavings,
732 PairableInsts, FixedOrderPairs, PairConnectionTypes,
733 ConnectedPairs, ConnectedPairDeps,
734 PairableInstUsers, ChosenPairs);
736 if (ChosenPairs.empty()) continue;
737 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(),
738 PairableInsts.end());
739 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end());
741 // Only for the chosen pairs, propagate information on fixed-order pairs,
742 // pair connections, and their types to the data structures used by the
743 // pair fusion procedures.
744 for (DenseMap<Value *, Value *>::iterator I = ChosenPairs.begin(),
745 IE = ChosenPairs.end(); I != IE; ++I) {
746 if (FixedOrderPairs.count(*I))
747 AllFixedOrderPairs.insert(*I);
748 else if (FixedOrderPairs.count(ValuePair(I->second, I->first)))
749 AllFixedOrderPairs.insert(ValuePair(I->second, I->first));
751 for (DenseMap<Value *, Value *>::iterator J = ChosenPairs.begin();
753 DenseMap<VPPair, unsigned>::iterator K =
754 PairConnectionTypes.find(VPPair(*I, *J));
755 if (K != PairConnectionTypes.end()) {
756 AllPairConnectionTypes.insert(*K);
758 K = PairConnectionTypes.find(VPPair(*J, *I));
759 if (K != PairConnectionTypes.end())
760 AllPairConnectionTypes.insert(*K);
765 for (std::multimap<ValuePair, ValuePair>::iterator
766 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
768 if (AllPairConnectionTypes.count(*I)) {
769 AllConnectedPairs.insert(*I);
770 AllConnectedPairDeps.insert(VPPair(I->second, I->first));
773 } while (ShouldContinue);
775 if (AllChosenPairs.empty()) return false;
776 NumFusedOps += AllChosenPairs.size();
778 // A set of pairs has now been selected. It is now necessary to replace the
779 // paired instructions with vector instructions. For this procedure each
780 // operand must be replaced with a vector operand. This vector is formed
781 // by using build_vector on the old operands. The replaced values are then
782 // replaced with a vector_extract on the result. Subsequent optimization
783 // passes should coalesce the build/extract combinations.
785 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs, AllFixedOrderPairs,
786 AllPairConnectionTypes,
787 AllConnectedPairs, AllConnectedPairDeps);
789 // It is important to cleanup here so that future iterations of this
790 // function have less work to do.
791 (void) SimplifyInstructionsInBlock(&BB, TD, AA->getTargetLibraryInfo());
795 // This function returns true if the provided instruction is capable of being
796 // fused into a vector instruction. This determination is based only on the
797 // type and other attributes of the instruction.
798 bool BBVectorize::isInstVectorizable(Instruction *I,
799 bool &IsSimpleLoadStore) {
800 IsSimpleLoadStore = false;
802 if (CallInst *C = dyn_cast<CallInst>(I)) {
803 if (!isVectorizableIntrinsic(C))
805 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) {
806 // Vectorize simple loads if possbile:
807 IsSimpleLoadStore = L->isSimple();
808 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
810 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) {
811 // Vectorize simple stores if possbile:
812 IsSimpleLoadStore = S->isSimple();
813 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
815 } else if (CastInst *C = dyn_cast<CastInst>(I)) {
816 // We can vectorize casts, but not casts of pointer types, etc.
817 if (!Config.VectorizeCasts)
820 Type *SrcTy = C->getSrcTy();
821 if (!SrcTy->isSingleValueType())
824 Type *DestTy = C->getDestTy();
825 if (!DestTy->isSingleValueType())
827 } else if (isa<SelectInst>(I)) {
828 if (!Config.VectorizeSelect)
830 } else if (isa<CmpInst>(I)) {
831 if (!Config.VectorizeCmp)
833 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) {
834 if (!Config.VectorizeGEP)
837 // Currently, vector GEPs exist only with one index.
838 if (G->getNumIndices() != 1)
840 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) ||
841 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) {
845 // We can't vectorize memory operations without target data
846 if (TD == 0 && IsSimpleLoadStore)
850 getInstructionTypes(I, T1, T2);
852 // Not every type can be vectorized...
853 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) ||
854 !(VectorType::isValidElementType(T2) || T2->isVectorTy()))
857 if (T1->getScalarSizeInBits() == 1) {
858 if (!Config.VectorizeBools)
861 if (!Config.VectorizeInts && T1->isIntOrIntVectorTy())
865 if (T2->getScalarSizeInBits() == 1) {
866 if (!Config.VectorizeBools)
869 if (!Config.VectorizeInts && T2->isIntOrIntVectorTy())
873 if (!Config.VectorizeFloats
874 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy()))
877 // Don't vectorize target-specific types.
878 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy())
880 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy())
883 if ((!Config.VectorizePointers || TD == 0) &&
884 (T1->getScalarType()->isPointerTy() ||
885 T2->getScalarType()->isPointerTy()))
888 if (!VTTI && (T1->getPrimitiveSizeInBits() >= Config.VectorBits ||
889 T2->getPrimitiveSizeInBits() >= Config.VectorBits))
895 // This function returns true if the two provided instructions are compatible
896 // (meaning that they can be fused into a vector instruction). This assumes
897 // that I has already been determined to be vectorizable and that J is not
898 // in the use tree of I.
899 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J,
900 bool IsSimpleLoadStore, bool NonPow2Len,
901 int &CostSavings, int &FixedOrder) {
902 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I <<
903 " <-> " << *J << "\n");
908 // Loads and stores can be merged if they have different alignments,
909 // but are otherwise the same.
910 if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment |
911 (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0)))
914 Type *IT1, *IT2, *JT1, *JT2;
915 getInstructionTypes(I, IT1, IT2);
916 getInstructionTypes(J, JT1, JT2);
917 unsigned MaxTypeBits = std::max(
918 IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(),
919 IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits());
920 if (!VTTI && MaxTypeBits > Config.VectorBits)
923 // FIXME: handle addsub-type operations!
925 if (IsSimpleLoadStore) {
927 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
928 int64_t OffsetInElmts = 0;
929 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
930 IAddressSpace, JAddressSpace,
931 OffsetInElmts) && abs64(OffsetInElmts) == 1) {
932 FixedOrder = (int) OffsetInElmts;
933 unsigned BottomAlignment = IAlignment;
934 if (OffsetInElmts < 0) BottomAlignment = JAlignment;
936 Type *aTypeI = isa<StoreInst>(I) ?
937 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType();
938 Type *aTypeJ = isa<StoreInst>(J) ?
939 cast<StoreInst>(J)->getValueOperand()->getType() : J->getType();
940 Type *VType = getVecTypeForPair(aTypeI, aTypeJ);
942 if (Config.AlignedOnly) {
943 // An aligned load or store is possible only if the instruction
944 // with the lower offset has an alignment suitable for the
947 unsigned VecAlignment = TD->getPrefTypeAlignment(VType);
948 if (BottomAlignment < VecAlignment)
953 unsigned ICost = VTTI->getMemoryOpCost(I->getOpcode(), I->getType(),
954 IAlignment, IAddressSpace);
955 unsigned JCost = VTTI->getMemoryOpCost(J->getOpcode(), J->getType(),
956 JAlignment, JAddressSpace);
957 unsigned VCost = VTTI->getMemoryOpCost(I->getOpcode(), VType,
960 if (VCost > ICost + JCost)
963 // We don't want to fuse to a type that will be split, even
964 // if the two input types will also be split and there is no other
966 unsigned VParts = VTTI->getNumberOfParts(VType);
969 else if (!VParts && VCost == ICost + JCost)
972 CostSavings = ICost + JCost - VCost;
978 unsigned ICost = getInstrCost(I->getOpcode(), IT1, IT2);
979 unsigned JCost = getInstrCost(J->getOpcode(), JT1, JT2);
980 Type *VT1 = getVecTypeForPair(IT1, JT1),
981 *VT2 = getVecTypeForPair(IT2, JT2);
982 unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2);
984 if (VCost > ICost + JCost)
987 // We don't want to fuse to a type that will be split, even
988 // if the two input types will also be split and there is no other
990 unsigned VParts = VTTI->getNumberOfParts(VT1);
993 else if (!VParts && VCost == ICost + JCost)
996 CostSavings = ICost + JCost - VCost;
999 // The powi intrinsic is special because only the first argument is
1000 // vectorized, the second arguments must be equal.
1001 CallInst *CI = dyn_cast<CallInst>(I);
1003 if (CI && (FI = CI->getCalledFunction()) &&
1004 FI->getIntrinsicID() == Intrinsic::powi) {
1006 Value *A1I = CI->getArgOperand(1),
1007 *A1J = cast<CallInst>(J)->getArgOperand(1);
1008 const SCEV *A1ISCEV = SE->getSCEV(A1I),
1009 *A1JSCEV = SE->getSCEV(A1J);
1010 return (A1ISCEV == A1JSCEV);
1016 // Figure out whether or not J uses I and update the users and write-set
1017 // structures associated with I. Specifically, Users represents the set of
1018 // instructions that depend on I. WriteSet represents the set
1019 // of memory locations that are dependent on I. If UpdateUsers is true,
1020 // and J uses I, then Users is updated to contain J and WriteSet is updated
1021 // to contain any memory locations to which J writes. The function returns
1022 // true if J uses I. By default, alias analysis is used to determine
1023 // whether J reads from memory that overlaps with a location in WriteSet.
