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 #include "llvm/Transforms/Vectorize.h"
19 #include "llvm/ADT/DenseMap.h"
20 #include "llvm/ADT/DenseSet.h"
21 #include "llvm/ADT/STLExtras.h"
22 #include "llvm/ADT/SmallSet.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/ADT/StringExtras.h"
26 #include "llvm/Analysis/AliasAnalysis.h"
27 #include "llvm/Analysis/AliasSetTracker.h"
28 #include "llvm/Analysis/ScalarEvolution.h"
29 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
30 #include "llvm/Analysis/TargetTransformInfo.h"
31 #include "llvm/Analysis/ValueTracking.h"
32 #include "llvm/IR/Constants.h"
33 #include "llvm/IR/DataLayout.h"
34 #include "llvm/IR/DerivedTypes.h"
35 #include "llvm/IR/Dominators.h"
36 #include "llvm/IR/Function.h"
37 #include "llvm/IR/Instructions.h"
38 #include "llvm/IR/IntrinsicInst.h"
39 #include "llvm/IR/Intrinsics.h"
40 #include "llvm/IR/LLVMContext.h"
41 #include "llvm/IR/Metadata.h"
42 #include "llvm/IR/Type.h"
43 #include "llvm/IR/ValueHandle.h"
44 #include "llvm/Pass.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/raw_ostream.h"
48 #include "llvm/Transforms/Utils/Local.h"
52 #define DEBUG_TYPE BBV_NAME
55 IgnoreTargetInfo("bb-vectorize-ignore-target-info", cl::init(false),
56 cl::Hidden, cl::desc("Ignore target information"));
58 static cl::opt<unsigned>
59 ReqChainDepth("bb-vectorize-req-chain-depth", cl::init(6), cl::Hidden,
60 cl::desc("The required chain depth for vectorization"));
63 UseChainDepthWithTI("bb-vectorize-use-chain-depth", cl::init(false),
64 cl::Hidden, cl::desc("Use the chain depth requirement with"
65 " target information"));
67 static cl::opt<unsigned>
68 SearchLimit("bb-vectorize-search-limit", cl::init(400), cl::Hidden,
69 cl::desc("The maximum search distance for instruction pairs"));
72 SplatBreaksChain("bb-vectorize-splat-breaks-chain", cl::init(false), cl::Hidden,
73 cl::desc("Replicating one element to a pair breaks the chain"));
75 static cl::opt<unsigned>
76 VectorBits("bb-vectorize-vector-bits", cl::init(128), cl::Hidden,
77 cl::desc("The size of the native vector registers"));
79 static cl::opt<unsigned>
80 MaxIter("bb-vectorize-max-iter", cl::init(0), cl::Hidden,
81 cl::desc("The maximum number of pairing iterations"));
84 Pow2LenOnly("bb-vectorize-pow2-len-only", cl::init(false), cl::Hidden,
85 cl::desc("Don't try to form non-2^n-length vectors"));
87 static cl::opt<unsigned>
88 MaxInsts("bb-vectorize-max-instr-per-group", cl::init(500), cl::Hidden,
89 cl::desc("The maximum number of pairable instructions per group"));
91 static cl::opt<unsigned>
92 MaxPairs("bb-vectorize-max-pairs-per-group", cl::init(3000), cl::Hidden,
93 cl::desc("The maximum number of candidate instruction pairs per group"));
95 static cl::opt<unsigned>
96 MaxCandPairsForCycleCheck("bb-vectorize-max-cycle-check-pairs", cl::init(200),
97 cl::Hidden, cl::desc("The maximum number of candidate pairs with which to use"
98 " a full cycle check"));
101 NoBools("bb-vectorize-no-bools", cl::init(false), cl::Hidden,
102 cl::desc("Don't try to vectorize boolean (i1) values"));
105 NoInts("bb-vectorize-no-ints", cl::init(false), cl::Hidden,
106 cl::desc("Don't try to vectorize integer values"));
109 NoFloats("bb-vectorize-no-floats", cl::init(false), cl::Hidden,
110 cl::desc("Don't try to vectorize floating-point values"));
112 // FIXME: This should default to false once pointer vector support works.
114 NoPointers("bb-vectorize-no-pointers", cl::init(/*false*/ true), cl::Hidden,
115 cl::desc("Don't try to vectorize pointer values"));
118 NoCasts("bb-vectorize-no-casts", cl::init(false), cl::Hidden,
119 cl::desc("Don't try to vectorize casting (conversion) operations"));
122 NoMath("bb-vectorize-no-math", cl::init(false), cl::Hidden,
123 cl::desc("Don't try to vectorize floating-point math intrinsics"));
126 NoFMA("bb-vectorize-no-fma", cl::init(false), cl::Hidden,
127 cl::desc("Don't try to vectorize the fused-multiply-add intrinsic"));
130 NoSelect("bb-vectorize-no-select", cl::init(false), cl::Hidden,
131 cl::desc("Don't try to vectorize select instructions"));
134 NoCmp("bb-vectorize-no-cmp", cl::init(false), cl::Hidden,
135 cl::desc("Don't try to vectorize comparison instructions"));
138 NoGEP("bb-vectorize-no-gep", cl::init(false), cl::Hidden,
139 cl::desc("Don't try to vectorize getelementptr instructions"));
142 NoMemOps("bb-vectorize-no-mem-ops", cl::init(false), cl::Hidden,
143 cl::desc("Don't try to vectorize loads and stores"));
146 AlignedOnly("bb-vectorize-aligned-only", cl::init(false), cl::Hidden,
147 cl::desc("Only generate aligned loads and stores"));
150 NoMemOpBoost("bb-vectorize-no-mem-op-boost",
151 cl::init(false), cl::Hidden,
152 cl::desc("Don't boost the chain-depth contribution of loads and stores"));
155 FastDep("bb-vectorize-fast-dep", cl::init(false), cl::Hidden,
156 cl::desc("Use a fast instruction dependency analysis"));
160 DebugInstructionExamination("bb-vectorize-debug-instruction-examination",
161 cl::init(false), cl::Hidden,
162 cl::desc("When debugging is enabled, output information on the"
163 " instruction-examination process"));
165 DebugCandidateSelection("bb-vectorize-debug-candidate-selection",
166 cl::init(false), cl::Hidden,
167 cl::desc("When debugging is enabled, output information on the"
168 " candidate-selection process"));
170 DebugPairSelection("bb-vectorize-debug-pair-selection",
171 cl::init(false), cl::Hidden,
172 cl::desc("When debugging is enabled, output information on the"
173 " pair-selection process"));
175 DebugCycleCheck("bb-vectorize-debug-cycle-check",
176 cl::init(false), cl::Hidden,
177 cl::desc("When debugging is enabled, output information on the"
178 " cycle-checking process"));
181 PrintAfterEveryPair("bb-vectorize-debug-print-after-every-pair",
182 cl::init(false), cl::Hidden,
183 cl::desc("When debugging is enabled, dump the basic block after"
184 " every pair is fused"));
187 STATISTIC(NumFusedOps, "Number of operations fused by bb-vectorize");
190 struct BBVectorize : public BasicBlockPass {
191 static char ID; // Pass identification, replacement for typeid
193 const VectorizeConfig Config;
195 BBVectorize(const VectorizeConfig &C = VectorizeConfig())
196 : BasicBlockPass(ID), Config(C) {
197 initializeBBVectorizePass(*PassRegistry::getPassRegistry());
200 BBVectorize(Pass *P, const VectorizeConfig &C)
201 : BasicBlockPass(ID), Config(C) {
202 AA = &P->getAnalysis<AliasAnalysis>();
203 DT = &P->getAnalysis<DominatorTreeWrapperPass>().getDomTree();
204 SE = &P->getAnalysis<ScalarEvolution>();
205 DataLayoutPass *DLP = P->getAnalysisIfAvailable<DataLayoutPass>();
206 DL = DLP ? &DLP->getDataLayout() : 0;
207 TTI = IgnoreTargetInfo ? 0 : &P->getAnalysis<TargetTransformInfo>();
210 typedef std::pair<Value *, Value *> ValuePair;
211 typedef std::pair<ValuePair, int> ValuePairWithCost;
212 typedef std::pair<ValuePair, size_t> ValuePairWithDepth;
213 typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair
214 typedef std::pair<VPPair, unsigned> VPPairWithType;
219 const DataLayout *DL;
220 const TargetTransformInfo *TTI;
222 // FIXME: const correct?
224 bool vectorizePairs(BasicBlock &BB, bool NonPow2Len = false);
226 bool getCandidatePairs(BasicBlock &BB,
227 BasicBlock::iterator &Start,
228 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
229 DenseSet<ValuePair> &FixedOrderPairs,
230 DenseMap<ValuePair, int> &CandidatePairCostSavings,
231 std::vector<Value *> &PairableInsts, bool NonPow2Len);
233 // FIXME: The current implementation does not account for pairs that
234 // are connected in multiple ways. For example:
235 // C1 = A1 / A2; C2 = A2 / A1 (which may be both direct and a swap)
236 enum PairConnectionType {
237 PairConnectionDirect,
242 void computeConnectedPairs(
243 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
244 DenseSet<ValuePair> &CandidatePairsSet,
245 std::vector<Value *> &PairableInsts,
246 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
247 DenseMap<VPPair, unsigned> &PairConnectionTypes);
249 void buildDepMap(BasicBlock &BB,
250 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
251 std::vector<Value *> &PairableInsts,
252 DenseSet<ValuePair> &PairableInstUsers);
254 void choosePairs(DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
255 DenseSet<ValuePair> &CandidatePairsSet,
256 DenseMap<ValuePair, int> &CandidatePairCostSavings,
257 std::vector<Value *> &PairableInsts,
258 DenseSet<ValuePair> &FixedOrderPairs,
259 DenseMap<VPPair, unsigned> &PairConnectionTypes,
260 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
261 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
262 DenseSet<ValuePair> &PairableInstUsers,
263 DenseMap<Value *, Value *>& ChosenPairs);
265 void fuseChosenPairs(BasicBlock &BB,
266 std::vector<Value *> &PairableInsts,
267 DenseMap<Value *, Value *>& ChosenPairs,
268 DenseSet<ValuePair> &FixedOrderPairs,
269 DenseMap<VPPair, unsigned> &PairConnectionTypes,
270 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
271 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps);
274 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore);
276 bool areInstsCompatible(Instruction *I, Instruction *J,
277 bool IsSimpleLoadStore, bool NonPow2Len,
278 int &CostSavings, int &FixedOrder);
280 bool trackUsesOfI(DenseSet<Value *> &Users,
281 AliasSetTracker &WriteSet, Instruction *I,
282 Instruction *J, bool UpdateUsers = true,
283 DenseSet<ValuePair> *LoadMoveSetPairs = 0);
285 void computePairsConnectedTo(
286 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
287 DenseSet<ValuePair> &CandidatePairsSet,
288 std::vector<Value *> &PairableInsts,
289 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
290 DenseMap<VPPair, unsigned> &PairConnectionTypes,
293 bool pairsConflict(ValuePair P, ValuePair Q,
294 DenseSet<ValuePair> &PairableInstUsers,
295 DenseMap<ValuePair, std::vector<ValuePair> >
296 *PairableInstUserMap = 0,
297 DenseSet<VPPair> *PairableInstUserPairSet = 0);
299 bool pairWillFormCycle(ValuePair P,
300 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUsers,
301 DenseSet<ValuePair> &CurrentPairs);
304 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
305 std::vector<Value *> &PairableInsts,
306 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
307 DenseSet<ValuePair> &PairableInstUsers,
308 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
309 DenseSet<VPPair> &PairableInstUserPairSet,
310 DenseMap<Value *, Value *> &ChosenPairs,
311 DenseMap<ValuePair, size_t> &DAG,
312 DenseSet<ValuePair> &PrunedDAG, ValuePair J,
315 void buildInitialDAGFor(
316 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
317 DenseSet<ValuePair> &CandidatePairsSet,
318 std::vector<Value *> &PairableInsts,
319 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
320 DenseSet<ValuePair> &PairableInstUsers,
321 DenseMap<Value *, Value *> &ChosenPairs,
322 DenseMap<ValuePair, size_t> &DAG, ValuePair J);
325 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
326 DenseSet<ValuePair> &CandidatePairsSet,
327 DenseMap<ValuePair, int> &CandidatePairCostSavings,
328 std::vector<Value *> &PairableInsts,
329 DenseSet<ValuePair> &FixedOrderPairs,
330 DenseMap<VPPair, unsigned> &PairConnectionTypes,
331 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
332 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
333 DenseSet<ValuePair> &PairableInstUsers,
334 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
335 DenseSet<VPPair> &PairableInstUserPairSet,
336 DenseMap<Value *, Value *> &ChosenPairs,
337 DenseSet<ValuePair> &BestDAG, size_t &BestMaxDepth,
338 int &BestEffSize, Value *II, std::vector<Value *>&JJ,
341 Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I,
342 Instruction *J, unsigned o);
344 void fillNewShuffleMask(LLVMContext& Context, Instruction *J,
345 unsigned MaskOffset, unsigned NumInElem,
346 unsigned NumInElem1, unsigned IdxOffset,
347 std::vector<Constant*> &Mask);
349 Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I,
352 bool expandIEChain(LLVMContext& Context, Instruction *I, Instruction *J,
353 unsigned o, Value *&LOp, unsigned numElemL,
354 Type *ArgTypeL, Type *ArgTypeR, bool IBeforeJ,
355 unsigned IdxOff = 0);
357 Value *getReplacementInput(LLVMContext& Context, Instruction *I,
358 Instruction *J, unsigned o, bool IBeforeJ);
360 void getReplacementInputsForPair(LLVMContext& Context, Instruction *I,
361 Instruction *J, SmallVectorImpl<Value *> &ReplacedOperands,
364 void replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
365 Instruction *J, Instruction *K,
366 Instruction *&InsertionPt, Instruction *&K1,
369 void collectPairLoadMoveSet(BasicBlock &BB,
370 DenseMap<Value *, Value *> &ChosenPairs,
371 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
372 DenseSet<ValuePair> &LoadMoveSetPairs,
375 void collectLoadMoveSet(BasicBlock &BB,
376 std::vector<Value *> &PairableInsts,
377 DenseMap<Value *, Value *> &ChosenPairs,
378 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
379 DenseSet<ValuePair> &LoadMoveSetPairs);
381 bool canMoveUsesOfIAfterJ(BasicBlock &BB,
382 DenseSet<ValuePair> &LoadMoveSetPairs,
383 Instruction *I, Instruction *J);
385 void moveUsesOfIAfterJ(BasicBlock &BB,
386 DenseSet<ValuePair> &LoadMoveSetPairs,
387 Instruction *&InsertionPt,
388 Instruction *I, Instruction *J);
390 void combineMetadata(Instruction *K, const Instruction *J);
392 bool vectorizeBB(BasicBlock &BB) {
393 if (skipOptnoneFunction(BB))
395 if (!DT->isReachableFromEntry(&BB)) {
396 DEBUG(dbgs() << "BBV: skipping unreachable " << BB.getName() <<
397 " in " << BB.getParent()->getName() << "\n");
401 DEBUG(if (TTI) dbgs() << "BBV: using target information\n");
403 bool changed = false;
404 // Iterate a sufficient number of times to merge types of size 1 bit,
405 // then 2 bits, then 4, etc. up to half of the target vector width of the
406 // target vector register.
