1 //===- BBVectorize.cpp - A Basic-Block Vectorizer -------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements a basic-block vectorization pass. The algorithm was
11 // inspired by that used by the Vienna MAP Vectorizor by Franchetti and Kral,
12 // et al. It works by looking for chains of pairable operations and then
15 //===----------------------------------------------------------------------===//
17 #define BBV_NAME "bb-vectorize"
18 #define DEBUG_TYPE BBV_NAME
19 #include "llvm/Transforms/Vectorize.h"
20 #include "llvm/ADT/DenseMap.h"
21 #include "llvm/ADT/DenseSet.h"
22 #include "llvm/ADT/STLExtras.h"
23 #include "llvm/ADT/SmallSet.h"
24 #include "llvm/ADT/SmallVector.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/ADT/StringExtras.h"
27 #include "llvm/Analysis/AliasAnalysis.h"
28 #include "llvm/Analysis/AliasSetTracker.h"
29 #include "llvm/Analysis/ScalarEvolution.h"
30 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
31 #include "llvm/Analysis/TargetTransformInfo.h"
32 #include "llvm/Analysis/ValueTracking.h"
33 #include "llvm/IR/Constants.h"
34 #include "llvm/IR/DataLayout.h"
35 #include "llvm/IR/DerivedTypes.h"
36 #include "llvm/IR/Dominators.h"
37 #include "llvm/IR/Function.h"
38 #include "llvm/IR/Instructions.h"
39 #include "llvm/IR/IntrinsicInst.h"
40 #include "llvm/IR/Intrinsics.h"
41 #include "llvm/IR/LLVMContext.h"
42 #include "llvm/IR/Metadata.h"
43 #include "llvm/IR/Type.h"
44 #include "llvm/Pass.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/ValueHandle.h"
48 #include "llvm/Support/raw_ostream.h"
49 #include "llvm/Transforms/Utils/Local.h"
54 IgnoreTargetInfo("bb-vectorize-ignore-target-info", cl::init(false),
55 cl::Hidden, cl::desc("Ignore target information"));
57 static cl::opt<unsigned>
58 ReqChainDepth("bb-vectorize-req-chain-depth", cl::init(6), cl::Hidden,
59 cl::desc("The required chain depth for vectorization"));
62 UseChainDepthWithTI("bb-vectorize-use-chain-depth", cl::init(false),
63 cl::Hidden, cl::desc("Use the chain depth requirement with"
64 " target information"));
66 static cl::opt<unsigned>
67 SearchLimit("bb-vectorize-search-limit", cl::init(400), cl::Hidden,
68 cl::desc("The maximum search distance for instruction pairs"));
71 SplatBreaksChain("bb-vectorize-splat-breaks-chain", cl::init(false), cl::Hidden,
72 cl::desc("Replicating one element to a pair breaks the chain"));
74 static cl::opt<unsigned>
75 VectorBits("bb-vectorize-vector-bits", cl::init(128), cl::Hidden,
76 cl::desc("The size of the native vector registers"));
78 static cl::opt<unsigned>
79 MaxIter("bb-vectorize-max-iter", cl::init(0), cl::Hidden,
80 cl::desc("The maximum number of pairing iterations"));
83 Pow2LenOnly("bb-vectorize-pow2-len-only", cl::init(false), cl::Hidden,
84 cl::desc("Don't try to form non-2^n-length vectors"));
86 static cl::opt<unsigned>
87 MaxInsts("bb-vectorize-max-instr-per-group", cl::init(500), cl::Hidden,
88 cl::desc("The maximum number of pairable instructions per group"));
90 static cl::opt<unsigned>
91 MaxPairs("bb-vectorize-max-pairs-per-group", cl::init(3000), cl::Hidden,
92 cl::desc("The maximum number of candidate instruction pairs per group"));
94 static cl::opt<unsigned>
95 MaxCandPairsForCycleCheck("bb-vectorize-max-cycle-check-pairs", cl::init(200),
96 cl::Hidden, cl::desc("The maximum number of candidate pairs with which to use"
97 " a full cycle check"));
100 NoBools("bb-vectorize-no-bools", cl::init(false), cl::Hidden,
101 cl::desc("Don't try to vectorize boolean (i1) values"));
104 NoInts("bb-vectorize-no-ints", cl::init(false), cl::Hidden,
105 cl::desc("Don't try to vectorize integer values"));
108 NoFloats("bb-vectorize-no-floats", cl::init(false), cl::Hidden,
109 cl::desc("Don't try to vectorize floating-point values"));
111 // FIXME: This should default to false once pointer vector support works.
113 NoPointers("bb-vectorize-no-pointers", cl::init(/*false*/ true), cl::Hidden,
114 cl::desc("Don't try to vectorize pointer values"));
117 NoCasts("bb-vectorize-no-casts", cl::init(false), cl::Hidden,
118 cl::desc("Don't try to vectorize casting (conversion) operations"));
121 NoMath("bb-vectorize-no-math", cl::init(false), cl::Hidden,
122 cl::desc("Don't try to vectorize floating-point math intrinsics"));
125 NoFMA("bb-vectorize-no-fma", cl::init(false), cl::Hidden,
126 cl::desc("Don't try to vectorize the fused-multiply-add intrinsic"));
129 NoSelect("bb-vectorize-no-select", cl::init(false), cl::Hidden,
130 cl::desc("Don't try to vectorize select instructions"));
133 NoCmp("bb-vectorize-no-cmp", cl::init(false), cl::Hidden,
134 cl::desc("Don't try to vectorize comparison instructions"));
137 NoGEP("bb-vectorize-no-gep", cl::init(false), cl::Hidden,
138 cl::desc("Don't try to vectorize getelementptr instructions"));
141 NoMemOps("bb-vectorize-no-mem-ops", cl::init(false), cl::Hidden,
142 cl::desc("Don't try to vectorize loads and stores"));
145 AlignedOnly("bb-vectorize-aligned-only", cl::init(false), cl::Hidden,
146 cl::desc("Only generate aligned loads and stores"));
149 NoMemOpBoost("bb-vectorize-no-mem-op-boost",
150 cl::init(false), cl::Hidden,
151 cl::desc("Don't boost the chain-depth contribution of loads and stores"));
154 FastDep("bb-vectorize-fast-dep", cl::init(false), cl::Hidden,
155 cl::desc("Use a fast instruction dependency analysis"));
159 DebugInstructionExamination("bb-vectorize-debug-instruction-examination",
160 cl::init(false), cl::Hidden,
161 cl::desc("When debugging is enabled, output information on the"
162 " instruction-examination process"));
164 DebugCandidateSelection("bb-vectorize-debug-candidate-selection",
165 cl::init(false), cl::Hidden,
166 cl::desc("When debugging is enabled, output information on the"
167 " candidate-selection process"));
169 DebugPairSelection("bb-vectorize-debug-pair-selection",
170 cl::init(false), cl::Hidden,
171 cl::desc("When debugging is enabled, output information on the"
172 " pair-selection process"));
174 DebugCycleCheck("bb-vectorize-debug-cycle-check",
175 cl::init(false), cl::Hidden,
176 cl::desc("When debugging is enabled, output information on the"
177 " cycle-checking process"));
180 PrintAfterEveryPair("bb-vectorize-debug-print-after-every-pair",
181 cl::init(false), cl::Hidden,
182 cl::desc("When debugging is enabled, dump the basic block after"
183 " every pair is fused"));
186 STATISTIC(NumFusedOps, "Number of operations fused by bb-vectorize");
189 struct BBVectorize : public BasicBlockPass {
190 static char ID; // Pass identification, replacement for typeid
192 const VectorizeConfig Config;
194 BBVectorize(const VectorizeConfig &C = VectorizeConfig())
195 : BasicBlockPass(ID), Config(C) {
196 initializeBBVectorizePass(*PassRegistry::getPassRegistry());
199 BBVectorize(Pass *P, const VectorizeConfig &C)
200 : BasicBlockPass(ID), Config(C) {
201 AA = &P->getAnalysis<AliasAnalysis>();
202 DT = &P->getAnalysis<DominatorTreeWrapperPass>().getDomTree();
203 SE = &P->getAnalysis<ScalarEvolution>();
204 DataLayoutPass *DLP = P->getAnalysisIfAvailable<DataLayoutPass>();
205 DL = DLP ? &DLP->getDataLayout() : 0;
206 TTI = IgnoreTargetInfo ? 0 : &P->getAnalysis<TargetTransformInfo>();
209 typedef std::pair<Value *, Value *> ValuePair;
210 typedef std::pair<ValuePair, int> ValuePairWithCost;
211 typedef std::pair<ValuePair, size_t> ValuePairWithDepth;
212 typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair
213 typedef std::pair<VPPair, unsigned> VPPairWithType;
218 const DataLayout *DL;
219 const TargetTransformInfo *TTI;
221 // FIXME: const correct?
223 bool vectorizePairs(BasicBlock &BB, bool NonPow2Len = false);
225 bool getCandidatePairs(BasicBlock &BB,
226 BasicBlock::iterator &Start,
227 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
228 DenseSet<ValuePair> &FixedOrderPairs,
229 DenseMap<ValuePair, int> &CandidatePairCostSavings,
230 std::vector<Value *> &PairableInsts, bool NonPow2Len);
232 // FIXME: The current implementation does not account for pairs that
233 // are connected in multiple ways. For example:
234 // C1 = A1 / A2; C2 = A2 / A1 (which may be both direct and a swap)
235 enum PairConnectionType {
236 PairConnectionDirect,
241 void computeConnectedPairs(
242 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
243 DenseSet<ValuePair> &CandidatePairsSet,
244 std::vector<Value *> &PairableInsts,
245 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
246 DenseMap<VPPair, unsigned> &PairConnectionTypes);
248 void buildDepMap(BasicBlock &BB,
249 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
250 std::vector<Value *> &PairableInsts,
251 DenseSet<ValuePair> &PairableInstUsers);
253 void choosePairs(DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
254 DenseSet<ValuePair> &CandidatePairsSet,
255 DenseMap<ValuePair, int> &CandidatePairCostSavings,
256 std::vector<Value *> &PairableInsts,
257 DenseSet<ValuePair> &FixedOrderPairs,
258 DenseMap<VPPair, unsigned> &PairConnectionTypes,
259 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
260 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
261 DenseSet<ValuePair> &PairableInstUsers,
262 DenseMap<Value *, Value *>& ChosenPairs);
264 void fuseChosenPairs(BasicBlock &BB,
265 std::vector<Value *> &PairableInsts,
266 DenseMap<Value *, Value *>& ChosenPairs,
267 DenseSet<ValuePair> &FixedOrderPairs,
268 DenseMap<VPPair, unsigned> &PairConnectionTypes,
269 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
270 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps);
273 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore);
275 bool areInstsCompatible(Instruction *I, Instruction *J,
276 bool IsSimpleLoadStore, bool NonPow2Len,
277 int &CostSavings, int &FixedOrder);
279 bool trackUsesOfI(DenseSet<Value *> &Users,
280 AliasSetTracker &WriteSet, Instruction *I,
281 Instruction *J, bool UpdateUsers = true,
282 DenseSet<ValuePair> *LoadMoveSetPairs = 0);
284 void computePairsConnectedTo(
285 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
286 DenseSet<ValuePair> &CandidatePairsSet,
287 std::vector<Value *> &PairableInsts,
288 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
289 DenseMap<VPPair, unsigned> &PairConnectionTypes,
292 bool pairsConflict(ValuePair P, ValuePair Q,
293 DenseSet<ValuePair> &PairableInstUsers,
294 DenseMap<ValuePair, std::vector<ValuePair> >
295 *PairableInstUserMap = 0,
296 DenseSet<VPPair> *PairableInstUserPairSet = 0);
298 bool pairWillFormCycle(ValuePair P,
299 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUsers,
300 DenseSet<ValuePair> &CurrentPairs);
303 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
304 std::vector<Value *> &PairableInsts,
305 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
306 DenseSet<ValuePair> &PairableInstUsers,
307 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
308 DenseSet<VPPair> &PairableInstUserPairSet,
309 DenseMap<Value *, Value *> &ChosenPairs,
310 DenseMap<ValuePair, size_t> &DAG,
311 DenseSet<ValuePair> &PrunedDAG, ValuePair J,
314 void buildInitialDAGFor(
315 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
316 DenseSet<ValuePair> &CandidatePairsSet,
317 std::vector<Value *> &PairableInsts,
318 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
319 DenseSet<ValuePair> &PairableInstUsers,
320 DenseMap<Value *, Value *> &ChosenPairs,
321 DenseMap<ValuePair, size_t> &DAG, ValuePair J);
324 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
325 DenseSet<ValuePair> &CandidatePairsSet,
326 DenseMap<ValuePair, int> &CandidatePairCostSavings,
327 std::vector<Value *> &PairableInsts,
328 DenseSet<ValuePair> &FixedOrderPairs,
329 DenseMap<VPPair, unsigned> &PairConnectionTypes,
330 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
331 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
332 DenseSet<ValuePair> &PairableInstUsers,
333 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
334 DenseSet<VPPair> &PairableInstUserPairSet,
335 DenseMap<Value *, Value *> &ChosenPairs,
336 DenseSet<ValuePair> &BestDAG, size_t &BestMaxDepth,
337 int &BestEffSize, Value *II, std::vector<Value *>&JJ,
340 Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I,
341 Instruction *J, unsigned o);
343 void fillNewShuffleMask(LLVMContext& Context, Instruction *J,
344 unsigned MaskOffset, unsigned NumInElem,
345 unsigned NumInElem1, unsigned IdxOffset,
346 std::vector<Constant*> &Mask);
348 Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I,
351 bool expandIEChain(LLVMContext& Context, Instruction *I, Instruction *J,
352 unsigned o, Value *&LOp, unsigned numElemL,
353 Type *ArgTypeL, Type *ArgTypeR, bool IBeforeJ,
354 unsigned IdxOff = 0);
356 Value *getReplacementInput(LLVMContext& Context, Instruction *I,
357 Instruction *J, unsigned o, bool IBeforeJ);
359 void getReplacementInputsForPair(LLVMContext& Context, Instruction *I,
360 Instruction *J, SmallVectorImpl<Value *> &ReplacedOperands,
363 void replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
364 Instruction *J, Instruction *K,
365 Instruction *&InsertionPt, Instruction *&K1,
368 void collectPairLoadMoveSet(BasicBlock &BB,
369 DenseMap<Value *, Value *> &ChosenPairs,
370 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
371 DenseSet<ValuePair> &LoadMoveSetPairs,
374 void collectLoadMoveSet(BasicBlock &BB,
375 std::vector<Value *> &PairableInsts,
376 DenseMap<Value *, Value *> &ChosenPairs,
377 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
378 DenseSet<ValuePair> &LoadMoveSetPairs);
380 bool canMoveUsesOfIAfterJ(BasicBlock &BB,
381 DenseSet<ValuePair> &LoadMoveSetPairs,
382 Instruction *I, Instruction *J);
384 void moveUsesOfIAfterJ(BasicBlock &BB,
385 DenseSet<ValuePair> &LoadMoveSetPairs,
386 Instruction *&InsertionPt,
387 Instruction *I, Instruction *J);
389 void combineMetadata(Instruction *K, const Instruction *J);
391 bool vectorizeBB(BasicBlock &BB) {
392 if (skipOptnoneFunction(BB))
394 if (!DT->isReachableFromEntry(&BB)) {
395 DEBUG(dbgs() << "BBV: skipping unreachable " << BB.getName() <<
396 " in " << BB.getParent()->getName() << "\n");
400 DEBUG(if (TTI) dbgs() << "BBV: using target information\n");
402 bool changed = false;
403 // Iterate a sufficient number of times to merge types of size 1 bit,
404 // then 2 bits, then 4, etc. up to half of the target vector width of the
405 // target vector register.
