1 //===- BBVectorize.cpp - A Basic-Block Vectorizer -------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements a basic-block vectorization pass. The algorithm was
11 // inspired by that used by the Vienna MAP Vectorizor by Franchetti and Kral,
12 // et al. It works by looking for chains of pairable operations and then
15 //===----------------------------------------------------------------------===//
17 #define BBV_NAME "bb-vectorize"
18 #include "llvm/Transforms/Vectorize.h"
19 #include "llvm/ADT/DenseMap.h"
20 #include "llvm/ADT/DenseSet.h"
21 #include "llvm/ADT/STLExtras.h"
22 #include "llvm/ADT/SmallSet.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/ADT/StringExtras.h"
26 #include "llvm/Analysis/AliasAnalysis.h"
27 #include "llvm/Analysis/AliasSetTracker.h"
28 #include "llvm/Analysis/ScalarEvolution.h"
29 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
30 #include "llvm/Analysis/TargetTransformInfo.h"
31 #include "llvm/Analysis/ValueTracking.h"
32 #include "llvm/IR/Constants.h"
33 #include "llvm/IR/DataLayout.h"
34 #include "llvm/IR/DerivedTypes.h"
35 #include "llvm/IR/Dominators.h"
36 #include "llvm/IR/Function.h"
37 #include "llvm/IR/Instructions.h"
38 #include "llvm/IR/IntrinsicInst.h"
39 #include "llvm/IR/Intrinsics.h"
40 #include "llvm/IR/LLVMContext.h"
41 #include "llvm/IR/Metadata.h"
42 #include "llvm/IR/Type.h"
43 #include "llvm/IR/ValueHandle.h"
44 #include "llvm/Pass.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/raw_ostream.h"
48 #include "llvm/Transforms/Utils/Local.h"
52 #define DEBUG_TYPE BBV_NAME
55 IgnoreTargetInfo("bb-vectorize-ignore-target-info", cl::init(false),
56 cl::Hidden, cl::desc("Ignore target information"));
58 static cl::opt<unsigned>
59 ReqChainDepth("bb-vectorize-req-chain-depth", cl::init(6), cl::Hidden,
60 cl::desc("The required chain depth for vectorization"));
63 UseChainDepthWithTI("bb-vectorize-use-chain-depth", cl::init(false),
64 cl::Hidden, cl::desc("Use the chain depth requirement with"
65 " target information"));
67 static cl::opt<unsigned>
68 SearchLimit("bb-vectorize-search-limit", cl::init(400), cl::Hidden,
69 cl::desc("The maximum search distance for instruction pairs"));
72 SplatBreaksChain("bb-vectorize-splat-breaks-chain", cl::init(false), cl::Hidden,
73 cl::desc("Replicating one element to a pair breaks the chain"));
75 static cl::opt<unsigned>
76 VectorBits("bb-vectorize-vector-bits", cl::init(128), cl::Hidden,
77 cl::desc("The size of the native vector registers"));
79 static cl::opt<unsigned>
80 MaxIter("bb-vectorize-max-iter", cl::init(0), cl::Hidden,
81 cl::desc("The maximum number of pairing iterations"));
84 Pow2LenOnly("bb-vectorize-pow2-len-only", cl::init(false), cl::Hidden,
85 cl::desc("Don't try to form non-2^n-length vectors"));
87 static cl::opt<unsigned>
88 MaxInsts("bb-vectorize-max-instr-per-group", cl::init(500), cl::Hidden,
89 cl::desc("The maximum number of pairable instructions per group"));
91 static cl::opt<unsigned>
92 MaxPairs("bb-vectorize-max-pairs-per-group", cl::init(3000), cl::Hidden,
93 cl::desc("The maximum number of candidate instruction pairs per group"));
95 static cl::opt<unsigned>
96 MaxCandPairsForCycleCheck("bb-vectorize-max-cycle-check-pairs", cl::init(200),
97 cl::Hidden, cl::desc("The maximum number of candidate pairs with which to use"
98 " a full cycle check"));
101 NoBools("bb-vectorize-no-bools", cl::init(false), cl::Hidden,
102 cl::desc("Don't try to vectorize boolean (i1) values"));
105 NoInts("bb-vectorize-no-ints", cl::init(false), cl::Hidden,
106 cl::desc("Don't try to vectorize integer values"));
109 NoFloats("bb-vectorize-no-floats", cl::init(false), cl::Hidden,
110 cl::desc("Don't try to vectorize floating-point values"));
112 // FIXME: This should default to false once pointer vector support works.
114 NoPointers("bb-vectorize-no-pointers", cl::init(/*false*/ true), cl::Hidden,
115 cl::desc("Don't try to vectorize pointer values"));
118 NoCasts("bb-vectorize-no-casts", cl::init(false), cl::Hidden,
119 cl::desc("Don't try to vectorize casting (conversion) operations"));
122 NoMath("bb-vectorize-no-math", cl::init(false), cl::Hidden,
123 cl::desc("Don't try to vectorize floating-point math intrinsics"));
126 NoBitManipulation("bb-vectorize-no-bitmanip", cl::init(false), cl::Hidden,
127 cl::desc("Don't try to vectorize BitManipulation intrinsics"));
130 NoFMA("bb-vectorize-no-fma", cl::init(false), cl::Hidden,
131 cl::desc("Don't try to vectorize the fused-multiply-add intrinsic"));
134 NoSelect("bb-vectorize-no-select", cl::init(false), cl::Hidden,
135 cl::desc("Don't try to vectorize select instructions"));
138 NoCmp("bb-vectorize-no-cmp", cl::init(false), cl::Hidden,
139 cl::desc("Don't try to vectorize comparison instructions"));
142 NoGEP("bb-vectorize-no-gep", cl::init(false), cl::Hidden,
143 cl::desc("Don't try to vectorize getelementptr instructions"));
146 NoMemOps("bb-vectorize-no-mem-ops", cl::init(false), cl::Hidden,
147 cl::desc("Don't try to vectorize loads and stores"));
150 AlignedOnly("bb-vectorize-aligned-only", cl::init(false), cl::Hidden,
151 cl::desc("Only generate aligned loads and stores"));
154 NoMemOpBoost("bb-vectorize-no-mem-op-boost",
155 cl::init(false), cl::Hidden,
156 cl::desc("Don't boost the chain-depth contribution of loads and stores"));
159 FastDep("bb-vectorize-fast-dep", cl::init(false), cl::Hidden,
160 cl::desc("Use a fast instruction dependency analysis"));
164 DebugInstructionExamination("bb-vectorize-debug-instruction-examination",
165 cl::init(false), cl::Hidden,
166 cl::desc("When debugging is enabled, output information on the"
167 " instruction-examination process"));
169 DebugCandidateSelection("bb-vectorize-debug-candidate-selection",
170 cl::init(false), cl::Hidden,
171 cl::desc("When debugging is enabled, output information on the"
172 " candidate-selection process"));
174 DebugPairSelection("bb-vectorize-debug-pair-selection",
175 cl::init(false), cl::Hidden,
176 cl::desc("When debugging is enabled, output information on the"
177 " pair-selection process"));
179 DebugCycleCheck("bb-vectorize-debug-cycle-check",
180 cl::init(false), cl::Hidden,
181 cl::desc("When debugging is enabled, output information on the"
182 " cycle-checking process"));
185 PrintAfterEveryPair("bb-vectorize-debug-print-after-every-pair",
186 cl::init(false), cl::Hidden,
187 cl::desc("When debugging is enabled, dump the basic block after"
188 " every pair is fused"));
191 STATISTIC(NumFusedOps, "Number of operations fused by bb-vectorize");
194 struct BBVectorize : public BasicBlockPass {
195 static char ID; // Pass identification, replacement for typeid
197 const VectorizeConfig Config;
199 BBVectorize(const VectorizeConfig &C = VectorizeConfig())
200 : BasicBlockPass(ID), Config(C) {
201 initializeBBVectorizePass(*PassRegistry::getPassRegistry());
204 BBVectorize(Pass *P, const VectorizeConfig &C)
205 : BasicBlockPass(ID), Config(C) {
206 AA = &P->getAnalysis<AliasAnalysis>();
207 DT = &P->getAnalysis<DominatorTreeWrapperPass>().getDomTree();
208 SE = &P->getAnalysis<ScalarEvolution>();
209 DataLayoutPass *DLP = P->getAnalysisIfAvailable<DataLayoutPass>();
210 DL = DLP ? &DLP->getDataLayout() : nullptr;
211 TTI = IgnoreTargetInfo ? nullptr : &P->getAnalysis<TargetTransformInfo>();
214 typedef std::pair<Value *, Value *> ValuePair;
215 typedef std::pair<ValuePair, int> ValuePairWithCost;
216 typedef std::pair<ValuePair, size_t> ValuePairWithDepth;
217 typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair
218 typedef std::pair<VPPair, unsigned> VPPairWithType;
223 const DataLayout *DL;
224 const TargetTransformInfo *TTI;
226 // FIXME: const correct?
228 bool vectorizePairs(BasicBlock &BB, bool NonPow2Len = false);
230 bool getCandidatePairs(BasicBlock &BB,
231 BasicBlock::iterator &Start,
232 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
233 DenseSet<ValuePair> &FixedOrderPairs,
234 DenseMap<ValuePair, int> &CandidatePairCostSavings,
235 std::vector<Value *> &PairableInsts, bool NonPow2Len);
237 // FIXME: The current implementation does not account for pairs that
238 // are connected in multiple ways. For example:
239 // C1 = A1 / A2; C2 = A2 / A1 (which may be both direct and a swap)
240 enum PairConnectionType {
241 PairConnectionDirect,
246 void computeConnectedPairs(
247 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
248 DenseSet<ValuePair> &CandidatePairsSet,
249 std::vector<Value *> &PairableInsts,
250 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
251 DenseMap<VPPair, unsigned> &PairConnectionTypes);
253 void buildDepMap(BasicBlock &BB,
254 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
255 std::vector<Value *> &PairableInsts,
256 DenseSet<ValuePair> &PairableInstUsers);
258 void choosePairs(DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
259 DenseSet<ValuePair> &CandidatePairsSet,
260 DenseMap<ValuePair, int> &CandidatePairCostSavings,
261 std::vector<Value *> &PairableInsts,
262 DenseSet<ValuePair> &FixedOrderPairs,
263 DenseMap<VPPair, unsigned> &PairConnectionTypes,
264 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
265 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
266 DenseSet<ValuePair> &PairableInstUsers,
267 DenseMap<Value *, Value *>& ChosenPairs);
269 void fuseChosenPairs(BasicBlock &BB,
270 std::vector<Value *> &PairableInsts,
271 DenseMap<Value *, Value *>& ChosenPairs,
272 DenseSet<ValuePair> &FixedOrderPairs,
273 DenseMap<VPPair, unsigned> &PairConnectionTypes,
274 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
275 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps);
278 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore);
280 bool areInstsCompatible(Instruction *I, Instruction *J,
281 bool IsSimpleLoadStore, bool NonPow2Len,
282 int &CostSavings, int &FixedOrder);
284 bool trackUsesOfI(DenseSet<Value *> &Users,
285 AliasSetTracker &WriteSet, Instruction *I,
286 Instruction *J, bool UpdateUsers = true,
287 DenseSet<ValuePair> *LoadMoveSetPairs = nullptr);
289 void computePairsConnectedTo(
290 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
291 DenseSet<ValuePair> &CandidatePairsSet,
292 std::vector<Value *> &PairableInsts,
293 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
294 DenseMap<VPPair, unsigned> &PairConnectionTypes,
297 bool pairsConflict(ValuePair P, ValuePair Q,
298 DenseSet<ValuePair> &PairableInstUsers,
299 DenseMap<ValuePair, std::vector<ValuePair> >
300 *PairableInstUserMap = nullptr,
301 DenseSet<VPPair> *PairableInstUserPairSet = nullptr);
303 bool pairWillFormCycle(ValuePair P,
304 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUsers,
305 DenseSet<ValuePair> &CurrentPairs);
308 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
309 std::vector<Value *> &PairableInsts,
310 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
311 DenseSet<ValuePair> &PairableInstUsers,
312 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
313 DenseSet<VPPair> &PairableInstUserPairSet,
314 DenseMap<Value *, Value *> &ChosenPairs,
315 DenseMap<ValuePair, size_t> &DAG,
316 DenseSet<ValuePair> &PrunedDAG, ValuePair J,
319 void buildInitialDAGFor(
320 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
321 DenseSet<ValuePair> &CandidatePairsSet,
322 std::vector<Value *> &PairableInsts,
323 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
324 DenseSet<ValuePair> &PairableInstUsers,
325 DenseMap<Value *, Value *> &ChosenPairs,
326 DenseMap<ValuePair, size_t> &DAG, ValuePair J);
329 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
330 DenseSet<ValuePair> &CandidatePairsSet,
331 DenseMap<ValuePair, int> &CandidatePairCostSavings,
332 std::vector<Value *> &PairableInsts,
333 DenseSet<ValuePair> &FixedOrderPairs,
334 DenseMap<VPPair, unsigned> &PairConnectionTypes,
335 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
336 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
337 DenseSet<ValuePair> &PairableInstUsers,
338 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
339 DenseSet<VPPair> &PairableInstUserPairSet,
340 DenseMap<Value *, Value *> &ChosenPairs,
341 DenseSet<ValuePair> &BestDAG, size_t &BestMaxDepth,
342 int &BestEffSize, Value *II, std::vector<Value *>&JJ,
345 Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I,
346 Instruction *J, unsigned o);
348 void fillNewShuffleMask(LLVMContext& Context, Instruction *J,
349 unsigned MaskOffset, unsigned NumInElem,
350 unsigned NumInElem1, unsigned IdxOffset,
351 std::vector<Constant*> &Mask);
353 Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I,
356 bool expandIEChain(LLVMContext& Context, Instruction *I, Instruction *J,
357 unsigned o, Value *&LOp, unsigned numElemL,
358 Type *ArgTypeL, Type *ArgTypeR, bool IBeforeJ,
359 unsigned IdxOff = 0);
361 Value *getReplacementInput(LLVMContext& Context, Instruction *I,
362 Instruction *J, unsigned o, bool IBeforeJ);
364 void getReplacementInputsForPair(LLVMContext& Context, Instruction *I,
365 Instruction *J, SmallVectorImpl<Value *> &ReplacedOperands,
368 void replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
369 Instruction *J, Instruction *K,
370 Instruction *&InsertionPt, Instruction *&K1,
373 void collectPairLoadMoveSet(BasicBlock &BB,
374 DenseMap<Value *, Value *> &ChosenPairs,
375 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
376 DenseSet<ValuePair> &LoadMoveSetPairs,
379 void collectLoadMoveSet(BasicBlock &BB,
380 std::vector<Value *> &PairableInsts,
381 DenseMap<Value *, Value *> &ChosenPairs,
382 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
383 DenseSet<ValuePair> &LoadMoveSetPairs);
385 bool canMoveUsesOfIAfterJ(BasicBlock &BB,
386 DenseSet<ValuePair> &LoadMoveSetPairs,
387 Instruction *I, Instruction *J);
389 void moveUsesOfIAfterJ(BasicBlock &BB,
390 DenseSet<ValuePair> &LoadMoveSetPairs,
391 Instruction *&InsertionPt,
392 Instruction *I, Instruction *J);
394 bool vectorizeBB(BasicBlock &BB) {
395 if (skipOptnoneFunction(BB))
397 if (!DT->isReachableFromEntry(&BB)) {
398 DEBUG(dbgs() << "BBV: skipping unreachable " << BB.getName() <<
399 " in " << BB.getParent()->getName() << "\n");
403 DEBUG(if (TTI) dbgs() << "BBV: using target information\n");
405 bool changed = false;
406 // Iterate a sufficient number of times to merge types of size 1 bit,
407 // then 2 bits, then 4, etc. up to half of the target vector width of the
408 // target vector register.