409 (TTI || v <= Config.VectorBits) &&
410 (!Config.MaxIter || n <= Config.MaxIter);
412 DEBUG(dbgs() << "BBV: fusing loop #" << n <<
413 " for " << BB.getName() << " in " <<
414 BB.getParent()->getName() << "...\n");
415 if (vectorizePairs(BB))
421 if (changed && !Pow2LenOnly) {
423 for (; !Config.MaxIter || n <= Config.MaxIter; ++n) {
424 DEBUG(dbgs() << "BBV: fusing for non-2^n-length vectors loop #: " <<
425 n << " for " << BB.getName() << " in " <<
426 BB.getParent()->getName() << "...\n");
427 if (!vectorizePairs(BB, true)) break;
431 DEBUG(dbgs() << "BBV: done!\n");
435 bool runOnBasicBlock(BasicBlock &BB) override {
436 // OptimizeNone check deferred to vectorizeBB().
438 AA = &getAnalysis<AliasAnalysis>();
439 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
440 SE = &getAnalysis<ScalarEvolution>();
441 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
442 DL = DLP ? &DLP->getDataLayout() : 0;
443 TTI = IgnoreTargetInfo ? 0 : &getAnalysis<TargetTransformInfo>();
445 return vectorizeBB(BB);
448 void getAnalysisUsage(AnalysisUsage &AU) const override {
449 BasicBlockPass::getAnalysisUsage(AU);
450 AU.addRequired<AliasAnalysis>();
451 AU.addRequired<DominatorTreeWrapperPass>();
452 AU.addRequired<ScalarEvolution>();
453 AU.addRequired<TargetTransformInfo>();
454 AU.addPreserved<AliasAnalysis>();
455 AU.addPreserved<DominatorTreeWrapperPass>();
456 AU.addPreserved<ScalarEvolution>();
457 AU.setPreservesCFG();
460 static inline VectorType *getVecTypeForPair(Type *ElemTy, Type *Elem2Ty) {
461 assert(ElemTy->getScalarType() == Elem2Ty->getScalarType() &&
462 "Cannot form vector from incompatible scalar types");
463 Type *STy = ElemTy->getScalarType();
466 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) {
467 numElem = VTy->getNumElements();
472 if (VectorType *VTy = dyn_cast<VectorType>(Elem2Ty)) {
473 numElem += VTy->getNumElements();
478 return VectorType::get(STy, numElem);
481 static inline void getInstructionTypes(Instruction *I,
482 Type *&T1, Type *&T2) {
483 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
484 // For stores, it is the value type, not the pointer type that matters
485 // because the value is what will come from a vector register.
487 Value *IVal = SI->getValueOperand();
488 T1 = IVal->getType();
493 if (CastInst *CI = dyn_cast<CastInst>(I))
498 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
499 T2 = SI->getCondition()->getType();
500 } else if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(I)) {
501 T2 = SI->getOperand(0)->getType();
502 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) {
503 T2 = CI->getOperand(0)->getType();
507 // Returns the weight associated with the provided value. A chain of
508 // candidate pairs has a length given by the sum of the weights of its
509 // members (one weight per pair; the weight of each member of the pair
510 // is assumed to be the same). This length is then compared to the
511 // chain-length threshold to determine if a given chain is significant
512 // enough to be vectorized. The length is also used in comparing
513 // candidate chains where longer chains are considered to be better.
514 // Note: when this function returns 0, the resulting instructions are
515 // not actually fused.
516 inline size_t getDepthFactor(Value *V) {
517 // InsertElement and ExtractElement have a depth factor of zero. This is
518 // for two reasons: First, they cannot be usefully fused. Second, because
519 // the pass generates a lot of these, they can confuse the simple metric
520 // used to compare the dags in the next iteration. Thus, giving them a
521 // weight of zero allows the pass to essentially ignore them in
522 // subsequent iterations when looking for vectorization opportunities
523 // while still tracking dependency chains that flow through those
525 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V))
528 // Give a load or store half of the required depth so that load/store
529 // pairs will vectorize.
530 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V)))
531 return Config.ReqChainDepth/2;
536 // Returns the cost of the provided instruction using TTI.
537 // This does not handle loads and stores.
538 unsigned getInstrCost(unsigned Opcode, Type *T1, Type *T2,
539 TargetTransformInfo::OperandValueKind Op1VK =
540 TargetTransformInfo::OK_AnyValue,
541 TargetTransformInfo::OperandValueKind Op2VK =
542 TargetTransformInfo::OK_AnyValue) {
545 case Instruction::GetElementPtr:
546 // We mark this instruction as zero-cost because scalar GEPs are usually
547 // lowered to the instruction addressing mode. At the moment we don't
548 // generate vector GEPs.
550 case Instruction::Br:
551 return TTI->getCFInstrCost(Opcode);
552 case Instruction::PHI:
554 case Instruction::Add:
555 case Instruction::FAdd:
556 case Instruction::Sub:
557 case Instruction::FSub:
558 case Instruction::Mul:
559 case Instruction::FMul:
560 case Instruction::UDiv:
561 case Instruction::SDiv:
562 case Instruction::FDiv:
563 case Instruction::URem:
564 case Instruction::SRem:
565 case Instruction::FRem:
566 case Instruction::Shl:
567 case Instruction::LShr:
568 case Instruction::AShr:
569 case Instruction::And:
570 case Instruction::Or:
571 case Instruction::Xor:
572 return TTI->getArithmeticInstrCost(Opcode, T1, Op1VK, Op2VK);
573 case Instruction::Select:
574 case Instruction::ICmp:
575 case Instruction::FCmp:
576 return TTI->getCmpSelInstrCost(Opcode, T1, T2);
577 case Instruction::ZExt:
578 case Instruction::SExt:
579 case Instruction::FPToUI:
580 case Instruction::FPToSI:
581 case Instruction::FPExt:
582 case Instruction::PtrToInt:
583 case Instruction::IntToPtr:
584 case Instruction::SIToFP:
585 case Instruction::UIToFP:
586 case Instruction::Trunc:
587 case Instruction::FPTrunc:
588 case Instruction::BitCast:
589 case Instruction::ShuffleVector:
590 return TTI->getCastInstrCost(Opcode, T1, T2);
596 // This determines the relative offset of two loads or stores, returning
597 // true if the offset could be determined to be some constant value.
598 // For example, if OffsetInElmts == 1, then J accesses the memory directly
599 // after I; if OffsetInElmts == -1 then I accesses the memory
601 bool getPairPtrInfo(Instruction *I, Instruction *J,
602 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment,
603 unsigned &IAddressSpace, unsigned &JAddressSpace,
604 int64_t &OffsetInElmts, bool ComputeOffset = true) {
606 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
607 LoadInst *LJ = cast<LoadInst>(J);
608 IPtr = LI->getPointerOperand();
609 JPtr = LJ->getPointerOperand();
610 IAlignment = LI->getAlignment();
611 JAlignment = LJ->getAlignment();
612 IAddressSpace = LI->getPointerAddressSpace();
613 JAddressSpace = LJ->getPointerAddressSpace();
615 StoreInst *SI = cast<StoreInst>(I), *SJ = cast<StoreInst>(J);
616 IPtr = SI->getPointerOperand();
617 JPtr = SJ->getPointerOperand();
618 IAlignment = SI->getAlignment();
619 JAlignment = SJ->getAlignment();
620 IAddressSpace = SI->getPointerAddressSpace();
621 JAddressSpace = SJ->getPointerAddressSpace();
627 const SCEV *IPtrSCEV = SE->getSCEV(IPtr);
628 const SCEV *JPtrSCEV = SE->getSCEV(JPtr);
630 // If this is a trivial offset, then we'll get something like
631 // 1*sizeof(type). With target data, which we need anyway, this will get
632 // constant folded into a number.
633 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV);
634 if (const SCEVConstant *ConstOffSCEV =
635 dyn_cast<SCEVConstant>(OffsetSCEV)) {
636 ConstantInt *IntOff = ConstOffSCEV->getValue();
637 int64_t Offset = IntOff->getSExtValue();
639 Type *VTy = IPtr->getType()->getPointerElementType();
640 int64_t VTyTSS = (int64_t) DL->getTypeStoreSize(VTy);
642 Type *VTy2 = JPtr->getType()->getPointerElementType();
643 if (VTy != VTy2 && Offset < 0) {
644 int64_t VTy2TSS = (int64_t) DL->getTypeStoreSize(VTy2);
645 OffsetInElmts = Offset/VTy2TSS;
646 return (abs64(Offset) % VTy2TSS) == 0;
649 OffsetInElmts = Offset/VTyTSS;
650 return (abs64(Offset) % VTyTSS) == 0;
656 // Returns true if the provided CallInst represents an intrinsic that can
658 bool isVectorizableIntrinsic(CallInst* I) {
659 Function *F = I->getCalledFunction();
660 if (!F) return false;
662 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID();
663 if (!IID) return false;
668 case Intrinsic::sqrt:
669 case Intrinsic::powi:
673 case Intrinsic::log2:
674 case Intrinsic::log10:
676 case Intrinsic::exp2:
678 return Config.VectorizeMath;
680 case Intrinsic::fmuladd:
681 return Config.VectorizeFMA;
685 bool isPureIEChain(InsertElementInst *IE) {
686 InsertElementInst *IENext = IE;
688 if (!isa<UndefValue>(IENext->getOperand(0)) &&
689 !isa<InsertElementInst>(IENext->getOperand(0))) {
693 dyn_cast<InsertElementInst>(IENext->getOperand(0))));
699 // This function implements one vectorization iteration on the provided
700 // basic block. It returns true if the block is changed.
701 bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) {
703 BasicBlock::iterator Start = BB.getFirstInsertionPt();
705 std::vector<Value *> AllPairableInsts;
706 DenseMap<Value *, Value *> AllChosenPairs;
707 DenseSet<ValuePair> AllFixedOrderPairs;
708 DenseMap<VPPair, unsigned> AllPairConnectionTypes;
709 DenseMap<ValuePair, std::vector<ValuePair> > AllConnectedPairs,
710 AllConnectedPairDeps;
713 std::vector<Value *> PairableInsts;
714 DenseMap<Value *, std::vector<Value *> > CandidatePairs;
715 DenseSet<ValuePair> FixedOrderPairs;
716 DenseMap<ValuePair, int> CandidatePairCostSavings;
717 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs,
719 CandidatePairCostSavings,
720 PairableInsts, NonPow2Len);
721 if (PairableInsts.empty()) continue;
723 // Build the candidate pair set for faster lookups.
724 DenseSet<ValuePair> CandidatePairsSet;
725 for (DenseMap<Value *, std::vector<Value *> >::iterator I =
726 CandidatePairs.begin(), E = CandidatePairs.end(); I != E; ++I)
727 for (std::vector<Value *>::iterator J = I->second.begin(),
728 JE = I->second.end(); J != JE; ++J)
729 CandidatePairsSet.insert(ValuePair(I->first, *J));
731 // Now we have a map of all of the pairable instructions and we need to
732 // select the best possible pairing. A good pairing is one such that the
733 // users of the pair are also paired. This defines a (directed) forest
734 // over the pairs such that two pairs are connected iff the second pair
737 // Note that it only matters that both members of the second pair use some
738 // element of the first pair (to allow for splatting).
740 DenseMap<ValuePair, std::vector<ValuePair> > ConnectedPairs,
742 DenseMap<VPPair, unsigned> PairConnectionTypes;
743 computeConnectedPairs(CandidatePairs, CandidatePairsSet,
744 PairableInsts, ConnectedPairs, PairConnectionTypes);
745 if (ConnectedPairs.empty()) continue;
747 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator
748 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
750 for (std::vector<ValuePair>::iterator J = I->second.begin(),
751 JE = I->second.end(); J != JE; ++J)
752 ConnectedPairDeps[*J].push_back(I->first);
754 // Build the pairable-instruction dependency map
755 DenseSet<ValuePair> PairableInstUsers;
756 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers);
758 // There is now a graph of the connected pairs. For each variable, pick
759 // the pairing with the largest dag meeting the depth requirement on at
760 // least one branch. Then select all pairings that are part of that dag
761 // and remove them from the list of available pairings and pairable
764 DenseMap<Value *, Value *> ChosenPairs;
765 choosePairs(CandidatePairs, CandidatePairsSet,
766 CandidatePairCostSavings,
767 PairableInsts, FixedOrderPairs, PairConnectionTypes,
768 ConnectedPairs, ConnectedPairDeps,
769 PairableInstUsers, ChosenPairs);
771 if (ChosenPairs.empty()) continue;
772 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(),
773 PairableInsts.end());
774 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end());
776 // Only for the chosen pairs, propagate information on fixed-order pairs,
777 // pair connections, and their types to the data structures used by the
778 // pair fusion procedures.