408 (TTI || v <= Config.VectorBits) &&
409 (!Config.MaxIter || n <= Config.MaxIter);
411 DEBUG(dbgs() << "BBV: fusing loop #" << n <<
412 " for " << BB.getName() << " in " <<
413 BB.getParent()->getName() << "...\n");
414 if (vectorizePairs(BB))
420 if (changed && !Pow2LenOnly) {
422 for (; !Config.MaxIter || n <= Config.MaxIter; ++n) {
423 DEBUG(dbgs() << "BBV: fusing for non-2^n-length vectors loop #: " <<
424 n << " for " << BB.getName() << " in " <<
425 BB.getParent()->getName() << "...\n");
426 if (!vectorizePairs(BB, true)) break;
430 DEBUG(dbgs() << "BBV: done!\n");
434 virtual bool runOnBasicBlock(BasicBlock &BB) {
435 // OptimizeNone check deferred to vectorizeBB().
437 AA = &getAnalysis<AliasAnalysis>();
438 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
439 SE = &getAnalysis<ScalarEvolution>();
440 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
441 DL = DLP ? &DLP->getDataLayout() : 0;
442 TTI = IgnoreTargetInfo ? 0 : &getAnalysis<TargetTransformInfo>();
444 return vectorizeBB(BB);
447 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
448 BasicBlockPass::getAnalysisUsage(AU);
449 AU.addRequired<AliasAnalysis>();
450 AU.addRequired<DominatorTreeWrapperPass>();
451 AU.addRequired<ScalarEvolution>();
452 AU.addRequired<TargetTransformInfo>();
453 AU.addPreserved<AliasAnalysis>();
454 AU.addPreserved<DominatorTreeWrapperPass>();
455 AU.addPreserved<ScalarEvolution>();
456 AU.setPreservesCFG();
459 static inline VectorType *getVecTypeForPair(Type *ElemTy, Type *Elem2Ty) {
460 assert(ElemTy->getScalarType() == Elem2Ty->getScalarType() &&
461 "Cannot form vector from incompatible scalar types");
462 Type *STy = ElemTy->getScalarType();
465 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) {
466 numElem = VTy->getNumElements();
471 if (VectorType *VTy = dyn_cast<VectorType>(Elem2Ty)) {
472 numElem += VTy->getNumElements();
477 return VectorType::get(STy, numElem);
480 static inline void getInstructionTypes(Instruction *I,
481 Type *&T1, Type *&T2) {
482 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
483 // For stores, it is the value type, not the pointer type that matters
484 // because the value is what will come from a vector register.
486 Value *IVal = SI->getValueOperand();
487 T1 = IVal->getType();
492 if (CastInst *CI = dyn_cast<CastInst>(I))
497 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
498 T2 = SI->getCondition()->getType();
499 } else if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(I)) {
500 T2 = SI->getOperand(0)->getType();
501 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) {
502 T2 = CI->getOperand(0)->getType();
506 // Returns the weight associated with the provided value. A chain of
507 // candidate pairs has a length given by the sum of the weights of its
508 // members (one weight per pair; the weight of each member of the pair
509 // is assumed to be the same). This length is then compared to the
510 // chain-length threshold to determine if a given chain is significant
511 // enough to be vectorized. The length is also used in comparing
512 // candidate chains where longer chains are considered to be better.
513 // Note: when this function returns 0, the resulting instructions are
514 // not actually fused.
515 inline size_t getDepthFactor(Value *V) {
516 // InsertElement and ExtractElement have a depth factor of zero. This is
517 // for two reasons: First, they cannot be usefully fused. Second, because
518 // the pass generates a lot of these, they can confuse the simple metric
519 // used to compare the dags in the next iteration. Thus, giving them a
520 // weight of zero allows the pass to essentially ignore them in
521 // subsequent iterations when looking for vectorization opportunities
522 // while still tracking dependency chains that flow through those
524 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V))
527 // Give a load or store half of the required depth so that load/store
528 // pairs will vectorize.
529 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V)))
530 return Config.ReqChainDepth/2;
535 // Returns the cost of the provided instruction using TTI.
536 // This does not handle loads and stores.
537 unsigned getInstrCost(unsigned Opcode, Type *T1, Type *T2,
538 TargetTransformInfo::OperandValueKind Op1VK =
539 TargetTransformInfo::OK_AnyValue,
540 TargetTransformInfo::OperandValueKind Op2VK =
541 TargetTransformInfo::OK_AnyValue) {
544 case Instruction::GetElementPtr:
545 // We mark this instruction as zero-cost because scalar GEPs are usually
546 // lowered to the instruction addressing mode. At the moment we don't
547 // generate vector GEPs.
549 case Instruction::Br:
550 return TTI->getCFInstrCost(Opcode);
551 case Instruction::PHI:
553 case Instruction::Add:
554 case Instruction::FAdd:
555 case Instruction::Sub:
556 case Instruction::FSub:
557 case Instruction::Mul:
558 case Instruction::FMul:
559 case Instruction::UDiv:
560 case Instruction::SDiv:
561 case Instruction::FDiv:
562 case Instruction::URem:
563 case Instruction::SRem:
564 case Instruction::FRem:
565 case Instruction::Shl:
566 case Instruction::LShr:
567 case Instruction::AShr:
568 case Instruction::And:
569 case Instruction::Or:
570 case Instruction::Xor:
571 return TTI->getArithmeticInstrCost(Opcode, T1, Op1VK, Op2VK);
572 case Instruction::Select:
573 case Instruction::ICmp:
574 case Instruction::FCmp:
575 return TTI->getCmpSelInstrCost(Opcode, T1, T2);
576 case Instruction::ZExt:
577 case Instruction::SExt:
578 case Instruction::FPToUI:
579 case Instruction::FPToSI:
580 case Instruction::FPExt:
581 case Instruction::PtrToInt:
582 case Instruction::IntToPtr:
583 case Instruction::SIToFP:
584 case Instruction::UIToFP:
585 case Instruction::Trunc:
586 case Instruction::FPTrunc:
587 case Instruction::BitCast:
588 case Instruction::ShuffleVector:
589 return TTI->getCastInstrCost(Opcode, T1, T2);
595 // This determines the relative offset of two loads or stores, returning
596 // true if the offset could be determined to be some constant value.
597 // For example, if OffsetInElmts == 1, then J accesses the memory directly
598 // after I; if OffsetInElmts == -1 then I accesses the memory
600 bool getPairPtrInfo(Instruction *I, Instruction *J,
601 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment,
602 unsigned &IAddressSpace, unsigned &JAddressSpace,
603 int64_t &OffsetInElmts, bool ComputeOffset = true) {
605 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
606 LoadInst *LJ = cast<LoadInst>(J);
607 IPtr = LI->getPointerOperand();
608 JPtr = LJ->getPointerOperand();
609 IAlignment = LI->getAlignment();
610 JAlignment = LJ->getAlignment();
611 IAddressSpace = LI->getPointerAddressSpace();
612 JAddressSpace = LJ->getPointerAddressSpace();
614 StoreInst *SI = cast<StoreInst>(I), *SJ = cast<StoreInst>(J);
615 IPtr = SI->getPointerOperand();
616 JPtr = SJ->getPointerOperand();
617 IAlignment = SI->getAlignment();
618 JAlignment = SJ->getAlignment();
619 IAddressSpace = SI->getPointerAddressSpace();
620 JAddressSpace = SJ->getPointerAddressSpace();
626 const SCEV *IPtrSCEV = SE->getSCEV(IPtr);
627 const SCEV *JPtrSCEV = SE->getSCEV(JPtr);
629 // If this is a trivial offset, then we'll get something like
630 // 1*sizeof(type). With target data, which we need anyway, this will get
631 // constant folded into a number.
632 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV);
633 if (const SCEVConstant *ConstOffSCEV =
634 dyn_cast<SCEVConstant>(OffsetSCEV)) {
635 ConstantInt *IntOff = ConstOffSCEV->getValue();
636 int64_t Offset = IntOff->getSExtValue();
638 Type *VTy = IPtr->getType()->getPointerElementType();
639 int64_t VTyTSS = (int64_t) DL->getTypeStoreSize(VTy);
641 Type *VTy2 = JPtr->getType()->getPointerElementType();
642 if (VTy != VTy2 && Offset < 0) {
643 int64_t VTy2TSS = (int64_t) DL->getTypeStoreSize(VTy2);
644 OffsetInElmts = Offset/VTy2TSS;
645 return (abs64(Offset) % VTy2TSS) == 0;
648 OffsetInElmts = Offset/VTyTSS;
649 return (abs64(Offset) % VTyTSS) == 0;
655 // Returns true if the provided CallInst represents an intrinsic that can
657 bool isVectorizableIntrinsic(CallInst* I) {
658 Function *F = I->getCalledFunction();
659 if (!F) return false;
661 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID();
662 if (!IID) return false;
667 case Intrinsic::sqrt:
668 case Intrinsic::powi:
672 case Intrinsic::log2:
673 case Intrinsic::log10:
675 case Intrinsic::exp2:
677 return Config.VectorizeMath;
679 case Intrinsic::fmuladd:
680 return Config.VectorizeFMA;
684 bool isPureIEChain(InsertElementInst *IE) {
685 InsertElementInst *IENext = IE;
687 if (!isa<UndefValue>(IENext->getOperand(0)) &&
688 !isa<InsertElementInst>(IENext->getOperand(0))) {
692 dyn_cast<InsertElementInst>(IENext->getOperand(0))));
698 // This function implements one vectorization iteration on the provided
699 // basic block. It returns true if the block is changed.
700 bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) {
702 BasicBlock::iterator Start = BB.getFirstInsertionPt();
704 std::vector<Value *> AllPairableInsts;
705 DenseMap<Value *, Value *> AllChosenPairs;
706 DenseSet<ValuePair> AllFixedOrderPairs;
707 DenseMap<VPPair, unsigned> AllPairConnectionTypes;
708 DenseMap<ValuePair, std::vector<ValuePair> > AllConnectedPairs,
709 AllConnectedPairDeps;
712 std::vector<Value *> PairableInsts;
713 DenseMap<Value *, std::vector<Value *> > CandidatePairs;
714 DenseSet<ValuePair> FixedOrderPairs;
715 DenseMap<ValuePair, int> CandidatePairCostSavings;
716 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs,
718 CandidatePairCostSavings,
719 PairableInsts, NonPow2Len);
720 if (PairableInsts.empty()) continue;
722 // Build the candidate pair set for faster lookups.
723 DenseSet<ValuePair> CandidatePairsSet;
724 for (DenseMap<Value *, std::vector<Value *> >::iterator I =
725 CandidatePairs.begin(), E = CandidatePairs.end(); I != E; ++I)
726 for (std::vector<Value *>::iterator J = I->second.begin(),
727 JE = I->second.end(); J != JE; ++J)
728 CandidatePairsSet.insert(ValuePair(I->first, *J));
730 // Now we have a map of all of the pairable instructions and we need to
731 // select the best possible pairing. A good pairing is one such that the
732 // users of the pair are also paired. This defines a (directed) forest
733 // over the pairs such that two pairs are connected iff the second pair
736 // Note that it only matters that both members of the second pair use some
737 // element of the first pair (to allow for splatting).
739 DenseMap<ValuePair, std::vector<ValuePair> > ConnectedPairs,
741 DenseMap<VPPair, unsigned> PairConnectionTypes;
742 computeConnectedPairs(CandidatePairs, CandidatePairsSet,
743 PairableInsts, ConnectedPairs, PairConnectionTypes);
744 if (ConnectedPairs.empty()) continue;
746 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator
747 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
749 for (std::vector<ValuePair>::iterator J = I->second.begin(),
750 JE = I->second.end(); J != JE; ++J)
751 ConnectedPairDeps[*J].push_back(I->first);
753 // Build the pairable-instruction dependency map
754 DenseSet<ValuePair> PairableInstUsers;
755 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers);
757 // There is now a graph of the connected pairs. For each variable, pick
758 // the pairing with the largest dag meeting the depth requirement on at
759 // least one branch. Then select all pairings that are part of that dag
760 // and remove them from the list of available pairings and pairable
763 DenseMap<Value *, Value *> ChosenPairs;
764 choosePairs(CandidatePairs, CandidatePairsSet,
765 CandidatePairCostSavings,
766 PairableInsts, FixedOrderPairs, PairConnectionTypes,
767 ConnectedPairs, ConnectedPairDeps,
768 PairableInstUsers, ChosenPairs);
770 if (ChosenPairs.empty()) continue;
771 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(),
772 PairableInsts.end());
773 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end());
775 // Only for the chosen pairs, propagate information on fixed-order pairs,
776 // pair connections, and their types to the data structures used by the
777 // pair fusion procedures.