411 (TTI || v <= Config.VectorBits) &&
412 (!Config.MaxIter || n <= Config.MaxIter);
414 DEBUG(dbgs() << "BBV: fusing loop #" << n <<
415 " for " << BB.getName() << " in " <<
416 BB.getParent()->getName() << "...\n");
417 if (vectorizePairs(BB))
423 if (changed && !Pow2LenOnly) {
425 for (; !Config.MaxIter || n <= Config.MaxIter; ++n) {
426 DEBUG(dbgs() << "BBV: fusing for non-2^n-length vectors loop #: " <<
427 n << " for " << BB.getName() << " in " <<
428 BB.getParent()->getName() << "...\n");
429 if (!vectorizePairs(BB, true)) break;
433 DEBUG(dbgs() << "BBV: done!\n");
437 bool runOnBasicBlock(BasicBlock &BB) override {
438 // OptimizeNone check deferred to vectorizeBB().
440 AA = &getAnalysis<AliasAnalysis>();
441 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
442 SE = &getAnalysis<ScalarEvolution>();
443 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
444 DL = DLP ? &DLP->getDataLayout() : nullptr;
445 TTI = IgnoreTargetInfo ? nullptr : &getAnalysis<TargetTransformInfo>();
447 return vectorizeBB(BB);
450 void getAnalysisUsage(AnalysisUsage &AU) const override {
451 BasicBlockPass::getAnalysisUsage(AU);
452 AU.addRequired<AliasAnalysis>();
453 AU.addRequired<DominatorTreeWrapperPass>();
454 AU.addRequired<ScalarEvolution>();
455 AU.addRequired<TargetTransformInfo>();
456 AU.addPreserved<AliasAnalysis>();
457 AU.addPreserved<DominatorTreeWrapperPass>();
458 AU.addPreserved<ScalarEvolution>();
459 AU.setPreservesCFG();
462 static inline VectorType *getVecTypeForPair(Type *ElemTy, Type *Elem2Ty) {
463 assert(ElemTy->getScalarType() == Elem2Ty->getScalarType() &&
464 "Cannot form vector from incompatible scalar types");
465 Type *STy = ElemTy->getScalarType();
468 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) {
469 numElem = VTy->getNumElements();
474 if (VectorType *VTy = dyn_cast<VectorType>(Elem2Ty)) {
475 numElem += VTy->getNumElements();
480 return VectorType::get(STy, numElem);
483 static inline void getInstructionTypes(Instruction *I,
484 Type *&T1, Type *&T2) {
485 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
486 // For stores, it is the value type, not the pointer type that matters
487 // because the value is what will come from a vector register.
489 Value *IVal = SI->getValueOperand();
490 T1 = IVal->getType();
495 if (CastInst *CI = dyn_cast<CastInst>(I))
500 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
501 T2 = SI->getCondition()->getType();
502 } else if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(I)) {
503 T2 = SI->getOperand(0)->getType();
504 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) {
505 T2 = CI->getOperand(0)->getType();
509 // Returns the weight associated with the provided value. A chain of
510 // candidate pairs has a length given by the sum of the weights of its
511 // members (one weight per pair; the weight of each member of the pair
512 // is assumed to be the same). This length is then compared to the
513 // chain-length threshold to determine if a given chain is significant
514 // enough to be vectorized. The length is also used in comparing
515 // candidate chains where longer chains are considered to be better.
516 // Note: when this function returns 0, the resulting instructions are
517 // not actually fused.
518 inline size_t getDepthFactor(Value *V) {
519 // InsertElement and ExtractElement have a depth factor of zero. This is
520 // for two reasons: First, they cannot be usefully fused. Second, because
521 // the pass generates a lot of these, they can confuse the simple metric
522 // used to compare the dags in the next iteration. Thus, giving them a
523 // weight of zero allows the pass to essentially ignore them in
524 // subsequent iterations when looking for vectorization opportunities
525 // while still tracking dependency chains that flow through those
527 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V))
530 // Give a load or store half of the required depth so that load/store
531 // pairs will vectorize.
532 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V)))
533 return Config.ReqChainDepth/2;
538 // Returns the cost of the provided instruction using TTI.
539 // This does not handle loads and stores.
540 unsigned getInstrCost(unsigned Opcode, Type *T1, Type *T2,
541 TargetTransformInfo::OperandValueKind Op1VK =
542 TargetTransformInfo::OK_AnyValue,
543 TargetTransformInfo::OperandValueKind Op2VK =
544 TargetTransformInfo::OK_AnyValue) {
547 case Instruction::GetElementPtr:
548 // We mark this instruction as zero-cost because scalar GEPs are usually
549 // lowered to the instruction addressing mode. At the moment we don't
550 // generate vector GEPs.
552 case Instruction::Br:
553 return TTI->getCFInstrCost(Opcode);
554 case Instruction::PHI:
556 case Instruction::Add:
557 case Instruction::FAdd:
558 case Instruction::Sub:
559 case Instruction::FSub:
560 case Instruction::Mul:
561 case Instruction::FMul:
562 case Instruction::UDiv:
563 case Instruction::SDiv:
564 case Instruction::FDiv:
565 case Instruction::URem:
566 case Instruction::SRem:
567 case Instruction::FRem:
568 case Instruction::Shl:
569 case Instruction::LShr:
570 case Instruction::AShr:
571 case Instruction::And:
572 case Instruction::Or:
573 case Instruction::Xor:
574 return TTI->getArithmeticInstrCost(Opcode, T1, Op1VK, Op2VK);
575 case Instruction::Select:
576 case Instruction::ICmp:
577 case Instruction::FCmp:
578 return TTI->getCmpSelInstrCost(Opcode, T1, T2);
579 case Instruction::ZExt:
580 case Instruction::SExt:
581 case Instruction::FPToUI:
582 case Instruction::FPToSI:
583 case Instruction::FPExt:
584 case Instruction::PtrToInt:
585 case Instruction::IntToPtr:
586 case Instruction::SIToFP:
587 case Instruction::UIToFP:
588 case Instruction::Trunc:
589 case Instruction::FPTrunc:
590 case Instruction::BitCast:
591 case Instruction::ShuffleVector:
592 return TTI->getCastInstrCost(Opcode, T1, T2);
598 // This determines the relative offset of two loads or stores, returning
599 // true if the offset could be determined to be some constant value.
600 // For example, if OffsetInElmts == 1, then J accesses the memory directly
601 // after I; if OffsetInElmts == -1 then I accesses the memory
603 bool getPairPtrInfo(Instruction *I, Instruction *J,
604 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment,
605 unsigned &IAddressSpace, unsigned &JAddressSpace,
606 int64_t &OffsetInElmts, bool ComputeOffset = true) {
608 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
609 LoadInst *LJ = cast<LoadInst>(J);
610 IPtr = LI->getPointerOperand();
611 JPtr = LJ->getPointerOperand();
612 IAlignment = LI->getAlignment();
613 JAlignment = LJ->getAlignment();
614 IAddressSpace = LI->getPointerAddressSpace();
615 JAddressSpace = LJ->getPointerAddressSpace();
617 StoreInst *SI = cast<StoreInst>(I), *SJ = cast<StoreInst>(J);
618 IPtr = SI->getPointerOperand();
619 JPtr = SJ->getPointerOperand();
620 IAlignment = SI->getAlignment();
621 JAlignment = SJ->getAlignment();
622 IAddressSpace = SI->getPointerAddressSpace();
623 JAddressSpace = SJ->getPointerAddressSpace();
629 const SCEV *IPtrSCEV = SE->getSCEV(IPtr);
630 const SCEV *JPtrSCEV = SE->getSCEV(JPtr);
632 // If this is a trivial offset, then we'll get something like
633 // 1*sizeof(type). With target data, which we need anyway, this will get
634 // constant folded into a number.
635 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV);
636 if (const SCEVConstant *ConstOffSCEV =
637 dyn_cast<SCEVConstant>(OffsetSCEV)) {
638 ConstantInt *IntOff = ConstOffSCEV->getValue();
639 int64_t Offset = IntOff->getSExtValue();
641 Type *VTy = IPtr->getType()->getPointerElementType();
642 int64_t VTyTSS = (int64_t) DL->getTypeStoreSize(VTy);
644 Type *VTy2 = JPtr->getType()->getPointerElementType();
645 if (VTy != VTy2 && Offset < 0) {
646 int64_t VTy2TSS = (int64_t) DL->getTypeStoreSize(VTy2);
647 OffsetInElmts = Offset/VTy2TSS;
648 return (abs64(Offset) % VTy2TSS) == 0;
651 OffsetInElmts = Offset/VTyTSS;
652 return (abs64(Offset) % VTyTSS) == 0;
658 // Returns true if the provided CallInst represents an intrinsic that can
660 bool isVectorizableIntrinsic(CallInst* I) {
661 Function *F = I->getCalledFunction();
662 if (!F) return false;
664 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID();
665 if (!IID) return false;
670 case Intrinsic::sqrt:
671 case Intrinsic::powi:
675 case Intrinsic::log2:
676 case Intrinsic::log10:
678 case Intrinsic::exp2:
680 case Intrinsic::round:
681 case Intrinsic::copysign:
682 case Intrinsic::ceil:
683 case Intrinsic::nearbyint:
684 case Intrinsic::rint:
685 case Intrinsic::trunc:
686 case Intrinsic::floor:
687 case Intrinsic::fabs:
688 return Config.VectorizeMath;
689 case Intrinsic::bswap:
690 case Intrinsic::ctpop:
691 case Intrinsic::ctlz:
692 case Intrinsic::cttz:
693 return Config.VectorizeBitManipulations;
695 case Intrinsic::fmuladd:
696 return Config.VectorizeFMA;
700 bool isPureIEChain(InsertElementInst *IE) {
701 InsertElementInst *IENext = IE;
703 if (!isa<UndefValue>(IENext->getOperand(0)) &&
704 !isa<InsertElementInst>(IENext->getOperand(0))) {
708 dyn_cast<InsertElementInst>(IENext->getOperand(0))));
714 // This function implements one vectorization iteration on the provided
715 // basic block. It returns true if the block is changed.
716 bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) {
718 BasicBlock::iterator Start = BB.getFirstInsertionPt();
720 std::vector<Value *> AllPairableInsts;
721 DenseMap<Value *, Value *> AllChosenPairs;
722 DenseSet<ValuePair> AllFixedOrderPairs;
723 DenseMap<VPPair, unsigned> AllPairConnectionTypes;
724 DenseMap<ValuePair, std::vector<ValuePair> > AllConnectedPairs,
725 AllConnectedPairDeps;
728 std::vector<Value *> PairableInsts;
729 DenseMap<Value *, std::vector<Value *> > CandidatePairs;
730 DenseSet<ValuePair> FixedOrderPairs;
731 DenseMap<ValuePair, int> CandidatePairCostSavings;
732 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs,
734 CandidatePairCostSavings,
735 PairableInsts, NonPow2Len);
736 if (PairableInsts.empty()) continue;
738 // Build the candidate pair set for faster lookups.
739 DenseSet<ValuePair> CandidatePairsSet;
740 for (DenseMap<Value *, std::vector<Value *> >::iterator I =
741 CandidatePairs.begin(), E = CandidatePairs.end(); I != E; ++I)
742 for (std::vector<Value *>::iterator J = I->second.begin(),
743 JE = I->second.end(); J != JE; ++J)
744 CandidatePairsSet.insert(ValuePair(I->first, *J));
746 // Now we have a map of all of the pairable instructions and we need to
747 // select the best possible pairing. A good pairing is one such that the
748 // users of the pair are also paired. This defines a (directed) forest
749 // over the pairs such that two pairs are connected iff the second pair
752 // Note that it only matters that both members of the second pair use some
753 // element of the first pair (to allow for splatting).
755 DenseMap<ValuePair, std::vector<ValuePair> > ConnectedPairs,
757 DenseMap<VPPair, unsigned> PairConnectionTypes;
758 computeConnectedPairs(CandidatePairs, CandidatePairsSet,
759 PairableInsts, ConnectedPairs, PairConnectionTypes);
760 if (ConnectedPairs.empty()) continue;
762 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator
763 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
765 for (std::vector<ValuePair>::iterator J = I->second.begin(),
766 JE = I->second.end(); J != JE; ++J)
767 ConnectedPairDeps[*J].push_back(I->first);
769 // Build the pairable-instruction dependency map
770 DenseSet<ValuePair> PairableInstUsers;
771 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers);
773 // There is now a graph of the connected pairs. For each variable, pick
774 // the pairing with the largest dag meeting the depth requirement on at
775 // least one branch. Then select all pairings that are part of that dag
776 // and remove them from the list of available pairings and pairable
779 DenseMap<Value *, Value *> ChosenPairs;
780 choosePairs(CandidatePairs, CandidatePairsSet,
781 CandidatePairCostSavings,
782 PairableInsts, FixedOrderPairs, PairConnectionTypes,
783 ConnectedPairs, ConnectedPairDeps,
784 PairableInstUsers, ChosenPairs);
786 if (ChosenPairs.empty()) continue;
787 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(),
788 PairableInsts.end());
789 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end());
791 // Only for the chosen pairs, propagate information on fixed-order pairs,
792 // pair connections, and their types to the data structures used by the
793 // pair fusion procedures.