779 for (DenseMap<Value *, Value *>::iterator I = ChosenPairs.begin(),
780 IE = ChosenPairs.end(); I != IE; ++I) {
781 if (FixedOrderPairs.count(*I))
782 AllFixedOrderPairs.insert(*I);
783 else if (FixedOrderPairs.count(ValuePair(I->second, I->first)))
784 AllFixedOrderPairs.insert(ValuePair(I->second, I->first));
786 for (DenseMap<Value *, Value *>::iterator J = ChosenPairs.begin();
788 DenseMap<VPPair, unsigned>::iterator K =
789 PairConnectionTypes.find(VPPair(*I, *J));
790 if (K != PairConnectionTypes.end()) {
791 AllPairConnectionTypes.insert(*K);
793 K = PairConnectionTypes.find(VPPair(*J, *I));
794 if (K != PairConnectionTypes.end())
795 AllPairConnectionTypes.insert(*K);
800 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator
801 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
803 for (std::vector<ValuePair>::iterator J = I->second.begin(),
804 JE = I->second.end(); J != JE; ++J)
805 if (AllPairConnectionTypes.count(VPPair(I->first, *J))) {
806 AllConnectedPairs[I->first].push_back(*J);
807 AllConnectedPairDeps[*J].push_back(I->first);
809 } while (ShouldContinue);
811 if (AllChosenPairs.empty()) return false;
812 NumFusedOps += AllChosenPairs.size();
814 // A set of pairs has now been selected. It is now necessary to replace the
815 // paired instructions with vector instructions. For this procedure each
816 // operand must be replaced with a vector operand. This vector is formed
817 // by using build_vector on the old operands. The replaced values are then
818 // replaced with a vector_extract on the result. Subsequent optimization
819 // passes should coalesce the build/extract combinations.
821 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs, AllFixedOrderPairs,
822 AllPairConnectionTypes,
823 AllConnectedPairs, AllConnectedPairDeps);
825 // It is important to cleanup here so that future iterations of this
826 // function have less work to do.
827 (void) SimplifyInstructionsInBlock(&BB, DL, AA->getTargetLibraryInfo());
831 // This function returns true if the provided instruction is capable of being
832 // fused into a vector instruction. This determination is based only on the
833 // type and other attributes of the instruction.
834 bool BBVectorize::isInstVectorizable(Instruction *I,
835 bool &IsSimpleLoadStore) {
836 IsSimpleLoadStore = false;
838 if (CallInst *C = dyn_cast<CallInst>(I)) {
839 if (!isVectorizableIntrinsic(C))
841 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) {
842 // Vectorize simple loads if possbile:
843 IsSimpleLoadStore = L->isSimple();
844 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
846 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) {
847 // Vectorize simple stores if possbile:
848 IsSimpleLoadStore = S->isSimple();
849 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
851 } else if (CastInst *C = dyn_cast<CastInst>(I)) {
852 // We can vectorize casts, but not casts of pointer types, etc.
853 if (!Config.VectorizeCasts)
856 Type *SrcTy = C->getSrcTy();
857 if (!SrcTy->isSingleValueType())
860 Type *DestTy = C->getDestTy();
861 if (!DestTy->isSingleValueType())
863 } else if (isa<SelectInst>(I)) {
864 if (!Config.VectorizeSelect)
866 } else if (isa<CmpInst>(I)) {
867 if (!Config.VectorizeCmp)
869 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) {
870 if (!Config.VectorizeGEP)
873 // Currently, vector GEPs exist only with one index.
874 if (G->getNumIndices() != 1)
876 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) ||
877 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) {
881 // We can't vectorize memory operations without target data
882 if (DL == 0 && IsSimpleLoadStore)
886 getInstructionTypes(I, T1, T2);
888 // Not every type can be vectorized...
889 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) ||
890 !(VectorType::isValidElementType(T2) || T2->isVectorTy()))
893 if (T1->getScalarSizeInBits() == 1) {
894 if (!Config.VectorizeBools)
897 if (!Config.VectorizeInts && T1->isIntOrIntVectorTy())
901 if (T2->getScalarSizeInBits() == 1) {
902 if (!Config.VectorizeBools)
905 if (!Config.VectorizeInts && T2->isIntOrIntVectorTy())
909 if (!Config.VectorizeFloats
910 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy()))
913 // Don't vectorize target-specific types.
914 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy())
916 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy())
919 if ((!Config.VectorizePointers || DL == 0) &&
920 (T1->getScalarType()->isPointerTy() ||
921 T2->getScalarType()->isPointerTy()))
924 if (!TTI && (T1->getPrimitiveSizeInBits() >= Config.VectorBits ||
925 T2->getPrimitiveSizeInBits() >= Config.VectorBits))
931 // This function returns true if the two provided instructions are compatible
932 // (meaning that they can be fused into a vector instruction). This assumes
933 // that I has already been determined to be vectorizable and that J is not
934 // in the use dag of I.
935 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J,
936 bool IsSimpleLoadStore, bool NonPow2Len,
937 int &CostSavings, int &FixedOrder) {
938 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I <<
939 " <-> " << *J << "\n");
944 // Loads and stores can be merged if they have different alignments,
945 // but are otherwise the same.
946 if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment |
947 (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0)))
950 Type *IT1, *IT2, *JT1, *JT2;
951 getInstructionTypes(I, IT1, IT2);
952 getInstructionTypes(J, JT1, JT2);
953 unsigned MaxTypeBits = std::max(
954 IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(),
955 IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits());
956 if (!TTI && MaxTypeBits > Config.VectorBits)
959 // FIXME: handle addsub-type operations!
961 if (IsSimpleLoadStore) {
963 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
964 int64_t OffsetInElmts = 0;
965 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
966 IAddressSpace, JAddressSpace,
967 OffsetInElmts) && abs64(OffsetInElmts) == 1) {
968 FixedOrder = (int) OffsetInElmts;
969 unsigned BottomAlignment = IAlignment;
970 if (OffsetInElmts < 0) BottomAlignment = JAlignment;
972 Type *aTypeI = isa<StoreInst>(I) ?
973 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType();
974 Type *aTypeJ = isa<StoreInst>(J) ?
975 cast<StoreInst>(J)->getValueOperand()->getType() : J->getType();
976 Type *VType = getVecTypeForPair(aTypeI, aTypeJ);
978 if (Config.AlignedOnly) {
979 // An aligned load or store is possible only if the instruction
980 // with the lower offset has an alignment suitable for the
983 unsigned VecAlignment = DL->getPrefTypeAlignment(VType);
984 if (BottomAlignment < VecAlignment)
989 unsigned ICost = TTI->getMemoryOpCost(I->getOpcode(), aTypeI,
990 IAlignment, IAddressSpace);
991 unsigned JCost = TTI->getMemoryOpCost(J->getOpcode(), aTypeJ,
992 JAlignment, JAddressSpace);
993 unsigned VCost = TTI->getMemoryOpCost(I->getOpcode(), VType,
997 ICost += TTI->getAddressComputationCost(aTypeI);
998 JCost += TTI->getAddressComputationCost(aTypeJ);
999 VCost += TTI->getAddressComputationCost(VType);
1001 if (VCost > ICost + JCost)
1004 // We don't want to fuse to a type that will be split, even
1005 // if the two input types will also be split and there is no other
1007 unsigned VParts = TTI->getNumberOfParts(VType);
1010 else if (!VParts && VCost == ICost + JCost)
1013 CostSavings = ICost + JCost - VCost;
1019 unsigned ICost = getInstrCost(I->getOpcode(), IT1, IT2);
1020 unsigned JCost = getInstrCost(J->getOpcode(), JT1, JT2);
1021 Type *VT1 = getVecTypeForPair(IT1, JT1),
1022 *VT2 = getVecTypeForPair(IT2, JT2);
1023 TargetTransformInfo::OperandValueKind Op1VK =
1024 TargetTransformInfo::OK_AnyValue;
1025 TargetTransformInfo::OperandValueKind Op2VK =
1026 TargetTransformInfo::OK_AnyValue;
1028 // On some targets (example X86) the cost of a vector shift may vary
1029 // depending on whether the second operand is a Uniform or
1030 // NonUniform Constant.
1031 switch (I->getOpcode()) {
1033 case Instruction::Shl:
1034 case Instruction::LShr:
1035 case Instruction::AShr:
1037 // If both I and J are scalar shifts by constant, then the
1038 // merged vector shift count would be either a constant splat value
1039 // or a non-uniform vector of constants.
1040 if (ConstantInt *CII = dyn_cast<ConstantInt>(I->getOperand(1))) {
1041 if (ConstantInt *CIJ = dyn_cast<ConstantInt>(J->getOperand(1)))
1042 Op2VK = CII == CIJ ? TargetTransformInfo::OK_UniformConstantValue :
1043 TargetTransformInfo::OK_NonUniformConstantValue;
1045 // Check for a splat of a constant or for a non uniform vector
1047 Value *IOp = I->getOperand(1);
1048 Value *JOp = J->getOperand(1);
1049 if ((isa<ConstantVector>(IOp) || isa<ConstantDataVector>(IOp)) &&
1050 (isa<ConstantVector>(JOp) || isa<ConstantDataVector>(JOp))) {
1051 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
1052 Constant *SplatValue = cast<Constant>(IOp)->getSplatValue();
1053 if (SplatValue != NULL &&
1054 SplatValue == cast<Constant>(JOp)->getSplatValue())
1055 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
1060 // Note that this procedure is incorrect for insert and extract element
1061 // instructions (because combining these often results in a shuffle),
1062 // but this cost is ignored (because insert and extract element
1063 // instructions are assigned a zero depth factor and are not really
1064 // fused in general).
1065 unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2, Op1VK, Op2VK);
1067 if (VCost > ICost + JCost)
1070 // We don't want to fuse to a type that will be split, even
1071 // if the two input types will also be split and there is no other
1073 unsigned VParts1 = TTI->getNumberOfParts(VT1),
1074 VParts2 = TTI->getNumberOfParts(VT2);
1075 if (VParts1 > 1 || VParts2 > 1)
1077 else if ((!VParts1 || !VParts2) && VCost == ICost + JCost)
1080 CostSavings = ICost + JCost - VCost;
1083 // The powi intrinsic is special because only the first argument is
1084 // vectorized, the second arguments must be equal.
1085 CallInst *CI = dyn_cast<CallInst>(I);
1087 if (CI && (FI = CI->getCalledFunction())) {
1088 Intrinsic::ID IID = (Intrinsic::ID) FI->getIntrinsicID();
1089 if (IID == Intrinsic::powi) {
1090 Value *A1I = CI->getArgOperand(1),
1091 *A1J = cast<CallInst>(J)->getArgOperand(1);
1092 const SCEV *A1ISCEV = SE->getSCEV(A1I),
1093 *A1JSCEV = SE->getSCEV(A1J);
1094 return (A1ISCEV == A1JSCEV);
1098 SmallVector<Type*, 4> Tys;
1099 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
1100 Tys.push_back(CI->getArgOperand(i)->getType());
1101 unsigned ICost = TTI->getIntrinsicInstrCost(IID, IT1, Tys);
1104 CallInst *CJ = cast<CallInst>(J);
1105 for (unsigned i = 0, ie = CJ->getNumArgOperands(); i != ie; ++i)
1106 Tys.push_back(CJ->getArgOperand(i)->getType());
1107 unsigned JCost = TTI->getIntrinsicInstrCost(IID, JT1, Tys);
1110 assert(CI->getNumArgOperands() == CJ->getNumArgOperands() &&
1111 "Intrinsic argument counts differ");
1112 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
1113 if (IID == Intrinsic::powi && i == 1)
1114 Tys.push_back(CI->getArgOperand(i)->getType());
1116 Tys.push_back(getVecTypeForPair(CI->getArgOperand(i)->getType(),
1117 CJ->getArgOperand(i)->getType()));
1120 Type *RetTy = getVecTypeForPair(IT1, JT1);
1121 unsigned VCost = TTI->getIntrinsicInstrCost(IID, RetTy, Tys);
1123 if (VCost > ICost + JCost)
1126 // We don't want to fuse to a type that will be split, even
1127 // if the two input types will also be split and there is no other
1129 unsigned RetParts = TTI->getNumberOfParts(RetTy);
1132 else if (!RetParts && VCost == ICost + JCost)
1135 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
1136 if (!Tys[i]->isVectorTy())
1139 unsigned NumParts = TTI->getNumberOfParts(Tys[i]);
1142 else if (!NumParts && VCost == ICost + JCost)
1146 CostSavings = ICost + JCost - VCost;
1153 // Figure out whether or not J uses I and update the users and write-set
1154 // structures associated with I. Specifically, Users represents the set of
1155 // instructions that depend on I. WriteSet represents the set
1156 // of memory locations that are dependent on I. If UpdateUsers is true,
1157 // and J uses I, then Users is updated to contain J and WriteSet is updated
1158 // to contain any memory locations to which J writes. The function returns
1159 // true if J uses I. By default, alias analysis is used to determine
1160 // whether J reads from memory that overlaps with a location in WriteSet.
1161 // If LoadMoveSet is not null, then it is a previously-computed map
1162 // where the key is the memory-based user instruction and the value is
1163 // the instruction to be compared with I. So, if LoadMoveSet is provided,
1164 // then the alias analysis is not used. This is necessary because this
1165 // function is called during the process of moving instructions during
1166 // vectorization and the results of the alias analysis are not stable during
1168 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users,
1169 AliasSetTracker &WriteSet, Instruction *I,
1170 Instruction *J, bool UpdateUsers,
1171 DenseSet<ValuePair> *LoadMoveSetPairs) {
1174 // This instruction may already be marked as a user due, for example, to
1175 // being a member of a selected pair.