778 for (DenseMap<Value *, Value *>::iterator I = ChosenPairs.begin(),
779 IE = ChosenPairs.end(); I != IE; ++I) {
780 if (FixedOrderPairs.count(*I))
781 AllFixedOrderPairs.insert(*I);
782 else if (FixedOrderPairs.count(ValuePair(I->second, I->first)))
783 AllFixedOrderPairs.insert(ValuePair(I->second, I->first));
785 for (DenseMap<Value *, Value *>::iterator J = ChosenPairs.begin();
787 DenseMap<VPPair, unsigned>::iterator K =
788 PairConnectionTypes.find(VPPair(*I, *J));
789 if (K != PairConnectionTypes.end()) {
790 AllPairConnectionTypes.insert(*K);
792 K = PairConnectionTypes.find(VPPair(*J, *I));
793 if (K != PairConnectionTypes.end())
794 AllPairConnectionTypes.insert(*K);
799 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator
800 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
802 for (std::vector<ValuePair>::iterator J = I->second.begin(),
803 JE = I->second.end(); J != JE; ++J)
804 if (AllPairConnectionTypes.count(VPPair(I->first, *J))) {
805 AllConnectedPairs[I->first].push_back(*J);
806 AllConnectedPairDeps[*J].push_back(I->first);
808 } while (ShouldContinue);
810 if (AllChosenPairs.empty()) return false;
811 NumFusedOps += AllChosenPairs.size();
813 // A set of pairs has now been selected. It is now necessary to replace the
814 // paired instructions with vector instructions. For this procedure each
815 // operand must be replaced with a vector operand. This vector is formed
816 // by using build_vector on the old operands. The replaced values are then
817 // replaced with a vector_extract on the result. Subsequent optimization
818 // passes should coalesce the build/extract combinations.
820 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs, AllFixedOrderPairs,
821 AllPairConnectionTypes,
822 AllConnectedPairs, AllConnectedPairDeps);
824 // It is important to cleanup here so that future iterations of this
825 // function have less work to do.
826 (void) SimplifyInstructionsInBlock(&BB, DL, AA->getTargetLibraryInfo());
830 // This function returns true if the provided instruction is capable of being
831 // fused into a vector instruction. This determination is based only on the
832 // type and other attributes of the instruction.
833 bool BBVectorize::isInstVectorizable(Instruction *I,
834 bool &IsSimpleLoadStore) {
835 IsSimpleLoadStore = false;
837 if (CallInst *C = dyn_cast<CallInst>(I)) {
838 if (!isVectorizableIntrinsic(C))
840 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) {
841 // Vectorize simple loads if possbile:
842 IsSimpleLoadStore = L->isSimple();
843 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
845 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) {
846 // Vectorize simple stores if possbile:
847 IsSimpleLoadStore = S->isSimple();
848 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
850 } else if (CastInst *C = dyn_cast<CastInst>(I)) {
851 // We can vectorize casts, but not casts of pointer types, etc.
852 if (!Config.VectorizeCasts)
855 Type *SrcTy = C->getSrcTy();
856 if (!SrcTy->isSingleValueType())
859 Type *DestTy = C->getDestTy();
860 if (!DestTy->isSingleValueType())
862 } else if (isa<SelectInst>(I)) {
863 if (!Config.VectorizeSelect)
865 } else if (isa<CmpInst>(I)) {
866 if (!Config.VectorizeCmp)
868 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) {
869 if (!Config.VectorizeGEP)
872 // Currently, vector GEPs exist only with one index.
873 if (G->getNumIndices() != 1)
875 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) ||
876 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) {
880 // We can't vectorize memory operations without target data
881 if (DL == 0 && IsSimpleLoadStore)
885 getInstructionTypes(I, T1, T2);
887 // Not every type can be vectorized...
888 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) ||
889 !(VectorType::isValidElementType(T2) || T2->isVectorTy()))
892 if (T1->getScalarSizeInBits() == 1) {
893 if (!Config.VectorizeBools)
896 if (!Config.VectorizeInts && T1->isIntOrIntVectorTy())
900 if (T2->getScalarSizeInBits() == 1) {
901 if (!Config.VectorizeBools)
904 if (!Config.VectorizeInts && T2->isIntOrIntVectorTy())
908 if (!Config.VectorizeFloats
909 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy()))
912 // Don't vectorize target-specific types.
913 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy())
915 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy())
918 if ((!Config.VectorizePointers || DL == 0) &&
919 (T1->getScalarType()->isPointerTy() ||
920 T2->getScalarType()->isPointerTy()))
923 if (!TTI && (T1->getPrimitiveSizeInBits() >= Config.VectorBits ||
924 T2->getPrimitiveSizeInBits() >= Config.VectorBits))
930 // This function returns true if the two provided instructions are compatible
931 // (meaning that they can be fused into a vector instruction). This assumes
932 // that I has already been determined to be vectorizable and that J is not
933 // in the use dag of I.
934 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J,
935 bool IsSimpleLoadStore, bool NonPow2Len,
936 int &CostSavings, int &FixedOrder) {
937 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I <<
938 " <-> " << *J << "\n");
943 // Loads and stores can be merged if they have different alignments,
944 // but are otherwise the same.
945 if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment |
946 (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0)))
949 Type *IT1, *IT2, *JT1, *JT2;
950 getInstructionTypes(I, IT1, IT2);
951 getInstructionTypes(J, JT1, JT2);
952 unsigned MaxTypeBits = std::max(
953 IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(),
954 IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits());
955 if (!TTI && MaxTypeBits > Config.VectorBits)
958 // FIXME: handle addsub-type operations!
960 if (IsSimpleLoadStore) {
962 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
963 int64_t OffsetInElmts = 0;
964 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
965 IAddressSpace, JAddressSpace,
966 OffsetInElmts) && abs64(OffsetInElmts) == 1) {
967 FixedOrder = (int) OffsetInElmts;
968 unsigned BottomAlignment = IAlignment;
969 if (OffsetInElmts < 0) BottomAlignment = JAlignment;
971 Type *aTypeI = isa<StoreInst>(I) ?
972 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType();
973 Type *aTypeJ = isa<StoreInst>(J) ?
974 cast<StoreInst>(J)->getValueOperand()->getType() : J->getType();
975 Type *VType = getVecTypeForPair(aTypeI, aTypeJ);
977 if (Config.AlignedOnly) {
978 // An aligned load or store is possible only if the instruction
979 // with the lower offset has an alignment suitable for the
982 unsigned VecAlignment = DL->getPrefTypeAlignment(VType);
983 if (BottomAlignment < VecAlignment)
988 unsigned ICost = TTI->getMemoryOpCost(I->getOpcode(), aTypeI,
989 IAlignment, IAddressSpace);
990 unsigned JCost = TTI->getMemoryOpCost(J->getOpcode(), aTypeJ,
991 JAlignment, JAddressSpace);
992 unsigned VCost = TTI->getMemoryOpCost(I->getOpcode(), VType,
996 ICost += TTI->getAddressComputationCost(aTypeI);
997 JCost += TTI->getAddressComputationCost(aTypeJ);
998 VCost += TTI->getAddressComputationCost(VType);
1000 if (VCost > ICost + JCost)
1003 // We don't want to fuse to a type that will be split, even
1004 // if the two input types will also be split and there is no other
1006 unsigned VParts = TTI->getNumberOfParts(VType);
1009 else if (!VParts && VCost == ICost + JCost)
1012 CostSavings = ICost + JCost - VCost;
1018 unsigned ICost = getInstrCost(I->getOpcode(), IT1, IT2);
1019 unsigned JCost = getInstrCost(J->getOpcode(), JT1, JT2);
1020 Type *VT1 = getVecTypeForPair(IT1, JT1),
1021 *VT2 = getVecTypeForPair(IT2, JT2);
1022 TargetTransformInfo::OperandValueKind Op1VK =
1023 TargetTransformInfo::OK_AnyValue;
1024 TargetTransformInfo::OperandValueKind Op2VK =
1025 TargetTransformInfo::OK_AnyValue;
1027 // On some targets (example X86) the cost of a vector shift may vary
1028 // depending on whether the second operand is a Uniform or
1029 // NonUniform Constant.
1030 switch (I->getOpcode()) {
1032 case Instruction::Shl:
1033 case Instruction::LShr:
1034 case Instruction::AShr:
1036 // If both I and J are scalar shifts by constant, then the
1037 // merged vector shift count would be either a constant splat value
1038 // or a non-uniform vector of constants.
1039 if (ConstantInt *CII = dyn_cast<ConstantInt>(I->getOperand(1))) {
1040 if (ConstantInt *CIJ = dyn_cast<ConstantInt>(J->getOperand(1)))
1041 Op2VK = CII == CIJ ? TargetTransformInfo::OK_UniformConstantValue :
1042 TargetTransformInfo::OK_NonUniformConstantValue;
1044 // Check for a splat of a constant or for a non uniform vector
1046 Value *IOp = I->getOperand(1);
1047 Value *JOp = J->getOperand(1);
1048 if ((isa<ConstantVector>(IOp) || isa<ConstantDataVector>(IOp)) &&
1049 (isa<ConstantVector>(JOp) || isa<ConstantDataVector>(JOp))) {
1050 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
1051 Constant *SplatValue = cast<Constant>(IOp)->getSplatValue();
1052 if (SplatValue != NULL &&
1053 SplatValue == cast<Constant>(JOp)->getSplatValue())
1054 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
1059 // Note that this procedure is incorrect for insert and extract element
1060 // instructions (because combining these often results in a shuffle),
1061 // but this cost is ignored (because insert and extract element
1062 // instructions are assigned a zero depth factor and are not really
1063 // fused in general).
1064 unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2, Op1VK, Op2VK);
1066 if (VCost > ICost + JCost)
1069 // We don't want to fuse to a type that will be split, even
1070 // if the two input types will also be split and there is no other
1072 unsigned VParts1 = TTI->getNumberOfParts(VT1),
1073 VParts2 = TTI->getNumberOfParts(VT2);
1074 if (VParts1 > 1 || VParts2 > 1)
1076 else if ((!VParts1 || !VParts2) && VCost == ICost + JCost)
1079 CostSavings = ICost + JCost - VCost;
1082 // The powi intrinsic is special because only the first argument is
1083 // vectorized, the second arguments must be equal.
1084 CallInst *CI = dyn_cast<CallInst>(I);
1086 if (CI && (FI = CI->getCalledFunction())) {
1087 Intrinsic::ID IID = (Intrinsic::ID) FI->getIntrinsicID();
1088 if (IID == Intrinsic::powi) {
1089 Value *A1I = CI->getArgOperand(1),
1090 *A1J = cast<CallInst>(J)->getArgOperand(1);
1091 const SCEV *A1ISCEV = SE->getSCEV(A1I),
1092 *A1JSCEV = SE->getSCEV(A1J);
1093 return (A1ISCEV == A1JSCEV);
1097 SmallVector<Type*, 4> Tys;
1098 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
1099 Tys.push_back(CI->getArgOperand(i)->getType());
1100 unsigned ICost = TTI->getIntrinsicInstrCost(IID, IT1, Tys);
1103 CallInst *CJ = cast<CallInst>(J);
1104 for (unsigned i = 0, ie = CJ->getNumArgOperands(); i != ie; ++i)
1105 Tys.push_back(CJ->getArgOperand(i)->getType());
1106 unsigned JCost = TTI->getIntrinsicInstrCost(IID, JT1, Tys);
1109 assert(CI->getNumArgOperands() == CJ->getNumArgOperands() &&
1110 "Intrinsic argument counts differ");
1111 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
1112 if (IID == Intrinsic::powi && i == 1)
1113 Tys.push_back(CI->getArgOperand(i)->getType());
1115 Tys.push_back(getVecTypeForPair(CI->getArgOperand(i)->getType(),
1116 CJ->getArgOperand(i)->getType()));
1119 Type *RetTy = getVecTypeForPair(IT1, JT1);
1120 unsigned VCost = TTI->getIntrinsicInstrCost(IID, RetTy, Tys);
1122 if (VCost > ICost + JCost)
1125 // We don't want to fuse to a type that will be split, even
1126 // if the two input types will also be split and there is no other
1128 unsigned RetParts = TTI->getNumberOfParts(RetTy);
1131 else if (!RetParts && VCost == ICost + JCost)
1134 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
1135 if (!Tys[i]->isVectorTy())
1138 unsigned NumParts = TTI->getNumberOfParts(Tys[i]);
1141 else if (!NumParts && VCost == ICost + JCost)
1145 CostSavings = ICost + JCost - VCost;
1152 // Figure out whether or not J uses I and update the users and write-set
1153 // structures associated with I. Specifically, Users represents the set of
1154 // instructions that depend on I. WriteSet represents the set
1155 // of memory locations that are dependent on I. If UpdateUsers is true,
1156 // and J uses I, then Users is updated to contain J and WriteSet is updated
1157 // to contain any memory locations to which J writes. The function returns
1158 // true if J uses I. By default, alias analysis is used to determine
1159 // whether J reads from memory that overlaps with a location in WriteSet.
1160 // If LoadMoveSet is not null, then it is a previously-computed map
1161 // where the key is the memory-based user instruction and the value is
1162 // the instruction to be compared with I. So, if LoadMoveSet is provided,
1163 // then the alias analysis is not used. This is necessary because this
1164 // function is called during the process of moving instructions during
1165 // vectorization and the results of the alias analysis are not stable during
1167 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users,
1168 AliasSetTracker &WriteSet, Instruction *I,
1169 Instruction *J, bool UpdateUsers,
1170 DenseSet<ValuePair> *LoadMoveSetPairs) {
1173 // This instruction may already be marked as a user due, for example, to
1174 // being a member of a selected pair.