794 for (DenseMap<Value *, Value *>::iterator I = ChosenPairs.begin(),
795 IE = ChosenPairs.end(); I != IE; ++I) {
796 if (FixedOrderPairs.count(*I))
797 AllFixedOrderPairs.insert(*I);
798 else if (FixedOrderPairs.count(ValuePair(I->second, I->first)))
799 AllFixedOrderPairs.insert(ValuePair(I->second, I->first));
801 for (DenseMap<Value *, Value *>::iterator J = ChosenPairs.begin();
803 DenseMap<VPPair, unsigned>::iterator K =
804 PairConnectionTypes.find(VPPair(*I, *J));
805 if (K != PairConnectionTypes.end()) {
806 AllPairConnectionTypes.insert(*K);
808 K = PairConnectionTypes.find(VPPair(*J, *I));
809 if (K != PairConnectionTypes.end())
810 AllPairConnectionTypes.insert(*K);
815 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator
816 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
818 for (std::vector<ValuePair>::iterator J = I->second.begin(),
819 JE = I->second.end(); J != JE; ++J)
820 if (AllPairConnectionTypes.count(VPPair(I->first, *J))) {
821 AllConnectedPairs[I->first].push_back(*J);
822 AllConnectedPairDeps[*J].push_back(I->first);
824 } while (ShouldContinue);
826 if (AllChosenPairs.empty()) return false;
827 NumFusedOps += AllChosenPairs.size();
829 // A set of pairs has now been selected. It is now necessary to replace the
830 // paired instructions with vector instructions. For this procedure each
831 // operand must be replaced with a vector operand. This vector is formed
832 // by using build_vector on the old operands. The replaced values are then
833 // replaced with a vector_extract on the result. Subsequent optimization
834 // passes should coalesce the build/extract combinations.
836 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs, AllFixedOrderPairs,
837 AllPairConnectionTypes,
838 AllConnectedPairs, AllConnectedPairDeps);
840 // It is important to cleanup here so that future iterations of this
841 // function have less work to do.
842 (void) SimplifyInstructionsInBlock(&BB, DL, AA->getTargetLibraryInfo());
846 // This function returns true if the provided instruction is capable of being
847 // fused into a vector instruction. This determination is based only on the
848 // type and other attributes of the instruction.
849 bool BBVectorize::isInstVectorizable(Instruction *I,
850 bool &IsSimpleLoadStore) {
851 IsSimpleLoadStore = false;
853 if (CallInst *C = dyn_cast<CallInst>(I)) {
854 if (!isVectorizableIntrinsic(C))
856 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) {
857 // Vectorize simple loads if possbile:
858 IsSimpleLoadStore = L->isSimple();
859 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
861 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) {
862 // Vectorize simple stores if possbile:
863 IsSimpleLoadStore = S->isSimple();
864 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
866 } else if (CastInst *C = dyn_cast<CastInst>(I)) {
867 // We can vectorize casts, but not casts of pointer types, etc.
868 if (!Config.VectorizeCasts)
871 Type *SrcTy = C->getSrcTy();
872 if (!SrcTy->isSingleValueType())
875 Type *DestTy = C->getDestTy();
876 if (!DestTy->isSingleValueType())
878 } else if (isa<SelectInst>(I)) {
879 if (!Config.VectorizeSelect)
881 } else if (isa<CmpInst>(I)) {
882 if (!Config.VectorizeCmp)
884 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) {
885 if (!Config.VectorizeGEP)
888 // Currently, vector GEPs exist only with one index.
889 if (G->getNumIndices() != 1)
891 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) ||
892 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) {
896 // We can't vectorize memory operations without target data
897 if (!DL && IsSimpleLoadStore)
901 getInstructionTypes(I, T1, T2);
903 // Not every type can be vectorized...
904 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) ||
905 !(VectorType::isValidElementType(T2) || T2->isVectorTy()))
908 if (T1->getScalarSizeInBits() == 1) {
909 if (!Config.VectorizeBools)
912 if (!Config.VectorizeInts && T1->isIntOrIntVectorTy())
916 if (T2->getScalarSizeInBits() == 1) {
917 if (!Config.VectorizeBools)
920 if (!Config.VectorizeInts && T2->isIntOrIntVectorTy())
924 if (!Config.VectorizeFloats
925 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy()))
928 // Don't vectorize target-specific types.
929 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy())
931 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy())
934 if ((!Config.VectorizePointers || !DL) &&
935 (T1->getScalarType()->isPointerTy() ||
936 T2->getScalarType()->isPointerTy()))
939 if (!TTI && (T1->getPrimitiveSizeInBits() >= Config.VectorBits ||
940 T2->getPrimitiveSizeInBits() >= Config.VectorBits))
946 // This function returns true if the two provided instructions are compatible
947 // (meaning that they can be fused into a vector instruction). This assumes
948 // that I has already been determined to be vectorizable and that J is not
949 // in the use dag of I.
950 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J,
951 bool IsSimpleLoadStore, bool NonPow2Len,
952 int &CostSavings, int &FixedOrder) {
953 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I <<
954 " <-> " << *J << "\n");
959 // Loads and stores can be merged if they have different alignments,
960 // but are otherwise the same.
961 if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment |
962 (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0)))
965 Type *IT1, *IT2, *JT1, *JT2;
966 getInstructionTypes(I, IT1, IT2);
967 getInstructionTypes(J, JT1, JT2);
968 unsigned MaxTypeBits = std::max(
969 IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(),
970 IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits());
971 if (!TTI && MaxTypeBits > Config.VectorBits)
974 // FIXME: handle addsub-type operations!
976 if (IsSimpleLoadStore) {
978 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
979 int64_t OffsetInElmts = 0;
980 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
981 IAddressSpace, JAddressSpace,
982 OffsetInElmts) && abs64(OffsetInElmts) == 1) {
983 FixedOrder = (int) OffsetInElmts;
984 unsigned BottomAlignment = IAlignment;
985 if (OffsetInElmts < 0) BottomAlignment = JAlignment;
987 Type *aTypeI = isa<StoreInst>(I) ?
988 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType();
989 Type *aTypeJ = isa<StoreInst>(J) ?
990 cast<StoreInst>(J)->getValueOperand()->getType() : J->getType();
991 Type *VType = getVecTypeForPair(aTypeI, aTypeJ);
993 if (Config.AlignedOnly) {
994 // An aligned load or store is possible only if the instruction
995 // with the lower offset has an alignment suitable for the
998 unsigned VecAlignment = DL->getPrefTypeAlignment(VType);
999 if (BottomAlignment < VecAlignment)
1004 unsigned ICost = TTI->getMemoryOpCost(I->getOpcode(), aTypeI,
1005 IAlignment, IAddressSpace);
1006 unsigned JCost = TTI->getMemoryOpCost(J->getOpcode(), aTypeJ,
1007 JAlignment, JAddressSpace);
1008 unsigned VCost = TTI->getMemoryOpCost(I->getOpcode(), VType,
1012 ICost += TTI->getAddressComputationCost(aTypeI);
1013 JCost += TTI->getAddressComputationCost(aTypeJ);
1014 VCost += TTI->getAddressComputationCost(VType);
1016 if (VCost > ICost + JCost)
1019 // We don't want to fuse to a type that will be split, even
1020 // if the two input types will also be split and there is no other
1022 unsigned VParts = TTI->getNumberOfParts(VType);
1025 else if (!VParts && VCost == ICost + JCost)
1028 CostSavings = ICost + JCost - VCost;
1034 unsigned ICost = getInstrCost(I->getOpcode(), IT1, IT2);
1035 unsigned JCost = getInstrCost(J->getOpcode(), JT1, JT2);
1036 Type *VT1 = getVecTypeForPair(IT1, JT1),
1037 *VT2 = getVecTypeForPair(IT2, JT2);
1038 TargetTransformInfo::OperandValueKind Op1VK =
1039 TargetTransformInfo::OK_AnyValue;
1040 TargetTransformInfo::OperandValueKind Op2VK =
1041 TargetTransformInfo::OK_AnyValue;
1043 // On some targets (example X86) the cost of a vector shift may vary
1044 // depending on whether the second operand is a Uniform or
1045 // NonUniform Constant.
1046 switch (I->getOpcode()) {
1048 case Instruction::Shl:
1049 case Instruction::LShr:
1050 case Instruction::AShr:
1052 // If both I and J are scalar shifts by constant, then the
1053 // merged vector shift count would be either a constant splat value
1054 // or a non-uniform vector of constants.
1055 if (ConstantInt *CII = dyn_cast<ConstantInt>(I->getOperand(1))) {
1056 if (ConstantInt *CIJ = dyn_cast<ConstantInt>(J->getOperand(1)))
1057 Op2VK = CII == CIJ ? TargetTransformInfo::OK_UniformConstantValue :
1058 TargetTransformInfo::OK_NonUniformConstantValue;
1060 // Check for a splat of a constant or for a non uniform vector
1062 Value *IOp = I->getOperand(1);
1063 Value *JOp = J->getOperand(1);
1064 if ((isa<ConstantVector>(IOp) || isa<ConstantDataVector>(IOp)) &&
1065 (isa<ConstantVector>(JOp) || isa<ConstantDataVector>(JOp))) {
1066 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
1067 Constant *SplatValue = cast<Constant>(IOp)->getSplatValue();
1068 if (SplatValue != nullptr &&
1069 SplatValue == cast<Constant>(JOp)->getSplatValue())
1070 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
1075 // Note that this procedure is incorrect for insert and extract element
1076 // instructions (because combining these often results in a shuffle),
1077 // but this cost is ignored (because insert and extract element
1078 // instructions are assigned a zero depth factor and are not really
1079 // fused in general).
1080 unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2, Op1VK, Op2VK);
1082 if (VCost > ICost + JCost)
1085 // We don't want to fuse to a type that will be split, even
1086 // if the two input types will also be split and there is no other
1088 unsigned VParts1 = TTI->getNumberOfParts(VT1),
1089 VParts2 = TTI->getNumberOfParts(VT2);
1090 if (VParts1 > 1 || VParts2 > 1)
1092 else if ((!VParts1 || !VParts2) && VCost == ICost + JCost)
1095 CostSavings = ICost + JCost - VCost;
1098 // The powi,ctlz,cttz intrinsics are special because only the first
1099 // argument is vectorized, the second arguments must be equal.
1100 CallInst *CI = dyn_cast<CallInst>(I);
1102 if (CI && (FI = CI->getCalledFunction())) {
1103 Intrinsic::ID IID = (Intrinsic::ID) FI->getIntrinsicID();
1104 if (IID == Intrinsic::powi || IID == Intrinsic::ctlz ||
1105 IID == Intrinsic::cttz) {
1106 Value *A1I = CI->getArgOperand(1),
1107 *A1J = cast<CallInst>(J)->getArgOperand(1);
1108 const SCEV *A1ISCEV = SE->getSCEV(A1I),
1109 *A1JSCEV = SE->getSCEV(A1J);
1110 return (A1ISCEV == A1JSCEV);
1114 SmallVector<Type*, 4> Tys;
1115 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
1116 Tys.push_back(CI->getArgOperand(i)->getType());
1117 unsigned ICost = TTI->getIntrinsicInstrCost(IID, IT1, Tys);
1120 CallInst *CJ = cast<CallInst>(J);
1121 for (unsigned i = 0, ie = CJ->getNumArgOperands(); i != ie; ++i)
1122 Tys.push_back(CJ->getArgOperand(i)->getType());
1123 unsigned JCost = TTI->getIntrinsicInstrCost(IID, JT1, Tys);
1126 assert(CI->getNumArgOperands() == CJ->getNumArgOperands() &&
1127 "Intrinsic argument counts differ");
1128 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
1129 if ((IID == Intrinsic::powi || IID == Intrinsic::ctlz ||
1130 IID == Intrinsic::cttz) && i == 1)
1131 Tys.push_back(CI->getArgOperand(i)->getType());
1133 Tys.push_back(getVecTypeForPair(CI->getArgOperand(i)->getType(),
1134 CJ->getArgOperand(i)->getType()));
1137 Type *RetTy = getVecTypeForPair(IT1, JT1);
1138 unsigned VCost = TTI->getIntrinsicInstrCost(IID, RetTy, Tys);
1140 if (VCost > ICost + JCost)
1143 // We don't want to fuse to a type that will be split, even
1144 // if the two input types will also be split and there is no other
1146 unsigned RetParts = TTI->getNumberOfParts(RetTy);
1149 else if (!RetParts && VCost == ICost + JCost)
1152 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
1153 if (!Tys[i]->isVectorTy())
1156 unsigned NumParts = TTI->getNumberOfParts(Tys[i]);
1159 else if (!NumParts && VCost == ICost + JCost)
1163 CostSavings = ICost + JCost - VCost;
1170 // Figure out whether or not J uses I and update the users and write-set
1171 // structures associated with I. Specifically, Users represents the set of
1172 // instructions that depend on I. WriteSet represents the set
1173 // of memory locations that are dependent on I. If UpdateUsers is true,
1174 // and J uses I, then Users is updated to contain J and WriteSet is updated
1175 // to contain any memory locations to which J writes. The function returns
1176 // true if J uses I. By default, alias analysis is used to determine
1177 // whether J reads from memory that overlaps with a location in WriteSet.
