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, Function &F, 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
213 : &P->getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
216 typedef std::pair<Value *, Value *> ValuePair;
217 typedef std::pair<ValuePair, int> ValuePairWithCost;
218 typedef std::pair<ValuePair, size_t> ValuePairWithDepth;
219 typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair
220 typedef std::pair<VPPair, unsigned> VPPairWithType;
225 const DataLayout *DL;
226 const TargetTransformInfo *TTI;
228 // FIXME: const correct?
230 bool vectorizePairs(BasicBlock &BB, bool NonPow2Len = false);
232 bool getCandidatePairs(BasicBlock &BB,
233 BasicBlock::iterator &Start,
234 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
235 DenseSet<ValuePair> &FixedOrderPairs,
236 DenseMap<ValuePair, int> &CandidatePairCostSavings,
237 std::vector<Value *> &PairableInsts, bool NonPow2Len);
239 // FIXME: The current implementation does not account for pairs that
240 // are connected in multiple ways. For example:
241 // C1 = A1 / A2; C2 = A2 / A1 (which may be both direct and a swap)
242 enum PairConnectionType {
243 PairConnectionDirect,
248 void computeConnectedPairs(
249 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
250 DenseSet<ValuePair> &CandidatePairsSet,
251 std::vector<Value *> &PairableInsts,
252 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
253 DenseMap<VPPair, unsigned> &PairConnectionTypes);
255 void buildDepMap(BasicBlock &BB,
256 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
257 std::vector<Value *> &PairableInsts,
258 DenseSet<ValuePair> &PairableInstUsers);
260 void choosePairs(DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
261 DenseSet<ValuePair> &CandidatePairsSet,
262 DenseMap<ValuePair, int> &CandidatePairCostSavings,
263 std::vector<Value *> &PairableInsts,
264 DenseSet<ValuePair> &FixedOrderPairs,
265 DenseMap<VPPair, unsigned> &PairConnectionTypes,
266 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
267 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
268 DenseSet<ValuePair> &PairableInstUsers,
269 DenseMap<Value *, Value *>& ChosenPairs);
271 void fuseChosenPairs(BasicBlock &BB,
272 std::vector<Value *> &PairableInsts,
273 DenseMap<Value *, Value *>& ChosenPairs,
274 DenseSet<ValuePair> &FixedOrderPairs,
275 DenseMap<VPPair, unsigned> &PairConnectionTypes,
276 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
277 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps);
280 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore);
282 bool areInstsCompatible(Instruction *I, Instruction *J,
283 bool IsSimpleLoadStore, bool NonPow2Len,
284 int &CostSavings, int &FixedOrder);
286 bool trackUsesOfI(DenseSet<Value *> &Users,
287 AliasSetTracker &WriteSet, Instruction *I,
288 Instruction *J, bool UpdateUsers = true,
289 DenseSet<ValuePair> *LoadMoveSetPairs = nullptr);
291 void computePairsConnectedTo(
292 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
293 DenseSet<ValuePair> &CandidatePairsSet,
294 std::vector<Value *> &PairableInsts,
295 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
296 DenseMap<VPPair, unsigned> &PairConnectionTypes,
299 bool pairsConflict(ValuePair P, ValuePair Q,
300 DenseSet<ValuePair> &PairableInstUsers,
301 DenseMap<ValuePair, std::vector<ValuePair> >
302 *PairableInstUserMap = nullptr,
303 DenseSet<VPPair> *PairableInstUserPairSet = nullptr);
305 bool pairWillFormCycle(ValuePair P,
306 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUsers,
307 DenseSet<ValuePair> &CurrentPairs);
310 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
311 std::vector<Value *> &PairableInsts,
312 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
313 DenseSet<ValuePair> &PairableInstUsers,
314 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
315 DenseSet<VPPair> &PairableInstUserPairSet,
316 DenseMap<Value *, Value *> &ChosenPairs,
317 DenseMap<ValuePair, size_t> &DAG,
318 DenseSet<ValuePair> &PrunedDAG, ValuePair J,
321 void buildInitialDAGFor(
322 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
323 DenseSet<ValuePair> &CandidatePairsSet,
324 std::vector<Value *> &PairableInsts,
325 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
326 DenseSet<ValuePair> &PairableInstUsers,
327 DenseMap<Value *, Value *> &ChosenPairs,
328 DenseMap<ValuePair, size_t> &DAG, ValuePair J);
331 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
332 DenseSet<ValuePair> &CandidatePairsSet,
333 DenseMap<ValuePair, int> &CandidatePairCostSavings,
334 std::vector<Value *> &PairableInsts,
335 DenseSet<ValuePair> &FixedOrderPairs,
336 DenseMap<VPPair, unsigned> &PairConnectionTypes,
337 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