1024 // If LoadMoveSet is not null, then it is a previously-computed multimap
1025 // where the key is the memory-based user instruction and the value is
1026 // the instruction to be compared with I. So, if LoadMoveSet is provided,
1027 // then the alias analysis is not used. This is necessary because this
1028 // function is called during the process of moving instructions during
1029 // vectorization and the results of the alias analysis are not stable during
1031 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users,
1032 AliasSetTracker &WriteSet, Instruction *I,
1033 Instruction *J, bool UpdateUsers,
1034 std::multimap<Value *, Value *> *LoadMoveSet) {
1037 // This instruction may already be marked as a user due, for example, to
1038 // being a member of a selected pair.
1043 for (User::op_iterator JU = J->op_begin(), JE = J->op_end();
1046 if (I == V || Users.count(V)) {
1051 if (!UsesI && J->mayReadFromMemory()) {
1053 VPIteratorPair JPairRange = LoadMoveSet->equal_range(J);
1054 UsesI = isSecondInIteratorPair<Value*>(I, JPairRange);
1056 for (AliasSetTracker::iterator W = WriteSet.begin(),
1057 WE = WriteSet.end(); W != WE; ++W) {
1058 if (W->aliasesUnknownInst(J, *AA)) {
1066 if (UsesI && UpdateUsers) {
1067 if (J->mayWriteToMemory()) WriteSet.add(J);
1074 // This function iterates over all instruction pairs in the provided
1075 // basic block and collects all candidate pairs for vectorization.
1076 bool BBVectorize::getCandidatePairs(BasicBlock &BB,
1077 BasicBlock::iterator &Start,
1078 std::multimap<Value *, Value *> &CandidatePairs,
1079 DenseSet<ValuePair> &FixedOrderPairs,
1080 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1081 std::vector<Value *> &PairableInsts, bool NonPow2Len) {
1082 BasicBlock::iterator E = BB.end();
1083 if (Start == E) return false;
1085 bool ShouldContinue = false, IAfterStart = false;
1086 for (BasicBlock::iterator I = Start++; I != E; ++I) {
1087 if (I == Start) IAfterStart = true;
1089 bool IsSimpleLoadStore;
1090 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue;
1092 // Look for an instruction with which to pair instruction *I...
1093 DenseSet<Value *> Users;
1094 AliasSetTracker WriteSet(*AA);
1095 bool JAfterStart = IAfterStart;
1096 BasicBlock::iterator J = llvm::next(I);
1097 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) {
1098 if (J == Start) JAfterStart = true;
1100 // Determine if J uses I, if so, exit the loop.
1101 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep);
1102 if (Config.FastDep) {
1103 // Note: For this heuristic to be effective, independent operations
1104 // must tend to be intermixed. This is likely to be true from some
1105 // kinds of grouped loop unrolling (but not the generic LLVM pass),
1106 // but otherwise may require some kind of reordering pass.
1108 // When using fast dependency analysis,
1109 // stop searching after first use:
1112 if (UsesI) continue;
1115 // J does not use I, and comes before the first use of I, so it can be
1116 // merged with I if the instructions are compatible.
1117 int CostSavings, FixedOrder;
1118 if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len,
1119 CostSavings, FixedOrder)) continue;
1121 // J is a candidate for merging with I.
1122 if (!PairableInsts.size() ||
1123 PairableInsts[PairableInsts.size()-1] != I) {
1124 PairableInsts.push_back(I);
1127 CandidatePairs.insert(ValuePair(I, J));
1129 CandidatePairCostSavings.insert(ValuePairWithCost(ValuePair(I, J),
1132 if (FixedOrder == 1)
1133 FixedOrderPairs.insert(ValuePair(I, J));
1134 else if (FixedOrder == -1)
1135 FixedOrderPairs.insert(ValuePair(J, I));
1137 // The next call to this function must start after the last instruction
1138 // selected during this invocation.
1140 Start = llvm::next(J);
1141 IAfterStart = JAfterStart = false;
1144 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair "
1145 << *I << " <-> " << *J << " (cost savings: " <<
1146 CostSavings << ")\n");
1148 // If we have already found too many pairs, break here and this function
1149 // will be called again starting after the last instruction selected
1150 // during this invocation.
1151 if (PairableInsts.size() >= Config.MaxInsts) {
1152 ShouldContinue = true;
1161 DEBUG(dbgs() << "BBV: found " << PairableInsts.size()
1162 << " instructions with candidate pairs\n");
1164 return ShouldContinue;
1167 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that
1168 // it looks for pairs such that both members have an input which is an
1169 // output of PI or PJ.
1170 void BBVectorize::computePairsConnectedTo(
1171 std::multimap<Value *, Value *> &CandidatePairs,
1172 std::vector<Value *> &PairableInsts,
1173 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1174 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1178 // For each possible pairing for this variable, look at the uses of
1179 // the first value...
1180 for (Value::use_iterator I = P.first->use_begin(),
1181 E = P.first->use_end(); I != E; ++I) {
1182 if (isa<LoadInst>(*I)) {
1183 // A pair cannot be connected to a load because the load only takes one
1184 // operand (the address) and it is a scalar even after vectorization.
1186 } else if ((SI = dyn_cast<StoreInst>(*I)) &&
1187 P.first == SI->getPointerOperand()) {
1188 // Similarly, a pair cannot be connected to a store through its
1193 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I);
1195 // For each use of the first variable, look for uses of the second
1197 for (Value::use_iterator J = P.second->use_begin(),
1198 E2 = P.second->use_end(); J != E2; ++J) {
1199 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1200 P.second == SJ->getPointerOperand())
1203 VPIteratorPair JPairRange = CandidatePairs.equal_range(*J);
1206 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) {
1207 VPPair VP(P, ValuePair(*I, *J));
1208 ConnectedPairs.insert(VP);
1209 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionDirect));
1213 if (isSecondInIteratorPair<Value*>(*I, JPairRange)) {
1214 VPPair VP(P, ValuePair(*J, *I));
1215 ConnectedPairs.insert(VP);
1216 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSwap));
1220 if (Config.SplatBreaksChain) continue;
1221 // Look for cases where just the first value in the pair is used by
1222 // both members of another pair (splatting).
1223 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) {
1224 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1225 P.first == SJ->getPointerOperand())
1228 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) {
1229 VPPair VP(P, ValuePair(*I, *J));
1230 ConnectedPairs.insert(VP);
1231 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1236 if (Config.SplatBreaksChain) return;
1237 // Look for cases where just the second value in the pair is used by
1238 // both members of another pair (splatting).
1239 for (Value::use_iterator I = P.second->use_begin(),
1240 E = P.second->use_end(); I != E; ++I) {
1241 if (isa<LoadInst>(*I))
1243 else if ((SI = dyn_cast<StoreInst>(*I)) &&
1244 P.second == SI->getPointerOperand())
1247 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I);
1249 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) {
1250 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1251 P.second == SJ->getPointerOperand())
1254 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) {
1255 VPPair VP(P, ValuePair(*I, *J));
1256 ConnectedPairs.insert(VP);
1257 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1263 // This function figures out which pairs are connected. Two pairs are
1264 // connected if some output of the first pair forms an input to both members
1265 // of the second pair.
1266 void BBVectorize::computeConnectedPairs(
1267 std::multimap<Value *, Value *> &CandidatePairs,
1268 std::vector<Value *> &PairableInsts,
1269 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1270 DenseMap<VPPair, unsigned> &PairConnectionTypes) {
1272 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
1273 PE = PairableInsts.end(); PI != PE; ++PI) {
1274 VPIteratorPair choiceRange = CandidatePairs.equal_range(*PI);
1276 for (std::multimap<Value *, Value *>::iterator P = choiceRange.first;
1277 P != choiceRange.second; ++P)
1278 computePairsConnectedTo(CandidatePairs, PairableInsts,
1279 ConnectedPairs, PairConnectionTypes, *P);
1282 DEBUG(dbgs() << "BBV: found " << ConnectedPairs.size()
1283 << " pair connections.\n");
1286 // This function builds a set of use tuples such that <A, B> is in the set
1287 // if B is in the use tree of A. If B is in the use tree of A, then B
1288 // depends on the output of A.
1289 void BBVectorize::buildDepMap(
1291 std::multimap<Value *, Value *> &CandidatePairs,
1292 std::vector<Value *> &PairableInsts,
1293 DenseSet<ValuePair> &PairableInstUsers) {
1294 DenseSet<Value *> IsInPair;
1295 for (std::multimap<Value *, Value *>::iterator C = CandidatePairs.begin(),
1296 E = CandidatePairs.end(); C != E; ++C) {
1297 IsInPair.insert(C->first);
1298 IsInPair.insert(C->second);
1301 // Iterate through the basic block, recording all Users of each
1302 // pairable instruction.
1304 BasicBlock::iterator E = BB.end();
1305 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) {
1306 if (IsInPair.find(I) == IsInPair.end()) continue;
1308 DenseSet<Value *> Users;
1309 AliasSetTracker WriteSet(*AA);
1310 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J)
1311 (void) trackUsesOfI(Users, WriteSet, I, J);
1313 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end();
1315 PairableInstUsers.insert(ValuePair(I, *U));
1319 // Returns true if an input to pair P is an output of pair Q and also an
1320 // input of pair Q is an output of pair P. If this is the case, then these
1321 // two pairs cannot be simultaneously fused.