1178 // If LoadMoveSet is not null, then it is a previously-computed map
1179 // where the key is the memory-based user instruction and the value is
1180 // the instruction to be compared with I. So, if LoadMoveSet is provided,
1181 // then the alias analysis is not used. This is necessary because this
1182 // function is called during the process of moving instructions during
1183 // vectorization and the results of the alias analysis are not stable during
1185 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users,
1186 AliasSetTracker &WriteSet, Instruction *I,
1187 Instruction *J, bool UpdateUsers,
1188 DenseSet<ValuePair> *LoadMoveSetPairs) {
1191 // This instruction may already be marked as a user due, for example, to
1192 // being a member of a selected pair.
1197 for (User::op_iterator JU = J->op_begin(), JE = J->op_end();
1200 if (I == V || Users.count(V)) {
1205 if (!UsesI && J->mayReadFromMemory()) {
1206 if (LoadMoveSetPairs) {
1207 UsesI = LoadMoveSetPairs->count(ValuePair(J, I));
1209 for (AliasSetTracker::iterator W = WriteSet.begin(),
1210 WE = WriteSet.end(); W != WE; ++W) {
1211 if (W->aliasesUnknownInst(J, *AA)) {
1219 if (UsesI && UpdateUsers) {
1220 if (J->mayWriteToMemory()) WriteSet.add(J);
1227 // This function iterates over all instruction pairs in the provided
1228 // basic block and collects all candidate pairs for vectorization.
1229 bool BBVectorize::getCandidatePairs(BasicBlock &BB,
1230 BasicBlock::iterator &Start,
1231 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1232 DenseSet<ValuePair> &FixedOrderPairs,
1233 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1234 std::vector<Value *> &PairableInsts, bool NonPow2Len) {
1235 size_t TotalPairs = 0;
1236 BasicBlock::iterator E = BB.end();
1237 if (Start == E) return false;
1239 bool ShouldContinue = false, IAfterStart = false;
1240 for (BasicBlock::iterator I = Start++; I != E; ++I) {
1241 if (I == Start) IAfterStart = true;
1243 bool IsSimpleLoadStore;
1244 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue;
1246 // Look for an instruction with which to pair instruction *I...
1247 DenseSet<Value *> Users;
1248 AliasSetTracker WriteSet(*AA);
1249 if (I->mayWriteToMemory()) WriteSet.add(I);
1251 bool JAfterStart = IAfterStart;
1252 BasicBlock::iterator J = std::next(I);
1253 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) {
1254 if (J == Start) JAfterStart = true;
1256 // Determine if J uses I, if so, exit the loop.
1257 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep);
1258 if (Config.FastDep) {
1259 // Note: For this heuristic to be effective, independent operations
1260 // must tend to be intermixed. This is likely to be true from some
1261 // kinds of grouped loop unrolling (but not the generic LLVM pass),
1262 // but otherwise may require some kind of reordering pass.
1264 // When using fast dependency analysis,
1265 // stop searching after first use:
1268 if (UsesI) continue;
1271 // J does not use I, and comes before the first use of I, so it can be
1272 // merged with I if the instructions are compatible.
1273 int CostSavings, FixedOrder;
1274 if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len,
1275 CostSavings, FixedOrder)) continue;
1277 // J is a candidate for merging with I.
1278 if (!PairableInsts.size() ||
1279 PairableInsts[PairableInsts.size()-1] != I) {
1280 PairableInsts.push_back(I);
1283 CandidatePairs[I].push_back(J);
1286 CandidatePairCostSavings.insert(ValuePairWithCost(ValuePair(I, J),
1289 if (FixedOrder == 1)
1290 FixedOrderPairs.insert(ValuePair(I, J));
1291 else if (FixedOrder == -1)
1292 FixedOrderPairs.insert(ValuePair(J, I));
1294 // The next call to this function must start after the last instruction
1295 // selected during this invocation.
1297 Start = std::next(J);
1298 IAfterStart = JAfterStart = false;
1301 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair "
1302 << *I << " <-> " << *J << " (cost savings: " <<
1303 CostSavings << ")\n");
1305 // If we have already found too many pairs, break here and this function
1306 // will be called again starting after the last instruction selected
1307 // during this invocation.
1308 if (PairableInsts.size() >= Config.MaxInsts ||
1309 TotalPairs >= Config.MaxPairs) {
1310 ShouldContinue = true;
1319 DEBUG(dbgs() << "BBV: found " << PairableInsts.size()
1320 << " instructions with candidate pairs\n");
1322 return ShouldContinue;
1325 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that
1326 // it looks for pairs such that both members have an input which is an
1327 // output of PI or PJ.
1328 void BBVectorize::computePairsConnectedTo(
1329 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1330 DenseSet<ValuePair> &CandidatePairsSet,
1331 std::vector<Value *> &PairableInsts,
1332 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1333 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1337 // For each possible pairing for this variable, look at the uses of
1338 // the first value...
1339 for (Value::user_iterator I = P.first->user_begin(),
1340 E = P.first->user_end();
1343 if (isa<LoadInst>(UI)) {
1344 // A pair cannot be connected to a load because the load only takes one
1345 // operand (the address) and it is a scalar even after vectorization.
1347 } else if ((SI = dyn_cast<StoreInst>(UI)) &&
1348 P.first == SI->getPointerOperand()) {
1349 // Similarly, a pair cannot be connected to a store through its
1354 // For each use of the first variable, look for uses of the second
1356 for (User *UJ : P.second->users()) {
1357 if ((SJ = dyn_cast<StoreInst>(UJ)) &&
1358 P.second == SJ->getPointerOperand())
1362 if (CandidatePairsSet.count(ValuePair(UI, UJ))) {
1363 VPPair VP(P, ValuePair(UI, UJ));
1364 ConnectedPairs[VP.first].push_back(VP.second);
1365 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionDirect));
1369 if (CandidatePairsSet.count(ValuePair(UJ, UI))) {
1370 VPPair VP(P, ValuePair(UJ, UI));
1371 ConnectedPairs[VP.first].push_back(VP.second);
1372 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSwap));
1376 if (Config.SplatBreaksChain) continue;
1377 // Look for cases where just the first value in the pair is used by
1378 // both members of another pair (splatting).
1379 for (Value::user_iterator J = P.first->user_begin(); J != E; ++J) {
1381 if ((SJ = dyn_cast<StoreInst>(UJ)) &&
1382 P.first == SJ->getPointerOperand())
1385 if (CandidatePairsSet.count(ValuePair(UI, UJ))) {
1386 VPPair VP(P, ValuePair(UI, UJ));
1387 ConnectedPairs[VP.first].push_back(VP.second);
1388 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1393 if (Config.SplatBreaksChain) return;
1394 // Look for cases where just the second value in the pair is used by
1395 // both members of another pair (splatting).
1396 for (Value::user_iterator I = P.second->user_begin(),
1397 E = P.second->user_end();
1400 if (isa<LoadInst>(UI))
1402 else if ((SI = dyn_cast<StoreInst>(UI)) &&
1403 P.second == SI->getPointerOperand())
1406 for (Value::user_iterator J = P.second->user_begin(); J != E; ++J) {
1408 if ((SJ = dyn_cast<StoreInst>(UJ)) &&
1409 P.second == SJ->getPointerOperand())
1412 if (CandidatePairsSet.count(ValuePair(UI, UJ))) {
1413 VPPair VP(P, ValuePair(UI, UJ));
1414 ConnectedPairs[VP.first].push_back(VP.second);
1415 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1421 // This function figures out which pairs are connected. Two pairs are
1422 // connected if some output of the first pair forms an input to both members
1423 // of the second pair.
1424 void BBVectorize::computeConnectedPairs(
1425 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1426 DenseSet<ValuePair> &CandidatePairsSet,
1427 std::vector<Value *> &PairableInsts,
1428 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1429 DenseMap<VPPair, unsigned> &PairConnectionTypes) {
1430 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
1431 PE = PairableInsts.end(); PI != PE; ++PI) {
1432 DenseMap<Value *, std::vector<Value *> >::iterator PP =
1433 CandidatePairs.find(*PI);
1434 if (PP == CandidatePairs.end())
1437 for (std::vector<Value *>::iterator P = PP->second.begin(),
1438 E = PP->second.end(); P != E; ++P)
1439 computePairsConnectedTo(CandidatePairs, CandidatePairsSet,
1440 PairableInsts, ConnectedPairs,
1441 PairConnectionTypes, ValuePair(*PI, *P));
1444 DEBUG(size_t TotalPairs = 0;
1445 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator I =
1446 ConnectedPairs.begin(), IE = ConnectedPairs.end(); I != IE; ++I)
1447 TotalPairs += I->second.size();
1448 dbgs() << "BBV: found " << TotalPairs
1449 << " pair connections.\n");
1452 // This function builds a set of use tuples such that <A, B> is in the set
1453 // if B is in the use dag of A. If B is in the use dag of A, then B
1454 // depends on the output of A.
1455 void BBVectorize::buildDepMap(
1457 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1458 std::vector<Value *> &PairableInsts,
1459 DenseSet<ValuePair> &PairableInstUsers) {
1460 DenseSet<Value *> IsInPair;
1461 for (DenseMap<Value *, std::vector<Value *> >::iterator C =
1462 CandidatePairs.begin(), E = CandidatePairs.end(); C != E; ++C) {
1463 IsInPair.insert(C->first);
1464 IsInPair.insert(C->second.begin(), C->second.end());
1467 // Iterate through the basic block, recording all users of each
1468 // pairable instruction.
1470 BasicBlock::iterator E = BB.end(), EL =
1471 BasicBlock::iterator(cast<Instruction>(PairableInsts.back()));
1472 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) {
1473 if (IsInPair.find(I) == IsInPair.end()) continue;
1475 DenseSet<Value *> Users;
1476 AliasSetTracker WriteSet(*AA);
1477 if (I->mayWriteToMemory()) WriteSet.add(I);
1479 for (BasicBlock::iterator J = std::next(I); J != E; ++J) {
1480 (void) trackUsesOfI(Users, WriteSet, I, J);
1486 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end();
1488 if (IsInPair.find(*U) == IsInPair.end()) continue;
1489 PairableInstUsers.insert(ValuePair(I, *U));
1497 // Returns true if an input to pair P is an output of pair Q and also an
1498 // input of pair Q is an output of pair P. If this is the case, then these
1499 // two pairs cannot be simultaneously fused.
1500 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q,
1501 DenseSet<ValuePair> &PairableInstUsers,
1502 DenseMap<ValuePair, std::vector<ValuePair> > *PairableInstUserMap,
1503 DenseSet<VPPair> *PairableInstUserPairSet) {
1504 // Two pairs are in conflict if they are mutual Users of eachother.
1505 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) ||
1506 PairableInstUsers.count(ValuePair(P.first, Q.second)) ||
1507 PairableInstUsers.count(ValuePair(P.second, Q.first)) ||
1508 PairableInstUsers.count(ValuePair(P.second, Q.second));
1509 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) ||
1510 PairableInstUsers.count(ValuePair(Q.first, P.second)) ||
1511 PairableInstUsers.count(ValuePair(Q.second, P.first)) ||
1512 PairableInstUsers.count(ValuePair(Q.second, P.second));
1513 if (PairableInstUserMap) {
1514 // FIXME: The expensive part of the cycle check is not so much the cycle
1515 // check itself but this edge insertion procedure. This needs some
1516 // profiling and probably a different data structure.
1518 if (PairableInstUserPairSet->insert(VPPair(Q, P)).second)
1519 (*PairableInstUserMap)[Q].push_back(P);
1522 if (PairableInstUserPairSet->insert(VPPair(P, Q)).second)
1523 (*PairableInstUserMap)[P].push_back(Q);
1527 return (QUsesP && PUsesQ);
1530 // This function walks the use graph of current pairs to see if, starting
1531 // from P, the walk returns to P.
1532 bool BBVectorize::pairWillFormCycle(ValuePair P,
1533 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
1534 DenseSet<ValuePair> &CurrentPairs) {
1535 DEBUG(if (DebugCycleCheck)
1536 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> "
1537 << *P.second << "\n");
1538 // A lookup table of visisted pairs is kept because the PairableInstUserMap
1539 // contains non-direct associations.
1540 DenseSet<ValuePair> Visited;
1541 SmallVector<ValuePair, 32> Q;
1542 // General depth-first post-order traversal:
1545 ValuePair QTop = Q.pop_back_val();
1546 Visited.insert(QTop);
1548 DEBUG(if (DebugCycleCheck)
1549 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> "
1550 << *QTop.second << "\n");
1551 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
1552 PairableInstUserMap.find(QTop);
1553 if (QQ == PairableInstUserMap.end())
1556 for (std::vector<ValuePair>::iterator C = QQ->second.begin(),
1557 CE = QQ->second.end(); C != CE; ++C) {
1560 << "BBV: rejected to prevent non-trivial cycle formation: "
1561 << QTop.first << " <-> " << C->second << "\n");
1565 if (CurrentPairs.count(*C) && !Visited.count(*C))
1568 } while (!Q.empty());
1573 // This function builds the initial dag of connected pairs with the
1574 // pair J at the root.
1575 void BBVectorize::buildInitialDAGFor(
1576 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1577 DenseSet<ValuePair> &CandidatePairsSet,
1578 std::vector<Value *> &PairableInsts,
1579 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1580 DenseSet<ValuePair> &PairableInstUsers,
1581 DenseMap<Value *, Value *> &ChosenPairs,
1582 DenseMap<ValuePair, size_t> &DAG, ValuePair J) {
1583 // Each of these pairs is viewed as the root node of a DAG. The DAG
1584 // is then walked (depth-first). As this happens, we keep track of
1585 // the pairs that compose the DAG and the maximum depth of the DAG.