338 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
339 DenseSet<ValuePair> &PairableInstUsers,
340 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
341 DenseSet<VPPair> &PairableInstUserPairSet,
342 DenseMap<Value *, Value *> &ChosenPairs,
343 DenseSet<ValuePair> &BestDAG, size_t &BestMaxDepth,
344 int &BestEffSize, Value *II, std::vector<Value *>&JJ,
347 Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I,
348 Instruction *J, unsigned o);
350 void fillNewShuffleMask(LLVMContext& Context, Instruction *J,
351 unsigned MaskOffset, unsigned NumInElem,
352 unsigned NumInElem1, unsigned IdxOffset,
353 std::vector<Constant*> &Mask);
355 Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I,
358 bool expandIEChain(LLVMContext& Context, Instruction *I, Instruction *J,
359 unsigned o, Value *&LOp, unsigned numElemL,
360 Type *ArgTypeL, Type *ArgTypeR, bool IBeforeJ,
361 unsigned IdxOff = 0);
363 Value *getReplacementInput(LLVMContext& Context, Instruction *I,
364 Instruction *J, unsigned o, bool IBeforeJ);
366 void getReplacementInputsForPair(LLVMContext& Context, Instruction *I,
367 Instruction *J, SmallVectorImpl<Value *> &ReplacedOperands,
370 void replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
371 Instruction *J, Instruction *K,
372 Instruction *&InsertionPt, Instruction *&K1,
375 void collectPairLoadMoveSet(BasicBlock &BB,
376 DenseMap<Value *, Value *> &ChosenPairs,
377 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
378 DenseSet<ValuePair> &LoadMoveSetPairs,
381 void collectLoadMoveSet(BasicBlock &BB,
382 std::vector<Value *> &PairableInsts,
383 DenseMap<Value *, Value *> &ChosenPairs,
384 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
385 DenseSet<ValuePair> &LoadMoveSetPairs);
387 bool canMoveUsesOfIAfterJ(BasicBlock &BB,
388 DenseSet<ValuePair> &LoadMoveSetPairs,
389 Instruction *I, Instruction *J);
391 void moveUsesOfIAfterJ(BasicBlock &BB,
392 DenseSet<ValuePair> &LoadMoveSetPairs,
393 Instruction *&InsertionPt,
394 Instruction *I, Instruction *J);
396 bool vectorizeBB(BasicBlock &BB) {
397 if (skipOptnoneFunction(BB))
399 if (!DT->isReachableFromEntry(&BB)) {
400 DEBUG(dbgs() << "BBV: skipping unreachable " << BB.getName() <<
401 " in " << BB.getParent()->getName() << "\n");
405 DEBUG(if (TTI) dbgs() << "BBV: using target information\n");
407 bool changed = false;
408 // Iterate a sufficient number of times to merge types of size 1 bit,
409 // then 2 bits, then 4, etc. up to half of the target vector width of the
410 // target vector register.
413 (TTI || v <= Config.VectorBits) &&
414 (!Config.MaxIter || n <= Config.MaxIter);
416 DEBUG(dbgs() << "BBV: fusing loop #" << n <<
417 " for " << BB.getName() << " in " <<
418 BB.getParent()->getName() << "...\n");
419 if (vectorizePairs(BB))
425 if (changed && !Pow2LenOnly) {
427 for (; !Config.MaxIter || n <= Config.MaxIter; ++n) {
428 DEBUG(dbgs() << "BBV: fusing for non-2^n-length vectors loop #: " <<
429 n << " for " << BB.getName() << " in " <<
430 BB.getParent()->getName() << "...\n");
431 if (!vectorizePairs(BB, true)) break;
435 DEBUG(dbgs() << "BBV: done!\n");
439 bool runOnBasicBlock(BasicBlock &BB) override {
440 // OptimizeNone check deferred to vectorizeBB().
442 AA = &getAnalysis<AliasAnalysis>();
443 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
444 SE = &getAnalysis<ScalarEvolution>();
445 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
446 DL = DLP ? &DLP->getDataLayout() : nullptr;
447 TTI = IgnoreTargetInfo
449 : &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
452 return vectorizeBB(BB);
455 void getAnalysisUsage(AnalysisUsage &AU) const override {
456 BasicBlockPass::getAnalysisUsage(AU);
457 AU.addRequired<AliasAnalysis>();
458 AU.addRequired<DominatorTreeWrapperPass>();
459 AU.addRequired<ScalarEvolution>();
460 AU.addRequired<TargetTransformInfoWrapperPass>();
461 AU.addPreserved<AliasAnalysis>();
462 AU.addPreserved<DominatorTreeWrapperPass>();
463 AU.addPreserved<ScalarEvolution>();
464 AU.setPreservesCFG();
467 static inline VectorType *getVecTypeForPair(Type *ElemTy, Type *Elem2Ty) {
468 assert(ElemTy->getScalarType() == Elem2Ty->getScalarType() &&
469 "Cannot form vector from incompatible scalar types");
470 Type *STy = ElemTy->getScalarType();
473 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) {
474 numElem = VTy->getNumElements();
479 if (VectorType *VTy = dyn_cast<VectorType>(Elem2Ty)) {
480 numElem += VTy->getNumElements();
485 return VectorType::get(STy, numElem);
488 static inline void getInstructionTypes(Instruction *I,
489 Type *&T1, Type *&T2) {
490 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
491 // For stores, it is the value type, not the pointer type that matters
492 // because the value is what will come from a vector register.