1322 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q,
1323 DenseSet<ValuePair> &PairableInstUsers,
1324 std::multimap<ValuePair, ValuePair> *PairableInstUserMap) {
1325 // Two pairs are in conflict if they are mutual Users of eachother.
1326 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) ||
1327 PairableInstUsers.count(ValuePair(P.first, Q.second)) ||
1328 PairableInstUsers.count(ValuePair(P.second, Q.first)) ||
1329 PairableInstUsers.count(ValuePair(P.second, Q.second));
1330 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) ||
1331 PairableInstUsers.count(ValuePair(Q.first, P.second)) ||
1332 PairableInstUsers.count(ValuePair(Q.second, P.first)) ||
1333 PairableInstUsers.count(ValuePair(Q.second, P.second));
1334 if (PairableInstUserMap) {
1335 // FIXME: The expensive part of the cycle check is not so much the cycle
1336 // check itself but this edge insertion procedure. This needs some
1337 // profiling and probably a different data structure (same is true of
1338 // most uses of std::multimap).
1340 VPPIteratorPair QPairRange = PairableInstUserMap->equal_range(Q);
1341 if (!isSecondInIteratorPair(P, QPairRange))
1342 PairableInstUserMap->insert(VPPair(Q, P));
1345 VPPIteratorPair PPairRange = PairableInstUserMap->equal_range(P);
1346 if (!isSecondInIteratorPair(Q, PPairRange))
1347 PairableInstUserMap->insert(VPPair(P, Q));
1351 return (QUsesP && PUsesQ);
1354 // This function walks the use graph of current pairs to see if, starting
1355 // from P, the walk returns to P.
1356 bool BBVectorize::pairWillFormCycle(ValuePair P,
1357 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1358 DenseSet<ValuePair> &CurrentPairs) {
1359 DEBUG(if (DebugCycleCheck)
1360 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> "
1361 << *P.second << "\n");
1362 // A lookup table of visisted pairs is kept because the PairableInstUserMap
1363 // contains non-direct associations.
1364 DenseSet<ValuePair> Visited;
1365 SmallVector<ValuePair, 32> Q;
1366 // General depth-first post-order traversal:
1369 ValuePair QTop = Q.pop_back_val();
1370 Visited.insert(QTop);
1372 DEBUG(if (DebugCycleCheck)
1373 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> "
1374 << *QTop.second << "\n");
1375 VPPIteratorPair QPairRange = PairableInstUserMap.equal_range(QTop);
1376 for (std::multimap<ValuePair, ValuePair>::iterator C = QPairRange.first;
1377 C != QPairRange.second; ++C) {
1378 if (C->second == P) {
1380 << "BBV: rejected to prevent non-trivial cycle formation: "
1381 << *C->first.first << " <-> " << *C->first.second << "\n");
1385 if (CurrentPairs.count(C->second) && !Visited.count(C->second))
1386 Q.push_back(C->second);
1388 } while (!Q.empty());
1393 // This function builds the initial tree of connected pairs with the
1394 // pair J at the root.
1395 void BBVectorize::buildInitialTreeFor(
1396 std::multimap<Value *, Value *> &CandidatePairs,
1397 std::vector<Value *> &PairableInsts,
1398 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1399 DenseSet<ValuePair> &PairableInstUsers,
1400 DenseMap<Value *, Value *> &ChosenPairs,
1401 DenseMap<ValuePair, size_t> &Tree, ValuePair J) {
1402 // Each of these pairs is viewed as the root node of a Tree. The Tree
1403 // is then walked (depth-first). As this happens, we keep track of
1404 // the pairs that compose the Tree and the maximum depth of the Tree.
1405 SmallVector<ValuePairWithDepth, 32> Q;
1406 // General depth-first post-order traversal:
1407 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1409 ValuePairWithDepth QTop = Q.back();
1411 // Push each child onto the queue:
1412 bool MoreChildren = false;
1413 size_t MaxChildDepth = QTop.second;
1414 VPPIteratorPair qtRange = ConnectedPairs.equal_range(QTop.first);
1415 for (std::multimap<ValuePair, ValuePair>::iterator k = qtRange.first;
1416 k != qtRange.second; ++k) {
1417 // Make sure that this child pair is still a candidate:
1418 bool IsStillCand = false;
1419 VPIteratorPair checkRange =
1420 CandidatePairs.equal_range(k->second.first);
1421 for (std::multimap<Value *, Value *>::iterator m = checkRange.first;
1422 m != checkRange.second; ++m) {
1423 if (m->second == k->second.second) {
1430 DenseMap<ValuePair, size_t>::iterator C = Tree.find(k->second);
1431 if (C == Tree.end()) {
1432 size_t d = getDepthFactor(k->second.first);
1433 Q.push_back(ValuePairWithDepth(k->second, QTop.second+d));
1434 MoreChildren = true;
1436 MaxChildDepth = std::max(MaxChildDepth, C->second);
1441 if (!MoreChildren) {
1442 // Record the current pair as part of the Tree:
1443 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth));
1446 } while (!Q.empty());
1449 // Given some initial tree, prune it by removing conflicting pairs (pairs
1450 // that cannot be simultaneously chosen for vectorization).
1451 void BBVectorize::pruneTreeFor(
1452 std::multimap<Value *, Value *> &CandidatePairs,
1453 std::vector<Value *> &PairableInsts,
1454 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1455 DenseSet<ValuePair> &PairableInstUsers,
1456 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1457 DenseMap<Value *, Value *> &ChosenPairs,
1458 DenseMap<ValuePair, size_t> &Tree,
1459 DenseSet<ValuePair> &PrunedTree, ValuePair J,
1460 bool UseCycleCheck) {
1461 SmallVector<ValuePairWithDepth, 32> Q;
1462 // General depth-first post-order traversal:
1463 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1465 ValuePairWithDepth QTop = Q.pop_back_val();
1466 PrunedTree.insert(QTop.first);
1468 // Visit each child, pruning as necessary...
1469 DenseMap<ValuePair, size_t> BestChildren;
1470 VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first);
1471 for (std::multimap<ValuePair, ValuePair>::iterator K = QTopRange.first;
1472 K != QTopRange.second; ++K) {
1473 DenseMap<ValuePair, size_t>::iterator C = Tree.find(K->second);
1474 if (C == Tree.end()) continue;
1476 // This child is in the Tree, now we need to make sure it is the
1477 // best of any conflicting children. There could be multiple
1478 // conflicting children, so first, determine if we're keeping
1479 // this child, then delete conflicting children as necessary.
1481 // It is also necessary to guard against pairing-induced
1482 // dependencies. Consider instructions a .. x .. y .. b
1483 // such that (a,b) are to be fused and (x,y) are to be fused
1484 // but a is an input to x and b is an output from y. This
1485 // means that y cannot be moved after b but x must be moved
1486 // after b for (a,b) to be fused. In other words, after
1487 // fusing (a,b) we have y .. a/b .. x where y is an input
1488 // to a/b and x is an output to a/b: x and y can no longer
1489 // be legally fused. To prevent this condition, we must
1490 // make sure that a child pair added to the Tree is not
1491 // both an input and output of an already-selected pair.
1493 // Pairing-induced dependencies can also form from more complicated
1494 // cycles. The pair vs. pair conflicts are easy to check, and so
1495 // that is done explicitly for "fast rejection", and because for
1496 // child vs. child conflicts, we may prefer to keep the current
1497 // pair in preference to the already-selected child.
1498 DenseSet<ValuePair> CurrentPairs;
1501 for (DenseMap<ValuePair, size_t>::iterator C2
1502 = BestChildren.begin(), E2 = BestChildren.end();
1504 if (C2->first.first == C->first.first ||
1505 C2->first.first == C->first.second ||
1506 C2->first.second == C->first.first ||
1507 C2->first.second == C->first.second ||
1508 pairsConflict(C2->first, C->first, PairableInstUsers,
1509 UseCycleCheck ? &PairableInstUserMap : 0)) {
1510 if (C2->second >= C->second) {
1515 CurrentPairs.insert(C2->first);
1518 if (!CanAdd) continue;
1520 // Even worse, this child could conflict with another node already
1521 // selected for the Tree. If that is the case, ignore this child.
1522 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(),
1523 E2 = PrunedTree.end(); T != E2; ++T) {
1524 if (T->first == C->first.first ||
1525 T->first == C->first.second ||
1526 T->second == C->first.first ||
1527 T->second == C->first.second ||
1528 pairsConflict(*T, C->first, PairableInstUsers,
1529 UseCycleCheck ? &PairableInstUserMap : 0)) {
1534 CurrentPairs.insert(*T);
1536 if (!CanAdd) continue;
1538 // And check the queue too...
1539 for (SmallVector<ValuePairWithDepth, 32>::iterator C2 = Q.begin(),
1540 E2 = Q.end(); C2 != E2; ++C2) {
1541 if (C2->first.first == C->first.first ||
1542 C2->first.first == C->first.second ||
1543 C2->first.second == C->first.first ||
1544 C2->first.second == C->first.second ||
1545 pairsConflict(C2->first, C->first, PairableInstUsers,
1546 UseCycleCheck ? &PairableInstUserMap : 0)) {
1551 CurrentPairs.insert(C2->first);
1553 if (!CanAdd) continue;
1555 // Last but not least, check for a conflict with any of the
1556 // already-chosen pairs.