1180 for (User::op_iterator JU = J->op_begin(), JE = J->op_end();
1183 if (I == V || Users.count(V)) {
1188 if (!UsesI && J->mayReadFromMemory()) {
1189 if (LoadMoveSetPairs) {
1190 UsesI = LoadMoveSetPairs->count(ValuePair(J, I));
1192 for (AliasSetTracker::iterator W = WriteSet.begin(),
1193 WE = WriteSet.end(); W != WE; ++W) {
1194 if (W->aliasesUnknownInst(J, *AA)) {
1202 if (UsesI && UpdateUsers) {
1203 if (J->mayWriteToMemory()) WriteSet.add(J);
1210 // This function iterates over all instruction pairs in the provided
1211 // basic block and collects all candidate pairs for vectorization.
1212 bool BBVectorize::getCandidatePairs(BasicBlock &BB,
1213 BasicBlock::iterator &Start,
1214 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1215 DenseSet<ValuePair> &FixedOrderPairs,
1216 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1217 std::vector<Value *> &PairableInsts, bool NonPow2Len) {
1218 size_t TotalPairs = 0;
1219 BasicBlock::iterator E = BB.end();
1220 if (Start == E) return false;
1222 bool ShouldContinue = false, IAfterStart = false;
1223 for (BasicBlock::iterator I = Start++; I != E; ++I) {
1224 if (I == Start) IAfterStart = true;
1226 bool IsSimpleLoadStore;
1227 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue;
1229 // Look for an instruction with which to pair instruction *I...
1230 DenseSet<Value *> Users;
1231 AliasSetTracker WriteSet(*AA);
1232 if (I->mayWriteToMemory()) WriteSet.add(I);
1234 bool JAfterStart = IAfterStart;
1235 BasicBlock::iterator J = std::next(I);
1236 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) {
1237 if (J == Start) JAfterStart = true;
1239 // Determine if J uses I, if so, exit the loop.
1240 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep);
1241 if (Config.FastDep) {
1242 // Note: For this heuristic to be effective, independent operations
1243 // must tend to be intermixed. This is likely to be true from some
1244 // kinds of grouped loop unrolling (but not the generic LLVM pass),
1245 // but otherwise may require some kind of reordering pass.
1247 // When using fast dependency analysis,
1248 // stop searching after first use:
1251 if (UsesI) continue;
1254 // J does not use I, and comes before the first use of I, so it can be
1255 // merged with I if the instructions are compatible.
1256 int CostSavings, FixedOrder;
1257 if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len,
1258 CostSavings, FixedOrder)) continue;
1260 // J is a candidate for merging with I.
1261 if (!PairableInsts.size() ||
1262 PairableInsts[PairableInsts.size()-1] != I) {
1263 PairableInsts.push_back(I);
1266 CandidatePairs[I].push_back(J);
1269 CandidatePairCostSavings.insert(ValuePairWithCost(ValuePair(I, J),
1272 if (FixedOrder == 1)
1273 FixedOrderPairs.insert(ValuePair(I, J));
1274 else if (FixedOrder == -1)
1275 FixedOrderPairs.insert(ValuePair(J, I));
1277 // The next call to this function must start after the last instruction
1278 // selected during this invocation.
1280 Start = std::next(J);
1281 IAfterStart = JAfterStart = false;
1284 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair "
1285 << *I << " <-> " << *J << " (cost savings: " <<
1286 CostSavings << ")\n");
1288 // If we have already found too many pairs, break here and this function
1289 // will be called again starting after the last instruction selected
1290 // during this invocation.
1291 if (PairableInsts.size() >= Config.MaxInsts ||
1292 TotalPairs >= Config.MaxPairs) {
1293 ShouldContinue = true;
1302 DEBUG(dbgs() << "BBV: found " << PairableInsts.size()
1303 << " instructions with candidate pairs\n");
1305 return ShouldContinue;
1308 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that
1309 // it looks for pairs such that both members have an input which is an
1310 // output of PI or PJ.
1311 void BBVectorize::computePairsConnectedTo(
1312 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1313 DenseSet<ValuePair> &CandidatePairsSet,
1314 std::vector<Value *> &PairableInsts,
1315 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1316 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1320 // For each possible pairing for this variable, look at the uses of
1321 // the first value...
1322 for (Value::user_iterator I = P.first->user_begin(),
1323 E = P.first->user_end();
1326 if (isa<LoadInst>(UI)) {
1327 // A pair cannot be connected to a load because the load only takes one
1328 // operand (the address) and it is a scalar even after vectorization.
1330 } else if ((SI = dyn_cast<StoreInst>(UI)) &&
1331 P.first == SI->getPointerOperand()) {
1332 // Similarly, a pair cannot be connected to a store through its
1337 // For each use of the first variable, look for uses of the second
1339 for (User *UJ : P.second->users()) {
1340 if ((SJ = dyn_cast<StoreInst>(UJ)) &&
1341 P.second == SJ->getPointerOperand())
1345 if (CandidatePairsSet.count(ValuePair(UI, UJ))) {
1346 VPPair VP(P, ValuePair(UI, UJ));
1347 ConnectedPairs[VP.first].push_back(VP.second);
1348 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionDirect));
1352 if (CandidatePairsSet.count(ValuePair(UJ, UI))) {
1353 VPPair VP(P, ValuePair(UJ, UI));
1354 ConnectedPairs[VP.first].push_back(VP.second);
1355 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSwap));
1359 if (Config.SplatBreaksChain) continue;
1360 // Look for cases where just the first value in the pair is used by
1361 // both members of another pair (splatting).
1362 for (Value::user_iterator J = P.first->user_begin(); J != E; ++J) {
1364 if ((SJ = dyn_cast<StoreInst>(UJ)) &&
1365 P.first == SJ->getPointerOperand())
1368 if (CandidatePairsSet.count(ValuePair(UI, UJ))) {
1369 VPPair VP(P, ValuePair(UI, UJ));
1370 ConnectedPairs[VP.first].push_back(VP.second);
1371 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1376 if (Config.SplatBreaksChain) return;
1377 // Look for cases where just the second value in the pair is used by
1378 // both members of another pair (splatting).
1379 for (Value::user_iterator I = P.second->user_begin(),
1380 E = P.second->user_end();
1383 if (isa<LoadInst>(UI))
1385 else if ((SI = dyn_cast<StoreInst>(UI)) &&
1386 P.second == SI->getPointerOperand())
1389 for (Value::user_iterator J = P.second->user_begin(); J != E; ++J) {
1391 if ((SJ = dyn_cast<StoreInst>(UJ)) &&
1392 P.second == SJ->getPointerOperand())
1395 if (CandidatePairsSet.count(ValuePair(UI, UJ))) {
1396 VPPair VP(P, ValuePair(UI, UJ));
1397 ConnectedPairs[VP.first].push_back(VP.second);
1398 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1404 // This function figures out which pairs are connected. Two pairs are
1405 // connected if some output of the first pair forms an input to both members
1406 // of the second pair.
1407 void BBVectorize::computeConnectedPairs(
1408 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1409 DenseSet<ValuePair> &CandidatePairsSet,
1410 std::vector<Value *> &PairableInsts,
1411 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1412 DenseMap<VPPair, unsigned> &PairConnectionTypes) {
1413 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
1414 PE = PairableInsts.end(); PI != PE; ++PI) {
1415 DenseMap<Value *, std::vector<Value *> >::iterator PP =
1416 CandidatePairs.find(*PI);
1417 if (PP == CandidatePairs.end())
1420 for (std::vector<Value *>::iterator P = PP->second.begin(),
1421 E = PP->second.end(); P != E; ++P)
1422 computePairsConnectedTo(CandidatePairs, CandidatePairsSet,
1423 PairableInsts, ConnectedPairs,
1424 PairConnectionTypes, ValuePair(*PI, *P));
1427 DEBUG(size_t TotalPairs = 0;
1428 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator I =
1429 ConnectedPairs.begin(), IE = ConnectedPairs.end(); I != IE; ++I)
1430 TotalPairs += I->second.size();
1431 dbgs() << "BBV: found " << TotalPairs
1432 << " pair connections.\n");
1435 // This function builds a set of use tuples such that <A, B> is in the set
1436 // if B is in the use dag of A. If B is in the use dag of A, then B
1437 // depends on the output of A.
1438 void BBVectorize::buildDepMap(
1440 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1441 std::vector<Value *> &PairableInsts,
1442 DenseSet<ValuePair> &PairableInstUsers) {
1443 DenseSet<Value *> IsInPair;
1444 for (DenseMap<Value *, std::vector<Value *> >::iterator C =
1445 CandidatePairs.begin(), E = CandidatePairs.end(); C != E; ++C) {
1446 IsInPair.insert(C->first);
1447 IsInPair.insert(C->second.begin(), C->second.end());
1450 // Iterate through the basic block, recording all users of each
1451 // pairable instruction.
1453 BasicBlock::iterator E = BB.end(), EL =
1454 BasicBlock::iterator(cast<Instruction>(PairableInsts.back()));
1455 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) {
1456 if (IsInPair.find(I) == IsInPair.end()) continue;
1458 DenseSet<Value *> Users;
1459 AliasSetTracker WriteSet(*AA);
1460 if (I->mayWriteToMemory()) WriteSet.add(I);
1462 for (BasicBlock::iterator J = std::next(I); J != E; ++J) {
1463 (void) trackUsesOfI(Users, WriteSet, I, J);
1469 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end();
1471 if (IsInPair.find(*U) == IsInPair.end()) continue;
1472 PairableInstUsers.insert(ValuePair(I, *U));
1480 // Returns true if an input to pair P is an output of pair Q and also an
1481 // input of pair Q is an output of pair P. If this is the case, then these
1482 // two pairs cannot be simultaneously fused.
1483 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q,
1484 DenseSet<ValuePair> &PairableInstUsers,
1485 DenseMap<ValuePair, std::vector<ValuePair> > *PairableInstUserMap,
1486 DenseSet<VPPair> *PairableInstUserPairSet) {
1487 // Two pairs are in conflict if they are mutual Users of eachother.
1488 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) ||
1489 PairableInstUsers.count(ValuePair(P.first, Q.second)) ||
1490 PairableInstUsers.count(ValuePair(P.second, Q.first)) ||
1491 PairableInstUsers.count(ValuePair(P.second, Q.second));
1492 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) ||
1493 PairableInstUsers.count(ValuePair(Q.first, P.second)) ||
1494 PairableInstUsers.count(ValuePair(Q.second, P.first)) ||
1495 PairableInstUsers.count(ValuePair(Q.second, P.second));
1496 if (PairableInstUserMap) {
1497 // FIXME: The expensive part of the cycle check is not so much the cycle
1498 // check itself but this edge insertion procedure. This needs some
1499 // profiling and probably a different data structure.
1501 if (PairableInstUserPairSet->insert(VPPair(Q, P)).second)
1502 (*PairableInstUserMap)[Q].push_back(P);
1505 if (PairableInstUserPairSet->insert(VPPair(P, Q)).second)
1506 (*PairableInstUserMap)[P].push_back(Q);
1510 return (QUsesP && PUsesQ);
1513 // This function walks the use graph of current pairs to see if, starting
1514 // from P, the walk returns to P.
1515 bool BBVectorize::pairWillFormCycle(ValuePair P,
1516 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
1517 DenseSet<ValuePair> &CurrentPairs) {
1518 DEBUG(if (DebugCycleCheck)
1519 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> "
1520 << *P.second << "\n");
1521 // A lookup table of visisted pairs is kept because the PairableInstUserMap
1522 // contains non-direct associations.
1523 DenseSet<ValuePair> Visited;
1524 SmallVector<ValuePair, 32> Q;
1525 // General depth-first post-order traversal:
1528 ValuePair QTop = Q.pop_back_val();
1529 Visited.insert(QTop);
1531 DEBUG(if (DebugCycleCheck)
1532 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> "
1533 << *QTop.second << "\n");
1534 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
1535 PairableInstUserMap.find(QTop);
1536 if (QQ == PairableInstUserMap.end())
1539 for (std::vector<ValuePair>::iterator C = QQ->second.begin(),
1540 CE = QQ->second.end(); C != CE; ++C) {
1543 << "BBV: rejected to prevent non-trivial cycle formation: "
1544 << QTop.first << " <-> " << C->second << "\n");
1548 if (CurrentPairs.count(*C) && !Visited.count(*C))
1551 } while (!Q.empty());
1556 // This function builds the initial dag of connected pairs with the
1557 // pair J at the root.
1558 void BBVectorize::buildInitialDAGFor(
1559 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1560 DenseSet<ValuePair> &CandidatePairsSet,
1561 std::vector<Value *> &PairableInsts,
1562 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1563 DenseSet<ValuePair> &PairableInstUsers,
1564 DenseMap<Value *, Value *> &ChosenPairs,
1565 DenseMap<ValuePair, size_t> &DAG, ValuePair J) {
1566 // Each of these pairs is viewed as the root node of a DAG. The DAG
1567 // is then walked (depth-first). As this happens, we keep track of
1568 // the pairs that compose the DAG and the maximum depth of the DAG.
1569 SmallVector<ValuePairWithDepth, 32> Q;
1570 // General depth-first post-order traversal:
1571 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1573 ValuePairWithDepth QTop = Q.back();
1575 // Push each child onto the queue:
1576 bool MoreChildren = false;
1577 size_t MaxChildDepth = QTop.second;
1578 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
1579 ConnectedPairs.find(QTop.first);
1580 if (QQ != ConnectedPairs.end())
1581 for (std::vector<ValuePair>::iterator k = QQ->second.begin(),
1582 ke = QQ->second.end(); k != ke; ++k) {
1583 // Make sure that this child pair is still a candidate:
1584 if (CandidatePairsSet.count(*k)) {
1585 DenseMap<ValuePair, size_t>::iterator C = DAG.find(*k);
1586 if (C == DAG.end()) {
1587 size_t d = getDepthFactor(k->first);
1588 Q.push_back(ValuePairWithDepth(*k, QTop.second+d));
1589 MoreChildren = true;
1591 MaxChildDepth = std::max(MaxChildDepth, C->second);
1596 if (!MoreChildren) {
1597 // Record the current pair as part of the DAG:
1598 DAG.insert(ValuePairWithDepth(QTop.first, MaxChildDepth));
1601 } while (!Q.empty());
1604 // Given some initial dag, prune it by removing conflicting pairs (pairs
1605 // that cannot be simultaneously chosen for vectorization).