1586 SmallVector<ValuePairWithDepth, 32> Q;
1587 // General depth-first post-order traversal:
1588 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1590 ValuePairWithDepth QTop = Q.back();
1592 // Push each child onto the queue:
1593 bool MoreChildren = false;
1594 size_t MaxChildDepth = QTop.second;
1595 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
1596 ConnectedPairs.find(QTop.first);
1597 if (QQ != ConnectedPairs.end())
1598 for (std::vector<ValuePair>::iterator k = QQ->second.begin(),
1599 ke = QQ->second.end(); k != ke; ++k) {
1600 // Make sure that this child pair is still a candidate:
1601 if (CandidatePairsSet.count(*k)) {
1602 DenseMap<ValuePair, size_t>::iterator C = DAG.find(*k);
1603 if (C == DAG.end()) {
1604 size_t d = getDepthFactor(k->first);
1605 Q.push_back(ValuePairWithDepth(*k, QTop.second+d));
1606 MoreChildren = true;
1608 MaxChildDepth = std::max(MaxChildDepth, C->second);
1613 if (!MoreChildren) {
1614 // Record the current pair as part of the DAG:
1615 DAG.insert(ValuePairWithDepth(QTop.first, MaxChildDepth));
1618 } while (!Q.empty());
1621 // Given some initial dag, prune it by removing conflicting pairs (pairs
1622 // that cannot be simultaneously chosen for vectorization).
1623 void BBVectorize::pruneDAGFor(
1624 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1625 std::vector<Value *> &PairableInsts,
1626 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1627 DenseSet<ValuePair> &PairableInstUsers,
1628 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
1629 DenseSet<VPPair> &PairableInstUserPairSet,
1630 DenseMap<Value *, Value *> &ChosenPairs,
1631 DenseMap<ValuePair, size_t> &DAG,
1632 DenseSet<ValuePair> &PrunedDAG, ValuePair J,
1633 bool UseCycleCheck) {
1634 SmallVector<ValuePairWithDepth, 32> Q;
1635 // General depth-first post-order traversal:
1636 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1638 ValuePairWithDepth QTop = Q.pop_back_val();
1639 PrunedDAG.insert(QTop.first);
1641 // Visit each child, pruning as necessary...
1642 SmallVector<ValuePairWithDepth, 8> BestChildren;
1643 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
1644 ConnectedPairs.find(QTop.first);
1645 if (QQ == ConnectedPairs.end())
1648 for (std::vector<ValuePair>::iterator K = QQ->second.begin(),
1649 KE = QQ->second.end(); K != KE; ++K) {
1650 DenseMap<ValuePair, size_t>::iterator C = DAG.find(*K);
1651 if (C == DAG.end()) continue;
1653 // This child is in the DAG, now we need to make sure it is the
1654 // best of any conflicting children. There could be multiple
1655 // conflicting children, so first, determine if we're keeping
1656 // this child, then delete conflicting children as necessary.
1658 // It is also necessary to guard against pairing-induced
1659 // dependencies. Consider instructions a .. x .. y .. b
1660 // such that (a,b) are to be fused and (x,y) are to be fused
1661 // but a is an input to x and b is an output from y. This
1662 // means that y cannot be moved after b but x must be moved
1663 // after b for (a,b) to be fused. In other words, after
1664 // fusing (a,b) we have y .. a/b .. x where y is an input
1665 // to a/b and x is an output to a/b: x and y can no longer
1666 // be legally fused. To prevent this condition, we must
1667 // make sure that a child pair added to the DAG is not
1668 // both an input and output of an already-selected pair.
1670 // Pairing-induced dependencies can also form from more complicated
1671 // cycles. The pair vs. pair conflicts are easy to check, and so
1672 // that is done explicitly for "fast rejection", and because for
1673 // child vs. child conflicts, we may prefer to keep the current
1674 // pair in preference to the already-selected child.
1675 DenseSet<ValuePair> CurrentPairs;
1678 for (SmallVectorImpl<ValuePairWithDepth>::iterator C2
1679 = BestChildren.begin(), E2 = BestChildren.end();
1681 if (C2->first.first == C->first.first ||
1682 C2->first.first == C->first.second ||
1683 C2->first.second == C->first.first ||
1684 C2->first.second == C->first.second ||
1685 pairsConflict(C2->first, C->first, PairableInstUsers,
1686 UseCycleCheck ? &PairableInstUserMap : nullptr,
1687 UseCycleCheck ? &PairableInstUserPairSet
1689 if (C2->second >= C->second) {
1694 CurrentPairs.insert(C2->first);
1697 if (!CanAdd) continue;
1699 // Even worse, this child could conflict with another node already
1700 // selected for the DAG. If that is the case, ignore this child.
1701 for (DenseSet<ValuePair>::iterator T = PrunedDAG.begin(),
1702 E2 = PrunedDAG.end(); T != E2; ++T) {
1703 if (T->first == C->first.first ||
1704 T->first == C->first.second ||
1705 T->second == C->first.first ||
1706 T->second == C->first.second ||
1707 pairsConflict(*T, C->first, PairableInstUsers,
1708 UseCycleCheck ? &PairableInstUserMap : nullptr,
1709 UseCycleCheck ? &PairableInstUserPairSet
1715 CurrentPairs.insert(*T);
1717 if (!CanAdd) continue;
1719 // And check the queue too...
1720 for (SmallVectorImpl<ValuePairWithDepth>::iterator C2 = Q.begin(),
1721 E2 = Q.end(); C2 != E2; ++C2) {
1722 if (C2->first.first == C->first.first ||
1723 C2->first.first == C->first.second ||
1724 C2->first.second == C->first.first ||
1725 C2->first.second == C->first.second ||
1726 pairsConflict(C2->first, C->first, PairableInstUsers,
1727 UseCycleCheck ? &PairableInstUserMap : nullptr,
1728 UseCycleCheck ? &PairableInstUserPairSet
1734 CurrentPairs.insert(C2->first);
1736 if (!CanAdd) continue;
1738 // Last but not least, check for a conflict with any of the
1739 // already-chosen pairs.
1740 for (DenseMap<Value *, Value *>::iterator C2 =
1741 ChosenPairs.begin(), E2 = ChosenPairs.end();
1743 if (pairsConflict(*C2, C->first, PairableInstUsers,
1744 UseCycleCheck ? &PairableInstUserMap : nullptr,
1745 UseCycleCheck ? &PairableInstUserPairSet
1751 CurrentPairs.insert(*C2);
1753 if (!CanAdd) continue;
1755 // To check for non-trivial cycles formed by the addition of the
1756 // current pair we've formed a list of all relevant pairs, now use a
1757 // graph walk to check for a cycle. We start from the current pair and
1758 // walk the use dag to see if we again reach the current pair. If we
1759 // do, then the current pair is rejected.
1761 // FIXME: It may be more efficient to use a topological-ordering
1762 // algorithm to improve the cycle check. This should be investigated.
1763 if (UseCycleCheck &&
1764 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs))
1767 // This child can be added, but we may have chosen it in preference
1768 // to an already-selected child. Check for this here, and if a
1769 // conflict is found, then remove the previously-selected child
1770 // before adding this one in its place.
1771 for (SmallVectorImpl<ValuePairWithDepth>::iterator C2
1772 = BestChildren.begin(); C2 != BestChildren.end();) {
1773 if (C2->first.first == C->first.first ||
1774 C2->first.first == C->first.second ||
1775 C2->first.second == C->first.first ||
1776 C2->first.second == C->first.second ||
1777 pairsConflict(C2->first, C->first, PairableInstUsers))
1778 C2 = BestChildren.erase(C2);
1783 BestChildren.push_back(ValuePairWithDepth(C->first, C->second));
1786 for (SmallVectorImpl<ValuePairWithDepth>::iterator C
1787 = BestChildren.begin(), E2 = BestChildren.end();
1789 size_t DepthF = getDepthFactor(C->first.first);
1790 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF));
1792 } while (!Q.empty());
1795 // This function finds the best dag of mututally-compatible connected
1796 // pairs, given the choice of root pairs as an iterator range.
1797 void BBVectorize::findBestDAGFor(
1798 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1799 DenseSet<ValuePair> &CandidatePairsSet,
1800 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1801 std::vector<Value *> &PairableInsts,
1802 DenseSet<ValuePair> &FixedOrderPairs,
1803 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1804 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1805 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
1806 DenseSet<ValuePair> &PairableInstUsers,
1807 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
1808 DenseSet<VPPair> &PairableInstUserPairSet,
1809 DenseMap<Value *, Value *> &ChosenPairs,
1810 DenseSet<ValuePair> &BestDAG, size_t &BestMaxDepth,
1811 int &BestEffSize, Value *II, std::vector<Value *>&JJ,
1812 bool UseCycleCheck) {
1813 for (std::vector<Value *>::iterator J = JJ.begin(), JE = JJ.end();
1815 ValuePair IJ(II, *J);
1816 if (!CandidatePairsSet.count(IJ))
1819 // Before going any further, make sure that this pair does not
1820 // conflict with any already-selected pairs (see comment below
1821 // near the DAG pruning for more details).
1822 DenseSet<ValuePair> ChosenPairSet;
1823 bool DoesConflict = false;
1824 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(),
1825 E = ChosenPairs.end(); C != E; ++C) {
1826 if (pairsConflict(*C, IJ, PairableInstUsers,
1827 UseCycleCheck ? &PairableInstUserMap : nullptr,
1828 UseCycleCheck ? &PairableInstUserPairSet : nullptr)) {
1829 DoesConflict = true;
1833 ChosenPairSet.insert(*C);
1835 if (DoesConflict) continue;
1837 if (UseCycleCheck &&
1838 pairWillFormCycle(IJ, PairableInstUserMap, ChosenPairSet))
1841 DenseMap<ValuePair, size_t> DAG;
1842 buildInitialDAGFor(CandidatePairs, CandidatePairsSet,
1843 PairableInsts, ConnectedPairs,
1844 PairableInstUsers, ChosenPairs, DAG, IJ);
1846 // Because we'll keep the child with the largest depth, the largest
1847 // depth is still the same in the unpruned DAG.
1848 size_t MaxDepth = DAG.lookup(IJ);
1850 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found DAG for pair {"
1851 << *IJ.first << " <-> " << *IJ.second << "} of depth " <<
1852 MaxDepth << " and size " << DAG.size() << "\n");
1854 // At this point the DAG has been constructed, but, may contain
1855 // contradictory children (meaning that different children of
1856 // some dag node may be attempting to fuse the same instruction).
1857 // So now we walk the dag again, in the case of a conflict,
1858 // keep only the child with the largest depth. To break a tie,
1859 // favor the first child.
1861 DenseSet<ValuePair> PrunedDAG;
1862 pruneDAGFor(CandidatePairs, PairableInsts, ConnectedPairs,
1863 PairableInstUsers, PairableInstUserMap,
1864 PairableInstUserPairSet,
1865 ChosenPairs, DAG, PrunedDAG, IJ, UseCycleCheck);
1869 DenseSet<Value *> PrunedDAGInstrs;
1870 for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(),
1871 E = PrunedDAG.end(); S != E; ++S) {
1872 PrunedDAGInstrs.insert(S->first);
1873 PrunedDAGInstrs.insert(S->second);
1876 // The set of pairs that have already contributed to the total cost.
1877 DenseSet<ValuePair> IncomingPairs;
1879 // If the cost model were perfect, this might not be necessary; but we
1880 // need to make sure that we don't get stuck vectorizing our own
1882 bool HasNontrivialInsts = false;
1884 // The node weights represent the cost savings associated with
1885 // fusing the pair of instructions.
1886 for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(),
1887 E = PrunedDAG.end(); S != E; ++S) {
1888 if (!isa<ShuffleVectorInst>(S->first) &&
1889 !isa<InsertElementInst>(S->first) &&
1890 !isa<ExtractElementInst>(S->first))
1891 HasNontrivialInsts = true;
1893 bool FlipOrder = false;
1895 if (getDepthFactor(S->first)) {
1896 int ESContrib = CandidatePairCostSavings.find(*S)->second;
1897 DEBUG(if (DebugPairSelection) dbgs() << "\tweight {"
1898 << *S->first << " <-> " << *S->second << "} = " <<
1900 EffSize += ESContrib;
1903 // The edge weights contribute in a negative sense: they represent
1904 // the cost of shuffles.
1905 DenseMap<ValuePair, std::vector<ValuePair> >::iterator SS =
1906 ConnectedPairDeps.find(*S);
1907 if (SS != ConnectedPairDeps.end()) {
1908 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
1909 for (std::vector<ValuePair>::iterator T = SS->second.begin(),
1910 TE = SS->second.end(); T != TE; ++T) {
1912 if (!PrunedDAG.count(Q.second))
1914 DenseMap<VPPair, unsigned>::iterator R =
1915 PairConnectionTypes.find(VPPair(Q.second, Q.first));
1916 assert(R != PairConnectionTypes.end() &&
1917 "Cannot find pair connection type");
1918 if (R->second == PairConnectionDirect)
1920 else if (R->second == PairConnectionSwap)
1924 // If there are more swaps than direct connections, then
1925 // the pair order will be flipped during fusion. So the real
1926 // number of swaps is the minimum number.
1927 FlipOrder = !FixedOrderPairs.count(*S) &&
1928 ((NumDepsSwap > NumDepsDirect) ||
1929 FixedOrderPairs.count(ValuePair(S->second, S->first)));
1931 for (std::vector<ValuePair>::iterator T = SS->second.begin(),
1932 TE = SS->second.end(); T != TE; ++T) {
1934 if (!PrunedDAG.count(Q.second))
1936 DenseMap<VPPair, unsigned>::iterator R =
1937 PairConnectionTypes.find(VPPair(Q.second, Q.first));
1938 assert(R != PairConnectionTypes.end() &&
1939 "Cannot find pair connection type");
1940 Type *Ty1 = Q.second.first->getType(),
1941 *Ty2 = Q.second.second->getType();
1942 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1943 if ((R->second == PairConnectionDirect && FlipOrder) ||
1944 (R->second == PairConnectionSwap && !FlipOrder) ||
1945 R->second == PairConnectionSplat) {
1946 int ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1949 if (VTy->getVectorNumElements() == 2) {
1950 if (R->second == PairConnectionSplat)
1951 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1952 TargetTransformInfo::SK_Broadcast, VTy));
1954 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1955 TargetTransformInfo::SK_Reverse, VTy));
1958 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1959 *Q.second.first << " <-> " << *Q.second.second <<
1961 *S->first << " <-> " << *S->second << "} = " <<
1963 EffSize -= ESContrib;
1968 // Compute the cost of outgoing edges. We assume that edges outgoing
1969 // to shuffles, inserts or extracts can be merged, and so contribute
1970 // no additional cost.