1557 for (DenseMap<Value *, Value *>::iterator C2 =
1558 ChosenPairs.begin(), E2 = ChosenPairs.end();
1560 if (pairsConflict(*C2, C->first, PairableInstUsers,
1561 UseCycleCheck ? &PairableInstUserMap : 0)) {
1566 CurrentPairs.insert(*C2);
1568 if (!CanAdd) continue;
1570 // To check for non-trivial cycles formed by the addition of the
1571 // current pair we've formed a list of all relevant pairs, now use a
1572 // graph walk to check for a cycle. We start from the current pair and
1573 // walk the use tree to see if we again reach the current pair. If we
1574 // do, then the current pair is rejected.
1576 // FIXME: It may be more efficient to use a topological-ordering
1577 // algorithm to improve the cycle check. This should be investigated.
1578 if (UseCycleCheck &&
1579 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs))
1582 // This child can be added, but we may have chosen it in preference
1583 // to an already-selected child. Check for this here, and if a
1584 // conflict is found, then remove the previously-selected child
1585 // before adding this one in its place.
1586 for (DenseMap<ValuePair, size_t>::iterator C2
1587 = BestChildren.begin(); C2 != BestChildren.end();) {
1588 if (C2->first.first == C->first.first ||
1589 C2->first.first == C->first.second ||
1590 C2->first.second == C->first.first ||
1591 C2->first.second == C->first.second ||
1592 pairsConflict(C2->first, C->first, PairableInstUsers))
1593 BestChildren.erase(C2++);
1598 BestChildren.insert(ValuePairWithDepth(C->first, C->second));
1601 for (DenseMap<ValuePair, size_t>::iterator C
1602 = BestChildren.begin(), E2 = BestChildren.end();
1604 size_t DepthF = getDepthFactor(C->first.first);
1605 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF));
1607 } while (!Q.empty());
1610 // This function finds the best tree of mututally-compatible connected
1611 // pairs, given the choice of root pairs as an iterator range.
1612 void BBVectorize::findBestTreeFor(
1613 std::multimap<Value *, Value *> &CandidatePairs,
1614 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1615 std::vector<Value *> &PairableInsts,
1616 DenseSet<ValuePair> &FixedOrderPairs,
1617 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1618 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1619 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps,
1620 DenseSet<ValuePair> &PairableInstUsers,
1621 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1622 DenseMap<Value *, Value *> &ChosenPairs,
1623 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth,
1624 int &BestEffSize, VPIteratorPair ChoiceRange,
1625 bool UseCycleCheck) {
1626 for (std::multimap<Value *, Value *>::iterator J = ChoiceRange.first;
1627 J != ChoiceRange.second; ++J) {
1629 // Before going any further, make sure that this pair does not
1630 // conflict with any already-selected pairs (see comment below
1631 // near the Tree pruning for more details).
1632 DenseSet<ValuePair> ChosenPairSet;
1633 bool DoesConflict = false;
1634 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(),
1635 E = ChosenPairs.end(); C != E; ++C) {
1636 if (pairsConflict(*C, *J, PairableInstUsers,
1637 UseCycleCheck ? &PairableInstUserMap : 0)) {
1638 DoesConflict = true;
1642 ChosenPairSet.insert(*C);
1644 if (DoesConflict) continue;
1646 if (UseCycleCheck &&
1647 pairWillFormCycle(*J, PairableInstUserMap, ChosenPairSet))
1650 DenseMap<ValuePair, size_t> Tree;
1651 buildInitialTreeFor(CandidatePairs, PairableInsts, ConnectedPairs,
1652 PairableInstUsers, ChosenPairs, Tree, *J);
1654 // Because we'll keep the child with the largest depth, the largest
1655 // depth is still the same in the unpruned Tree.
1656 size_t MaxDepth = Tree.lookup(*J);
1658 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found Tree for pair {"
1659 << *J->first << " <-> " << *J->second << "} of depth " <<
1660 MaxDepth << " and size " << Tree.size() << "\n");
1662 // At this point the Tree has been constructed, but, may contain
1663 // contradictory children (meaning that different children of
1664 // some tree node may be attempting to fuse the same instruction).
1665 // So now we walk the tree again, in the case of a conflict,
1666 // keep only the child with the largest depth. To break a tie,
1667 // favor the first child.
1669 DenseSet<ValuePair> PrunedTree;
1670 pruneTreeFor(CandidatePairs, PairableInsts, ConnectedPairs,
1671 PairableInstUsers, PairableInstUserMap, ChosenPairs, Tree,
1672 PrunedTree, *J, UseCycleCheck);
1676 DenseSet<Value *> PrunedTreeInstrs;
1677 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
1678 E = PrunedTree.end(); S != E; ++S) {
1679 PrunedTreeInstrs.insert(S->first);
1680 PrunedTreeInstrs.insert(S->second);
1683 // The set of pairs that have already contributed to the total cost.
1684 DenseSet<ValuePair> IncomingPairs;
1686 // The node weights represent the cost savings associated with
1687 // fusing the pair of instructions.
1688 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
1689 E = PrunedTree.end(); S != E; ++S) {
1690 bool FlipOrder = false;
1692 if (getDepthFactor(S->first)) {
1693 int ESContrib = CandidatePairCostSavings.find(*S)->second;
1694 DEBUG(if (DebugPairSelection) dbgs() << "\tweight {"
1695 << *S->first << " <-> " << *S->second << "} = " <<
1697 EffSize += ESContrib;
1700 // The edge weights contribute in a negative sense: they represent
1701 // the cost of shuffles.
1702 VPPIteratorPair IP = ConnectedPairDeps.equal_range(*S);
1703 if (IP.first != ConnectedPairDeps.end()) {
1704 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
1705 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first;
1706 Q != IP.second; ++Q) {
1707 if (!PrunedTree.count(Q->second))
1709 DenseMap<VPPair, unsigned>::iterator R =
1710 PairConnectionTypes.find(VPPair(Q->second, Q->first));
1711 assert(R != PairConnectionTypes.end() &&
1712 "Cannot find pair connection type");
1713 if (R->second == PairConnectionDirect)
1715 else if (R->second == PairConnectionSwap)
1719 // If there are more swaps than direct connections, then
1720 // the pair order will be flipped during fusion. So the real
1721 // number of swaps is the minimum number.
1722 FlipOrder = !FixedOrderPairs.count(*S) &&
1723 ((NumDepsSwap > NumDepsDirect) ||
1724 FixedOrderPairs.count(ValuePair(S->second, S->first)));
1726 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first;
1727 Q != IP.second; ++Q) {
1728 if (!PrunedTree.count(Q->second))
1730 DenseMap<VPPair, unsigned>::iterator R =
1731 PairConnectionTypes.find(VPPair(Q->second, Q->first));
1732 assert(R != PairConnectionTypes.end() &&
1733 "Cannot find pair connection type");
1734 Type *Ty1 = Q->second.first->getType(),
1735 *Ty2 = Q->second.second->getType();
1736 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1737 if ((R->second == PairConnectionDirect && FlipOrder) ||
1738 (R->second == PairConnectionSwap && !FlipOrder) ||
1739 R->second == PairConnectionSplat) {
1740 int ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1742 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1743 *Q->second.first << " <-> " << *Q->second.second <<
1745 *S->first << " <-> " << *S->second << "} = " <<
1747 EffSize -= ESContrib;
1752 // Compute the cost of outgoing edges. We assume that edges outgoing
1753 // to shuffles, inserts or extracts can be merged, and so contribute
1754 // no additional cost.
1755 if (!S->first->getType()->isVoidTy()) {
1756 Type *Ty1 = S->first->getType(),
1757 *Ty2 = S->second->getType();
1758 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1760 bool NeedsExtraction = false;
1761 for (Value::use_iterator I = S->first->use_begin(),
1762 IE = S->first->use_end(); I != IE; ++I) {
1763 if (isa<ShuffleVectorInst>(*I) ||
1764 isa<InsertElementInst>(*I) ||
1765 isa<ExtractElementInst>(*I))
1767 if (PrunedTreeInstrs.count(*I))
1769 NeedsExtraction = true;
1773 if (NeedsExtraction) {
1775 if (Ty1->isVectorTy())
1776 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1779 ESContrib = (int) VTTI->getVectorInstrCost(
1780 Instruction::ExtractElement, VTy, 0);
1782 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1783 *S->first << "} = " << ESContrib << "\n");
1784 EffSize -= ESContrib;
1787 NeedsExtraction = false;
1788 for (Value::use_iterator I = S->second->use_begin(),
1789 IE = S->second->use_end(); I != IE; ++I) {
1790 if (isa<ShuffleVectorInst>(*I) ||
1791 isa<InsertElementInst>(*I) ||
1792 isa<ExtractElementInst>(*I))
1794 if (PrunedTreeInstrs.count(*I))
1796 NeedsExtraction = true;
1800 if (NeedsExtraction) {
1802 if (Ty2->isVectorTy())
1803 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1806 ESContrib = (int) VTTI->getVectorInstrCost(
1807 Instruction::ExtractElement, VTy, 1);
1808 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1809 *S->second << "} = " << ESContrib << "\n");
1810 EffSize -= ESContrib;
1814 // Compute the cost of incoming edges.