1179 for (User::op_iterator JU = J->op_begin(), JE = J->op_end();
1182 if (I == V || Users.count(V)) {
1187 if (!UsesI && J->mayReadFromMemory()) {
1188 if (LoadMoveSetPairs) {
1189 UsesI = LoadMoveSetPairs->count(ValuePair(J, I));
1191 for (AliasSetTracker::iterator W = WriteSet.begin(),
1192 WE = WriteSet.end(); W != WE; ++W) {
1193 if (W->aliasesUnknownInst(J, *AA)) {
1201 if (UsesI && UpdateUsers) {
1202 if (J->mayWriteToMemory()) WriteSet.add(J);
1209 // This function iterates over all instruction pairs in the provided
1210 // basic block and collects all candidate pairs for vectorization.
1211 bool BBVectorize::getCandidatePairs(BasicBlock &BB,
1212 BasicBlock::iterator &Start,
1213 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1214 DenseSet<ValuePair> &FixedOrderPairs,
1215 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1216 std::vector<Value *> &PairableInsts, bool NonPow2Len) {
1217 size_t TotalPairs = 0;
1218 BasicBlock::iterator E = BB.end();
1219 if (Start == E) return false;
1221 bool ShouldContinue = false, IAfterStart = false;
1222 for (BasicBlock::iterator I = Start++; I != E; ++I) {
1223 if (I == Start) IAfterStart = true;
1225 bool IsSimpleLoadStore;
1226 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue;
1228 // Look for an instruction with which to pair instruction *I...
1229 DenseSet<Value *> Users;
1230 AliasSetTracker WriteSet(*AA);
1231 if (I->mayWriteToMemory()) WriteSet.add(I);
1233 bool JAfterStart = IAfterStart;
1234 BasicBlock::iterator J = std::next(I);
1235 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) {
1236 if (J == Start) JAfterStart = true;
1238 // Determine if J uses I, if so, exit the loop.
1239 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep);
1240 if (Config.FastDep) {
1241 // Note: For this heuristic to be effective, independent operations
1242 // must tend to be intermixed. This is likely to be true from some
1243 // kinds of grouped loop unrolling (but not the generic LLVM pass),
1244 // but otherwise may require some kind of reordering pass.
1246 // When using fast dependency analysis,
1247 // stop searching after first use:
1250 if (UsesI) continue;
1253 // J does not use I, and comes before the first use of I, so it can be
1254 // merged with I if the instructions are compatible.
1255 int CostSavings, FixedOrder;
1256 if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len,
1257 CostSavings, FixedOrder)) continue;
1259 // J is a candidate for merging with I.
1260 if (!PairableInsts.size() ||
1261 PairableInsts[PairableInsts.size()-1] != I) {
1262 PairableInsts.push_back(I);
1265 CandidatePairs[I].push_back(J);
1268 CandidatePairCostSavings.insert(ValuePairWithCost(ValuePair(I, J),
1271 if (FixedOrder == 1)
1272 FixedOrderPairs.insert(ValuePair(I, J));
1273 else if (FixedOrder == -1)
1274 FixedOrderPairs.insert(ValuePair(J, I));
1276 // The next call to this function must start after the last instruction
1277 // selected during this invocation.
1279 Start = std::next(J);
1280 IAfterStart = JAfterStart = false;
1283 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair "
1284 << *I << " <-> " << *J << " (cost savings: " <<
1285 CostSavings << ")\n");
1287 // If we have already found too many pairs, break here and this function
1288 // will be called again starting after the last instruction selected
1289 // during this invocation.
1290 if (PairableInsts.size() >= Config.MaxInsts ||
1291 TotalPairs >= Config.MaxPairs) {
1292 ShouldContinue = true;
1301 DEBUG(dbgs() << "BBV: found " << PairableInsts.size()
1302 << " instructions with candidate pairs\n");
1304 return ShouldContinue;
1307 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that
1308 // it looks for pairs such that both members have an input which is an
1309 // output of PI or PJ.
1310 void BBVectorize::computePairsConnectedTo(
1311 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1312 DenseSet<ValuePair> &CandidatePairsSet,
1313 std::vector<Value *> &PairableInsts,
1314 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1315 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1319 // For each possible pairing for this variable, look at the uses of
1320 // the first value...
1321 for (Value::use_iterator I = P.first->use_begin(),
1322 E = P.first->use_end(); I != E; ++I) {
1323 if (isa<LoadInst>(*I)) {
1324 // A pair cannot be connected to a load because the load only takes one
1325 // operand (the address) and it is a scalar even after vectorization.
1327 } else if ((SI = dyn_cast<StoreInst>(*I)) &&
1328 P.first == SI->getPointerOperand()) {
1329 // Similarly, a pair cannot be connected to a store through its
1334 // For each use of the first variable, look for uses of the second
1336 for (Value::use_iterator J = P.second->use_begin(),
1337 E2 = P.second->use_end(); J != E2; ++J) {
1338 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1339 P.second == SJ->getPointerOperand())
1343 if (CandidatePairsSet.count(ValuePair(*I, *J))) {
1344 VPPair VP(P, ValuePair(*I, *J));
1345 ConnectedPairs[VP.first].push_back(VP.second);
1346 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionDirect));
1350 if (CandidatePairsSet.count(ValuePair(*J, *I))) {
1351 VPPair VP(P, ValuePair(*J, *I));
1352 ConnectedPairs[VP.first].push_back(VP.second);
1353 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSwap));
1357 if (Config.SplatBreaksChain) continue;
1358 // Look for cases where just the first value in the pair is used by
1359 // both members of another pair (splatting).
1360 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) {
1361 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1362 P.first == SJ->getPointerOperand())
1365 if (CandidatePairsSet.count(ValuePair(*I, *J))) {
1366 VPPair VP(P, ValuePair(*I, *J));
1367 ConnectedPairs[VP.first].push_back(VP.second);
1368 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1373 if (Config.SplatBreaksChain) return;
1374 // Look for cases where just the second value in the pair is used by
1375 // both members of another pair (splatting).
1376 for (Value::use_iterator I = P.second->use_begin(),
1377 E = P.second->use_end(); I != E; ++I) {
1378 if (isa<LoadInst>(*I))
1380 else if ((SI = dyn_cast<StoreInst>(*I)) &&
1381 P.second == SI->getPointerOperand())
1384 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) {
1385 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1386 P.second == SJ->getPointerOperand())
1389 if (CandidatePairsSet.count(ValuePair(*I, *J))) {
1390 VPPair VP(P, ValuePair(*I, *J));
1391 ConnectedPairs[VP.first].push_back(VP.second);
1392 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1398 // This function figures out which pairs are connected. Two pairs are
1399 // connected if some output of the first pair forms an input to both members
1400 // of the second pair.
1401 void BBVectorize::computeConnectedPairs(
1402 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1403 DenseSet<ValuePair> &CandidatePairsSet,
1404 std::vector<Value *> &PairableInsts,
1405 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1406 DenseMap<VPPair, unsigned> &PairConnectionTypes) {
1407 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
1408 PE = PairableInsts.end(); PI != PE; ++PI) {
1409 DenseMap<Value *, std::vector<Value *> >::iterator PP =
1410 CandidatePairs.find(*PI);
1411 if (PP == CandidatePairs.end())
1414 for (std::vector<Value *>::iterator P = PP->second.begin(),
1415 E = PP->second.end(); P != E; ++P)
1416 computePairsConnectedTo(CandidatePairs, CandidatePairsSet,
1417 PairableInsts, ConnectedPairs,
1418 PairConnectionTypes, ValuePair(*PI, *P));
1421 DEBUG(size_t TotalPairs = 0;
1422 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator I =
1423 ConnectedPairs.begin(), IE = ConnectedPairs.end(); I != IE; ++I)
1424 TotalPairs += I->second.size();
1425 dbgs() << "BBV: found " << TotalPairs
1426 << " pair connections.\n");
1429 // This function builds a set of use tuples such that <A, B> is in the set
1430 // if B is in the use dag of A. If B is in the use dag of A, then B
1431 // depends on the output of A.
1432 void BBVectorize::buildDepMap(
1434 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1435 std::vector<Value *> &PairableInsts,
1436 DenseSet<ValuePair> &PairableInstUsers) {
1437 DenseSet<Value *> IsInPair;
1438 for (DenseMap<Value *, std::vector<Value *> >::iterator C =
1439 CandidatePairs.begin(), E = CandidatePairs.end(); C != E; ++C) {
1440 IsInPair.insert(C->first);
1441 IsInPair.insert(C->second.begin(), C->second.end());
1444 // Iterate through the basic block, recording all users of each
1445 // pairable instruction.
1447 BasicBlock::iterator E = BB.end(), EL =
1448 BasicBlock::iterator(cast<Instruction>(PairableInsts.back()));
1449 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) {
1450 if (IsInPair.find(I) == IsInPair.end()) continue;
1452 DenseSet<Value *> Users;
1453 AliasSetTracker WriteSet(*AA);
1454 if (I->mayWriteToMemory()) WriteSet.add(I);
1456 for (BasicBlock::iterator J = std::next(I); J != E; ++J) {
1457 (void) trackUsesOfI(Users, WriteSet, I, J);
1463 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end();
1465 if (IsInPair.find(*U) == IsInPair.end()) continue;
1466 PairableInstUsers.insert(ValuePair(I, *U));
1474 // Returns true if an input to pair P is an output of pair Q and also an
1475 // input of pair Q is an output of pair P. If this is the case, then these
1476 // two pairs cannot be simultaneously fused.
1477 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q,
1478 DenseSet<ValuePair> &PairableInstUsers,
1479 DenseMap<ValuePair, std::vector<ValuePair> > *PairableInstUserMap,
1480 DenseSet<VPPair> *PairableInstUserPairSet) {
1481 // Two pairs are in conflict if they are mutual Users of eachother.
1482 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) ||
1483 PairableInstUsers.count(ValuePair(P.first, Q.second)) ||
1484 PairableInstUsers.count(ValuePair(P.second, Q.first)) ||
1485 PairableInstUsers.count(ValuePair(P.second, Q.second));
1486 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) ||
1487 PairableInstUsers.count(ValuePair(Q.first, P.second)) ||
1488 PairableInstUsers.count(ValuePair(Q.second, P.first)) ||
1489 PairableInstUsers.count(ValuePair(Q.second, P.second));
1490 if (PairableInstUserMap) {
1491 // FIXME: The expensive part of the cycle check is not so much the cycle
1492 // check itself but this edge insertion procedure. This needs some
1493 // profiling and probably a different data structure.
1495 if (PairableInstUserPairSet->insert(VPPair(Q, P)).second)
1496 (*PairableInstUserMap)[Q].push_back(P);
1499 if (PairableInstUserPairSet->insert(VPPair(P, Q)).second)
1500 (*PairableInstUserMap)[P].push_back(Q);
1504 return (QUsesP && PUsesQ);
1507 // This function walks the use graph of current pairs to see if, starting
1508 // from P, the walk returns to P.
1509 bool BBVectorize::pairWillFormCycle(ValuePair P,
1510 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
1511 DenseSet<ValuePair> &CurrentPairs) {
1512 DEBUG(if (DebugCycleCheck)
1513 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> "
1514 << *P.second << "\n");
1515 // A lookup table of visisted pairs is kept because the PairableInstUserMap
1516 // contains non-direct associations.
1517 DenseSet<ValuePair> Visited;
1518 SmallVector<ValuePair, 32> Q;
1519 // General depth-first post-order traversal:
1522 ValuePair QTop = Q.pop_back_val();
1523 Visited.insert(QTop);
1525 DEBUG(if (DebugCycleCheck)
1526 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> "
1527 << *QTop.second << "\n");
1528 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
1529 PairableInstUserMap.find(QTop);
1530 if (QQ == PairableInstUserMap.end())
1533 for (std::vector<ValuePair>::iterator C = QQ->second.begin(),
1534 CE = QQ->second.end(); C != CE; ++C) {
1537 << "BBV: rejected to prevent non-trivial cycle formation: "
1538 << QTop.first << " <-> " << C->second << "\n");
1542 if (CurrentPairs.count(*C) && !Visited.count(*C))
1545 } while (!Q.empty());
1550 // This function builds the initial dag of connected pairs with the
1551 // pair J at the root.
1552 void BBVectorize::buildInitialDAGFor(
1553 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1554 DenseSet<ValuePair> &CandidatePairsSet,
1555 std::vector<Value *> &PairableInsts,
1556 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1557 DenseSet<ValuePair> &PairableInstUsers,
1558 DenseMap<Value *, Value *> &ChosenPairs,
1559 DenseMap<ValuePair, size_t> &DAG, ValuePair J) {
1560 // Each of these pairs is viewed as the root node of a DAG. The DAG
1561 // is then walked (depth-first). As this happens, we keep track of
1562 // the pairs that compose the DAG and the maximum depth of the DAG.
1563 SmallVector<ValuePairWithDepth, 32> Q;
1564 // General depth-first post-order traversal:
1565 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1567 ValuePairWithDepth QTop = Q.back();
1569 // Push each child onto the queue:
1570 bool MoreChildren = false;
1571 size_t MaxChildDepth = QTop.second;
1572 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
1573 ConnectedPairs.find(QTop.first);
1574 if (QQ != ConnectedPairs.end())
1575 for (std::vector<ValuePair>::iterator k = QQ->second.begin(),
1576 ke = QQ->second.end(); k != ke; ++k) {
1577 // Make sure that this child pair is still a candidate:
1578 if (CandidatePairsSet.count(*k)) {
1579 DenseMap<ValuePair, size_t>::iterator C = DAG.find(*k);
1580 if (C == DAG.end()) {
1581 size_t d = getDepthFactor(k->first);
1582 Q.push_back(ValuePairWithDepth(*k, QTop.second+d));
1583 MoreChildren = true;
1585 MaxChildDepth = std::max(MaxChildDepth, C->second);
1590 if (!MoreChildren) {
1591 // Record the current pair as part of the DAG:
1592 DAG.insert(ValuePairWithDepth(QTop.first, MaxChildDepth));
1595 } while (!Q.empty());
1598 // Given some initial dag, prune it by removing conflicting pairs (pairs
1599 // that cannot be simultaneously chosen for vectorization).