1606 void BBVectorize::pruneDAGFor(
1607 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1608 std::vector<Value *> &PairableInsts,
1609 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1610 DenseSet<ValuePair> &PairableInstUsers,
1611 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
1612 DenseSet<VPPair> &PairableInstUserPairSet,
1613 DenseMap<Value *, Value *> &ChosenPairs,
1614 DenseMap<ValuePair, size_t> &DAG,
1615 DenseSet<ValuePair> &PrunedDAG, ValuePair J,
1616 bool UseCycleCheck) {
1617 SmallVector<ValuePairWithDepth, 32> Q;
1618 // General depth-first post-order traversal:
1619 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1621 ValuePairWithDepth QTop = Q.pop_back_val();
1622 PrunedDAG.insert(QTop.first);
1624 // Visit each child, pruning as necessary...
1625 SmallVector<ValuePairWithDepth, 8> BestChildren;
1626 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
1627 ConnectedPairs.find(QTop.first);
1628 if (QQ == ConnectedPairs.end())
1631 for (std::vector<ValuePair>::iterator K = QQ->second.begin(),
1632 KE = QQ->second.end(); K != KE; ++K) {
1633 DenseMap<ValuePair, size_t>::iterator C = DAG.find(*K);
1634 if (C == DAG.end()) continue;
1636 // This child is in the DAG, now we need to make sure it is the
1637 // best of any conflicting children. There could be multiple
1638 // conflicting children, so first, determine if we're keeping
1639 // this child, then delete conflicting children as necessary.
1641 // It is also necessary to guard against pairing-induced
1642 // dependencies. Consider instructions a .. x .. y .. b
1643 // such that (a,b) are to be fused and (x,y) are to be fused
1644 // but a is an input to x and b is an output from y. This
1645 // means that y cannot be moved after b but x must be moved
1646 // after b for (a,b) to be fused. In other words, after
1647 // fusing (a,b) we have y .. a/b .. x where y is an input
1648 // to a/b and x is an output to a/b: x and y can no longer
1649 // be legally fused. To prevent this condition, we must
1650 // make sure that a child pair added to the DAG is not
1651 // both an input and output of an already-selected pair.
1653 // Pairing-induced dependencies can also form from more complicated
1654 // cycles. The pair vs. pair conflicts are easy to check, and so
1655 // that is done explicitly for "fast rejection", and because for
1656 // child vs. child conflicts, we may prefer to keep the current
1657 // pair in preference to the already-selected child.
1658 DenseSet<ValuePair> CurrentPairs;
1661 for (SmallVectorImpl<ValuePairWithDepth>::iterator C2
1662 = BestChildren.begin(), E2 = BestChildren.end();
1664 if (C2->first.first == C->first.first ||
1665 C2->first.first == C->first.second ||
1666 C2->first.second == C->first.first ||
1667 C2->first.second == C->first.second ||
1668 pairsConflict(C2->first, C->first, PairableInstUsers,
1669 UseCycleCheck ? &PairableInstUserMap : 0,
1670 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1671 if (C2->second >= C->second) {
1676 CurrentPairs.insert(C2->first);
1679 if (!CanAdd) continue;
1681 // Even worse, this child could conflict with another node already
1682 // selected for the DAG. If that is the case, ignore this child.
1683 for (DenseSet<ValuePair>::iterator T = PrunedDAG.begin(),
1684 E2 = PrunedDAG.end(); T != E2; ++T) {
1685 if (T->first == C->first.first ||
1686 T->first == C->first.second ||
1687 T->second == C->first.first ||
1688 T->second == C->first.second ||
1689 pairsConflict(*T, C->first, PairableInstUsers,
1690 UseCycleCheck ? &PairableInstUserMap : 0,
1691 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1696 CurrentPairs.insert(*T);
1698 if (!CanAdd) continue;
1700 // And check the queue too...
1701 for (SmallVectorImpl<ValuePairWithDepth>::iterator C2 = Q.begin(),
1702 E2 = Q.end(); C2 != E2; ++C2) {
1703 if (C2->first.first == C->first.first ||
1704 C2->first.first == C->first.second ||
1705 C2->first.second == C->first.first ||
1706 C2->first.second == C->first.second ||
1707 pairsConflict(C2->first, C->first, PairableInstUsers,
1708 UseCycleCheck ? &PairableInstUserMap : 0,
1709 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1714 CurrentPairs.insert(C2->first);
1716 if (!CanAdd) continue;
1718 // Last but not least, check for a conflict with any of the
1719 // already-chosen pairs.
1720 for (DenseMap<Value *, Value *>::iterator C2 =
1721 ChosenPairs.begin(), E2 = ChosenPairs.end();
1723 if (pairsConflict(*C2, C->first, PairableInstUsers,
1724 UseCycleCheck ? &PairableInstUserMap : 0,
1725 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1730 CurrentPairs.insert(*C2);
1732 if (!CanAdd) continue;
1734 // To check for non-trivial cycles formed by the addition of the
1735 // current pair we've formed a list of all relevant pairs, now use a
1736 // graph walk to check for a cycle. We start from the current pair and
1737 // walk the use dag to see if we again reach the current pair. If we
1738 // do, then the current pair is rejected.
1740 // FIXME: It may be more efficient to use a topological-ordering
1741 // algorithm to improve the cycle check. This should be investigated.
1742 if (UseCycleCheck &&
1743 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs))
1746 // This child can be added, but we may have chosen it in preference
1747 // to an already-selected child. Check for this here, and if a
1748 // conflict is found, then remove the previously-selected child
1749 // before adding this one in its place.
1750 for (SmallVectorImpl<ValuePairWithDepth>::iterator C2
1751 = BestChildren.begin(); C2 != BestChildren.end();) {
1752 if (C2->first.first == C->first.first ||
1753 C2->first.first == C->first.second ||
1754 C2->first.second == C->first.first ||
1755 C2->first.second == C->first.second ||
1756 pairsConflict(C2->first, C->first, PairableInstUsers))
1757 C2 = BestChildren.erase(C2);
1762 BestChildren.push_back(ValuePairWithDepth(C->first, C->second));
1765 for (SmallVectorImpl<ValuePairWithDepth>::iterator C
1766 = BestChildren.begin(), E2 = BestChildren.end();
1768 size_t DepthF = getDepthFactor(C->first.first);
1769 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF));
1771 } while (!Q.empty());
1774 // This function finds the best dag of mututally-compatible connected
1775 // pairs, given the choice of root pairs as an iterator range.
1776 void BBVectorize::findBestDAGFor(
1777 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1778 DenseSet<ValuePair> &CandidatePairsSet,
1779 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1780 std::vector<Value *> &PairableInsts,
1781 DenseSet<ValuePair> &FixedOrderPairs,
1782 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1783 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1784 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
1785 DenseSet<ValuePair> &PairableInstUsers,
1786 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
1787 DenseSet<VPPair> &PairableInstUserPairSet,
1788 DenseMap<Value *, Value *> &ChosenPairs,
1789 DenseSet<ValuePair> &BestDAG, size_t &BestMaxDepth,
1790 int &BestEffSize, Value *II, std::vector<Value *>&JJ,
1791 bool UseCycleCheck) {
1792 for (std::vector<Value *>::iterator J = JJ.begin(), JE = JJ.end();
1794 ValuePair IJ(II, *J);
1795 if (!CandidatePairsSet.count(IJ))
1798 // Before going any further, make sure that this pair does not
1799 // conflict with any already-selected pairs (see comment below
1800 // near the DAG pruning for more details).
1801 DenseSet<ValuePair> ChosenPairSet;
1802 bool DoesConflict = false;
1803 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(),
1804 E = ChosenPairs.end(); C != E; ++C) {
1805 if (pairsConflict(*C, IJ, PairableInstUsers,
1806 UseCycleCheck ? &PairableInstUserMap : 0,
1807 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1808 DoesConflict = true;
1812 ChosenPairSet.insert(*C);
1814 if (DoesConflict) continue;
1816 if (UseCycleCheck &&
1817 pairWillFormCycle(IJ, PairableInstUserMap, ChosenPairSet))
1820 DenseMap<ValuePair, size_t> DAG;
1821 buildInitialDAGFor(CandidatePairs, CandidatePairsSet,
1822 PairableInsts, ConnectedPairs,
1823 PairableInstUsers, ChosenPairs, DAG, IJ);
1825 // Because we'll keep the child with the largest depth, the largest
1826 // depth is still the same in the unpruned DAG.
1827 size_t MaxDepth = DAG.lookup(IJ);
1829 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found DAG for pair {"
1830 << *IJ.first << " <-> " << *IJ.second << "} of depth " <<
1831 MaxDepth << " and size " << DAG.size() << "\n");
1833 // At this point the DAG has been constructed, but, may contain
1834 // contradictory children (meaning that different children of
1835 // some dag node may be attempting to fuse the same instruction).
1836 // So now we walk the dag again, in the case of a conflict,
1837 // keep only the child with the largest depth. To break a tie,
1838 // favor the first child.
1840 DenseSet<ValuePair> PrunedDAG;
1841 pruneDAGFor(CandidatePairs, PairableInsts, ConnectedPairs,
1842 PairableInstUsers, PairableInstUserMap,
1843 PairableInstUserPairSet,
1844 ChosenPairs, DAG, PrunedDAG, IJ, UseCycleCheck);
1848 DenseSet<Value *> PrunedDAGInstrs;
1849 for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(),
1850 E = PrunedDAG.end(); S != E; ++S) {
1851 PrunedDAGInstrs.insert(S->first);
1852 PrunedDAGInstrs.insert(S->second);
1855 // The set of pairs that have already contributed to the total cost.
1856 DenseSet<ValuePair> IncomingPairs;
1858 // If the cost model were perfect, this might not be necessary; but we
1859 // need to make sure that we don't get stuck vectorizing our own
1861 bool HasNontrivialInsts = false;
1863 // The node weights represent the cost savings associated with
1864 // fusing the pair of instructions.
1865 for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(),
1866 E = PrunedDAG.end(); S != E; ++S) {
1867 if (!isa<ShuffleVectorInst>(S->first) &&
1868 !isa<InsertElementInst>(S->first) &&
1869 !isa<ExtractElementInst>(S->first))
1870 HasNontrivialInsts = true;
1872 bool FlipOrder = false;
1874 if (getDepthFactor(S->first)) {
1875 int ESContrib = CandidatePairCostSavings.find(*S)->second;
1876 DEBUG(if (DebugPairSelection) dbgs() << "\tweight {"
1877 << *S->first << " <-> " << *S->second << "} = " <<
1879 EffSize += ESContrib;
1882 // The edge weights contribute in a negative sense: they represent
1883 // the cost of shuffles.
1884 DenseMap<ValuePair, std::vector<ValuePair> >::iterator SS =
1885 ConnectedPairDeps.find(*S);
1886 if (SS != ConnectedPairDeps.end()) {
1887 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
1888 for (std::vector<ValuePair>::iterator T = SS->second.begin(),
1889 TE = SS->second.end(); T != TE; ++T) {
1891 if (!PrunedDAG.count(Q.second))
1893 DenseMap<VPPair, unsigned>::iterator R =
1894 PairConnectionTypes.find(VPPair(Q.second, Q.first));
1895 assert(R != PairConnectionTypes.end() &&
1896 "Cannot find pair connection type");
1897 if (R->second == PairConnectionDirect)
1899 else if (R->second == PairConnectionSwap)
1903 // If there are more swaps than direct connections, then
1904 // the pair order will be flipped during fusion. So the real
1905 // number of swaps is the minimum number.
1906 FlipOrder = !FixedOrderPairs.count(*S) &&
1907 ((NumDepsSwap > NumDepsDirect) ||
1908 FixedOrderPairs.count(ValuePair(S->second, S->first)));
1910 for (std::vector<ValuePair>::iterator T = SS->second.begin(),
1911 TE = SS->second.end(); T != TE; ++T) {
1913 if (!PrunedDAG.count(Q.second))
1915 DenseMap<VPPair, unsigned>::iterator R =
1916 PairConnectionTypes.find(VPPair(Q.second, Q.first));
1917 assert(R != PairConnectionTypes.end() &&
1918 "Cannot find pair connection type");
1919 Type *Ty1 = Q.second.first->getType(),
1920 *Ty2 = Q.second.second->getType();
1921 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1922 if ((R->second == PairConnectionDirect && FlipOrder) ||
1923 (R->second == PairConnectionSwap && !FlipOrder) ||
1924 R->second == PairConnectionSplat) {
1925 int ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1928 if (VTy->getVectorNumElements() == 2) {
1929 if (R->second == PairConnectionSplat)
1930 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1931 TargetTransformInfo::SK_Broadcast, VTy));
1933 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1934 TargetTransformInfo::SK_Reverse, VTy));
1937 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1938 *Q.second.first << " <-> " << *Q.second.second <<
1940 *S->first << " <-> " << *S->second << "} = " <<
1942 EffSize -= ESContrib;
1947 // Compute the cost of outgoing edges. We assume that edges outgoing
1948 // to shuffles, inserts or extracts can be merged, and so contribute
1949 // no additional cost.