1815 if (!isa<LoadInst>(S->first) && !isa<StoreInst>(S->first)) {
1816 Instruction *S1 = cast<Instruction>(S->first),
1817 *S2 = cast<Instruction>(S->second);
1818 for (unsigned o = 0; o < S1->getNumOperands(); ++o) {
1819 Value *O1 = S1->getOperand(o), *O2 = S2->getOperand(o);
1821 // Combining constants into vector constants (or small vector
1822 // constants into larger ones are assumed free).
1823 if (isa<Constant>(O1) && isa<Constant>(O2))
1829 ValuePair VP = ValuePair(O1, O2);
1830 ValuePair VPR = ValuePair(O2, O1);
1832 // Internal edges are not handled here.
1833 if (PrunedTree.count(VP) || PrunedTree.count(VPR))
1836 Type *Ty1 = O1->getType(),
1837 *Ty2 = O2->getType();
1838 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1840 // Combining vector operations of the same type is also assumed
1841 // folded with other operations.
1843 (isa<ShuffleVectorInst>(O1) ||
1844 isa<InsertElementInst>(O1) ||
1845 isa<InsertElementInst>(O1)) &&
1846 (isa<ShuffleVectorInst>(O2) ||
1847 isa<InsertElementInst>(O2) ||
1848 isa<InsertElementInst>(O2)))
1852 // This pair has already been formed.
1853 if (IncomingPairs.count(VP)) {
1855 } else if (IncomingPairs.count(VPR)) {
1856 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1858 } else if (!Ty1->isVectorTy() && !Ty2->isVectorTy()) {
1859 ESContrib = (int) VTTI->getVectorInstrCost(
1860 Instruction::InsertElement, VTy, 0);
1861 ESContrib += (int) VTTI->getVectorInstrCost(
1862 Instruction::InsertElement, VTy, 1);
1863 } else if (!Ty1->isVectorTy()) {
1864 // O1 needs to be inserted into a vector of size O2, and then
1865 // both need to be shuffled together.
1866 ESContrib = (int) VTTI->getVectorInstrCost(
1867 Instruction::InsertElement, Ty2, 0);
1868 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
1870 } else if (!Ty2->isVectorTy()) {
1871 // O2 needs to be inserted into a vector of size O1, and then
1872 // both need to be shuffled together.
1873 ESContrib = (int) VTTI->getVectorInstrCost(
1874 Instruction::InsertElement, Ty1, 0);
1875 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
1878 Type *TyBig = Ty1, *TySmall = Ty2;
1879 if (Ty2->getVectorNumElements() > Ty1->getVectorNumElements())
1880 std::swap(TyBig, TySmall);
1882 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1884 if (TyBig != TySmall)
1885 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
1889 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {"
1890 << *O1 << " <-> " << *O2 << "} = " <<
1892 EffSize -= ESContrib;
1893 IncomingPairs.insert(VP);
1898 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
1899 E = PrunedTree.end(); S != E; ++S)
1900 EffSize += (int) getDepthFactor(S->first);
1903 DEBUG(if (DebugPairSelection)
1904 dbgs() << "BBV: found pruned Tree for pair {"
1905 << *J->first << " <-> " << *J->second << "} of depth " <<
1906 MaxDepth << " and size " << PrunedTree.size() <<
1907 " (effective size: " << EffSize << ")\n");
1908 if (((VTTI && !UseChainDepthWithTI) ||
1909 MaxDepth >= Config.ReqChainDepth) &&
1910 EffSize > 0 && EffSize > BestEffSize) {
1911 BestMaxDepth = MaxDepth;
1912 BestEffSize = EffSize;
1913 BestTree = PrunedTree;
1918 // Given the list of candidate pairs, this function selects those
1919 // that will be fused into vector instructions.
1920 void BBVectorize::choosePairs(
1921 std::multimap<Value *, Value *> &CandidatePairs,
1922 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1923 std::vector<Value *> &PairableInsts,
1924 DenseSet<ValuePair> &FixedOrderPairs,
1925 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1926 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1927 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps,
1928 DenseSet<ValuePair> &PairableInstUsers,
1929 DenseMap<Value *, Value *>& ChosenPairs) {
1930 bool UseCycleCheck =
1931 CandidatePairs.size() <= Config.MaxCandPairsForCycleCheck;
1932 std::multimap<ValuePair, ValuePair> PairableInstUserMap;
1933 for (std::vector<Value *>::iterator I = PairableInsts.begin(),
1934 E = PairableInsts.end(); I != E; ++I) {
1935 // The number of possible pairings for this variable:
1936 size_t NumChoices = CandidatePairs.count(*I);
1937 if (!NumChoices) continue;
1939 VPIteratorPair ChoiceRange = CandidatePairs.equal_range(*I);
1941 // The best pair to choose and its tree:
1942 size_t BestMaxDepth = 0;
1943 int BestEffSize = 0;
1944 DenseSet<ValuePair> BestTree;
1945 findBestTreeFor(CandidatePairs, CandidatePairCostSavings,
1946 PairableInsts, FixedOrderPairs, PairConnectionTypes,
1947 ConnectedPairs, ConnectedPairDeps,
1948 PairableInstUsers, PairableInstUserMap, ChosenPairs,
1949 BestTree, BestMaxDepth, BestEffSize, ChoiceRange,
1952 // A tree has been chosen (or not) at this point. If no tree was
1953 // chosen, then this instruction, I, cannot be paired (and is no longer
1956 DEBUG(if (BestTree.size() > 0)
1957 dbgs() << "BBV: selected pairs in the best tree for: "
1958 << *cast<Instruction>(*I) << "\n");
1960 for (DenseSet<ValuePair>::iterator S = BestTree.begin(),
1961 SE2 = BestTree.end(); S != SE2; ++S) {
1962 // Insert the members of this tree into the list of chosen pairs.
1963 ChosenPairs.insert(ValuePair(S->first, S->second));
1964 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " <<
1965 *S->second << "\n");
1967 // Remove all candidate pairs that have values in the chosen tree.
1968 for (std::multimap<Value *, Value *>::iterator K =
1969 CandidatePairs.begin(); K != CandidatePairs.end();) {
1970 if (K->first == S->first || K->second == S->first ||
1971 K->second == S->second || K->first == S->second) {
1972 // Don't remove the actual pair chosen so that it can be used
1973 // in subsequent tree selections.
1974 if (!(K->first == S->first && K->second == S->second))
1975 CandidatePairs.erase(K++);
1985 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n");
1988 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o,
1993 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) +
1994 (n > 0 ? "." + utostr(n) : "")).str();
1997 // Returns the value that is to be used as the pointer input to the vector
1998 // instruction that fuses I with J.
1999 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context,
2000 Instruction *I, Instruction *J, unsigned o) {
2002 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
2003 int64_t OffsetInElmts;
2005 // Note: the analysis might fail here, that is why the pair order has
2006 // been precomputed (OffsetInElmts must be unused here).
2007 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
2008 IAddressSpace, JAddressSpace,
2009 OffsetInElmts, false);
2011 // The pointer value is taken to be the one with the lowest offset.
2014 Type *ArgTypeI = cast<PointerType>(IPtr->getType())->getElementType();
2015 Type *ArgTypeJ = cast<PointerType>(JPtr->getType())->getElementType();
2016 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2017 Type *VArgPtrType = PointerType::get(VArgType,
2018 cast<PointerType>(IPtr->getType())->getAddressSpace());
2019 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o),
2020 /* insert before */ I);
2023 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J,
2024 unsigned MaskOffset, unsigned NumInElem,
2025 unsigned NumInElem1, unsigned IdxOffset,
2026 std::vector<Constant*> &Mask) {
2027 unsigned NumElem1 = cast<VectorType>(J->getType())->getNumElements();
2028 for (unsigned v = 0; v < NumElem1; ++v) {
2029 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v);
2031 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context));
2033 unsigned mm = m + (int) IdxOffset;
2034 if (m >= (int) NumInElem1)
2035 mm += (int) NumInElem;
2037 Mask[v+MaskOffset] =
2038 ConstantInt::get(Type::getInt32Ty(Context), mm);
2043 // Returns the value that is to be used as the vector-shuffle mask to the
2044 // vector instruction that fuses I with J.
2045 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context,
2046 Instruction *I, Instruction *J) {
2047 // This is the shuffle mask. We need to append the second
2048 // mask to the first, and the numbers need to be adjusted.
2050 Type *ArgTypeI = I->getType();
2051 Type *ArgTypeJ = J->getType();
2052 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2054 unsigned NumElemI = cast<VectorType>(ArgTypeI)->getNumElements();
2056 // Get the total number of elements in the fused vector type.
2057 // By definition, this must equal the number of elements in
2059 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements();
2060 std::vector<Constant*> Mask(NumElem);
2062 Type *OpTypeI = I->getOperand(0)->getType();
2063 unsigned NumInElemI = cast<VectorType>(OpTypeI)->getNumElements();
2064 Type *OpTypeJ = J->getOperand(0)->getType();
2065 unsigned NumInElemJ = cast<VectorType>(OpTypeJ)->getNumElements();
2067 // The fused vector will be:
2068 // -----------------------------------------------------
2069 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ |
2070 // -----------------------------------------------------
2071 // from which we'll extract NumElem total elements (where the first NumElemI
2072 // of them come from the mask in I and the remainder come from the mask
2075 // For the mask from the first pair...