1971 if (!S->first->getType()->isVoidTy()) {
1972 Type *Ty1 = S->first->getType(),
1973 *Ty2 = S->second->getType();
1974 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1976 bool NeedsExtraction = false;
1977 for (User *U : S->first->users()) {
1978 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(U)) {
1979 // Shuffle can be folded if it has no other input
1980 if (isa<UndefValue>(SI->getOperand(1)))
1983 if (isa<ExtractElementInst>(U))
1985 if (PrunedDAGInstrs.count(U))
1987 NeedsExtraction = true;
1991 if (NeedsExtraction) {
1993 if (Ty1->isVectorTy()) {
1994 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1996 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1997 TargetTransformInfo::SK_ExtractSubvector, VTy, 0, Ty1));
1999 ESContrib = (int) TTI->getVectorInstrCost(
2000 Instruction::ExtractElement, VTy, 0);
2002 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
2003 *S->first << "} = " << ESContrib << "\n");
2004 EffSize -= ESContrib;
2007 NeedsExtraction = false;
2008 for (User *U : S->second->users()) {
2009 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(U)) {
2010 // Shuffle can be folded if it has no other input
2011 if (isa<UndefValue>(SI->getOperand(1)))
2014 if (isa<ExtractElementInst>(U))
2016 if (PrunedDAGInstrs.count(U))
2018 NeedsExtraction = true;
2022 if (NeedsExtraction) {
2024 if (Ty2->isVectorTy()) {
2025 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2027 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
2028 TargetTransformInfo::SK_ExtractSubvector, VTy,
2029 Ty1->isVectorTy() ? Ty1->getVectorNumElements() : 1, Ty2));
2031 ESContrib = (int) TTI->getVectorInstrCost(
2032 Instruction::ExtractElement, VTy, 1);
2033 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
2034 *S->second << "} = " << ESContrib << "\n");
2035 EffSize -= ESContrib;
2039 // Compute the cost of incoming edges.
2040 if (!isa<LoadInst>(S->first) && !isa<StoreInst>(S->first)) {
2041 Instruction *S1 = cast<Instruction>(S->first),
2042 *S2 = cast<Instruction>(S->second);
2043 for (unsigned o = 0; o < S1->getNumOperands(); ++o) {
2044 Value *O1 = S1->getOperand(o), *O2 = S2->getOperand(o);
2046 // Combining constants into vector constants (or small vector
2047 // constants into larger ones are assumed free).
2048 if (isa<Constant>(O1) && isa<Constant>(O2))
2054 ValuePair VP = ValuePair(O1, O2);
2055 ValuePair VPR = ValuePair(O2, O1);
2057 // Internal edges are not handled here.
2058 if (PrunedDAG.count(VP) || PrunedDAG.count(VPR))
2061 Type *Ty1 = O1->getType(),
2062 *Ty2 = O2->getType();
2063 Type *VTy = getVecTypeForPair(Ty1, Ty2);
2065 // Combining vector operations of the same type is also assumed
2066 // folded with other operations.
2068 // If both are insert elements, then both can be widened.
2069 InsertElementInst *IEO1 = dyn_cast<InsertElementInst>(O1),
2070 *IEO2 = dyn_cast<InsertElementInst>(O2);
2071 if (IEO1 && IEO2 && isPureIEChain(IEO1) && isPureIEChain(IEO2))
2073 // If both are extract elements, and both have the same input
2074 // type, then they can be replaced with a shuffle
2075 ExtractElementInst *EIO1 = dyn_cast<ExtractElementInst>(O1),
2076 *EIO2 = dyn_cast<ExtractElementInst>(O2);
2078 EIO1->getOperand(0)->getType() ==
2079 EIO2->getOperand(0)->getType())
2081 // If both are a shuffle with equal operand types and only two
2082 // unqiue operands, then they can be replaced with a single
2084 ShuffleVectorInst *SIO1 = dyn_cast<ShuffleVectorInst>(O1),
2085 *SIO2 = dyn_cast<ShuffleVectorInst>(O2);
2087 SIO1->getOperand(0)->getType() ==
2088 SIO2->getOperand(0)->getType()) {
2089 SmallSet<Value *, 4> SIOps;
2090 SIOps.insert(SIO1->getOperand(0));
2091 SIOps.insert(SIO1->getOperand(1));
2092 SIOps.insert(SIO2->getOperand(0));
2093 SIOps.insert(SIO2->getOperand(1));
2094 if (SIOps.size() <= 2)
2100 // This pair has already been formed.
2101 if (IncomingPairs.count(VP)) {
2103 } else if (IncomingPairs.count(VPR)) {
2104 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2107 if (VTy->getVectorNumElements() == 2)
2108 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
2109 TargetTransformInfo::SK_Reverse, VTy));
2110 } else if (!Ty1->isVectorTy() && !Ty2->isVectorTy()) {
2111 ESContrib = (int) TTI->getVectorInstrCost(
2112 Instruction::InsertElement, VTy, 0);
2113 ESContrib += (int) TTI->getVectorInstrCost(
2114 Instruction::InsertElement, VTy, 1);
2115 } else if (!Ty1->isVectorTy()) {
2116 // O1 needs to be inserted into a vector of size O2, and then
2117 // both need to be shuffled together.
2118 ESContrib = (int) TTI->getVectorInstrCost(
2119 Instruction::InsertElement, Ty2, 0);
2120 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2122 } else if (!Ty2->isVectorTy()) {
2123 // O2 needs to be inserted into a vector of size O1, and then
2124 // both need to be shuffled together.
2125 ESContrib = (int) TTI->getVectorInstrCost(
2126 Instruction::InsertElement, Ty1, 0);
2127 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2130 Type *TyBig = Ty1, *TySmall = Ty2;
2131 if (Ty2->getVectorNumElements() > Ty1->getVectorNumElements())
2132 std::swap(TyBig, TySmall);
2134 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2136 if (TyBig != TySmall)
2137 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2141 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {"
2142 << *O1 << " <-> " << *O2 << "} = " <<
2144 EffSize -= ESContrib;
2145 IncomingPairs.insert(VP);
2150 if (!HasNontrivialInsts) {
2151 DEBUG(if (DebugPairSelection) dbgs() <<
2152 "\tNo non-trivial instructions in DAG;"
2153 " override to zero effective size\n");
2157 for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(),
2158 E = PrunedDAG.end(); S != E; ++S)
2159 EffSize += (int) getDepthFactor(S->first);
2162 DEBUG(if (DebugPairSelection)
2163 dbgs() << "BBV: found pruned DAG for pair {"
2164 << *IJ.first << " <-> " << *IJ.second << "} of depth " <<
2165 MaxDepth << " and size " << PrunedDAG.size() <<
2166 " (effective size: " << EffSize << ")\n");
2167 if (((TTI && !UseChainDepthWithTI) ||
2168 MaxDepth >= Config.ReqChainDepth) &&
2169 EffSize > 0 && EffSize > BestEffSize) {
2170 BestMaxDepth = MaxDepth;
2171 BestEffSize = EffSize;
2172 BestDAG = PrunedDAG;
2177 // Given the list of candidate pairs, this function selects those
2178 // that will be fused into vector instructions.
2179 void BBVectorize::choosePairs(
2180 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
2181 DenseSet<ValuePair> &CandidatePairsSet,
2182 DenseMap<ValuePair, int> &CandidatePairCostSavings,
2183 std::vector<Value *> &PairableInsts,
2184 DenseSet<ValuePair> &FixedOrderPairs,
2185 DenseMap<VPPair, unsigned> &PairConnectionTypes,
2186 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
2187 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
2188 DenseSet<ValuePair> &PairableInstUsers,
2189 DenseMap<Value *, Value *>& ChosenPairs) {
2190 bool UseCycleCheck =
2191 CandidatePairsSet.size() <= Config.MaxCandPairsForCycleCheck;
2193 DenseMap<Value *, std::vector<Value *> > CandidatePairs2;
2194 for (DenseSet<ValuePair>::iterator I = CandidatePairsSet.begin(),
2195 E = CandidatePairsSet.end(); I != E; ++I) {
2196 std::vector<Value *> &JJ = CandidatePairs2[I->second];
2197 if (JJ.empty()) JJ.reserve(32);
2198 JJ.push_back(I->first);
2201 DenseMap<ValuePair, std::vector<ValuePair> > PairableInstUserMap;
2202 DenseSet<VPPair> PairableInstUserPairSet;
2203 for (std::vector<Value *>::iterator I = PairableInsts.begin(),
2204 E = PairableInsts.end(); I != E; ++I) {
2205 // The number of possible pairings for this variable:
2206 size_t NumChoices = CandidatePairs.lookup(*I).size();
2207 if (!NumChoices) continue;
2209 std::vector<Value *> &JJ = CandidatePairs[*I];
2211 // The best pair to choose and its dag:
2212 size_t BestMaxDepth = 0;
2213 int BestEffSize = 0;
2214 DenseSet<ValuePair> BestDAG;
2215 findBestDAGFor(CandidatePairs, CandidatePairsSet,
2216 CandidatePairCostSavings,
2217 PairableInsts, FixedOrderPairs, PairConnectionTypes,
2218 ConnectedPairs, ConnectedPairDeps,
2219 PairableInstUsers, PairableInstUserMap,
2220 PairableInstUserPairSet, ChosenPairs,
2221 BestDAG, BestMaxDepth, BestEffSize, *I, JJ,
2224 if (BestDAG.empty())
2227 // A dag has been chosen (or not) at this point. If no dag was
2228 // chosen, then this instruction, I, cannot be paired (and is no longer
2231 DEBUG(dbgs() << "BBV: selected pairs in the best DAG for: "
2232 << *cast<Instruction>(*I) << "\n");
2234 for (DenseSet<ValuePair>::iterator S = BestDAG.begin(),
2235 SE2 = BestDAG.end(); S != SE2; ++S) {
2236 // Insert the members of this dag into the list of chosen pairs.
2237 ChosenPairs.insert(ValuePair(S->first, S->second));
2238 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " <<
2239 *S->second << "\n");
2241 // Remove all candidate pairs that have values in the chosen dag.
2242 std::vector<Value *> &KK = CandidatePairs[S->first];
2243 for (std::vector<Value *>::iterator K = KK.begin(), KE = KK.end();
2245 if (*K == S->second)
2248 CandidatePairsSet.erase(ValuePair(S->first, *K));
2251 std::vector<Value *> &LL = CandidatePairs2[S->second];
2252 for (std::vector<Value *>::iterator L = LL.begin(), LE = LL.end();
2257 CandidatePairsSet.erase(ValuePair(*L, S->second));
2260 std::vector<Value *> &MM = CandidatePairs[S->second];
2261 for (std::vector<Value *>::iterator M = MM.begin(), ME = MM.end();
2263 assert(*M != S->first && "Flipped pair in candidate list?");
2264 CandidatePairsSet.erase(ValuePair(S->second, *M));
2267 std::vector<Value *> &NN = CandidatePairs2[S->first];
2268 for (std::vector<Value *>::iterator N = NN.begin(), NE = NN.end();
2270 assert(*N != S->second && "Flipped pair in candidate list?");
2271 CandidatePairsSet.erase(ValuePair(*N, S->first));
2276 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n");
2279 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o,
2284 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) +
2285 (n > 0 ? "." + utostr(n) : "")).str();
2288 // Returns the value that is to be used as the pointer input to the vector
2289 // instruction that fuses I with J.
2290 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context,
2291 Instruction *I, Instruction *J, unsigned o) {
2293 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
2294 int64_t OffsetInElmts;
2296 // Note: the analysis might fail here, that is why the pair order has
2297 // been precomputed (OffsetInElmts must be unused here).
2298 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
2299 IAddressSpace, JAddressSpace,
2300 OffsetInElmts, false);
2302 // The pointer value is taken to be the one with the lowest offset.
2305 Type *ArgTypeI = IPtr->getType()->getPointerElementType();
2306 Type *ArgTypeJ = JPtr->getType()->getPointerElementType();
2307 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2309 = PointerType::get(VArgType,
2310 IPtr->getType()->getPointerAddressSpace());
2311 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o),
2312 /* insert before */ I);
2315 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J,
2316 unsigned MaskOffset, unsigned NumInElem,
2317 unsigned NumInElem1, unsigned IdxOffset,
2318 std::vector<Constant*> &Mask) {
2319 unsigned NumElem1 = J->getType()->getVectorNumElements();
2320 for (unsigned v = 0; v < NumElem1; ++v) {
2321 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v);
2323 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context));
2325 unsigned mm = m + (int) IdxOffset;
2326 if (m >= (int) NumInElem1)
2327 mm += (int) NumInElem;
2329 Mask[v+MaskOffset] =
2330 ConstantInt::get(Type::getInt32Ty(Context), mm);
2335 // Returns the value that is to be used as the vector-shuffle mask to the
2336 // vector instruction that fuses I with J.
2337 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context,
2338 Instruction *I, Instruction *J) {
2339 // This is the shuffle mask. We need to append the second
2340 // mask to the first, and the numbers need to be adjusted.
2342 Type *ArgTypeI = I->getType();
2343 Type *ArgTypeJ = J->getType();
2344 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2346 unsigned NumElemI = ArgTypeI->getVectorNumElements();
2348 // Get the total number of elements in the fused vector type.
2349 // By definition, this must equal the number of elements in
2351 unsigned NumElem = VArgType->getVectorNumElements();
2352 std::vector<Constant*> Mask(NumElem);
2354 Type *OpTypeI = I->getOperand(0)->getType();
2355 unsigned NumInElemI = OpTypeI->getVectorNumElements();
2356 Type *OpTypeJ = J->getOperand(0)->getType();
2357 unsigned NumInElemJ = OpTypeJ->getVectorNumElements();
2359 // The fused vector will be:
2360 // -----------------------------------------------------
2361 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ |
2362 // -----------------------------------------------------
2363 // from which we'll extract NumElem total elements (where the first NumElemI
2364 // of them come from the mask in I and the remainder come from the mask
2367 // For the mask from the first pair...
2368 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI,
2371 // For the mask from the second pair...
2372 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ,
2375 return ConstantVector::get(Mask);
2378 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I,
2379 Instruction *J, unsigned o, Value *&LOp,
2381 Type *ArgTypeL, Type *ArgTypeH,
2382 bool IBeforeJ, unsigned IdxOff) {
2383 bool ExpandedIEChain = false;
2384 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) {
2385 // If we have a pure insertelement chain, then this can be rewritten
2386 // into a chain that directly builds the larger type.