1950 if (!S->first->getType()->isVoidTy()) {
1951 Type *Ty1 = S->first->getType(),
1952 *Ty2 = S->second->getType();
1953 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1955 bool NeedsExtraction = false;
1956 for (User *U : S->first->users()) {
1957 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(U)) {
1958 // Shuffle can be folded if it has no other input
1959 if (isa<UndefValue>(SI->getOperand(1)))
1962 if (isa<ExtractElementInst>(U))
1964 if (PrunedDAGInstrs.count(U))
1966 NeedsExtraction = true;
1970 if (NeedsExtraction) {
1972 if (Ty1->isVectorTy()) {
1973 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1975 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1976 TargetTransformInfo::SK_ExtractSubvector, VTy, 0, Ty1));
1978 ESContrib = (int) TTI->getVectorInstrCost(
1979 Instruction::ExtractElement, VTy, 0);
1981 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1982 *S->first << "} = " << ESContrib << "\n");
1983 EffSize -= ESContrib;
1986 NeedsExtraction = false;
1987 for (User *U : S->second->users()) {
1988 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(U)) {
1989 // Shuffle can be folded if it has no other input
1990 if (isa<UndefValue>(SI->getOperand(1)))
1993 if (isa<ExtractElementInst>(U))
1995 if (PrunedDAGInstrs.count(U))
1997 NeedsExtraction = true;
2001 if (NeedsExtraction) {
2003 if (Ty2->isVectorTy()) {
2004 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2006 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
2007 TargetTransformInfo::SK_ExtractSubvector, VTy,
2008 Ty1->isVectorTy() ? Ty1->getVectorNumElements() : 1, Ty2));
2010 ESContrib = (int) TTI->getVectorInstrCost(
2011 Instruction::ExtractElement, VTy, 1);
2012 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
2013 *S->second << "} = " << ESContrib << "\n");
2014 EffSize -= ESContrib;
2018 // Compute the cost of incoming edges.
2019 if (!isa<LoadInst>(S->first) && !isa<StoreInst>(S->first)) {
2020 Instruction *S1 = cast<Instruction>(S->first),
2021 *S2 = cast<Instruction>(S->second);
2022 for (unsigned o = 0; o < S1->getNumOperands(); ++o) {
2023 Value *O1 = S1->getOperand(o), *O2 = S2->getOperand(o);
2025 // Combining constants into vector constants (or small vector
2026 // constants into larger ones are assumed free).
2027 if (isa<Constant>(O1) && isa<Constant>(O2))
2033 ValuePair VP = ValuePair(O1, O2);
2034 ValuePair VPR = ValuePair(O2, O1);
2036 // Internal edges are not handled here.
2037 if (PrunedDAG.count(VP) || PrunedDAG.count(VPR))
2040 Type *Ty1 = O1->getType(),
2041 *Ty2 = O2->getType();
2042 Type *VTy = getVecTypeForPair(Ty1, Ty2);
2044 // Combining vector operations of the same type is also assumed
2045 // folded with other operations.
2047 // If both are insert elements, then both can be widened.
2048 InsertElementInst *IEO1 = dyn_cast<InsertElementInst>(O1),
2049 *IEO2 = dyn_cast<InsertElementInst>(O2);
2050 if (IEO1 && IEO2 && isPureIEChain(IEO1) && isPureIEChain(IEO2))
2052 // If both are extract elements, and both have the same input
2053 // type, then they can be replaced with a shuffle
2054 ExtractElementInst *EIO1 = dyn_cast<ExtractElementInst>(O1),
2055 *EIO2 = dyn_cast<ExtractElementInst>(O2);
2057 EIO1->getOperand(0)->getType() ==
2058 EIO2->getOperand(0)->getType())
2060 // If both are a shuffle with equal operand types and only two
2061 // unqiue operands, then they can be replaced with a single
2063 ShuffleVectorInst *SIO1 = dyn_cast<ShuffleVectorInst>(O1),
2064 *SIO2 = dyn_cast<ShuffleVectorInst>(O2);
2066 SIO1->getOperand(0)->getType() ==
2067 SIO2->getOperand(0)->getType()) {
2068 SmallSet<Value *, 4> SIOps;
2069 SIOps.insert(SIO1->getOperand(0));
2070 SIOps.insert(SIO1->getOperand(1));
2071 SIOps.insert(SIO2->getOperand(0));
2072 SIOps.insert(SIO2->getOperand(1));
2073 if (SIOps.size() <= 2)
2079 // This pair has already been formed.
2080 if (IncomingPairs.count(VP)) {
2082 } else if (IncomingPairs.count(VPR)) {
2083 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2086 if (VTy->getVectorNumElements() == 2)
2087 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
2088 TargetTransformInfo::SK_Reverse, VTy));
2089 } else if (!Ty1->isVectorTy() && !Ty2->isVectorTy()) {
2090 ESContrib = (int) TTI->getVectorInstrCost(
2091 Instruction::InsertElement, VTy, 0);
2092 ESContrib += (int) TTI->getVectorInstrCost(
2093 Instruction::InsertElement, VTy, 1);
2094 } else if (!Ty1->isVectorTy()) {
2095 // O1 needs to be inserted into a vector of size O2, and then
2096 // both need to be shuffled together.
2097 ESContrib = (int) TTI->getVectorInstrCost(
2098 Instruction::InsertElement, Ty2, 0);
2099 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2101 } else if (!Ty2->isVectorTy()) {
2102 // O2 needs to be inserted into a vector of size O1, and then
2103 // both need to be shuffled together.
2104 ESContrib = (int) TTI->getVectorInstrCost(
2105 Instruction::InsertElement, Ty1, 0);
2106 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2109 Type *TyBig = Ty1, *TySmall = Ty2;
2110 if (Ty2->getVectorNumElements() > Ty1->getVectorNumElements())
2111 std::swap(TyBig, TySmall);
2113 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2115 if (TyBig != TySmall)
2116 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2120 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {"
2121 << *O1 << " <-> " << *O2 << "} = " <<
2123 EffSize -= ESContrib;
2124 IncomingPairs.insert(VP);
2129 if (!HasNontrivialInsts) {
2130 DEBUG(if (DebugPairSelection) dbgs() <<
2131 "\tNo non-trivial instructions in DAG;"
2132 " override to zero effective size\n");
2136 for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(),
2137 E = PrunedDAG.end(); S != E; ++S)
2138 EffSize += (int) getDepthFactor(S->first);
2141 DEBUG(if (DebugPairSelection)
2142 dbgs() << "BBV: found pruned DAG for pair {"
2143 << *IJ.first << " <-> " << *IJ.second << "} of depth " <<
2144 MaxDepth << " and size " << PrunedDAG.size() <<
2145 " (effective size: " << EffSize << ")\n");
2146 if (((TTI && !UseChainDepthWithTI) ||
2147 MaxDepth >= Config.ReqChainDepth) &&
2148 EffSize > 0 && EffSize > BestEffSize) {
2149 BestMaxDepth = MaxDepth;
2150 BestEffSize = EffSize;
2151 BestDAG = PrunedDAG;
2156 // Given the list of candidate pairs, this function selects those
2157 // that will be fused into vector instructions.
2158 void BBVectorize::choosePairs(
2159 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
2160 DenseSet<ValuePair> &CandidatePairsSet,
2161 DenseMap<ValuePair, int> &CandidatePairCostSavings,
2162 std::vector<Value *> &PairableInsts,
2163 DenseSet<ValuePair> &FixedOrderPairs,
2164 DenseMap<VPPair, unsigned> &PairConnectionTypes,
2165 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
2166 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
2167 DenseSet<ValuePair> &PairableInstUsers,
2168 DenseMap<Value *, Value *>& ChosenPairs) {
2169 bool UseCycleCheck =
2170 CandidatePairsSet.size() <= Config.MaxCandPairsForCycleCheck;
2172 DenseMap<Value *, std::vector<Value *> > CandidatePairs2;
2173 for (DenseSet<ValuePair>::iterator I = CandidatePairsSet.begin(),
2174 E = CandidatePairsSet.end(); I != E; ++I) {
2175 std::vector<Value *> &JJ = CandidatePairs2[I->second];
2176 if (JJ.empty()) JJ.reserve(32);
2177 JJ.push_back(I->first);
2180 DenseMap<ValuePair, std::vector<ValuePair> > PairableInstUserMap;
2181 DenseSet<VPPair> PairableInstUserPairSet;
2182 for (std::vector<Value *>::iterator I = PairableInsts.begin(),
2183 E = PairableInsts.end(); I != E; ++I) {
2184 // The number of possible pairings for this variable:
2185 size_t NumChoices = CandidatePairs.lookup(*I).size();
2186 if (!NumChoices) continue;
2188 std::vector<Value *> &JJ = CandidatePairs[*I];
2190 // The best pair to choose and its dag:
2191 size_t BestMaxDepth = 0;
2192 int BestEffSize = 0;
2193 DenseSet<ValuePair> BestDAG;
2194 findBestDAGFor(CandidatePairs, CandidatePairsSet,
2195 CandidatePairCostSavings,
2196 PairableInsts, FixedOrderPairs, PairConnectionTypes,
2197 ConnectedPairs, ConnectedPairDeps,
2198 PairableInstUsers, PairableInstUserMap,
2199 PairableInstUserPairSet, ChosenPairs,
2200 BestDAG, BestMaxDepth, BestEffSize, *I, JJ,
2203 if (BestDAG.empty())
2206 // A dag has been chosen (or not) at this point. If no dag was
2207 // chosen, then this instruction, I, cannot be paired (and is no longer
2210 DEBUG(dbgs() << "BBV: selected pairs in the best DAG for: "
2211 << *cast<Instruction>(*I) << "\n");
2213 for (DenseSet<ValuePair>::iterator S = BestDAG.begin(),
2214 SE2 = BestDAG.end(); S != SE2; ++S) {
2215 // Insert the members of this dag into the list of chosen pairs.
2216 ChosenPairs.insert(ValuePair(S->first, S->second));
2217 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " <<
2218 *S->second << "\n");
2220 // Remove all candidate pairs that have values in the chosen dag.
2221 std::vector<Value *> &KK = CandidatePairs[S->first];
2222 for (std::vector<Value *>::iterator K = KK.begin(), KE = KK.end();
2224 if (*K == S->second)
2227 CandidatePairsSet.erase(ValuePair(S->first, *K));
2230 std::vector<Value *> &LL = CandidatePairs2[S->second];
2231 for (std::vector<Value *>::iterator L = LL.begin(), LE = LL.end();
2236 CandidatePairsSet.erase(ValuePair(*L, S->second));
2239 std::vector<Value *> &MM = CandidatePairs[S->second];
2240 for (std::vector<Value *>::iterator M = MM.begin(), ME = MM.end();
2242 assert(*M != S->first && "Flipped pair in candidate list?");
2243 CandidatePairsSet.erase(ValuePair(S->second, *M));
2246 std::vector<Value *> &NN = CandidatePairs2[S->first];
2247 for (std::vector<Value *>::iterator N = NN.begin(), NE = NN.end();
2249 assert(*N != S->second && "Flipped pair in candidate list?");
2250 CandidatePairsSet.erase(ValuePair(*N, S->first));
2255 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n");
2258 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o,
2263 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) +
2264 (n > 0 ? "." + utostr(n) : "")).str();
2267 // Returns the value that is to be used as the pointer input to the vector
2268 // instruction that fuses I with J.
2269 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context,
2270 Instruction *I, Instruction *J, unsigned o) {
2272 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
2273 int64_t OffsetInElmts;
2275 // Note: the analysis might fail here, that is why the pair order has
2276 // been precomputed (OffsetInElmts must be unused here).
2277 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
2278 IAddressSpace, JAddressSpace,
2279 OffsetInElmts, false);
2281 // The pointer value is taken to be the one with the lowest offset.
2284 Type *ArgTypeI = IPtr->getType()->getPointerElementType();
2285 Type *ArgTypeJ = JPtr->getType()->getPointerElementType();
2286 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2288 = PointerType::get(VArgType,
2289 IPtr->getType()->getPointerAddressSpace());
2290 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o),
2291 /* insert before */ I);
2294 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J,
2295 unsigned MaskOffset, unsigned NumInElem,
2296 unsigned NumInElem1, unsigned IdxOffset,
2297 std::vector<Constant*> &Mask) {
2298 unsigned NumElem1 = J->getType()->getVectorNumElements();
2299 for (unsigned v = 0; v < NumElem1; ++v) {
2300 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v);
2302 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context));
2304 unsigned mm = m + (int) IdxOffset;
2305 if (m >= (int) NumInElem1)
2306 mm += (int) NumInElem;
2308 Mask[v+MaskOffset] =
2309 ConstantInt::get(Type::getInt32Ty(Context), mm);
2314 // Returns the value that is to be used as the vector-shuffle mask to the
2315 // vector instruction that fuses I with J.
2316 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context,
2317 Instruction *I, Instruction *J) {
2318 // This is the shuffle mask. We need to append the second
2319 // mask to the first, and the numbers need to be adjusted.
2321 Type *ArgTypeI = I->getType();
2322 Type *ArgTypeJ = J->getType();
2323 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2325 unsigned NumElemI = ArgTypeI->getVectorNumElements();
2327 // Get the total number of elements in the fused vector type.
2328 // By definition, this must equal the number of elements in
2330 unsigned NumElem = VArgType->getVectorNumElements();
2331 std::vector<Constant*> Mask(NumElem);
2333 Type *OpTypeI = I->getOperand(0)->getType();
2334 unsigned NumInElemI = OpTypeI->getVectorNumElements();
2335 Type *OpTypeJ = J->getOperand(0)->getType();
2336 unsigned NumInElemJ = OpTypeJ->getVectorNumElements();
2338 // The fused vector will be:
2339 // -----------------------------------------------------
2340 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ |
2341 // -----------------------------------------------------
2342 // from which we'll extract NumElem total elements (where the first NumElemI
2343 // of them come from the mask in I and the remainder come from the mask
2346 // For the mask from the first pair...
2347 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI,
2350 // For the mask from the second pair...
2351 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ,
2354 return ConstantVector::get(Mask);
2357 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I,
2358 Instruction *J, unsigned o, Value *&LOp,
2360 Type *ArgTypeL, Type *ArgTypeH,
2361 bool IBeforeJ, unsigned IdxOff) {
2362 bool ExpandedIEChain = false;
2363 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) {
2364 // If we have a pure insertelement chain, then this can be rewritten
2365 // into a chain that directly builds the larger type.