2076 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI,
2079 // For the mask from the second pair...
2080 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ,
2083 return ConstantVector::get(Mask);
2086 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I,
2087 Instruction *J, unsigned o, Value *&LOp,
2089 Type *ArgTypeL, Type *ArgTypeH,
2090 bool IBeforeJ, unsigned IdxOff) {
2091 bool ExpandedIEChain = false;
2092 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) {
2093 // If we have a pure insertelement chain, then this can be rewritten
2094 // into a chain that directly builds the larger type.
2095 bool PureChain = true;
2096 InsertElementInst *LIENext = LIE;
2098 if (!isa<UndefValue>(LIENext->getOperand(0)) &&
2099 !isa<InsertElementInst>(LIENext->getOperand(0))) {
2104 dyn_cast<InsertElementInst>(LIENext->getOperand(0))));
2107 SmallVector<Value *, 8> VectElemts(numElemL,
2108 UndefValue::get(ArgTypeL->getScalarType()));
2109 InsertElementInst *LIENext = LIE;
2112 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue();
2113 VectElemts[Idx] = LIENext->getOperand(1);
2115 dyn_cast<InsertElementInst>(LIENext->getOperand(0))));
2118 Value *LIEPrev = UndefValue::get(ArgTypeH);
2119 for (unsigned i = 0; i < numElemL; ++i) {
2120 if (isa<UndefValue>(VectElemts[i])) continue;
2121 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i],
2122 ConstantInt::get(Type::getInt32Ty(Context),
2124 getReplacementName(IBeforeJ ? I : J,
2126 LIENext->insertBefore(IBeforeJ ? J : I);
2130 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH);
2131 ExpandedIEChain = true;
2135 return ExpandedIEChain;
2138 // Returns the value to be used as the specified operand of the vector
2139 // instruction that fuses I with J.
2140 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I,
2141 Instruction *J, unsigned o, bool IBeforeJ) {
2142 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2143 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1);
2145 // Compute the fused vector type for this operand
2146 Type *ArgTypeI = I->getOperand(o)->getType();
2147 Type *ArgTypeJ = J->getOperand(o)->getType();
2148 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2150 Instruction *L = I, *H = J;
2151 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ;
2154 if (ArgTypeL->isVectorTy())
2155 numElemL = cast<VectorType>(ArgTypeL)->getNumElements();
2160 if (ArgTypeH->isVectorTy())
2161 numElemH = cast<VectorType>(ArgTypeH)->getNumElements();
2165 Value *LOp = L->getOperand(o);
2166 Value *HOp = H->getOperand(o);
2167 unsigned numElem = VArgType->getNumElements();
2169 // First, we check if we can reuse the "original" vector outputs (if these
2170 // exist). We might need a shuffle.
2171 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp);
2172 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp);
2173 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp);
2174 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp);
2176 // FIXME: If we're fusing shuffle instructions, then we can't apply this
2177 // optimization. The input vectors to the shuffle might be a different
2178 // length from the shuffle outputs. Unfortunately, the replacement
2179 // shuffle mask has already been formed, and the mask entries are sensitive
2180 // to the sizes of the inputs.
2181 bool IsSizeChangeShuffle =
2182 isa<ShuffleVectorInst>(L) &&
2183 (LOp->getType() != L->getType() || HOp->getType() != H->getType());
2185 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) {
2186 // We can have at most two unique vector inputs.
2187 bool CanUseInputs = true;
2190 I1 = LEE->getOperand(0);
2192 I1 = LSV->getOperand(0);
2193 I2 = LSV->getOperand(1);
2194 if (I2 == I1 || isa<UndefValue>(I2))
2199 Value *I3 = HEE->getOperand(0);
2200 if (!I2 && I3 != I1)
2202 else if (I3 != I1 && I3 != I2)
2203 CanUseInputs = false;
2205 Value *I3 = HSV->getOperand(0);
2206 if (!I2 && I3 != I1)
2208 else if (I3 != I1 && I3 != I2)
2209 CanUseInputs = false;
2212 Value *I4 = HSV->getOperand(1);
2213 if (!isa<UndefValue>(I4)) {
2214 if (!I2 && I4 != I1)
2216 else if (I4 != I1 && I4 != I2)
2217 CanUseInputs = false;
2224 cast<VectorType>(cast<Instruction>(LOp)->getOperand(0)->getType())
2227 cast<VectorType>(cast<Instruction>(HOp)->getOperand(0)->getType())
2230 // We have one or two input vectors. We need to map each index of the
2231 // operands to the index of the original vector.
2232 SmallVector<std::pair<int, int>, 8> II(numElem);
2233 for (unsigned i = 0; i < numElemL; ++i) {
2237 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue();
2238 INum = LEE->getOperand(0) == I1 ? 0 : 1;
2240 Idx = LSV->getMaskValue(i);
2241 if (Idx < (int) LOpElem) {
2242 INum = LSV->getOperand(0) == I1 ? 0 : 1;
2245 INum = LSV->getOperand(1) == I1 ? 0 : 1;
2249 II[i] = std::pair<int, int>(Idx, INum);
2251 for (unsigned i = 0; i < numElemH; ++i) {
2255 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue();
2256 INum = HEE->getOperand(0) == I1 ? 0 : 1;
2258 Idx = HSV->getMaskValue(i);
2259 if (Idx < (int) HOpElem) {
2260 INum = HSV->getOperand(0) == I1 ? 0 : 1;
2263 INum = HSV->getOperand(1) == I1 ? 0 : 1;
2267 II[i + numElemL] = std::pair<int, int>(Idx, INum);
2270 // We now have an array which tells us from which index of which
2271 // input vector each element of the operand comes.
2272 VectorType *I1T = cast<VectorType>(I1->getType());
2273 unsigned I1Elem = I1T->getNumElements();
2276 // In this case there is only one underlying vector input. Check for
2277 // the trivial case where we can use the input directly.
2278 if (I1Elem == numElem) {
2279 bool ElemInOrder = true;
2280 for (unsigned i = 0; i < numElem; ++i) {
2281 if (II[i].first != (int) i && II[i].first != -1) {
2282 ElemInOrder = false;
2291 // A shuffle is needed.
2292 std::vector<Constant *> Mask(numElem);
2293 for (unsigned i = 0; i < numElem; ++i) {
2294 int Idx = II[i].first;
2296 Mask[i] = UndefValue::get(Type::getInt32Ty(Context));
2298 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2302 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2303 ConstantVector::get(Mask),
2304 getReplacementName(IBeforeJ ? I : J,
2306 S->insertBefore(IBeforeJ ? J : I);
2310 VectorType *I2T = cast<VectorType>(I2->getType());
2311 unsigned I2Elem = I2T->getNumElements();
2313 // This input comes from two distinct vectors. The first step is to
2314 // make sure that both vectors are the same length. If not, the
2315 // smaller one will need to grow before they can be shuffled together.
2316 if (I1Elem < I2Elem) {
2317 std::vector<Constant *> Mask(I2Elem);
2319 for (; v < I1Elem; ++v)
2320 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2321 for (; v < I2Elem; ++v)
2322 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2324 Instruction *NewI1 =
2325 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2326 ConstantVector::get(Mask),
2327 getReplacementName(IBeforeJ ? I : J,
2329 NewI1->insertBefore(IBeforeJ ? J : I);
2333 } else if (I1Elem > I2Elem) {
2334 std::vector<Constant *> Mask(I1Elem);
2336 for (; v < I2Elem; ++v)
2337 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2338 for (; v < I1Elem; ++v)
2339 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2341 Instruction *NewI2 =
2342 new ShuffleVectorInst(I2, UndefValue::get(I2T),
2343 ConstantVector::get(Mask),
2344 getReplacementName(IBeforeJ ? I : J,
2346 NewI2->insertBefore(IBeforeJ ? J : I);
2352 // Now that both I1 and I2 are the same length we can shuffle them
2353 // together (and use the result).
2354 std::vector<Constant *> Mask(numElem);
2355 for (unsigned v = 0; v < numElem; ++v) {
2356 if (II[v].first == -1) {
2357 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2359 int Idx = II[v].first + II[v].second * I1Elem;
2360 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2364 Instruction *NewOp =
2365 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask),
2366 getReplacementName(IBeforeJ ? I : J, true, o));
2367 NewOp->insertBefore(IBeforeJ ? J : I);
2372 Type *ArgType = ArgTypeL;
2373 if (numElemL < numElemH) {
2374 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH,
2375 ArgTypeL, VArgType, IBeforeJ, 1)) {
2376 // This is another short-circuit case: we're combining a scalar into
2377 // a vector that is formed by an IE chain. We've just expanded the IE
2378 // chain, now insert the scalar and we're done.
2380 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0,
2381 getReplacementName(IBeforeJ ? I : J, true, o));
2382 S->insertBefore(IBeforeJ ? J : I);
2384 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL,
2385 ArgTypeH, IBeforeJ)) {
2386 // The two vector inputs to the shuffle must be the same length,
2387 // so extend the smaller vector to be the same length as the larger one.