1600 void BBVectorize::pruneDAGFor(
1601 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1602 std::vector<Value *> &PairableInsts,
1603 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1604 DenseSet<ValuePair> &PairableInstUsers,
1605 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
1606 DenseSet<VPPair> &PairableInstUserPairSet,
1607 DenseMap<Value *, Value *> &ChosenPairs,
1608 DenseMap<ValuePair, size_t> &DAG,
1609 DenseSet<ValuePair> &PrunedDAG, ValuePair J,
1610 bool UseCycleCheck) {
1611 SmallVector<ValuePairWithDepth, 32> Q;
1612 // General depth-first post-order traversal:
1613 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1615 ValuePairWithDepth QTop = Q.pop_back_val();
1616 PrunedDAG.insert(QTop.first);
1618 // Visit each child, pruning as necessary...
1619 SmallVector<ValuePairWithDepth, 8> BestChildren;
1620 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
1621 ConnectedPairs.find(QTop.first);
1622 if (QQ == ConnectedPairs.end())
1625 for (std::vector<ValuePair>::iterator K = QQ->second.begin(),
1626 KE = QQ->second.end(); K != KE; ++K) {
1627 DenseMap<ValuePair, size_t>::iterator C = DAG.find(*K);
1628 if (C == DAG.end()) continue;
1630 // This child is in the DAG, now we need to make sure it is the
1631 // best of any conflicting children. There could be multiple
1632 // conflicting children, so first, determine if we're keeping
1633 // this child, then delete conflicting children as necessary.
1635 // It is also necessary to guard against pairing-induced
1636 // dependencies. Consider instructions a .. x .. y .. b
1637 // such that (a,b) are to be fused and (x,y) are to be fused
1638 // but a is an input to x and b is an output from y. This
1639 // means that y cannot be moved after b but x must be moved
1640 // after b for (a,b) to be fused. In other words, after
1641 // fusing (a,b) we have y .. a/b .. x where y is an input
1642 // to a/b and x is an output to a/b: x and y can no longer
1643 // be legally fused. To prevent this condition, we must
1644 // make sure that a child pair added to the DAG is not
1645 // both an input and output of an already-selected pair.
1647 // Pairing-induced dependencies can also form from more complicated
1648 // cycles. The pair vs. pair conflicts are easy to check, and so
1649 // that is done explicitly for "fast rejection", and because for
1650 // child vs. child conflicts, we may prefer to keep the current
1651 // pair in preference to the already-selected child.
1652 DenseSet<ValuePair> CurrentPairs;
1655 for (SmallVectorImpl<ValuePairWithDepth>::iterator C2
1656 = BestChildren.begin(), E2 = BestChildren.end();
1658 if (C2->first.first == C->first.first ||
1659 C2->first.first == C->first.second ||
1660 C2->first.second == C->first.first ||
1661 C2->first.second == C->first.second ||
1662 pairsConflict(C2->first, C->first, PairableInstUsers,
1663 UseCycleCheck ? &PairableInstUserMap : 0,
1664 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1665 if (C2->second >= C->second) {
1670 CurrentPairs.insert(C2->first);
1673 if (!CanAdd) continue;
1675 // Even worse, this child could conflict with another node already
1676 // selected for the DAG. If that is the case, ignore this child.
1677 for (DenseSet<ValuePair>::iterator T = PrunedDAG.begin(),
1678 E2 = PrunedDAG.end(); T != E2; ++T) {
1679 if (T->first == C->first.first ||
1680 T->first == C->first.second ||
1681 T->second == C->first.first ||
1682 T->second == C->first.second ||
1683 pairsConflict(*T, C->first, PairableInstUsers,
1684 UseCycleCheck ? &PairableInstUserMap : 0,
1685 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1690 CurrentPairs.insert(*T);
1692 if (!CanAdd) continue;
1694 // And check the queue too...
1695 for (SmallVectorImpl<ValuePairWithDepth>::iterator C2 = Q.begin(),
1696 E2 = Q.end(); C2 != E2; ++C2) {
1697 if (C2->first.first == C->first.first ||
1698 C2->first.first == C->first.second ||
1699 C2->first.second == C->first.first ||
1700 C2->first.second == C->first.second ||
1701 pairsConflict(C2->first, C->first, PairableInstUsers,
1702 UseCycleCheck ? &PairableInstUserMap : 0,
1703 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1708 CurrentPairs.insert(C2->first);
1710 if (!CanAdd) continue;
1712 // Last but not least, check for a conflict with any of the
1713 // already-chosen pairs.
1714 for (DenseMap<Value *, Value *>::iterator C2 =
1715 ChosenPairs.begin(), E2 = ChosenPairs.end();
1717 if (pairsConflict(*C2, C->first, PairableInstUsers,
1718 UseCycleCheck ? &PairableInstUserMap : 0,
1719 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1724 CurrentPairs.insert(*C2);
1726 if (!CanAdd) continue;
1728 // To check for non-trivial cycles formed by the addition of the
1729 // current pair we've formed a list of all relevant pairs, now use a
1730 // graph walk to check for a cycle. We start from the current pair and
1731 // walk the use dag to see if we again reach the current pair. If we
1732 // do, then the current pair is rejected.
1734 // FIXME: It may be more efficient to use a topological-ordering
1735 // algorithm to improve the cycle check. This should be investigated.
1736 if (UseCycleCheck &&
1737 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs))
1740 // This child can be added, but we may have chosen it in preference
1741 // to an already-selected child. Check for this here, and if a
1742 // conflict is found, then remove the previously-selected child
1743 // before adding this one in its place.
1744 for (SmallVectorImpl<ValuePairWithDepth>::iterator C2
1745 = BestChildren.begin(); C2 != BestChildren.end();) {
1746 if (C2->first.first == C->first.first ||
1747 C2->first.first == C->first.second ||
1748 C2->first.second == C->first.first ||
1749 C2->first.second == C->first.second ||
1750 pairsConflict(C2->first, C->first, PairableInstUsers))
1751 C2 = BestChildren.erase(C2);
1756 BestChildren.push_back(ValuePairWithDepth(C->first, C->second));
1759 for (SmallVectorImpl<ValuePairWithDepth>::iterator C
1760 = BestChildren.begin(), E2 = BestChildren.end();
1762 size_t DepthF = getDepthFactor(C->first.first);
1763 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF));
1765 } while (!Q.empty());
1768 // This function finds the best dag of mututally-compatible connected
1769 // pairs, given the choice of root pairs as an iterator range.
1770 void BBVectorize::findBestDAGFor(
1771 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1772 DenseSet<ValuePair> &CandidatePairsSet,
1773 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1774 std::vector<Value *> &PairableInsts,
1775 DenseSet<ValuePair> &FixedOrderPairs,
1776 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1777 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1778 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
1779 DenseSet<ValuePair> &PairableInstUsers,
1780 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
1781 DenseSet<VPPair> &PairableInstUserPairSet,
1782 DenseMap<Value *, Value *> &ChosenPairs,
1783 DenseSet<ValuePair> &BestDAG, size_t &BestMaxDepth,
1784 int &BestEffSize, Value *II, std::vector<Value *>&JJ,
1785 bool UseCycleCheck) {
1786 for (std::vector<Value *>::iterator J = JJ.begin(), JE = JJ.end();
1788 ValuePair IJ(II, *J);
1789 if (!CandidatePairsSet.count(IJ))
1792 // Before going any further, make sure that this pair does not
1793 // conflict with any already-selected pairs (see comment below
1794 // near the DAG pruning for more details).
1795 DenseSet<ValuePair> ChosenPairSet;
1796 bool DoesConflict = false;
1797 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(),
1798 E = ChosenPairs.end(); C != E; ++C) {
1799 if (pairsConflict(*C, IJ, PairableInstUsers,
1800 UseCycleCheck ? &PairableInstUserMap : 0,
1801 UseCycleCheck ? &PairableInstUserPairSet : 0)) {
1802 DoesConflict = true;
1806 ChosenPairSet.insert(*C);
1808 if (DoesConflict) continue;
1810 if (UseCycleCheck &&
1811 pairWillFormCycle(IJ, PairableInstUserMap, ChosenPairSet))
1814 DenseMap<ValuePair, size_t> DAG;
1815 buildInitialDAGFor(CandidatePairs, CandidatePairsSet,
1816 PairableInsts, ConnectedPairs,
1817 PairableInstUsers, ChosenPairs, DAG, IJ);
1819 // Because we'll keep the child with the largest depth, the largest
1820 // depth is still the same in the unpruned DAG.
1821 size_t MaxDepth = DAG.lookup(IJ);
1823 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found DAG for pair {"
1824 << *IJ.first << " <-> " << *IJ.second << "} of depth " <<
1825 MaxDepth << " and size " << DAG.size() << "\n");
1827 // At this point the DAG has been constructed, but, may contain
1828 // contradictory children (meaning that different children of
1829 // some dag node may be attempting to fuse the same instruction).
1830 // So now we walk the dag again, in the case of a conflict,
1831 // keep only the child with the largest depth. To break a tie,
1832 // favor the first child.
1834 DenseSet<ValuePair> PrunedDAG;
1835 pruneDAGFor(CandidatePairs, PairableInsts, ConnectedPairs,
1836 PairableInstUsers, PairableInstUserMap,
1837 PairableInstUserPairSet,
1838 ChosenPairs, DAG, PrunedDAG, IJ, UseCycleCheck);
1842 DenseSet<Value *> PrunedDAGInstrs;
1843 for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(),
1844 E = PrunedDAG.end(); S != E; ++S) {
1845 PrunedDAGInstrs.insert(S->first);
1846 PrunedDAGInstrs.insert(S->second);
1849 // The set of pairs that have already contributed to the total cost.
1850 DenseSet<ValuePair> IncomingPairs;
1852 // If the cost model were perfect, this might not be necessary; but we
1853 // need to make sure that we don't get stuck vectorizing our own
1855 bool HasNontrivialInsts = false;
1857 // The node weights represent the cost savings associated with
1858 // fusing the pair of instructions.
1859 for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(),
1860 E = PrunedDAG.end(); S != E; ++S) {
1861 if (!isa<ShuffleVectorInst>(S->first) &&
1862 !isa<InsertElementInst>(S->first) &&
1863 !isa<ExtractElementInst>(S->first))
1864 HasNontrivialInsts = true;
1866 bool FlipOrder = false;
1868 if (getDepthFactor(S->first)) {
1869 int ESContrib = CandidatePairCostSavings.find(*S)->second;
1870 DEBUG(if (DebugPairSelection) dbgs() << "\tweight {"
1871 << *S->first << " <-> " << *S->second << "} = " <<
1873 EffSize += ESContrib;
1876 // The edge weights contribute in a negative sense: they represent
1877 // the cost of shuffles.
1878 DenseMap<ValuePair, std::vector<ValuePair> >::iterator SS =
1879 ConnectedPairDeps.find(*S);
1880 if (SS != ConnectedPairDeps.end()) {
1881 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
1882 for (std::vector<ValuePair>::iterator T = SS->second.begin(),
1883 TE = SS->second.end(); T != TE; ++T) {
1885 if (!PrunedDAG.count(Q.second))
1887 DenseMap<VPPair, unsigned>::iterator R =
1888 PairConnectionTypes.find(VPPair(Q.second, Q.first));
1889 assert(R != PairConnectionTypes.end() &&
1890 "Cannot find pair connection type");
1891 if (R->second == PairConnectionDirect)
1893 else if (R->second == PairConnectionSwap)
1897 // If there are more swaps than direct connections, then
1898 // the pair order will be flipped during fusion. So the real
1899 // number of swaps is the minimum number.
1900 FlipOrder = !FixedOrderPairs.count(*S) &&
1901 ((NumDepsSwap > NumDepsDirect) ||
1902 FixedOrderPairs.count(ValuePair(S->second, S->first)));
1904 for (std::vector<ValuePair>::iterator T = SS->second.begin(),
1905 TE = SS->second.end(); T != TE; ++T) {
1907 if (!PrunedDAG.count(Q.second))
1909 DenseMap<VPPair, unsigned>::iterator R =
1910 PairConnectionTypes.find(VPPair(Q.second, Q.first));
1911 assert(R != PairConnectionTypes.end() &&
1912 "Cannot find pair connection type");
1913 Type *Ty1 = Q.second.first->getType(),
1914 *Ty2 = Q.second.second->getType();
1915 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1916 if ((R->second == PairConnectionDirect && FlipOrder) ||
1917 (R->second == PairConnectionSwap && !FlipOrder) ||
1918 R->second == PairConnectionSplat) {
1919 int ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1922 if (VTy->getVectorNumElements() == 2) {
1923 if (R->second == PairConnectionSplat)
1924 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1925 TargetTransformInfo::SK_Broadcast, VTy));
1927 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1928 TargetTransformInfo::SK_Reverse, VTy));
1931 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1932 *Q.second.first << " <-> " << *Q.second.second <<
1934 *S->first << " <-> " << *S->second << "} = " <<
1936 EffSize -= ESContrib;
1941 // Compute the cost of outgoing edges. We assume that edges outgoing
1942 // to shuffles, inserts or extracts can be merged, and so contribute
1943 // no additional cost.