2366 if (isPureIEChain(LIE)) {
2367 SmallVector<Value *, 8> VectElemts(numElemL,
2368 UndefValue::get(ArgTypeL->getScalarType()));
2369 InsertElementInst *LIENext = LIE;
2372 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue();
2373 VectElemts[Idx] = LIENext->getOperand(1);
2375 dyn_cast<InsertElementInst>(LIENext->getOperand(0))));
2378 Value *LIEPrev = UndefValue::get(ArgTypeH);
2379 for (unsigned i = 0; i < numElemL; ++i) {
2380 if (isa<UndefValue>(VectElemts[i])) continue;
2381 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i],
2382 ConstantInt::get(Type::getInt32Ty(Context),
2384 getReplacementName(IBeforeJ ? I : J,
2386 LIENext->insertBefore(IBeforeJ ? J : I);
2390 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH);
2391 ExpandedIEChain = true;
2395 return ExpandedIEChain;
2398 static unsigned getNumScalarElements(Type *Ty) {
2399 if (VectorType *VecTy = dyn_cast<VectorType>(Ty))
2400 return VecTy->getNumElements();
2404 // Returns the value to be used as the specified operand of the vector
2405 // instruction that fuses I with J.
2406 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I,
2407 Instruction *J, unsigned o, bool IBeforeJ) {
2408 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2409 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1);
2411 // Compute the fused vector type for this operand
2412 Type *ArgTypeI = I->getOperand(o)->getType();
2413 Type *ArgTypeJ = J->getOperand(o)->getType();
2414 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2416 Instruction *L = I, *H = J;
2417 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ;
2419 unsigned numElemL = getNumScalarElements(ArgTypeL);
2420 unsigned numElemH = getNumScalarElements(ArgTypeH);
2422 Value *LOp = L->getOperand(o);
2423 Value *HOp = H->getOperand(o);
2424 unsigned numElem = VArgType->getNumElements();
2426 // First, we check if we can reuse the "original" vector outputs (if these
2427 // exist). We might need a shuffle.
2428 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp);
2429 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp);
2430 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp);
2431 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp);
2433 // FIXME: If we're fusing shuffle instructions, then we can't apply this
2434 // optimization. The input vectors to the shuffle might be a different
2435 // length from the shuffle outputs. Unfortunately, the replacement
2436 // shuffle mask has already been formed, and the mask entries are sensitive
2437 // to the sizes of the inputs.
2438 bool IsSizeChangeShuffle =
2439 isa<ShuffleVectorInst>(L) &&
2440 (LOp->getType() != L->getType() || HOp->getType() != H->getType());
2442 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) {
2443 // We can have at most two unique vector inputs.
2444 bool CanUseInputs = true;
2447 I1 = LEE->getOperand(0);
2449 I1 = LSV->getOperand(0);
2450 I2 = LSV->getOperand(1);
2451 if (I2 == I1 || isa<UndefValue>(I2))
2456 Value *I3 = HEE->getOperand(0);
2457 if (!I2 && I3 != I1)
2459 else if (I3 != I1 && I3 != I2)
2460 CanUseInputs = false;
2462 Value *I3 = HSV->getOperand(0);
2463 if (!I2 && I3 != I1)
2465 else if (I3 != I1 && I3 != I2)
2466 CanUseInputs = false;
2469 Value *I4 = HSV->getOperand(1);
2470 if (!isa<UndefValue>(I4)) {
2471 if (!I2 && I4 != I1)
2473 else if (I4 != I1 && I4 != I2)
2474 CanUseInputs = false;
2481 cast<Instruction>(LOp)->getOperand(0)->getType()
2482 ->getVectorNumElements();
2485 cast<Instruction>(HOp)->getOperand(0)->getType()
2486 ->getVectorNumElements();
2488 // We have one or two input vectors. We need to map each index of the
2489 // operands to the index of the original vector.
2490 SmallVector<std::pair<int, int>, 8> II(numElem);
2491 for (unsigned i = 0; i < numElemL; ++i) {
2495 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue();
2496 INum = LEE->getOperand(0) == I1 ? 0 : 1;
2498 Idx = LSV->getMaskValue(i);
2499 if (Idx < (int) LOpElem) {
2500 INum = LSV->getOperand(0) == I1 ? 0 : 1;
2503 INum = LSV->getOperand(1) == I1 ? 0 : 1;
2507 II[i] = std::pair<int, int>(Idx, INum);
2509 for (unsigned i = 0; i < numElemH; ++i) {
2513 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue();
2514 INum = HEE->getOperand(0) == I1 ? 0 : 1;
2516 Idx = HSV->getMaskValue(i);
2517 if (Idx < (int) HOpElem) {
2518 INum = HSV->getOperand(0) == I1 ? 0 : 1;
2521 INum = HSV->getOperand(1) == I1 ? 0 : 1;
2525 II[i + numElemL] = std::pair<int, int>(Idx, INum);
2528 // We now have an array which tells us from which index of which
2529 // input vector each element of the operand comes.
2530 VectorType *I1T = cast<VectorType>(I1->getType());
2531 unsigned I1Elem = I1T->getNumElements();
2534 // In this case there is only one underlying vector input. Check for
2535 // the trivial case where we can use the input directly.
2536 if (I1Elem == numElem) {
2537 bool ElemInOrder = true;
2538 for (unsigned i = 0; i < numElem; ++i) {
2539 if (II[i].first != (int) i && II[i].first != -1) {
2540 ElemInOrder = false;
2549 // A shuffle is needed.
2550 std::vector<Constant *> Mask(numElem);
2551 for (unsigned i = 0; i < numElem; ++i) {
2552 int Idx = II[i].first;
2554 Mask[i] = UndefValue::get(Type::getInt32Ty(Context));
2556 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2560 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2561 ConstantVector::get(Mask),
2562 getReplacementName(IBeforeJ ? I : J,
2564 S->insertBefore(IBeforeJ ? J : I);
2568 VectorType *I2T = cast<VectorType>(I2->getType());
2569 unsigned I2Elem = I2T->getNumElements();
2571 // This input comes from two distinct vectors. The first step is to
2572 // make sure that both vectors are the same length. If not, the
2573 // smaller one will need to grow before they can be shuffled together.
2574 if (I1Elem < I2Elem) {
2575 std::vector<Constant *> Mask(I2Elem);
2577 for (; v < I1Elem; ++v)
2578 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2579 for (; v < I2Elem; ++v)
2580 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2582 Instruction *NewI1 =
2583 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2584 ConstantVector::get(Mask),
2585 getReplacementName(IBeforeJ ? I : J,
2587 NewI1->insertBefore(IBeforeJ ? J : I);
2591 } else if (I1Elem > I2Elem) {
2592 std::vector<Constant *> Mask(I1Elem);
2594 for (; v < I2Elem; ++v)
2595 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2596 for (; v < I1Elem; ++v)
2597 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2599 Instruction *NewI2 =
2600 new ShuffleVectorInst(I2, UndefValue::get(I2T),
2601 ConstantVector::get(Mask),
2602 getReplacementName(IBeforeJ ? I : J,
2604 NewI2->insertBefore(IBeforeJ ? J : I);
2610 // Now that both I1 and I2 are the same length we can shuffle them
2611 // together (and use the result).
2612 std::vector<Constant *> Mask(numElem);
2613 for (unsigned v = 0; v < numElem; ++v) {
2614 if (II[v].first == -1) {
2615 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2617 int Idx = II[v].first + II[v].second * I1Elem;
2618 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2622 Instruction *NewOp =
2623 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask),
2624 getReplacementName(IBeforeJ ? I : J, true, o));
2625 NewOp->insertBefore(IBeforeJ ? J : I);
2630 Type *ArgType = ArgTypeL;
2631 if (numElemL < numElemH) {
2632 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH,
2633 ArgTypeL, VArgType, IBeforeJ, 1)) {
2634 // This is another short-circuit case: we're combining a scalar into
2635 // a vector that is formed by an IE chain. We've just expanded the IE
2636 // chain, now insert the scalar and we're done.
2638 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0,
2639 getReplacementName(IBeforeJ ? I : J, true, o));
2640 S->insertBefore(IBeforeJ ? J : I);
2642 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL,
2643 ArgTypeH, IBeforeJ)) {
2644 // The two vector inputs to the shuffle must be the same length,
2645 // so extend the smaller vector to be the same length as the larger one.
2649 std::vector<Constant *> Mask(numElemH);
2651 for (; v < numElemL; ++v)
2652 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2653 for (; v < numElemH; ++v)
2654 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2656 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL),
2657 ConstantVector::get(Mask),
2658 getReplacementName(IBeforeJ ? I : J,
2661 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0,
2662 getReplacementName(IBeforeJ ? I : J,
2666 NLOp->insertBefore(IBeforeJ ? J : I);
2671 } else if (numElemL > numElemH) {
2672 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL,
2673 ArgTypeH, VArgType, IBeforeJ)) {
2675 InsertElementInst::Create(LOp, HOp,
2676 ConstantInt::get(Type::getInt32Ty(Context),
2678 getReplacementName(IBeforeJ ? I : J,
2680 S->insertBefore(IBeforeJ ? J : I);
2682 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH,
2683 ArgTypeL, IBeforeJ)) {
2686 std::vector<Constant *> Mask(numElemL);
2688 for (; v < numElemH; ++v)
2689 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2690 for (; v < numElemL; ++v)
2691 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2693 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH),
2694 ConstantVector::get(Mask),
2695 getReplacementName(IBeforeJ ? I : J,
2698 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0,
2699 getReplacementName(IBeforeJ ? I : J,
2703 NHOp->insertBefore(IBeforeJ ? J : I);
2708 if (ArgType->isVectorTy()) {
2709 unsigned numElem = VArgType->getVectorNumElements();
2710 std::vector<Constant*> Mask(numElem);
2711 for (unsigned v = 0; v < numElem; ++v) {
2713 // If the low vector was expanded, we need to skip the extra
2714 // undefined entries.
2715 if (v >= numElemL && numElemH > numElemL)
2716 Idx += (numElemH - numElemL);
2717 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2720 Instruction *BV = new ShuffleVectorInst(LOp, HOp,
2721 ConstantVector::get(Mask),
2722 getReplacementName(IBeforeJ ? I : J, true, o));
2723 BV->insertBefore(IBeforeJ ? J : I);
2727 Instruction *BV1 = InsertElementInst::Create(
2728 UndefValue::get(VArgType), LOp, CV0,
2729 getReplacementName(IBeforeJ ? I : J,
2731 BV1->insertBefore(IBeforeJ ? J : I);
2732 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1,
2733 getReplacementName(IBeforeJ ? I : J,
2735 BV2->insertBefore(IBeforeJ ? J : I);
2739 // This function creates an array of values that will be used as the inputs
2740 // to the vector instruction that fuses I with J.
2741 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context,
2742 Instruction *I, Instruction *J,
2743 SmallVectorImpl<Value *> &ReplacedOperands,
2745 unsigned NumOperands = I->getNumOperands();
2747 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) {
2748 // Iterate backward so that we look at the store pointer
2749 // first and know whether or not we need to flip the inputs.
2751 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) {
2752 // This is the pointer for a load/store instruction.
2753 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o);
2755 } else if (isa<CallInst>(I)) {
2756 Function *F = cast<CallInst>(I)->getCalledFunction();
2757 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID();
2758 if (o == NumOperands-1) {
2759 BasicBlock &BB = *I->getParent();
2761 Module *M = BB.getParent()->getParent();
2762 Type *ArgTypeI = I->getType();
2763 Type *ArgTypeJ = J->getType();
2764 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2766 ReplacedOperands[o] = Intrinsic::getDeclaration(M, IID, VArgType);
2768 } else if (IID == Intrinsic::powi && o == 1) {
2769 // The second argument of powi is a single integer and we've already
2770 // checked that both arguments are equal. As a result, we just keep
2771 // I's second argument.
2772 ReplacedOperands[o] = I->getOperand(o);
2775 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) {
2776 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J);
2780 ReplacedOperands[o] = getReplacementInput(Context, I, J, o, IBeforeJ);
2784 // This function creates two values that represent the outputs of the
2785 // original I and J instructions. These are generally vector shuffles
2786 // or extracts. In many cases, these will end up being unused and, thus,
2787 // eliminated by later passes.
2788 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
2789 Instruction *J, Instruction *K,
2790 Instruction *&InsertionPt,
2791 Instruction *&K1, Instruction *&K2) {
2792 if (isa<StoreInst>(I)) {
2793 AA->replaceWithNewValue(I, K);
2794 AA->replaceWithNewValue(J, K);
2796 Type *IType = I->getType();
2797 Type *JType = J->getType();
2799 VectorType *VType = getVecTypeForPair(IType, JType);
2800 unsigned numElem = VType->getNumElements();
2802 unsigned numElemI = getNumScalarElements(IType);
2803 unsigned numElemJ = getNumScalarElements(JType);
2805 if (IType->isVectorTy()) {
2806 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI);
2807 for (unsigned v = 0; v < numElemI; ++v) {
2808 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2809 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v);
2812 K1 = new ShuffleVectorInst(K, UndefValue::get(VType),
2813 ConstantVector::get( Mask1),
2814 getReplacementName(K, false, 1));
2816 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2817 K1 = ExtractElementInst::Create(K, CV0,
2818 getReplacementName(K, false, 1));
2821 if (JType->isVectorTy()) {
2822 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ);
2823 for (unsigned v = 0; v < numElemJ; ++v) {
2824 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2825 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v);
2828 K2 = new ShuffleVectorInst(K, UndefValue::get(VType),
2829 ConstantVector::get( Mask2),
2830 getReplacementName(K, false, 2));
2832 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1);
2833 K2 = ExtractElementInst::Create(K, CV1,
2834 getReplacementName(K, false, 2));
2838 K2->insertAfter(K1);
2843 // Move all uses of the function I (including pairing-induced uses) after J.
2844 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB,
2845 DenseSet<ValuePair> &LoadMoveSetPairs,
2846 Instruction *I, Instruction *J) {
2847 // Skip to the first instruction past I.
2848 BasicBlock::iterator L = std::next(BasicBlock::iterator(I));
2850 DenseSet<Value *> Users;
2851 AliasSetTracker WriteSet(*AA);
2852 if (I->mayWriteToMemory()) WriteSet.add(I);
2854 for (; cast<Instruction>(L) != J; ++L)
2855 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs);
2857 assert(cast<Instruction>(L) == J &&
2858 "Tracking has not proceeded far enough to check for dependencies");
2859 // If J is now in the use set of I, then trackUsesOfI will return true
2860 // and we have a dependency cycle (and the fusing operation must abort).