2391 std::vector<Constant *> Mask(numElemH);
2393 for (; v < numElemL; ++v)
2394 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2395 for (; v < numElemH; ++v)
2396 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2398 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL),
2399 ConstantVector::get(Mask),
2400 getReplacementName(IBeforeJ ? I : J,
2403 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0,
2404 getReplacementName(IBeforeJ ? I : J,
2408 NLOp->insertBefore(IBeforeJ ? J : I);
2413 } else if (numElemL > numElemH) {
2414 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL,
2415 ArgTypeH, VArgType, IBeforeJ)) {
2417 InsertElementInst::Create(LOp, HOp,
2418 ConstantInt::get(Type::getInt32Ty(Context),
2420 getReplacementName(IBeforeJ ? I : J,
2422 S->insertBefore(IBeforeJ ? J : I);
2424 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH,
2425 ArgTypeL, IBeforeJ)) {
2428 std::vector<Constant *> Mask(numElemL);
2430 for (; v < numElemH; ++v)
2431 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2432 for (; v < numElemL; ++v)
2433 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2435 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH),
2436 ConstantVector::get(Mask),
2437 getReplacementName(IBeforeJ ? I : J,
2440 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0,
2441 getReplacementName(IBeforeJ ? I : J,
2445 NHOp->insertBefore(IBeforeJ ? J : I);
2450 if (ArgType->isVectorTy()) {
2451 unsigned numElem = cast<VectorType>(VArgType)->getNumElements();
2452 std::vector<Constant*> Mask(numElem);
2453 for (unsigned v = 0; v < numElem; ++v) {
2455 // If the low vector was expanded, we need to skip the extra
2456 // undefined entries.
2457 if (v >= numElemL && numElemH > numElemL)
2458 Idx += (numElemH - numElemL);
2459 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2462 Instruction *BV = new ShuffleVectorInst(LOp, HOp,
2463 ConstantVector::get(Mask),
2464 getReplacementName(IBeforeJ ? I : J, true, o));
2465 BV->insertBefore(IBeforeJ ? J : I);
2469 Instruction *BV1 = InsertElementInst::Create(
2470 UndefValue::get(VArgType), LOp, CV0,
2471 getReplacementName(IBeforeJ ? I : J,
2473 BV1->insertBefore(IBeforeJ ? J : I);
2474 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1,
2475 getReplacementName(IBeforeJ ? I : J,
2477 BV2->insertBefore(IBeforeJ ? J : I);
2481 // This function creates an array of values that will be used as the inputs
2482 // to the vector instruction that fuses I with J.
2483 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context,
2484 Instruction *I, Instruction *J,
2485 SmallVector<Value *, 3> &ReplacedOperands,
2487 unsigned NumOperands = I->getNumOperands();
2489 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) {
2490 // Iterate backward so that we look at the store pointer
2491 // first and know whether or not we need to flip the inputs.
2493 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) {
2494 // This is the pointer for a load/store instruction.
2495 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o);
2497 } else if (isa<CallInst>(I)) {
2498 Function *F = cast<CallInst>(I)->getCalledFunction();
2499 unsigned IID = F->getIntrinsicID();
2500 if (o == NumOperands-1) {
2501 BasicBlock &BB = *I->getParent();
2503 Module *M = BB.getParent()->getParent();
2504 Type *ArgTypeI = I->getType();
2505 Type *ArgTypeJ = J->getType();
2506 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2508 ReplacedOperands[o] = Intrinsic::getDeclaration(M,
2509 (Intrinsic::ID) IID, VArgType);
2511 } else if (IID == Intrinsic::powi && o == 1) {
2512 // The second argument of powi is a single integer and we've already
2513 // checked that both arguments are equal. As a result, we just keep
2514 // I's second argument.
2515 ReplacedOperands[o] = I->getOperand(o);
2518 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) {
2519 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J);
2523 ReplacedOperands[o] = getReplacementInput(Context, I, J, o, IBeforeJ);
2527 // This function creates two values that represent the outputs of the
2528 // original I and J instructions. These are generally vector shuffles
2529 // or extracts. In many cases, these will end up being unused and, thus,
2530 // eliminated by later passes.
2531 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
2532 Instruction *J, Instruction *K,
2533 Instruction *&InsertionPt,
2534 Instruction *&K1, Instruction *&K2) {
2535 if (isa<StoreInst>(I)) {
2536 AA->replaceWithNewValue(I, K);
2537 AA->replaceWithNewValue(J, K);
2539 Type *IType = I->getType();
2540 Type *JType = J->getType();
2542 VectorType *VType = getVecTypeForPair(IType, JType);
2543 unsigned numElem = VType->getNumElements();
2545 unsigned numElemI, numElemJ;
2546 if (IType->isVectorTy())
2547 numElemI = cast<VectorType>(IType)->getNumElements();
2551 if (JType->isVectorTy())
2552 numElemJ = cast<VectorType>(JType)->getNumElements();
2556 if (IType->isVectorTy()) {
2557 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI);
2558 for (unsigned v = 0; v < numElemI; ++v) {
2559 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2560 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v);
2563 K1 = new ShuffleVectorInst(K, UndefValue::get(VType),
2564 ConstantVector::get( Mask1),
2565 getReplacementName(K, false, 1));
2567 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2568 K1 = ExtractElementInst::Create(K, CV0,
2569 getReplacementName(K, false, 1));
2572 if (JType->isVectorTy()) {
2573 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ);
2574 for (unsigned v = 0; v < numElemJ; ++v) {
2575 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2576 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v);
2579 K2 = new ShuffleVectorInst(K, UndefValue::get(VType),
2580 ConstantVector::get( Mask2),
2581 getReplacementName(K, false, 2));
2583 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1);
2584 K2 = ExtractElementInst::Create(K, CV1,
2585 getReplacementName(K, false, 2));
2589 K2->insertAfter(K1);
2594 // Move all uses of the function I (including pairing-induced uses) after J.
2595 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB,
2596 std::multimap<Value *, Value *> &LoadMoveSet,
2597 Instruction *I, Instruction *J) {
2598 // Skip to the first instruction past I.
2599 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2601 DenseSet<Value *> Users;
2602 AliasSetTracker WriteSet(*AA);
2603 for (; cast<Instruction>(L) != J; ++L)
2604 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet);
2606 assert(cast<Instruction>(L) == J &&
2607 "Tracking has not proceeded far enough to check for dependencies");
2608 // If J is now in the use set of I, then trackUsesOfI will return true
2609 // and we have a dependency cycle (and the fusing operation must abort).
2610 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSet);
2613 // Move all uses of the function I (including pairing-induced uses) after J.
2614 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB,
2615 std::multimap<Value *, Value *> &LoadMoveSet,
2616 Instruction *&InsertionPt,
2617 Instruction *I, Instruction *J) {
2618 // Skip to the first instruction past I.
2619 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2621 DenseSet<Value *> Users;
2622 AliasSetTracker WriteSet(*AA);
2623 for (; cast<Instruction>(L) != J;) {
2624 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet)) {
2625 // Move this instruction
2626 Instruction *InstToMove = L; ++L;
2628 DEBUG(dbgs() << "BBV: moving: " << *InstToMove <<
2629 " to after " << *InsertionPt << "\n");
2630 InstToMove->removeFromParent();
2631 InstToMove->insertAfter(InsertionPt);
2632 InsertionPt = InstToMove;
2639 // Collect all load instruction that are in the move set of a given first
2640 // pair member. These loads depend on the first instruction, I, and so need
2641 // to be moved after J (the second instruction) when the pair is fused.
2642 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB,
2643 DenseMap<Value *, Value *> &ChosenPairs,
2644 std::multimap<Value *, Value *> &LoadMoveSet,
2646 // Skip to the first instruction past I.
2647 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2649 DenseSet<Value *> Users;
2650 AliasSetTracker WriteSet(*AA);
2652 // Note: We cannot end the loop when we reach J because J could be moved
2653 // farther down the use chain by another instruction pairing. Also, J
2654 // could be before I if this is an inverted input.
2655 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) {
2656 if (trackUsesOfI(Users, WriteSet, I, L)) {
2657 if (L->mayReadFromMemory())
2658 LoadMoveSet.insert(ValuePair(L, I));
2663 // In cases where both load/stores and the computation of their pointers
2664 // are chosen for vectorization, we can end up in a situation where the
2665 // aliasing analysis starts returning different query results as the
2666 // process of fusing instruction pairs continues. Because the algorithm
2667 // relies on finding the same use trees here as were found earlier, we'll
2668 // need to precompute the necessary aliasing information here and then
2669 // manually update it during the fusion process.
2670 void BBVectorize::collectLoadMoveSet(BasicBlock &BB,
2671 std::vector<Value *> &PairableInsts,
2672 DenseMap<Value *, Value *> &ChosenPairs,
2673 std::multimap<Value *, Value *> &LoadMoveSet) {
2674 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
2675 PIE = PairableInsts.end(); PI != PIE; ++PI) {
2676 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI);
2677 if (P == ChosenPairs.end()) continue;
2679 Instruction *I = cast<Instruction>(P->first);
2680 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, I);
2684 // When the first instruction in each pair is cloned, it will inherit its
2685 // parent's metadata. This metadata must be combined with that of the other
2686 // instruction in a safe way.