2387 if (isPureIEChain(LIE)) {
2388 SmallVector<Value *, 8> VectElemts(numElemL,
2389 UndefValue::get(ArgTypeL->getScalarType()));
2390 InsertElementInst *LIENext = LIE;
2393 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue();
2394 VectElemts[Idx] = LIENext->getOperand(1);
2396 dyn_cast<InsertElementInst>(LIENext->getOperand(0))));
2399 Value *LIEPrev = UndefValue::get(ArgTypeH);
2400 for (unsigned i = 0; i < numElemL; ++i) {
2401 if (isa<UndefValue>(VectElemts[i])) continue;
2402 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i],
2403 ConstantInt::get(Type::getInt32Ty(Context),
2405 getReplacementName(IBeforeJ ? I : J,
2407 LIENext->insertBefore(IBeforeJ ? J : I);
2411 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH);
2412 ExpandedIEChain = true;
2416 return ExpandedIEChain;
2419 static unsigned getNumScalarElements(Type *Ty) {
2420 if (VectorType *VecTy = dyn_cast<VectorType>(Ty))
2421 return VecTy->getNumElements();
2425 // Returns the value to be used as the specified operand of the vector
2426 // instruction that fuses I with J.
2427 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I,
2428 Instruction *J, unsigned o, bool IBeforeJ) {
2429 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2430 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1);
2432 // Compute the fused vector type for this operand
2433 Type *ArgTypeI = I->getOperand(o)->getType();
2434 Type *ArgTypeJ = J->getOperand(o)->getType();
2435 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2437 Instruction *L = I, *H = J;
2438 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ;
2440 unsigned numElemL = getNumScalarElements(ArgTypeL);
2441 unsigned numElemH = getNumScalarElements(ArgTypeH);
2443 Value *LOp = L->getOperand(o);
2444 Value *HOp = H->getOperand(o);
2445 unsigned numElem = VArgType->getNumElements();
2447 // First, we check if we can reuse the "original" vector outputs (if these
2448 // exist). We might need a shuffle.
2449 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp);
2450 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp);
2451 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp);
2452 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp);
2454 // FIXME: If we're fusing shuffle instructions, then we can't apply this
2455 // optimization. The input vectors to the shuffle might be a different
2456 // length from the shuffle outputs. Unfortunately, the replacement
2457 // shuffle mask has already been formed, and the mask entries are sensitive
2458 // to the sizes of the inputs.
2459 bool IsSizeChangeShuffle =
2460 isa<ShuffleVectorInst>(L) &&
2461 (LOp->getType() != L->getType() || HOp->getType() != H->getType());
2463 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) {
2464 // We can have at most two unique vector inputs.
2465 bool CanUseInputs = true;
2466 Value *I1, *I2 = nullptr;
2468 I1 = LEE->getOperand(0);
2470 I1 = LSV->getOperand(0);
2471 I2 = LSV->getOperand(1);
2472 if (I2 == I1 || isa<UndefValue>(I2))
2477 Value *I3 = HEE->getOperand(0);
2478 if (!I2 && I3 != I1)
2480 else if (I3 != I1 && I3 != I2)
2481 CanUseInputs = false;
2483 Value *I3 = HSV->getOperand(0);
2484 if (!I2 && I3 != I1)
2486 else if (I3 != I1 && I3 != I2)
2487 CanUseInputs = false;
2490 Value *I4 = HSV->getOperand(1);
2491 if (!isa<UndefValue>(I4)) {
2492 if (!I2 && I4 != I1)
2494 else if (I4 != I1 && I4 != I2)
2495 CanUseInputs = false;
2502 cast<Instruction>(LOp)->getOperand(0)->getType()
2503 ->getVectorNumElements();
2506 cast<Instruction>(HOp)->getOperand(0)->getType()
2507 ->getVectorNumElements();
2509 // We have one or two input vectors. We need to map each index of the
2510 // operands to the index of the original vector.
2511 SmallVector<std::pair<int, int>, 8> II(numElem);
2512 for (unsigned i = 0; i < numElemL; ++i) {
2516 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue();
2517 INum = LEE->getOperand(0) == I1 ? 0 : 1;
2519 Idx = LSV->getMaskValue(i);
2520 if (Idx < (int) LOpElem) {
2521 INum = LSV->getOperand(0) == I1 ? 0 : 1;
2524 INum = LSV->getOperand(1) == I1 ? 0 : 1;
2528 II[i] = std::pair<int, int>(Idx, INum);
2530 for (unsigned i = 0; i < numElemH; ++i) {
2534 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue();
2535 INum = HEE->getOperand(0) == I1 ? 0 : 1;
2537 Idx = HSV->getMaskValue(i);
2538 if (Idx < (int) HOpElem) {
2539 INum = HSV->getOperand(0) == I1 ? 0 : 1;
2542 INum = HSV->getOperand(1) == I1 ? 0 : 1;
2546 II[i + numElemL] = std::pair<int, int>(Idx, INum);
2549 // We now have an array which tells us from which index of which
2550 // input vector each element of the operand comes.
2551 VectorType *I1T = cast<VectorType>(I1->getType());
2552 unsigned I1Elem = I1T->getNumElements();
2555 // In this case there is only one underlying vector input. Check for
2556 // the trivial case where we can use the input directly.
2557 if (I1Elem == numElem) {
2558 bool ElemInOrder = true;
2559 for (unsigned i = 0; i < numElem; ++i) {
2560 if (II[i].first != (int) i && II[i].first != -1) {
2561 ElemInOrder = false;
2570 // A shuffle is needed.
2571 std::vector<Constant *> Mask(numElem);
2572 for (unsigned i = 0; i < numElem; ++i) {
2573 int Idx = II[i].first;
2575 Mask[i] = UndefValue::get(Type::getInt32Ty(Context));
2577 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2581 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2582 ConstantVector::get(Mask),
2583 getReplacementName(IBeforeJ ? I : J,
2585 S->insertBefore(IBeforeJ ? J : I);
2589 VectorType *I2T = cast<VectorType>(I2->getType());
2590 unsigned I2Elem = I2T->getNumElements();
2592 // This input comes from two distinct vectors. The first step is to
2593 // make sure that both vectors are the same length. If not, the
2594 // smaller one will need to grow before they can be shuffled together.
2595 if (I1Elem < I2Elem) {
2596 std::vector<Constant *> Mask(I2Elem);
2598 for (; v < I1Elem; ++v)
2599 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2600 for (; v < I2Elem; ++v)
2601 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2603 Instruction *NewI1 =
2604 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2605 ConstantVector::get(Mask),
2606 getReplacementName(IBeforeJ ? I : J,
2608 NewI1->insertBefore(IBeforeJ ? J : I);
2612 } else if (I1Elem > I2Elem) {
2613 std::vector<Constant *> Mask(I1Elem);
2615 for (; v < I2Elem; ++v)
2616 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2617 for (; v < I1Elem; ++v)
2618 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2620 Instruction *NewI2 =
2621 new ShuffleVectorInst(I2, UndefValue::get(I2T),
2622 ConstantVector::get(Mask),
2623 getReplacementName(IBeforeJ ? I : J,
2625 NewI2->insertBefore(IBeforeJ ? J : I);
2631 // Now that both I1 and I2 are the same length we can shuffle them
2632 // together (and use the result).
2633 std::vector<Constant *> Mask(numElem);
2634 for (unsigned v = 0; v < numElem; ++v) {
2635 if (II[v].first == -1) {
2636 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2638 int Idx = II[v].first + II[v].second * I1Elem;
2639 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2643 Instruction *NewOp =
2644 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask),
2645 getReplacementName(IBeforeJ ? I : J, true, o));
2646 NewOp->insertBefore(IBeforeJ ? J : I);
2651 Type *ArgType = ArgTypeL;
2652 if (numElemL < numElemH) {
2653 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH,
2654 ArgTypeL, VArgType, IBeforeJ, 1)) {
2655 // This is another short-circuit case: we're combining a scalar into
2656 // a vector that is formed by an IE chain. We've just expanded the IE
2657 // chain, now insert the scalar and we're done.
2659 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0,
2660 getReplacementName(IBeforeJ ? I : J, true, o));
2661 S->insertBefore(IBeforeJ ? J : I);
2663 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL,
2664 ArgTypeH, IBeforeJ)) {
2665 // The two vector inputs to the shuffle must be the same length,
2666 // so extend the smaller vector to be the same length as the larger one.
2670 std::vector<Constant *> Mask(numElemH);
2672 for (; v < numElemL; ++v)
2673 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2674 for (; v < numElemH; ++v)
2675 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2677 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL),
2678 ConstantVector::get(Mask),
2679 getReplacementName(IBeforeJ ? I : J,
2682 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0,
2683 getReplacementName(IBeforeJ ? I : J,
2687 NLOp->insertBefore(IBeforeJ ? J : I);
2692 } else if (numElemL > numElemH) {
2693 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL,
2694 ArgTypeH, VArgType, IBeforeJ)) {
2696 InsertElementInst::Create(LOp, HOp,
2697 ConstantInt::get(Type::getInt32Ty(Context),
2699 getReplacementName(IBeforeJ ? I : J,
2701 S->insertBefore(IBeforeJ ? J : I);
2703 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH,
2704 ArgTypeL, IBeforeJ)) {
2707 std::vector<Constant *> Mask(numElemL);
2709 for (; v < numElemH; ++v)
2710 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2711 for (; v < numElemL; ++v)
2712 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2714 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH),
2715 ConstantVector::get(Mask),
2716 getReplacementName(IBeforeJ ? I : J,
2719 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0,
2720 getReplacementName(IBeforeJ ? I : J,
2724 NHOp->insertBefore(IBeforeJ ? J : I);
2729 if (ArgType->isVectorTy()) {
2730 unsigned numElem = VArgType->getVectorNumElements();
2731 std::vector<Constant*> Mask(numElem);
2732 for (unsigned v = 0; v < numElem; ++v) {
2734 // If the low vector was expanded, we need to skip the extra
2735 // undefined entries.
2736 if (v >= numElemL && numElemH > numElemL)
2737 Idx += (numElemH - numElemL);
2738 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2741 Instruction *BV = new ShuffleVectorInst(LOp, HOp,
2742 ConstantVector::get(Mask),
2743 getReplacementName(IBeforeJ ? I : J, true, o));
2744 BV->insertBefore(IBeforeJ ? J : I);
2748 Instruction *BV1 = InsertElementInst::Create(
2749 UndefValue::get(VArgType), LOp, CV0,
2750 getReplacementName(IBeforeJ ? I : J,
2752 BV1->insertBefore(IBeforeJ ? J : I);
2753 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1,
2754 getReplacementName(IBeforeJ ? I : J,
2756 BV2->insertBefore(IBeforeJ ? J : I);
2760 // This function creates an array of values that will be used as the inputs
2761 // to the vector instruction that fuses I with J.
2762 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context,
2763 Instruction *I, Instruction *J,
2764 SmallVectorImpl<Value *> &ReplacedOperands,
2766 unsigned NumOperands = I->getNumOperands();
2768 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) {
2769 // Iterate backward so that we look at the store pointer
2770 // first and know whether or not we need to flip the inputs.
2772 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) {
2773 // This is the pointer for a load/store instruction.
2774 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o);
2776 } else if (isa<CallInst>(I)) {
2777 Function *F = cast<CallInst>(I)->getCalledFunction();
2778 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID();
2779 if (o == NumOperands-1) {
2780 BasicBlock &BB = *I->getParent();
2782 Module *M = BB.getParent()->getParent();
2783 Type *ArgTypeI = I->getType();
2784 Type *ArgTypeJ = J->getType();
2785 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2787 ReplacedOperands[o] = Intrinsic::getDeclaration(M, IID, VArgType);
2789 } else if ((IID == Intrinsic::powi || IID == Intrinsic::ctlz ||
2790 IID == Intrinsic::cttz) && o == 1) {
2791 // The second argument of powi/ctlz/cttz is a single integer/constant
2792 // and we've already checked that both arguments are equal.
2793 // As a result, we just keep I's second argument.
2794 ReplacedOperands[o] = I->getOperand(o);
2797 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) {
2798 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J);
2802 ReplacedOperands[o] = getReplacementInput(Context, I, J, o, IBeforeJ);
2806 // This function creates two values that represent the outputs of the
2807 // original I and J instructions. These are generally vector shuffles
2808 // or extracts. In many cases, these will end up being unused and, thus,
2809 // eliminated by later passes.
2810 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
2811 Instruction *J, Instruction *K,
2812 Instruction *&InsertionPt,
2813 Instruction *&K1, Instruction *&K2) {
2814 if (isa<StoreInst>(I)) {
2815 AA->replaceWithNewValue(I, K);
2816 AA->replaceWithNewValue(J, K);
2818 Type *IType = I->getType();
2819 Type *JType = J->getType();
2821 VectorType *VType = getVecTypeForPair(IType, JType);
2822 unsigned numElem = VType->getNumElements();
2824 unsigned numElemI = getNumScalarElements(IType);
2825 unsigned numElemJ = getNumScalarElements(JType);
2827 if (IType->isVectorTy()) {
2828 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI);
2829 for (unsigned v = 0; v < numElemI; ++v) {
2830 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2831 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v);
2834 K1 = new ShuffleVectorInst(K, UndefValue::get(VType),
2835 ConstantVector::get( Mask1),
2836 getReplacementName(K, false, 1));
2838 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2839 K1 = ExtractElementInst::Create(K, CV0,
2840 getReplacementName(K, false, 1));
2843 if (JType->isVectorTy()) {
2844 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ);
2845 for (unsigned v = 0; v < numElemJ; ++v) {
2846 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2847 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v);
2850 K2 = new ShuffleVectorInst(K, UndefValue::get(VType),
2851 ConstantVector::get( Mask2),
2852 getReplacementName(K, false, 2));
2854 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1);
2855 K2 = ExtractElementInst::Create(K, CV1,
2856 getReplacementName(K, false, 2));
2860 K2->insertAfter(K1);
2865 // Move all uses of the function I (including pairing-induced uses) after J.