1944 if (!S->first->getType()->isVoidTy()) {
1945 Type *Ty1 = S->first->getType(),
1946 *Ty2 = S->second->getType();
1947 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1949 bool NeedsExtraction = false;
1950 for (Value::use_iterator I = S->first->use_begin(),
1951 IE = S->first->use_end(); I != IE; ++I) {
1952 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) {
1953 // Shuffle can be folded if it has no other input
1954 if (isa<UndefValue>(SI->getOperand(1)))
1957 if (isa<ExtractElementInst>(*I))
1959 if (PrunedDAGInstrs.count(*I))
1961 NeedsExtraction = true;
1965 if (NeedsExtraction) {
1967 if (Ty1->isVectorTy()) {
1968 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1970 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1971 TargetTransformInfo::SK_ExtractSubvector, VTy, 0, Ty1));
1973 ESContrib = (int) TTI->getVectorInstrCost(
1974 Instruction::ExtractElement, VTy, 0);
1976 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1977 *S->first << "} = " << ESContrib << "\n");
1978 EffSize -= ESContrib;
1981 NeedsExtraction = false;
1982 for (Value::use_iterator I = S->second->use_begin(),
1983 IE = S->second->use_end(); I != IE; ++I) {
1984 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) {
1985 // Shuffle can be folded if it has no other input
1986 if (isa<UndefValue>(SI->getOperand(1)))
1989 if (isa<ExtractElementInst>(*I))
1991 if (PrunedDAGInstrs.count(*I))
1993 NeedsExtraction = true;
1997 if (NeedsExtraction) {
1999 if (Ty2->isVectorTy()) {
2000 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2002 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
2003 TargetTransformInfo::SK_ExtractSubvector, VTy,
2004 Ty1->isVectorTy() ? Ty1->getVectorNumElements() : 1, Ty2));
2006 ESContrib = (int) TTI->getVectorInstrCost(
2007 Instruction::ExtractElement, VTy, 1);
2008 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
2009 *S->second << "} = " << ESContrib << "\n");
2010 EffSize -= ESContrib;
2014 // Compute the cost of incoming edges.
2015 if (!isa<LoadInst>(S->first) && !isa<StoreInst>(S->first)) {
2016 Instruction *S1 = cast<Instruction>(S->first),
2017 *S2 = cast<Instruction>(S->second);
2018 for (unsigned o = 0; o < S1->getNumOperands(); ++o) {
2019 Value *O1 = S1->getOperand(o), *O2 = S2->getOperand(o);
2021 // Combining constants into vector constants (or small vector
2022 // constants into larger ones are assumed free).
2023 if (isa<Constant>(O1) && isa<Constant>(O2))
2029 ValuePair VP = ValuePair(O1, O2);
2030 ValuePair VPR = ValuePair(O2, O1);
2032 // Internal edges are not handled here.
2033 if (PrunedDAG.count(VP) || PrunedDAG.count(VPR))
2036 Type *Ty1 = O1->getType(),
2037 *Ty2 = O2->getType();
2038 Type *VTy = getVecTypeForPair(Ty1, Ty2);
2040 // Combining vector operations of the same type is also assumed
2041 // folded with other operations.
2043 // If both are insert elements, then both can be widened.
2044 InsertElementInst *IEO1 = dyn_cast<InsertElementInst>(O1),
2045 *IEO2 = dyn_cast<InsertElementInst>(O2);
2046 if (IEO1 && IEO2 && isPureIEChain(IEO1) && isPureIEChain(IEO2))
2048 // If both are extract elements, and both have the same input
2049 // type, then they can be replaced with a shuffle
2050 ExtractElementInst *EIO1 = dyn_cast<ExtractElementInst>(O1),
2051 *EIO2 = dyn_cast<ExtractElementInst>(O2);
2053 EIO1->getOperand(0)->getType() ==
2054 EIO2->getOperand(0)->getType())
2056 // If both are a shuffle with equal operand types and only two
2057 // unqiue operands, then they can be replaced with a single
2059 ShuffleVectorInst *SIO1 = dyn_cast<ShuffleVectorInst>(O1),
2060 *SIO2 = dyn_cast<ShuffleVectorInst>(O2);
2062 SIO1->getOperand(0)->getType() ==
2063 SIO2->getOperand(0)->getType()) {
2064 SmallSet<Value *, 4> SIOps;
2065 SIOps.insert(SIO1->getOperand(0));
2066 SIOps.insert(SIO1->getOperand(1));
2067 SIOps.insert(SIO2->getOperand(0));
2068 SIOps.insert(SIO2->getOperand(1));
2069 if (SIOps.size() <= 2)
2075 // This pair has already been formed.
2076 if (IncomingPairs.count(VP)) {
2078 } else if (IncomingPairs.count(VPR)) {
2079 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2082 if (VTy->getVectorNumElements() == 2)
2083 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
2084 TargetTransformInfo::SK_Reverse, VTy));
2085 } else if (!Ty1->isVectorTy() && !Ty2->isVectorTy()) {
2086 ESContrib = (int) TTI->getVectorInstrCost(
2087 Instruction::InsertElement, VTy, 0);
2088 ESContrib += (int) TTI->getVectorInstrCost(
2089 Instruction::InsertElement, VTy, 1);
2090 } else if (!Ty1->isVectorTy()) {
2091 // O1 needs to be inserted into a vector of size O2, and then
2092 // both need to be shuffled together.
2093 ESContrib = (int) TTI->getVectorInstrCost(
2094 Instruction::InsertElement, Ty2, 0);
2095 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2097 } else if (!Ty2->isVectorTy()) {
2098 // O2 needs to be inserted into a vector of size O1, and then
2099 // both need to be shuffled together.
2100 ESContrib = (int) TTI->getVectorInstrCost(
2101 Instruction::InsertElement, Ty1, 0);
2102 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2105 Type *TyBig = Ty1, *TySmall = Ty2;
2106 if (Ty2->getVectorNumElements() > Ty1->getVectorNumElements())
2107 std::swap(TyBig, TySmall);
2109 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2111 if (TyBig != TySmall)
2112 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2116 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {"
2117 << *O1 << " <-> " << *O2 << "} = " <<
2119 EffSize -= ESContrib;
2120 IncomingPairs.insert(VP);
2125 if (!HasNontrivialInsts) {
2126 DEBUG(if (DebugPairSelection) dbgs() <<
2127 "\tNo non-trivial instructions in DAG;"
2128 " override to zero effective size\n");
2132 for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(),
2133 E = PrunedDAG.end(); S != E; ++S)
2134 EffSize += (int) getDepthFactor(S->first);
2137 DEBUG(if (DebugPairSelection)
2138 dbgs() << "BBV: found pruned DAG for pair {"
2139 << *IJ.first << " <-> " << *IJ.second << "} of depth " <<
2140 MaxDepth << " and size " << PrunedDAG.size() <<
2141 " (effective size: " << EffSize << ")\n");
2142 if (((TTI && !UseChainDepthWithTI) ||
2143 MaxDepth >= Config.ReqChainDepth) &&
2144 EffSize > 0 && EffSize > BestEffSize) {
2145 BestMaxDepth = MaxDepth;
2146 BestEffSize = EffSize;
2147 BestDAG = PrunedDAG;
2152 // Given the list of candidate pairs, this function selects those
2153 // that will be fused into vector instructions.
2154 void BBVectorize::choosePairs(
2155 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
2156 DenseSet<ValuePair> &CandidatePairsSet,
2157 DenseMap<ValuePair, int> &CandidatePairCostSavings,
2158 std::vector<Value *> &PairableInsts,
2159 DenseSet<ValuePair> &FixedOrderPairs,
2160 DenseMap<VPPair, unsigned> &PairConnectionTypes,
2161 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
2162 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
2163 DenseSet<ValuePair> &PairableInstUsers,
2164 DenseMap<Value *, Value *>& ChosenPairs) {
2165 bool UseCycleCheck =
2166 CandidatePairsSet.size() <= Config.MaxCandPairsForCycleCheck;
2168 DenseMap<Value *, std::vector<Value *> > CandidatePairs2;
2169 for (DenseSet<ValuePair>::iterator I = CandidatePairsSet.begin(),
2170 E = CandidatePairsSet.end(); I != E; ++I) {
2171 std::vector<Value *> &JJ = CandidatePairs2[I->second];
2172 if (JJ.empty()) JJ.reserve(32);
2173 JJ.push_back(I->first);
2176 DenseMap<ValuePair, std::vector<ValuePair> > PairableInstUserMap;
2177 DenseSet<VPPair> PairableInstUserPairSet;
2178 for (std::vector<Value *>::iterator I = PairableInsts.begin(),
2179 E = PairableInsts.end(); I != E; ++I) {
2180 // The number of possible pairings for this variable:
2181 size_t NumChoices = CandidatePairs.lookup(*I).size();
2182 if (!NumChoices) continue;
2184 std::vector<Value *> &JJ = CandidatePairs[*I];
2186 // The best pair to choose and its dag:
2187 size_t BestMaxDepth = 0;
2188 int BestEffSize = 0;
2189 DenseSet<ValuePair> BestDAG;
2190 findBestDAGFor(CandidatePairs, CandidatePairsSet,
2191 CandidatePairCostSavings,
2192 PairableInsts, FixedOrderPairs, PairConnectionTypes,
2193 ConnectedPairs, ConnectedPairDeps,
2194 PairableInstUsers, PairableInstUserMap,
2195 PairableInstUserPairSet, ChosenPairs,
2196 BestDAG, BestMaxDepth, BestEffSize, *I, JJ,
2199 if (BestDAG.empty())
2202 // A dag has been chosen (or not) at this point. If no dag was
2203 // chosen, then this instruction, I, cannot be paired (and is no longer
2206 DEBUG(dbgs() << "BBV: selected pairs in the best DAG for: "
2207 << *cast<Instruction>(*I) << "\n");
2209 for (DenseSet<ValuePair>::iterator S = BestDAG.begin(),
2210 SE2 = BestDAG.end(); S != SE2; ++S) {
2211 // Insert the members of this dag into the list of chosen pairs.
2212 ChosenPairs.insert(ValuePair(S->first, S->second));
2213 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " <<
2214 *S->second << "\n");
2216 // Remove all candidate pairs that have values in the chosen dag.
2217 std::vector<Value *> &KK = CandidatePairs[S->first];
2218 for (std::vector<Value *>::iterator K = KK.begin(), KE = KK.end();
2220 if (*K == S->second)
2223 CandidatePairsSet.erase(ValuePair(S->first, *K));
2226 std::vector<Value *> &LL = CandidatePairs2[S->second];
2227 for (std::vector<Value *>::iterator L = LL.begin(), LE = LL.end();
2232 CandidatePairsSet.erase(ValuePair(*L, S->second));
2235 std::vector<Value *> &MM = CandidatePairs[S->second];
2236 for (std::vector<Value *>::iterator M = MM.begin(), ME = MM.end();
2238 assert(*M != S->first && "Flipped pair in candidate list?");
2239 CandidatePairsSet.erase(ValuePair(S->second, *M));
2242 std::vector<Value *> &NN = CandidatePairs2[S->first];
2243 for (std::vector<Value *>::iterator N = NN.begin(), NE = NN.end();
2245 assert(*N != S->second && "Flipped pair in candidate list?");
2246 CandidatePairsSet.erase(ValuePair(*N, S->first));
2251 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n");
2254 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o,
2259 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) +
2260 (n > 0 ? "." + utostr(n) : "")).str();
2263 // Returns the value that is to be used as the pointer input to the vector
2264 // instruction that fuses I with J.
2265 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context,
2266 Instruction *I, Instruction *J, unsigned o) {
2268 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
2269 int64_t OffsetInElmts;
2271 // Note: the analysis might fail here, that is why the pair order has
2272 // been precomputed (OffsetInElmts must be unused here).
2273 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
2274 IAddressSpace, JAddressSpace,
2275 OffsetInElmts, false);
2277 // The pointer value is taken to be the one with the lowest offset.
2280 Type *ArgTypeI = IPtr->getType()->getPointerElementType();
2281 Type *ArgTypeJ = JPtr->getType()->getPointerElementType();
2282 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2284 = PointerType::get(VArgType,
2285 IPtr->getType()->getPointerAddressSpace());
2286 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o),
2287 /* insert before */ I);
2290 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J,
2291 unsigned MaskOffset, unsigned NumInElem,
2292 unsigned NumInElem1, unsigned IdxOffset,
2293 std::vector<Constant*> &Mask) {
2294 unsigned NumElem1 = J->getType()->getVectorNumElements();
2295 for (unsigned v = 0; v < NumElem1; ++v) {
2296 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v);
2298 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context));
2300 unsigned mm = m + (int) IdxOffset;
2301 if (m >= (int) NumInElem1)
2302 mm += (int) NumInElem;
2304 Mask[v+MaskOffset] =
2305 ConstantInt::get(Type::getInt32Ty(Context), mm);
2310 // Returns the value that is to be used as the vector-shuffle mask to the
2311 // vector instruction that fuses I with J.
2312 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context,
2313 Instruction *I, Instruction *J) {
2314 // This is the shuffle mask. We need to append the second
2315 // mask to the first, and the numbers need to be adjusted.
2317 Type *ArgTypeI = I->getType();
2318 Type *ArgTypeJ = J->getType();
2319 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2321 unsigned NumElemI = ArgTypeI->getVectorNumElements();
2323 // Get the total number of elements in the fused vector type.
2324 // By definition, this must equal the number of elements in
2326 unsigned NumElem = VArgType->getVectorNumElements();
2327 std::vector<Constant*> Mask(NumElem);
2329 Type *OpTypeI = I->getOperand(0)->getType();
2330 unsigned NumInElemI = OpTypeI->getVectorNumElements();
2331 Type *OpTypeJ = J->getOperand(0)->getType();
2332 unsigned NumInElemJ = OpTypeJ->getVectorNumElements();
2334 // The fused vector will be:
2335 // -----------------------------------------------------
2336 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ |
2337 // -----------------------------------------------------
2338 // from which we'll extract NumElem total elements (where the first NumElemI
2339 // of them come from the mask in I and the remainder come from the mask
2342 // For the mask from the first pair...
2343 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI,
2346 // For the mask from the second pair...
2347 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ,
2350 return ConstantVector::get(Mask);
2353 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I,
2354 Instruction *J, unsigned o, Value *&LOp,
2356 Type *ArgTypeL, Type *ArgTypeH,
2357 bool IBeforeJ, unsigned IdxOff) {
2358 bool ExpandedIEChain = false;
2359 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) {
2360 // If we have a pure insertelement chain, then this can be rewritten
2361 // into a chain that directly builds the larger type.