2687 void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) {
2688 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata;
2689 K->getAllMetadataOtherThanDebugLoc(Metadata);
2690 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) {
2691 unsigned Kind = Metadata[i].first;
2692 MDNode *JMD = J->getMetadata(Kind);
2693 MDNode *KMD = Metadata[i].second;
2697 K->setMetadata(Kind, 0); // Remove unknown metadata
2699 case LLVMContext::MD_tbaa:
2700 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD));
2702 case LLVMContext::MD_fpmath:
2703 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD));
2709 // This function fuses the chosen instruction pairs into vector instructions,
2710 // taking care preserve any needed scalar outputs and, then, it reorders the
2711 // remaining instructions as needed (users of the first member of the pair
2712 // need to be moved to after the location of the second member of the pair
2713 // because the vector instruction is inserted in the location of the pair's
2715 void BBVectorize::fuseChosenPairs(BasicBlock &BB,
2716 std::vector<Value *> &PairableInsts,
2717 DenseMap<Value *, Value *> &ChosenPairs,
2718 DenseSet<ValuePair> &FixedOrderPairs,
2719 DenseMap<VPPair, unsigned> &PairConnectionTypes,
2720 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
2721 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps) {
2722 LLVMContext& Context = BB.getContext();
2724 // During the vectorization process, the order of the pairs to be fused
2725 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs
2726 // list. After a pair is fused, the flipped pair is removed from the list.
2727 DenseSet<ValuePair> FlippedPairs;
2728 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(),
2729 E = ChosenPairs.end(); P != E; ++P)
2730 FlippedPairs.insert(ValuePair(P->second, P->first));
2731 for (DenseSet<ValuePair>::iterator P = FlippedPairs.begin(),
2732 E = FlippedPairs.end(); P != E; ++P)
2733 ChosenPairs.insert(*P);
2735 std::multimap<Value *, Value *> LoadMoveSet;
2736 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, LoadMoveSet);
2738 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n");
2740 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) {
2741 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI);
2742 if (P == ChosenPairs.end()) {
2747 if (getDepthFactor(P->first) == 0) {
2748 // These instructions are not really fused, but are tracked as though
2749 // they are. Any case in which it would be interesting to fuse them
2750 // will be taken care of by InstCombine.
2756 Instruction *I = cast<Instruction>(P->first),
2757 *J = cast<Instruction>(P->second);
2759 DEBUG(dbgs() << "BBV: fusing: " << *I <<
2760 " <-> " << *J << "\n");
2762 // Remove the pair and flipped pair from the list.
2763 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second);
2764 assert(FP != ChosenPairs.end() && "Flipped pair not found in list");
2765 ChosenPairs.erase(FP);
2766 ChosenPairs.erase(P);
2768 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSet, I, J)) {
2769 DEBUG(dbgs() << "BBV: fusion of: " << *I <<
2771 " aborted because of non-trivial dependency cycle\n");
2777 // If the pair must have the other order, then flip it.
2778 bool FlipPairOrder = FixedOrderPairs.count(ValuePair(J, I));
2779 if (!FlipPairOrder && !FixedOrderPairs.count(ValuePair(I, J))) {
2780 // This pair does not have a fixed order, and so we might want to
2781 // flip it if that will yield fewer shuffles. We count the number
2782 // of dependencies connected via swaps, and those directly connected,
2783 // and flip the order if the number of swaps is greater.
2784 bool OrigOrder = true;
2785 VPPIteratorPair IP = ConnectedPairDeps.equal_range(ValuePair(I, J));
2786 if (IP.first == ConnectedPairDeps.end()) {
2787 IP = ConnectedPairDeps.equal_range(ValuePair(J, I));
2791 if (IP.first != ConnectedPairDeps.end()) {
2792 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
2793 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first;
2794 Q != IP.second; ++Q) {
2795 DenseMap<VPPair, unsigned>::iterator R =
2796 PairConnectionTypes.find(VPPair(Q->second, Q->first));
2797 assert(R != PairConnectionTypes.end() &&
2798 "Cannot find pair connection type");
2799 if (R->second == PairConnectionDirect)
2801 else if (R->second == PairConnectionSwap)
2806 std::swap(NumDepsDirect, NumDepsSwap);
2808 if (NumDepsSwap > NumDepsDirect) {
2809 FlipPairOrder = true;
2810 DEBUG(dbgs() << "BBV: reordering pair: " << *I <<
2811 " <-> " << *J << "\n");
2816 Instruction *L = I, *H = J;
2820 // If the pair being fused uses the opposite order from that in the pair
2821 // connection map, then we need to flip the types.
2822 VPPIteratorPair IP = ConnectedPairs.equal_range(ValuePair(H, L));
2823 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first;
2824 Q != IP.second; ++Q) {
2825 DenseMap<VPPair, unsigned>::iterator R = PairConnectionTypes.find(*Q);
2826 assert(R != PairConnectionTypes.end() &&
2827 "Cannot find pair connection type");
2828 if (R->second == PairConnectionDirect)
2829 R->second = PairConnectionSwap;
2830 else if (R->second == PairConnectionSwap)
2831 R->second = PairConnectionDirect;
2834 bool LBeforeH = !FlipPairOrder;
2835 unsigned NumOperands = I->getNumOperands();
2836 SmallVector<Value *, 3> ReplacedOperands(NumOperands);
2837 getReplacementInputsForPair(Context, L, H, ReplacedOperands,
2840 // Make a copy of the original operation, change its type to the vector
2841 // type and replace its operands with the vector operands.
2842 Instruction *K = L->clone();
2845 else if (H->hasName())
2848 if (!isa<StoreInst>(K))
2849 K->mutateType(getVecTypeForPair(L->getType(), H->getType()));
2851 combineMetadata(K, H);
2853 for (unsigned o = 0; o < NumOperands; ++o)
2854 K->setOperand(o, ReplacedOperands[o]);
2858 // Instruction insertion point:
2859 Instruction *InsertionPt = K;
2860 Instruction *K1 = 0, *K2 = 0;
2861 replaceOutputsOfPair(Context, L, H, K, InsertionPt, K1, K2);
2863 // The use tree of the first original instruction must be moved to after
2864 // the location of the second instruction. The entire use tree of the
2865 // first instruction is disjoint from the input tree of the second
2866 // (by definition), and so commutes with it.
2868 moveUsesOfIAfterJ(BB, LoadMoveSet, InsertionPt, I, J);
2870 if (!isa<StoreInst>(I)) {
2871 L->replaceAllUsesWith(K1);
2872 H->replaceAllUsesWith(K2);
2873 AA->replaceWithNewValue(L, K1);
2874 AA->replaceWithNewValue(H, K2);
2877 // Instructions that may read from memory may be in the load move set.
2878 // Once an instruction is fused, we no longer need its move set, and so
2879 // the values of the map never need to be updated. However, when a load
2880 // is fused, we need to merge the entries from both instructions in the
2881 // pair in case those instructions were in the move set of some other
2882 // yet-to-be-fused pair. The loads in question are the keys of the map.
2883 if (I->mayReadFromMemory()) {
2884 std::vector<ValuePair> NewSetMembers;
2885 VPIteratorPair IPairRange = LoadMoveSet.equal_range(I);
2886 VPIteratorPair JPairRange = LoadMoveSet.equal_range(J);
2887 for (std::multimap<Value *, Value *>::iterator N = IPairRange.first;
2888 N != IPairRange.second; ++N)
2889 NewSetMembers.push_back(ValuePair(K, N->second));
2890 for (std::multimap<Value *, Value *>::iterator N = JPairRange.first;
2891 N != JPairRange.second; ++N)
2892 NewSetMembers.push_back(ValuePair(K, N->second));
2893 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(),
2894 AE = NewSetMembers.end(); A != AE; ++A)
2895 LoadMoveSet.insert(*A);
2898 // Before removing I, set the iterator to the next instruction.
2899 PI = llvm::next(BasicBlock::iterator(I));
2900 if (cast<Instruction>(PI) == J)
2905 I->eraseFromParent();
2906 J->eraseFromParent();
2908 DEBUG(if (PrintAfterEveryPair) dbgs() << "BBV: block is now: \n" <<
2912 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n");
2916 char BBVectorize::ID = 0;
2917 static const char bb_vectorize_name[] = "Basic-Block Vectorization";
2918 INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
2919 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
2920 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
2921 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
2922 INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
2924 BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) {
2925 return new BBVectorize(C);
2929 llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) {
2930 BBVectorize BBVectorizer(P, C);
2931 return BBVectorizer.vectorizeBB(BB);
2934 //===----------------------------------------------------------------------===//
2935 VectorizeConfig::VectorizeConfig() {
2936 VectorBits = ::VectorBits;
2937 VectorizeBools = !::NoBools;
2938 VectorizeInts = !::NoInts;
2939 VectorizeFloats = !::NoFloats;
2940 VectorizePointers = !::NoPointers;
2941 VectorizeCasts = !::NoCasts;
2942 VectorizeMath = !::NoMath;
2943 VectorizeFMA = !::NoFMA;
2944 VectorizeSelect = !::NoSelect;
2945 VectorizeCmp = !::NoCmp;
2946 VectorizeGEP = !::NoGEP;
2947 VectorizeMemOps = !::NoMemOps;
2948 AlignedOnly = ::AlignedOnly;
2949 ReqChainDepth= ::ReqChainDepth;
2950 SearchLimit = ::SearchLimit;
2951 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck;
2952 SplatBreaksChain = ::SplatBreaksChain;
2953 MaxInsts = ::MaxInsts;
2954 MaxIter = ::MaxIter;
2955 Pow2LenOnly = ::Pow2LenOnly;
2956 NoMemOpBoost = ::NoMemOpBoost;
2957 FastDep = ::FastDep;