2866 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB,
2867 DenseSet<ValuePair> &LoadMoveSetPairs,
2868 Instruction *I, Instruction *J) {
2869 // Skip to the first instruction past I.
2870 BasicBlock::iterator L = std::next(BasicBlock::iterator(I));
2872 DenseSet<Value *> Users;
2873 AliasSetTracker WriteSet(*AA);
2874 if (I->mayWriteToMemory()) WriteSet.add(I);
2876 for (; cast<Instruction>(L) != J; ++L)
2877 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs);
2879 assert(cast<Instruction>(L) == J &&
2880 "Tracking has not proceeded far enough to check for dependencies");
2881 // If J is now in the use set of I, then trackUsesOfI will return true
2882 // and we have a dependency cycle (and the fusing operation must abort).
2883 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSetPairs);
2886 // Move all uses of the function I (including pairing-induced uses) after J.
2887 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB,
2888 DenseSet<ValuePair> &LoadMoveSetPairs,
2889 Instruction *&InsertionPt,
2890 Instruction *I, Instruction *J) {
2891 // Skip to the first instruction past I.
2892 BasicBlock::iterator L = std::next(BasicBlock::iterator(I));
2894 DenseSet<Value *> Users;
2895 AliasSetTracker WriteSet(*AA);
2896 if (I->mayWriteToMemory()) WriteSet.add(I);
2898 for (; cast<Instruction>(L) != J;) {
2899 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs)) {
2900 // Move this instruction
2901 Instruction *InstToMove = L; ++L;
2903 DEBUG(dbgs() << "BBV: moving: " << *InstToMove <<
2904 " to after " << *InsertionPt << "\n");
2905 InstToMove->removeFromParent();
2906 InstToMove->insertAfter(InsertionPt);
2907 InsertionPt = InstToMove;
2914 // Collect all load instruction that are in the move set of a given first
2915 // pair member. These loads depend on the first instruction, I, and so need
2916 // to be moved after J (the second instruction) when the pair is fused.
2917 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB,
2918 DenseMap<Value *, Value *> &ChosenPairs,
2919 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
2920 DenseSet<ValuePair> &LoadMoveSetPairs,
2922 // Skip to the first instruction past I.
2923 BasicBlock::iterator L = std::next(BasicBlock::iterator(I));
2925 DenseSet<Value *> Users;
2926 AliasSetTracker WriteSet(*AA);
2927 if (I->mayWriteToMemory()) WriteSet.add(I);
2929 // Note: We cannot end the loop when we reach J because J could be moved
2930 // farther down the use chain by another instruction pairing. Also, J
2931 // could be before I if this is an inverted input.
2932 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) {
2933 if (trackUsesOfI(Users, WriteSet, I, L)) {
2934 if (L->mayReadFromMemory()) {
2935 LoadMoveSet[L].push_back(I);
2936 LoadMoveSetPairs.insert(ValuePair(L, I));
2942 // In cases where both load/stores and the computation of their pointers
2943 // are chosen for vectorization, we can end up in a situation where the
2944 // aliasing analysis starts returning different query results as the
2945 // process of fusing instruction pairs continues. Because the algorithm
2946 // relies on finding the same use dags here as were found earlier, we'll
2947 // need to precompute the necessary aliasing information here and then
2948 // manually update it during the fusion process.
2949 void BBVectorize::collectLoadMoveSet(BasicBlock &BB,
2950 std::vector<Value *> &PairableInsts,
2951 DenseMap<Value *, Value *> &ChosenPairs,
2952 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
2953 DenseSet<ValuePair> &LoadMoveSetPairs) {
2954 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
2955 PIE = PairableInsts.end(); PI != PIE; ++PI) {
2956 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI);
2957 if (P == ChosenPairs.end()) continue;
2959 Instruction *I = cast<Instruction>(P->first);
2960 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet,
2961 LoadMoveSetPairs, I);
2965 // This function fuses the chosen instruction pairs into vector instructions,
2966 // taking care preserve any needed scalar outputs and, then, it reorders the
2967 // remaining instructions as needed (users of the first member of the pair
2968 // need to be moved to after the location of the second member of the pair
2969 // because the vector instruction is inserted in the location of the pair's
2971 void BBVectorize::fuseChosenPairs(BasicBlock &BB,
2972 std::vector<Value *> &PairableInsts,
2973 DenseMap<Value *, Value *> &ChosenPairs,
2974 DenseSet<ValuePair> &FixedOrderPairs,
2975 DenseMap<VPPair, unsigned> &PairConnectionTypes,
2976 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
2977 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps) {
2978 LLVMContext& Context = BB.getContext();
2980 // During the vectorization process, the order of the pairs to be fused
2981 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs
2982 // list. After a pair is fused, the flipped pair is removed from the list.
2983 DenseSet<ValuePair> FlippedPairs;
2984 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(),
2985 E = ChosenPairs.end(); P != E; ++P)
2986 FlippedPairs.insert(ValuePair(P->second, P->first));
2987 for (DenseSet<ValuePair>::iterator P = FlippedPairs.begin(),
2988 E = FlippedPairs.end(); P != E; ++P)
2989 ChosenPairs.insert(*P);
2991 DenseMap<Value *, std::vector<Value *> > LoadMoveSet;
2992 DenseSet<ValuePair> LoadMoveSetPairs;
2993 collectLoadMoveSet(BB, PairableInsts, ChosenPairs,
2994 LoadMoveSet, LoadMoveSetPairs);
2996 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n");
2998 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) {
2999 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI);
3000 if (P == ChosenPairs.end()) {
3005 if (getDepthFactor(P->first) == 0) {
3006 // These instructions are not really fused, but are tracked as though
3007 // they are. Any case in which it would be interesting to fuse them
3008 // will be taken care of by InstCombine.
3014 Instruction *I = cast<Instruction>(P->first),
3015 *J = cast<Instruction>(P->second);
3017 DEBUG(dbgs() << "BBV: fusing: " << *I <<
3018 " <-> " << *J << "\n");
3020 // Remove the pair and flipped pair from the list.
3021 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second);
3022 assert(FP != ChosenPairs.end() && "Flipped pair not found in list");
3023 ChosenPairs.erase(FP);
3024 ChosenPairs.erase(P);
3026 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSetPairs, I, J)) {
3027 DEBUG(dbgs() << "BBV: fusion of: " << *I <<
3029 " aborted because of non-trivial dependency cycle\n");
3035 // If the pair must have the other order, then flip it.
3036 bool FlipPairOrder = FixedOrderPairs.count(ValuePair(J, I));
3037 if (!FlipPairOrder && !FixedOrderPairs.count(ValuePair(I, J))) {
3038 // This pair does not have a fixed order, and so we might want to
3039 // flip it if that will yield fewer shuffles. We count the number
3040 // of dependencies connected via swaps, and those directly connected,
3041 // and flip the order if the number of swaps is greater.
3042 bool OrigOrder = true;
3043 DenseMap<ValuePair, std::vector<ValuePair> >::iterator IJ =
3044 ConnectedPairDeps.find(ValuePair(I, J));
3045 if (IJ == ConnectedPairDeps.end()) {
3046 IJ = ConnectedPairDeps.find(ValuePair(J, I));
3050 if (IJ != ConnectedPairDeps.end()) {
3051 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
3052 for (std::vector<ValuePair>::iterator T = IJ->second.begin(),
3053 TE = IJ->second.end(); T != TE; ++T) {
3054 VPPair Q(IJ->first, *T);
3055 DenseMap<VPPair, unsigned>::iterator R =
3056 PairConnectionTypes.find(VPPair(Q.second, Q.first));
3057 assert(R != PairConnectionTypes.end() &&
3058 "Cannot find pair connection type");
3059 if (R->second == PairConnectionDirect)
3061 else if (R->second == PairConnectionSwap)
3066 std::swap(NumDepsDirect, NumDepsSwap);
3068 if (NumDepsSwap > NumDepsDirect) {
3069 FlipPairOrder = true;
3070 DEBUG(dbgs() << "BBV: reordering pair: " << *I <<
3071 " <-> " << *J << "\n");
3076 Instruction *L = I, *H = J;
3080 // If the pair being fused uses the opposite order from that in the pair
3081 // connection map, then we need to flip the types.
3082 DenseMap<ValuePair, std::vector<ValuePair> >::iterator HL =
3083 ConnectedPairs.find(ValuePair(H, L));
3084 if (HL != ConnectedPairs.end())
3085 for (std::vector<ValuePair>::iterator T = HL->second.begin(),
3086 TE = HL->second.end(); T != TE; ++T) {
3087 VPPair Q(HL->first, *T);
3088 DenseMap<VPPair, unsigned>::iterator R = PairConnectionTypes.find(Q);
3089 assert(R != PairConnectionTypes.end() &&
3090 "Cannot find pair connection type");
3091 if (R->second == PairConnectionDirect)
3092 R->second = PairConnectionSwap;
3093 else if (R->second == PairConnectionSwap)
3094 R->second = PairConnectionDirect;
3097 bool LBeforeH = !FlipPairOrder;
3098 unsigned NumOperands = I->getNumOperands();
3099 SmallVector<Value *, 3> ReplacedOperands(NumOperands);
3100 getReplacementInputsForPair(Context, L, H, ReplacedOperands,
3103 // Make a copy of the original operation, change its type to the vector
3104 // type and replace its operands with the vector operands.
3105 Instruction *K = L->clone();
3108 else if (H->hasName())
3111 if (!isa<StoreInst>(K))
3112 K->mutateType(getVecTypeForPair(L->getType(), H->getType()));
3114 unsigned KnownIDs[] = {
3115 LLVMContext::MD_tbaa,
3116 LLVMContext::MD_alias_scope,
3117 LLVMContext::MD_noalias,
3118 LLVMContext::MD_fpmath
3120 combineMetadata(K, H, KnownIDs);
3121 K->intersectOptionalDataWith(H);
3123 for (unsigned o = 0; o < NumOperands; ++o)
3124 K->setOperand(o, ReplacedOperands[o]);
3128 // Instruction insertion point:
3129 Instruction *InsertionPt = K;
3130 Instruction *K1 = nullptr, *K2 = nullptr;
3131 replaceOutputsOfPair(Context, L, H, K, InsertionPt, K1, K2);
3133 // The use dag of the first original instruction must be moved to after
3134 // the location of the second instruction. The entire use dag of the
3135 // first instruction is disjoint from the input dag of the second
3136 // (by definition), and so commutes with it.
3138 moveUsesOfIAfterJ(BB, LoadMoveSetPairs, InsertionPt, I, J);
3140 if (!isa<StoreInst>(I)) {
3141 L->replaceAllUsesWith(K1);
3142 H->replaceAllUsesWith(K2);
3143 AA->replaceWithNewValue(L, K1);
3144 AA->replaceWithNewValue(H, K2);
3147 // Instructions that may read from memory may be in the load move set.
3148 // Once an instruction is fused, we no longer need its move set, and so
3149 // the values of the map never need to be updated. However, when a load
3150 // is fused, we need to merge the entries from both instructions in the
3151 // pair in case those instructions were in the move set of some other
3152 // yet-to-be-fused pair. The loads in question are the keys of the map.
3153 if (I->mayReadFromMemory()) {
3154 std::vector<ValuePair> NewSetMembers;
3155 DenseMap<Value *, std::vector<Value *> >::iterator II =
3156 LoadMoveSet.find(I);
3157 if (II != LoadMoveSet.end())
3158 for (std::vector<Value *>::iterator N = II->second.begin(),
3159 NE = II->second.end(); N != NE; ++N)
3160 NewSetMembers.push_back(ValuePair(K, *N));
3161 DenseMap<Value *, std::vector<Value *> >::iterator JJ =
3162 LoadMoveSet.find(J);
3163 if (JJ != LoadMoveSet.end())
3164 for (std::vector<Value *>::iterator N = JJ->second.begin(),
3165 NE = JJ->second.end(); N != NE; ++N)
3166 NewSetMembers.push_back(ValuePair(K, *N));
3167 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(),
3168 AE = NewSetMembers.end(); A != AE; ++A) {
3169 LoadMoveSet[A->first].push_back(A->second);
3170 LoadMoveSetPairs.insert(*A);
3174 // Before removing I, set the iterator to the next instruction.
3175 PI = std::next(BasicBlock::iterator(I));
3176 if (cast<Instruction>(PI) == J)
3181 I->eraseFromParent();
3182 J->eraseFromParent();
3184 DEBUG(if (PrintAfterEveryPair) dbgs() << "BBV: block is now: \n" <<
3188 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n");
3192 char BBVectorize::ID = 0;
3193 static const char bb_vectorize_name[] = "Basic-Block Vectorization";
3194 INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
3195 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
3196 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
3197 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
3198 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3199 INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
3201 BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) {
3202 return new BBVectorize(C);
3206 llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) {
3207 BBVectorize BBVectorizer(P, C);
3208 return BBVectorizer.vectorizeBB(BB);
3211 //===----------------------------------------------------------------------===//
3212 VectorizeConfig::VectorizeConfig() {
3213 VectorBits = ::VectorBits;
3214 VectorizeBools = !::NoBools;
3215 VectorizeInts = !::NoInts;
3216 VectorizeFloats = !::NoFloats;
3217 VectorizePointers = !::NoPointers;
3218 VectorizeCasts = !::NoCasts;
3219 VectorizeMath = !::NoMath;
3220 VectorizeBitManipulations = !::NoBitManipulation;
3221 VectorizeFMA = !::NoFMA;
3222 VectorizeSelect = !::NoSelect;
3223 VectorizeCmp = !::NoCmp;
3224 VectorizeGEP = !::NoGEP;
3225 VectorizeMemOps = !::NoMemOps;
3226 AlignedOnly = ::AlignedOnly;
3227 ReqChainDepth= ::ReqChainDepth;
3228 SearchLimit = ::SearchLimit;
3229 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck;
3230 SplatBreaksChain = ::SplatBreaksChain;
3231 MaxInsts = ::MaxInsts;
3232 MaxPairs = ::MaxPairs;
3233 MaxIter = ::MaxIter;
3234 Pow2LenOnly = ::Pow2LenOnly;
3235 NoMemOpBoost = ::NoMemOpBoost;
3236 FastDep = ::FastDep;