2362 if (isPureIEChain(LIE)) {
2363 SmallVector<Value *, 8> VectElemts(numElemL,
2364 UndefValue::get(ArgTypeL->getScalarType()));
2365 InsertElementInst *LIENext = LIE;
2368 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue();
2369 VectElemts[Idx] = LIENext->getOperand(1);
2371 dyn_cast<InsertElementInst>(LIENext->getOperand(0))));
2374 Value *LIEPrev = UndefValue::get(ArgTypeH);
2375 for (unsigned i = 0; i < numElemL; ++i) {
2376 if (isa<UndefValue>(VectElemts[i])) continue;
2377 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i],
2378 ConstantInt::get(Type::getInt32Ty(Context),
2380 getReplacementName(IBeforeJ ? I : J,
2382 LIENext->insertBefore(IBeforeJ ? J : I);
2386 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH);
2387 ExpandedIEChain = true;
2391 return ExpandedIEChain;
2394 static unsigned getNumScalarElements(Type *Ty) {
2395 if (VectorType *VecTy = dyn_cast<VectorType>(Ty))
2396 return VecTy->getNumElements();
2400 // Returns the value to be used as the specified operand of the vector
2401 // instruction that fuses I with J.
2402 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I,
2403 Instruction *J, unsigned o, bool IBeforeJ) {
2404 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2405 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1);
2407 // Compute the fused vector type for this operand
2408 Type *ArgTypeI = I->getOperand(o)->getType();
2409 Type *ArgTypeJ = J->getOperand(o)->getType();
2410 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2412 Instruction *L = I, *H = J;
2413 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ;
2415 unsigned numElemL = getNumScalarElements(ArgTypeL);
2416 unsigned numElemH = getNumScalarElements(ArgTypeH);
2418 Value *LOp = L->getOperand(o);
2419 Value *HOp = H->getOperand(o);
2420 unsigned numElem = VArgType->getNumElements();
2422 // First, we check if we can reuse the "original" vector outputs (if these
2423 // exist). We might need a shuffle.
2424 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp);
2425 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp);
2426 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp);
2427 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp);
2429 // FIXME: If we're fusing shuffle instructions, then we can't apply this
2430 // optimization. The input vectors to the shuffle might be a different
2431 // length from the shuffle outputs. Unfortunately, the replacement
2432 // shuffle mask has already been formed, and the mask entries are sensitive
2433 // to the sizes of the inputs.
2434 bool IsSizeChangeShuffle =
2435 isa<ShuffleVectorInst>(L) &&
2436 (LOp->getType() != L->getType() || HOp->getType() != H->getType());
2438 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) {
2439 // We can have at most two unique vector inputs.
2440 bool CanUseInputs = true;
2443 I1 = LEE->getOperand(0);
2445 I1 = LSV->getOperand(0);
2446 I2 = LSV->getOperand(1);
2447 if (I2 == I1 || isa<UndefValue>(I2))
2452 Value *I3 = HEE->getOperand(0);
2453 if (!I2 && I3 != I1)
2455 else if (I3 != I1 && I3 != I2)
2456 CanUseInputs = false;
2458 Value *I3 = HSV->getOperand(0);
2459 if (!I2 && I3 != I1)
2461 else if (I3 != I1 && I3 != I2)
2462 CanUseInputs = false;
2465 Value *I4 = HSV->getOperand(1);
2466 if (!isa<UndefValue>(I4)) {
2467 if (!I2 && I4 != I1)
2469 else if (I4 != I1 && I4 != I2)
2470 CanUseInputs = false;
2477 cast<Instruction>(LOp)->getOperand(0)->getType()
2478 ->getVectorNumElements();
2481 cast<Instruction>(HOp)->getOperand(0)->getType()
2482 ->getVectorNumElements();
2484 // We have one or two input vectors. We need to map each index of the
2485 // operands to the index of the original vector.
2486 SmallVector<std::pair<int, int>, 8> II(numElem);
2487 for (unsigned i = 0; i < numElemL; ++i) {
2491 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue();
2492 INum = LEE->getOperand(0) == I1 ? 0 : 1;
2494 Idx = LSV->getMaskValue(i);
2495 if (Idx < (int) LOpElem) {
2496 INum = LSV->getOperand(0) == I1 ? 0 : 1;
2499 INum = LSV->getOperand(1) == I1 ? 0 : 1;
2503 II[i] = std::pair<int, int>(Idx, INum);
2505 for (unsigned i = 0; i < numElemH; ++i) {
2509 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue();
2510 INum = HEE->getOperand(0) == I1 ? 0 : 1;
2512 Idx = HSV->getMaskValue(i);
2513 if (Idx < (int) HOpElem) {
2514 INum = HSV->getOperand(0) == I1 ? 0 : 1;
2517 INum = HSV->getOperand(1) == I1 ? 0 : 1;
2521 II[i + numElemL] = std::pair<int, int>(Idx, INum);
2524 // We now have an array which tells us from which index of which
2525 // input vector each element of the operand comes.
2526 VectorType *I1T = cast<VectorType>(I1->getType());
2527 unsigned I1Elem = I1T->getNumElements();
2530 // In this case there is only one underlying vector input. Check for
2531 // the trivial case where we can use the input directly.
2532 if (I1Elem == numElem) {
2533 bool ElemInOrder = true;
2534 for (unsigned i = 0; i < numElem; ++i) {
2535 if (II[i].first != (int) i && II[i].first != -1) {
2536 ElemInOrder = false;
2545 // A shuffle is needed.
2546 std::vector<Constant *> Mask(numElem);
2547 for (unsigned i = 0; i < numElem; ++i) {
2548 int Idx = II[i].first;
2550 Mask[i] = UndefValue::get(Type::getInt32Ty(Context));
2552 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2556 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2557 ConstantVector::get(Mask),
2558 getReplacementName(IBeforeJ ? I : J,
2560 S->insertBefore(IBeforeJ ? J : I);
2564 VectorType *I2T = cast<VectorType>(I2->getType());
2565 unsigned I2Elem = I2T->getNumElements();
2567 // This input comes from two distinct vectors. The first step is to
2568 // make sure that both vectors are the same length. If not, the
2569 // smaller one will need to grow before they can be shuffled together.
2570 if (I1Elem < I2Elem) {
2571 std::vector<Constant *> Mask(I2Elem);
2573 for (; v < I1Elem; ++v)
2574 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2575 for (; v < I2Elem; ++v)
2576 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2578 Instruction *NewI1 =
2579 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2580 ConstantVector::get(Mask),
2581 getReplacementName(IBeforeJ ? I : J,
2583 NewI1->insertBefore(IBeforeJ ? J : I);
2587 } else if (I1Elem > I2Elem) {
2588 std::vector<Constant *> Mask(I1Elem);
2590 for (; v < I2Elem; ++v)
2591 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2592 for (; v < I1Elem; ++v)
2593 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2595 Instruction *NewI2 =
2596 new ShuffleVectorInst(I2, UndefValue::get(I2T),
2597 ConstantVector::get(Mask),
2598 getReplacementName(IBeforeJ ? I : J,
2600 NewI2->insertBefore(IBeforeJ ? J : I);
2606 // Now that both I1 and I2 are the same length we can shuffle them
2607 // together (and use the result).
2608 std::vector<Constant *> Mask(numElem);
2609 for (unsigned v = 0; v < numElem; ++v) {
2610 if (II[v].first == -1) {
2611 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2613 int Idx = II[v].first + II[v].second * I1Elem;
2614 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2618 Instruction *NewOp =
2619 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask),
2620 getReplacementName(IBeforeJ ? I : J, true, o));
2621 NewOp->insertBefore(IBeforeJ ? J : I);
2626 Type *ArgType = ArgTypeL;
2627 if (numElemL < numElemH) {
2628 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH,
2629 ArgTypeL, VArgType, IBeforeJ, 1)) {
2630 // This is another short-circuit case: we're combining a scalar into
2631 // a vector that is formed by an IE chain. We've just expanded the IE
2632 // chain, now insert the scalar and we're done.
2634 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0,
2635 getReplacementName(IBeforeJ ? I : J, true, o));
2636 S->insertBefore(IBeforeJ ? J : I);
2638 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL,
2639 ArgTypeH, IBeforeJ)) {
2640 // The two vector inputs to the shuffle must be the same length,
2641 // so extend the smaller vector to be the same length as the larger one.
2645 std::vector<Constant *> Mask(numElemH);
2647 for (; v < numElemL; ++v)
2648 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2649 for (; v < numElemH; ++v)
2650 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2652 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL),
2653 ConstantVector::get(Mask),
2654 getReplacementName(IBeforeJ ? I : J,
2657 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0,
2658 getReplacementName(IBeforeJ ? I : J,
2662 NLOp->insertBefore(IBeforeJ ? J : I);
2667 } else if (numElemL > numElemH) {
2668 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL,
2669 ArgTypeH, VArgType, IBeforeJ)) {
2671 InsertElementInst::Create(LOp, HOp,
2672 ConstantInt::get(Type::getInt32Ty(Context),
2674 getReplacementName(IBeforeJ ? I : J,
2676 S->insertBefore(IBeforeJ ? J : I);
2678 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH,
2679 ArgTypeL, IBeforeJ)) {
2682 std::vector<Constant *> Mask(numElemL);
2684 for (; v < numElemH; ++v)
2685 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2686 for (; v < numElemL; ++v)
2687 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2689 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH),
2690 ConstantVector::get(Mask),
2691 getReplacementName(IBeforeJ ? I : J,
2694 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0,
2695 getReplacementName(IBeforeJ ? I : J,
2699 NHOp->insertBefore(IBeforeJ ? J : I);
2704 if (ArgType->isVectorTy()) {
2705 unsigned numElem = VArgType->getVectorNumElements();
2706 std::vector<Constant*> Mask(numElem);
2707 for (unsigned v = 0; v < numElem; ++v) {
2709 // If the low vector was expanded, we need to skip the extra
2710 // undefined entries.
2711 if (v >= numElemL && numElemH > numElemL)
2712 Idx += (numElemH - numElemL);
2713 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2716 Instruction *BV = new ShuffleVectorInst(LOp, HOp,
2717 ConstantVector::get(Mask),
2718 getReplacementName(IBeforeJ ? I : J, true, o));
2719 BV->insertBefore(IBeforeJ ? J : I);
2723 Instruction *BV1 = InsertElementInst::Create(
2724 UndefValue::get(VArgType), LOp, CV0,
2725 getReplacementName(IBeforeJ ? I : J,
2727 BV1->insertBefore(IBeforeJ ? J : I);
2728 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1,
2729 getReplacementName(IBeforeJ ? I : J,
2731 BV2->insertBefore(IBeforeJ ? J : I);
2735 // This function creates an array of values that will be used as the inputs
2736 // to the vector instruction that fuses I with J.
2737 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context,
2738 Instruction *I, Instruction *J,
2739 SmallVectorImpl<Value *> &ReplacedOperands,
2741 unsigned NumOperands = I->getNumOperands();
2743 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) {
2744 // Iterate backward so that we look at the store pointer
2745 // first and know whether or not we need to flip the inputs.
2747 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) {
2748 // This is the pointer for a load/store instruction.
2749 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o);
2751 } else if (isa<CallInst>(I)) {
2752 Function *F = cast<CallInst>(I)->getCalledFunction();
2753 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID();
2754 if (o == NumOperands-1) {
2755 BasicBlock &BB = *I->getParent();
2757 Module *M = BB.getParent()->getParent();
2758 Type *ArgTypeI = I->getType();
2759 Type *ArgTypeJ = J->getType();
2760 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2762 ReplacedOperands[o] = Intrinsic::getDeclaration(M, IID, VArgType);
2764 } else if (IID == Intrinsic::powi && o == 1) {
2765 // The second argument of powi is a single integer and we've already
2766 // checked that both arguments are equal. As a result, we just keep
2767 // I's second argument.
2768 ReplacedOperands[o] = I->getOperand(o);
2771 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) {
2772 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J);
2776 ReplacedOperands[o] = getReplacementInput(Context, I, J, o, IBeforeJ);
2780 // This function creates two values that represent the outputs of the
2781 // original I and J instructions. These are generally vector shuffles
2782 // or extracts. In many cases, these will end up being unused and, thus,
2783 // eliminated by later passes.
2784 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
2785 Instruction *J, Instruction *K,
2786 Instruction *&InsertionPt,
2787 Instruction *&K1, Instruction *&K2) {
2788 if (isa<StoreInst>(I)) {
2789 AA->replaceWithNewValue(I, K);
2790 AA->replaceWithNewValue(J, K);
2792 Type *IType = I->getType();
2793 Type *JType = J->getType();
2795 VectorType *VType = getVecTypeForPair(IType, JType);
2796 unsigned numElem = VType->getNumElements();
2798 unsigned numElemI = getNumScalarElements(IType);
2799 unsigned numElemJ = getNumScalarElements(JType);
2801 if (IType->isVectorTy()) {
2802 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI);
2803 for (unsigned v = 0; v < numElemI; ++v) {
2804 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2805 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v);
2808 K1 = new ShuffleVectorInst(K, UndefValue::get(VType),
2809 ConstantVector::get( Mask1),
2810 getReplacementName(K, false, 1));
2812 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2813 K1 = ExtractElementInst::Create(K, CV0,
2814 getReplacementName(K, false, 1));
2817 if (JType->isVectorTy()) {
2818 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ);
2819 for (unsigned v = 0; v < numElemJ; ++v) {
2820 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2821 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v);
2824 K2 = new ShuffleVectorInst(K, UndefValue::get(VType),
2825 ConstantVector::get( Mask2),
2826 getReplacementName(K, false, 2));
2828 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1);
2829 K2 = ExtractElementInst::Create(K, CV1,
2830 getReplacementName(K, false, 2));
2834 K2->insertAfter(K1);
2839 // Move all uses of the function I (including pairing-induced uses) after J.
2840 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB,
2841 DenseSet<ValuePair> &LoadMoveSetPairs,
2842 Instruction *I, Instruction *J) {
2843 // Skip to the first instruction past I.
2844 BasicBlock::iterator L = std::next(BasicBlock::iterator(I));
2846 DenseSet<Value *> Users;
2847 AliasSetTracker WriteSet(*AA);
2848 if (I->mayWriteToMemory()) WriteSet.add(I);
2850 for (; cast<Instruction>(L) != J; ++L)
2851 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs);
2853 assert(cast<Instruction>(L) == J &&
2854 "Tracking has not proceeded far enough to check for dependencies");
2855 // If J is now in the use set of I, then trackUsesOfI will return true
2856 // and we have a dependency cycle (and the fusing operation must abort).