2861 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSetPairs);
2864 // Move all uses of the function I (including pairing-induced uses) after J.
2865 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB,
2866 DenseSet<ValuePair> &LoadMoveSetPairs,
2867 Instruction *&InsertionPt,
2868 Instruction *I, Instruction *J) {
2869 // Skip to the first instruction past I.
2870 BasicBlock::iterator L = std::next(BasicBlock::iterator(I));
2872 DenseSet<Value *> Users;
2873 AliasSetTracker WriteSet(*AA);
2874 if (I->mayWriteToMemory()) WriteSet.add(I);
2876 for (; cast<Instruction>(L) != J;) {
2877 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs)) {
2878 // Move this instruction
2879 Instruction *InstToMove = L; ++L;
2881 DEBUG(dbgs() << "BBV: moving: " << *InstToMove <<
2882 " to after " << *InsertionPt << "\n");
2883 InstToMove->removeFromParent();
2884 InstToMove->insertAfter(InsertionPt);
2885 InsertionPt = InstToMove;
2892 // Collect all load instruction that are in the move set of a given first
2893 // pair member. These loads depend on the first instruction, I, and so need
2894 // to be moved after J (the second instruction) when the pair is fused.
2895 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB,
2896 DenseMap<Value *, Value *> &ChosenPairs,
2897 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
2898 DenseSet<ValuePair> &LoadMoveSetPairs,
2900 // Skip to the first instruction past I.
2901 BasicBlock::iterator L = std::next(BasicBlock::iterator(I));
2903 DenseSet<Value *> Users;
2904 AliasSetTracker WriteSet(*AA);
2905 if (I->mayWriteToMemory()) WriteSet.add(I);
2907 // Note: We cannot end the loop when we reach J because J could be moved
2908 // farther down the use chain by another instruction pairing. Also, J
2909 // could be before I if this is an inverted input.
2910 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) {
2911 if (trackUsesOfI(Users, WriteSet, I, L)) {
2912 if (L->mayReadFromMemory()) {
2913 LoadMoveSet[L].push_back(I);
2914 LoadMoveSetPairs.insert(ValuePair(L, I));
2920 // In cases where both load/stores and the computation of their pointers
2921 // are chosen for vectorization, we can end up in a situation where the
2922 // aliasing analysis starts returning different query results as the
2923 // process of fusing instruction pairs continues. Because the algorithm
2924 // relies on finding the same use dags here as were found earlier, we'll
2925 // need to precompute the necessary aliasing information here and then
2926 // manually update it during the fusion process.
2927 void BBVectorize::collectLoadMoveSet(BasicBlock &BB,
2928 std::vector<Value *> &PairableInsts,
2929 DenseMap<Value *, Value *> &ChosenPairs,
2930 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
2931 DenseSet<ValuePair> &LoadMoveSetPairs) {
2932 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
2933 PIE = PairableInsts.end(); PI != PIE; ++PI) {
2934 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI);
2935 if (P == ChosenPairs.end()) continue;
2937 Instruction *I = cast<Instruction>(P->first);
2938 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet,
2939 LoadMoveSetPairs, I);
2943 // When the first instruction in each pair is cloned, it will inherit its
2944 // parent's metadata. This metadata must be combined with that of the other
2945 // instruction in a safe way.
2946 void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) {
2947 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata;
2948 K->getAllMetadataOtherThanDebugLoc(Metadata);
2949 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) {
2950 unsigned Kind = Metadata[i].first;
2951 MDNode *JMD = J->getMetadata(Kind);
2952 MDNode *KMD = Metadata[i].second;
2956 K->setMetadata(Kind, 0); // Remove unknown metadata
2958 case LLVMContext::MD_tbaa:
2959 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD));
2961 case LLVMContext::MD_fpmath:
2962 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD));
2968 // This function fuses the chosen instruction pairs into vector instructions,
2969 // taking care preserve any needed scalar outputs and, then, it reorders the
2970 // remaining instructions as needed (users of the first member of the pair
2971 // need to be moved to after the location of the second member of the pair
2972 // because the vector instruction is inserted in the location of the pair's
2974 void BBVectorize::fuseChosenPairs(BasicBlock &BB,
2975 std::vector<Value *> &PairableInsts,
2976 DenseMap<Value *, Value *> &ChosenPairs,
2977 DenseSet<ValuePair> &FixedOrderPairs,
2978 DenseMap<VPPair, unsigned> &PairConnectionTypes,
2979 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
2980 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps) {
2981 LLVMContext& Context = BB.getContext();
2983 // During the vectorization process, the order of the pairs to be fused
2984 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs
2985 // list. After a pair is fused, the flipped pair is removed from the list.
2986 DenseSet<ValuePair> FlippedPairs;
2987 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(),
2988 E = ChosenPairs.end(); P != E; ++P)
2989 FlippedPairs.insert(ValuePair(P->second, P->first));
2990 for (DenseSet<ValuePair>::iterator P = FlippedPairs.begin(),
2991 E = FlippedPairs.end(); P != E; ++P)
2992 ChosenPairs.insert(*P);
2994 DenseMap<Value *, std::vector<Value *> > LoadMoveSet;
2995 DenseSet<ValuePair> LoadMoveSetPairs;
2996 collectLoadMoveSet(BB, PairableInsts, ChosenPairs,
2997 LoadMoveSet, LoadMoveSetPairs);
2999 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n");
3001 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) {
3002 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI);
3003 if (P == ChosenPairs.end()) {
3008 if (getDepthFactor(P->first) == 0) {
3009 // These instructions are not really fused, but are tracked as though
3010 // they are. Any case in which it would be interesting to fuse them
3011 // will be taken care of by InstCombine.
3017 Instruction *I = cast<Instruction>(P->first),
3018 *J = cast<Instruction>(P->second);
3020 DEBUG(dbgs() << "BBV: fusing: " << *I <<
3021 " <-> " << *J << "\n");
3023 // Remove the pair and flipped pair from the list.
3024 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second);
3025 assert(FP != ChosenPairs.end() && "Flipped pair not found in list");
3026 ChosenPairs.erase(FP);
3027 ChosenPairs.erase(P);
3029 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSetPairs, I, J)) {
3030 DEBUG(dbgs() << "BBV: fusion of: " << *I <<
3032 " aborted because of non-trivial dependency cycle\n");
3038 // If the pair must have the other order, then flip it.
3039 bool FlipPairOrder = FixedOrderPairs.count(ValuePair(J, I));
3040 if (!FlipPairOrder && !FixedOrderPairs.count(ValuePair(I, J))) {
3041 // This pair does not have a fixed order, and so we might want to
3042 // flip it if that will yield fewer shuffles. We count the number
3043 // of dependencies connected via swaps, and those directly connected,
3044 // and flip the order if the number of swaps is greater.
3045 bool OrigOrder = true;
3046 DenseMap<ValuePair, std::vector<ValuePair> >::iterator IJ =
3047 ConnectedPairDeps.find(ValuePair(I, J));
3048 if (IJ == ConnectedPairDeps.end()) {
3049 IJ = ConnectedPairDeps.find(ValuePair(J, I));
3053 if (IJ != ConnectedPairDeps.end()) {
3054 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
3055 for (std::vector<ValuePair>::iterator T = IJ->second.begin(),
3056 TE = IJ->second.end(); T != TE; ++T) {
3057 VPPair Q(IJ->first, *T);
3058 DenseMap<VPPair, unsigned>::iterator R =
3059 PairConnectionTypes.find(VPPair(Q.second, Q.first));
3060 assert(R != PairConnectionTypes.end() &&
3061 "Cannot find pair connection type");
3062 if (R->second == PairConnectionDirect)
3064 else if (R->second == PairConnectionSwap)
3069 std::swap(NumDepsDirect, NumDepsSwap);
3071 if (NumDepsSwap > NumDepsDirect) {
3072 FlipPairOrder = true;
3073 DEBUG(dbgs() << "BBV: reordering pair: " << *I <<
3074 " <-> " << *J << "\n");
3079 Instruction *L = I, *H = J;
3083 // If the pair being fused uses the opposite order from that in the pair
3084 // connection map, then we need to flip the types.
3085 DenseMap<ValuePair, std::vector<ValuePair> >::iterator HL =
3086 ConnectedPairs.find(ValuePair(H, L));
3087 if (HL != ConnectedPairs.end())
3088 for (std::vector<ValuePair>::iterator T = HL->second.begin(),
3089 TE = HL->second.end(); T != TE; ++T) {
3090 VPPair Q(HL->first, *T);
3091 DenseMap<VPPair, unsigned>::iterator R = PairConnectionTypes.find(Q);
3092 assert(R != PairConnectionTypes.end() &&
3093 "Cannot find pair connection type");
3094 if (R->second == PairConnectionDirect)
3095 R->second = PairConnectionSwap;
3096 else if (R->second == PairConnectionSwap)
3097 R->second = PairConnectionDirect;
3100 bool LBeforeH = !FlipPairOrder;
3101 unsigned NumOperands = I->getNumOperands();
3102 SmallVector<Value *, 3> ReplacedOperands(NumOperands);
3103 getReplacementInputsForPair(Context, L, H, ReplacedOperands,
3106 // Make a copy of the original operation, change its type to the vector
3107 // type and replace its operands with the vector operands.
3108 Instruction *K = L->clone();
3111 else if (H->hasName())
3114 if (!isa<StoreInst>(K))
3115 K->mutateType(getVecTypeForPair(L->getType(), H->getType()));
3117 combineMetadata(K, H);
3118 K->intersectOptionalDataWith(H);
3120 for (unsigned o = 0; o < NumOperands; ++o)
3121 K->setOperand(o, ReplacedOperands[o]);
3125 // Instruction insertion point:
3126 Instruction *InsertionPt = K;
3127 Instruction *K1 = 0, *K2 = 0;
3128 replaceOutputsOfPair(Context, L, H, K, InsertionPt, K1, K2);
3130 // The use dag of the first original instruction must be moved to after
3131 // the location of the second instruction. The entire use dag of the
3132 // first instruction is disjoint from the input dag of the second
3133 // (by definition), and so commutes with it.
3135 moveUsesOfIAfterJ(BB, LoadMoveSetPairs, InsertionPt, I, J);
3137 if (!isa<StoreInst>(I)) {
3138 L->replaceAllUsesWith(K1);
3139 H->replaceAllUsesWith(K2);
3140 AA->replaceWithNewValue(L, K1);
3141 AA->replaceWithNewValue(H, K2);
3144 // Instructions that may read from memory may be in the load move set.
3145 // Once an instruction is fused, we no longer need its move set, and so
3146 // the values of the map never need to be updated. However, when a load
3147 // is fused, we need to merge the entries from both instructions in the
3148 // pair in case those instructions were in the move set of some other
3149 // yet-to-be-fused pair. The loads in question are the keys of the map.
3150 if (I->mayReadFromMemory()) {
3151 std::vector<ValuePair> NewSetMembers;
3152 DenseMap<Value *, std::vector<Value *> >::iterator II =
3153 LoadMoveSet.find(I);
3154 if (II != LoadMoveSet.end())
3155 for (std::vector<Value *>::iterator N = II->second.begin(),
3156 NE = II->second.end(); N != NE; ++N)
3157 NewSetMembers.push_back(ValuePair(K, *N));
3158 DenseMap<Value *, std::vector<Value *> >::iterator JJ =
3159 LoadMoveSet.find(J);
3160 if (JJ != LoadMoveSet.end())
3161 for (std::vector<Value *>::iterator N = JJ->second.begin(),
3162 NE = JJ->second.end(); N != NE; ++N)
3163 NewSetMembers.push_back(ValuePair(K, *N));
3164 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(),
3165 AE = NewSetMembers.end(); A != AE; ++A) {
3166 LoadMoveSet[A->first].push_back(A->second);
3167 LoadMoveSetPairs.insert(*A);
3171 // Before removing I, set the iterator to the next instruction.
3172 PI = std::next(BasicBlock::iterator(I));
3173 if (cast<Instruction>(PI) == J)
3178 I->eraseFromParent();
3179 J->eraseFromParent();
3181 DEBUG(if (PrintAfterEveryPair) dbgs() << "BBV: block is now: \n" <<
3185 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n");
3189 char BBVectorize::ID = 0;
3190 static const char bb_vectorize_name[] = "Basic-Block Vectorization";
3191 INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
3192 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
3193 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
3194 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
3195 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3196 INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
3198 BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) {
3199 return new BBVectorize(C);
3203 llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) {
3204 BBVectorize BBVectorizer(P, C);
3205 return BBVectorizer.vectorizeBB(BB);
3208 //===----------------------------------------------------------------------===//
3209 VectorizeConfig::VectorizeConfig() {
3210 VectorBits = ::VectorBits;
3211 VectorizeBools = !::NoBools;
3212 VectorizeInts = !::NoInts;
3213 VectorizeFloats = !::NoFloats;
3214 VectorizePointers = !::NoPointers;
3215 VectorizeCasts = !::NoCasts;
3216 VectorizeMath = !::NoMath;
3217 VectorizeFMA = !::NoFMA;
3218 VectorizeSelect = !::NoSelect;
3219 VectorizeCmp = !::NoCmp;
3220 VectorizeGEP = !::NoGEP;
3221 VectorizeMemOps = !::NoMemOps;
3222 AlignedOnly = ::AlignedOnly;
3223 ReqChainDepth= ::ReqChainDepth;
3224 SearchLimit = ::SearchLimit;
3225 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck;
3226 SplatBreaksChain = ::SplatBreaksChain;
3227 MaxInsts = ::MaxInsts;
3228 MaxPairs = ::MaxPairs;
3229 MaxIter = ::MaxIter;
3230 Pow2LenOnly = ::Pow2LenOnly;
3231 NoMemOpBoost = ::NoMemOpBoost;
3232 FastDep = ::FastDep;