2857 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSetPairs);
2860 // Move all uses of the function I (including pairing-induced uses) after J.
2861 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB,
2862 DenseSet<ValuePair> &LoadMoveSetPairs,
2863 Instruction *&InsertionPt,
2864 Instruction *I, Instruction *J) {
2865 // Skip to the first instruction past I.
2866 BasicBlock::iterator L = std::next(BasicBlock::iterator(I));
2868 DenseSet<Value *> Users;
2869 AliasSetTracker WriteSet(*AA);
2870 if (I->mayWriteToMemory()) WriteSet.add(I);
2872 for (; cast<Instruction>(L) != J;) {
2873 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs)) {
2874 // Move this instruction
2875 Instruction *InstToMove = L; ++L;
2877 DEBUG(dbgs() << "BBV: moving: " << *InstToMove <<
2878 " to after " << *InsertionPt << "\n");
2879 InstToMove->removeFromParent();
2880 InstToMove->insertAfter(InsertionPt);
2881 InsertionPt = InstToMove;
2888 // Collect all load instruction that are in the move set of a given first
2889 // pair member. These loads depend on the first instruction, I, and so need
2890 // to be moved after J (the second instruction) when the pair is fused.
2891 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB,
2892 DenseMap<Value *, Value *> &ChosenPairs,
2893 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
2894 DenseSet<ValuePair> &LoadMoveSetPairs,
2896 // Skip to the first instruction past I.
2897 BasicBlock::iterator L = std::next(BasicBlock::iterator(I));
2899 DenseSet<Value *> Users;
2900 AliasSetTracker WriteSet(*AA);
2901 if (I->mayWriteToMemory()) WriteSet.add(I);
2903 // Note: We cannot end the loop when we reach J because J could be moved
2904 // farther down the use chain by another instruction pairing. Also, J
2905 // could be before I if this is an inverted input.
2906 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) {
2907 if (trackUsesOfI(Users, WriteSet, I, L)) {
2908 if (L->mayReadFromMemory()) {
2909 LoadMoveSet[L].push_back(I);
2910 LoadMoveSetPairs.insert(ValuePair(L, I));
2916 // In cases where both load/stores and the computation of their pointers
2917 // are chosen for vectorization, we can end up in a situation where the
2918 // aliasing analysis starts returning different query results as the
2919 // process of fusing instruction pairs continues. Because the algorithm
2920 // relies on finding the same use dags here as were found earlier, we'll
2921 // need to precompute the necessary aliasing information here and then
2922 // manually update it during the fusion process.
2923 void BBVectorize::collectLoadMoveSet(BasicBlock &BB,
2924 std::vector<Value *> &PairableInsts,
2925 DenseMap<Value *, Value *> &ChosenPairs,
2926 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
2927 DenseSet<ValuePair> &LoadMoveSetPairs) {
2928 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
2929 PIE = PairableInsts.end(); PI != PIE; ++PI) {
2930 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI);
2931 if (P == ChosenPairs.end()) continue;
2933 Instruction *I = cast<Instruction>(P->first);
2934 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet,
2935 LoadMoveSetPairs, I);
2939 // When the first instruction in each pair is cloned, it will inherit its
2940 // parent's metadata. This metadata must be combined with that of the other
2941 // instruction in a safe way.
2942 void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) {
2943 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata;
2944 K->getAllMetadataOtherThanDebugLoc(Metadata);
2945 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) {
2946 unsigned Kind = Metadata[i].first;
2947 MDNode *JMD = J->getMetadata(Kind);
2948 MDNode *KMD = Metadata[i].second;
2952 K->setMetadata(Kind, 0); // Remove unknown metadata
2954 case LLVMContext::MD_tbaa:
2955 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD));
2957 case LLVMContext::MD_fpmath:
2958 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD));
2964 // This function fuses the chosen instruction pairs into vector instructions,
2965 // taking care preserve any needed scalar outputs and, then, it reorders the
2966 // remaining instructions as needed (users of the first member of the pair
2967 // need to be moved to after the location of the second member of the pair
2968 // because the vector instruction is inserted in the location of the pair's
2970 void BBVectorize::fuseChosenPairs(BasicBlock &BB,
2971 std::vector<Value *> &PairableInsts,
2972 DenseMap<Value *, Value *> &ChosenPairs,
2973 DenseSet<ValuePair> &FixedOrderPairs,
2974 DenseMap<VPPair, unsigned> &PairConnectionTypes,
2975 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
2976 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps) {
2977 LLVMContext& Context = BB.getContext();
2979 // During the vectorization process, the order of the pairs to be fused
2980 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs
2981 // list. After a pair is fused, the flipped pair is removed from the list.
2982 DenseSet<ValuePair> FlippedPairs;
2983 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(),
2984 E = ChosenPairs.end(); P != E; ++P)
2985 FlippedPairs.insert(ValuePair(P->second, P->first));
2986 for (DenseSet<ValuePair>::iterator P = FlippedPairs.begin(),
2987 E = FlippedPairs.end(); P != E; ++P)
2988 ChosenPairs.insert(*P);
2990 DenseMap<Value *, std::vector<Value *> > LoadMoveSet;
2991 DenseSet<ValuePair> LoadMoveSetPairs;
2992 collectLoadMoveSet(BB, PairableInsts, ChosenPairs,
2993 LoadMoveSet, LoadMoveSetPairs);
2995 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n");
2997 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) {
2998 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI);
2999 if (P == ChosenPairs.end()) {
3004 if (getDepthFactor(P->first) == 0) {
3005 // These instructions are not really fused, but are tracked as though
3006 // they are. Any case in which it would be interesting to fuse them
3007 // will be taken care of by InstCombine.
3013 Instruction *I = cast<Instruction>(P->first),
3014 *J = cast<Instruction>(P->second);
3016 DEBUG(dbgs() << "BBV: fusing: " << *I <<
3017 " <-> " << *J << "\n");
3019 // Remove the pair and flipped pair from the list.
3020 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second);
3021 assert(FP != ChosenPairs.end() && "Flipped pair not found in list");
3022 ChosenPairs.erase(FP);
3023 ChosenPairs.erase(P);
3025 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSetPairs, I, J)) {
3026 DEBUG(dbgs() << "BBV: fusion of: " << *I <<
3028 " aborted because of non-trivial dependency cycle\n");
3034 // If the pair must have the other order, then flip it.
3035 bool FlipPairOrder = FixedOrderPairs.count(ValuePair(J, I));
3036 if (!FlipPairOrder && !FixedOrderPairs.count(ValuePair(I, J))) {
3037 // This pair does not have a fixed order, and so we might want to
3038 // flip it if that will yield fewer shuffles. We count the number
3039 // of dependencies connected via swaps, and those directly connected,
3040 // and flip the order if the number of swaps is greater.
3041 bool OrigOrder = true;
3042 DenseMap<ValuePair, std::vector<ValuePair> >::iterator IJ =
3043 ConnectedPairDeps.find(ValuePair(I, J));
3044 if (IJ == ConnectedPairDeps.end()) {
3045 IJ = ConnectedPairDeps.find(ValuePair(J, I));
3049 if (IJ != ConnectedPairDeps.end()) {
3050 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
3051 for (std::vector<ValuePair>::iterator T = IJ->second.begin(),
3052 TE = IJ->second.end(); T != TE; ++T) {
3053 VPPair Q(IJ->first, *T);
3054 DenseMap<VPPair, unsigned>::iterator R =
3055 PairConnectionTypes.find(VPPair(Q.second, Q.first));
3056 assert(R != PairConnectionTypes.end() &&
3057 "Cannot find pair connection type");
3058 if (R->second == PairConnectionDirect)
3060 else if (R->second == PairConnectionSwap)
3065 std::swap(NumDepsDirect, NumDepsSwap);
3067 if (NumDepsSwap > NumDepsDirect) {
3068 FlipPairOrder = true;
3069 DEBUG(dbgs() << "BBV: reordering pair: " << *I <<
3070 " <-> " << *J << "\n");
3075 Instruction *L = I, *H = J;
3079 // If the pair being fused uses the opposite order from that in the pair
3080 // connection map, then we need to flip the types.
3081 DenseMap<ValuePair, std::vector<ValuePair> >::iterator HL =
3082 ConnectedPairs.find(ValuePair(H, L));
3083 if (HL != ConnectedPairs.end())
3084 for (std::vector<ValuePair>::iterator T = HL->second.begin(),
3085 TE = HL->second.end(); T != TE; ++T) {
3086 VPPair Q(HL->first, *T);
3087 DenseMap<VPPair, unsigned>::iterator R = PairConnectionTypes.find(Q);
3088 assert(R != PairConnectionTypes.end() &&
3089 "Cannot find pair connection type");
3090 if (R->second == PairConnectionDirect)
3091 R->second = PairConnectionSwap;
3092 else if (R->second == PairConnectionSwap)
3093 R->second = PairConnectionDirect;
3096 bool LBeforeH = !FlipPairOrder;
3097 unsigned NumOperands = I->getNumOperands();
3098 SmallVector<Value *, 3> ReplacedOperands(NumOperands);
3099 getReplacementInputsForPair(Context, L, H, ReplacedOperands,
3102 // Make a copy of the original operation, change its type to the vector
3103 // type and replace its operands with the vector operands.
3104 Instruction *K = L->clone();
3107 else if (H->hasName())
3110 if (!isa<StoreInst>(K))
3111 K->mutateType(getVecTypeForPair(L->getType(), H->getType()));
3113 combineMetadata(K, H);
3114 K->intersectOptionalDataWith(H);
3116 for (unsigned o = 0; o < NumOperands; ++o)
3117 K->setOperand(o, ReplacedOperands[o]);
3121 // Instruction insertion point:
3122 Instruction *InsertionPt = K;
3123 Instruction *K1 = 0, *K2 = 0;
3124 replaceOutputsOfPair(Context, L, H, K, InsertionPt, K1, K2);
3126 // The use dag of the first original instruction must be moved to after
3127 // the location of the second instruction. The entire use dag of the
3128 // first instruction is disjoint from the input dag of the second
3129 // (by definition), and so commutes with it.
3131 moveUsesOfIAfterJ(BB, LoadMoveSetPairs, InsertionPt, I, J);
3133 if (!isa<StoreInst>(I)) {
3134 L->replaceAllUsesWith(K1);
3135 H->replaceAllUsesWith(K2);
3136 AA->replaceWithNewValue(L, K1);
3137 AA->replaceWithNewValue(H, K2);
3140 // Instructions that may read from memory may be in the load move set.
3141 // Once an instruction is fused, we no longer need its move set, and so
3142 // the values of the map never need to be updated. However, when a load
3143 // is fused, we need to merge the entries from both instructions in the
3144 // pair in case those instructions were in the move set of some other
3145 // yet-to-be-fused pair. The loads in question are the keys of the map.
3146 if (I->mayReadFromMemory()) {
3147 std::vector<ValuePair> NewSetMembers;
3148 DenseMap<Value *, std::vector<Value *> >::iterator II =
3149 LoadMoveSet.find(I);
3150 if (II != LoadMoveSet.end())
3151 for (std::vector<Value *>::iterator N = II->second.begin(),
3152 NE = II->second.end(); N != NE; ++N)
3153 NewSetMembers.push_back(ValuePair(K, *N));
3154 DenseMap<Value *, std::vector<Value *> >::iterator JJ =
3155 LoadMoveSet.find(J);
3156 if (JJ != LoadMoveSet.end())
3157 for (std::vector<Value *>::iterator N = JJ->second.begin(),
3158 NE = JJ->second.end(); N != NE; ++N)
3159 NewSetMembers.push_back(ValuePair(K, *N));
3160 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(),
3161 AE = NewSetMembers.end(); A != AE; ++A) {
3162 LoadMoveSet[A->first].push_back(A->second);
3163 LoadMoveSetPairs.insert(*A);
3167 // Before removing I, set the iterator to the next instruction.
3168 PI = std::next(BasicBlock::iterator(I));
3169 if (cast<Instruction>(PI) == J)
3174 I->eraseFromParent();
3175 J->eraseFromParent();
3177 DEBUG(if (PrintAfterEveryPair) dbgs() << "BBV: block is now: \n" <<
3181 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n");
3185 char BBVectorize::ID = 0;
3186 static const char bb_vectorize_name[] = "Basic-Block Vectorization";
3187 INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
3188 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
3189 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
3190 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
3191 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3192 INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
3194 BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) {
3195 return new BBVectorize(C);
3199 llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) {
3200 BBVectorize BBVectorizer(P, C);
3201 return BBVectorizer.vectorizeBB(BB);
3204 //===----------------------------------------------------------------------===//
3205 VectorizeConfig::VectorizeConfig() {
3206 VectorBits = ::VectorBits;
3207 VectorizeBools = !::NoBools;
3208 VectorizeInts = !::NoInts;
3209 VectorizeFloats = !::NoFloats;
3210 VectorizePointers = !::NoPointers;
3211 VectorizeCasts = !::NoCasts;
3212 VectorizeMath = !::NoMath;
3213 VectorizeFMA = !::NoFMA;
3214 VectorizeSelect = !::NoSelect;
3215 VectorizeCmp = !::NoCmp;
3216 VectorizeGEP = !::NoGEP;
3217 VectorizeMemOps = !::NoMemOps;
3218 AlignedOnly = ::AlignedOnly;
3219 ReqChainDepth= ::ReqChainDepth;
3220 SearchLimit = ::SearchLimit;
3221 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck;
3222 SplatBreaksChain = ::SplatBreaksChain;
3223 MaxInsts = ::MaxInsts;
3224 MaxPairs = ::MaxPairs;
3225 MaxIter = ::MaxIter;
3226 Pow2LenOnly = ::Pow2LenOnly;
3227 NoMemOpBoost = ::NoMemOpBoost;
3228 FastDep = ::FastDep;