494 Value *IVal = SI->getValueOperand();
495 T1 = IVal->getType();
500 if (CastInst *CI = dyn_cast<CastInst>(I))
505 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
506 T2 = SI->getCondition()->getType();
507 } else if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(I)) {
508 T2 = SI->getOperand(0)->getType();
509 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) {
510 T2 = CI->getOperand(0)->getType();
514 // Returns the weight associated with the provided value. A chain of
515 // candidate pairs has a length given by the sum of the weights of its
516 // members (one weight per pair; the weight of each member of the pair
517 // is assumed to be the same). This length is then compared to the
518 // chain-length threshold to determine if a given chain is significant
519 // enough to be vectorized. The length is also used in comparing
520 // candidate chains where longer chains are considered to be better.
521 // Note: when this function returns 0, the resulting instructions are
522 // not actually fused.
523 inline size_t getDepthFactor(Value *V) {
524 // InsertElement and ExtractElement have a depth factor of zero. This is
525 // for two reasons: First, they cannot be usefully fused. Second, because
526 // the pass generates a lot of these, they can confuse the simple metric
527 // used to compare the dags in the next iteration. Thus, giving them a
528 // weight of zero allows the pass to essentially ignore them in
529 // subsequent iterations when looking for vectorization opportunities
530 // while still tracking dependency chains that flow through those
532 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V))
535 // Give a load or store half of the required depth so that load/store
536 // pairs will vectorize.
537 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V)))
538 return Config.ReqChainDepth/2;
543 // Returns the cost of the provided instruction using TTI.
544 // This does not handle loads and stores.
545 unsigned getInstrCost(unsigned Opcode, Type *T1, Type *T2,
546 TargetTransformInfo::OperandValueKind Op1VK =
547 TargetTransformInfo::OK_AnyValue,
548 TargetTransformInfo::OperandValueKind Op2VK =
549 TargetTransformInfo::OK_AnyValue) {
552 case Instruction::GetElementPtr:
553 // We mark this instruction as zero-cost because scalar GEPs are usually
554 // lowered to the instruction addressing mode. At the moment we don't
555 // generate vector GEPs.
557 case Instruction::Br:
558 return TTI->getCFInstrCost(Opcode);
559 case Instruction::PHI:
561 case Instruction::Add:
562 case Instruction::FAdd:
563 case Instruction::Sub:
564 case Instruction::FSub:
565 case Instruction::Mul:
566 case Instruction::FMul:
567 case Instruction::UDiv:
568 case Instruction::SDiv:
569 case Instruction::FDiv:
570 case Instruction::URem:
571 case Instruction::SRem:
572 case Instruction::FRem:
573 case Instruction::Shl:
574 case Instruction::LShr:
575 case Instruction::AShr:
576 case Instruction::And:
577 case Instruction::Or:
578 case Instruction::Xor:
579 return TTI->getArithmeticInstrCost(Opcode, T1, Op1VK, Op2VK);
580 case Instruction::Select:
581 case Instruction::ICmp:
582 case Instruction::FCmp:
583 return TTI->getCmpSelInstrCost(Opcode, T1, T2);
584 case Instruction::ZExt:
585 case Instruction::SExt:
586 case Instruction::FPToUI:
587 case Instruction::FPToSI:
588 case Instruction::FPExt:
589 case Instruction::PtrToInt:
590 case Instruction::IntToPtr:
591 case Instruction::SIToFP:
592 case Instruction::UIToFP:
593 case Instruction::Trunc:
594 case Instruction::FPTrunc:
595 case Instruction::BitCast:
596 case Instruction::ShuffleVector:
597 return TTI->getCastInstrCost(Opcode, T1, T2);
603 // This determines the relative offset of two loads or stores, returning
604 // true if the offset could be determined to be some constant value.
605 // For example, if OffsetInElmts == 1, then J accesses the memory directly
606 // after I; if OffsetInElmts == -1 then I accesses the memory
608 bool getPairPtrInfo(Instruction *I, Instruction *J,
609 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment,
610 unsigned &IAddressSpace, unsigned &JAddressSpace,
611 int64_t &OffsetInElmts, bool ComputeOffset = true) {
613 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
614 LoadInst *LJ = cast<LoadInst>(J);
615 IPtr = LI->getPointerOperand();
616 JPtr = LJ->getPointerOperand();
617 IAlignment = LI->getAlignment();
618 JAlignment = LJ->getAlignment();
619 IAddressSpace = LI->getPointerAddressSpace();
620 JAddressSpace = LJ->getPointerAddressSpace();
622 StoreInst *SI = cast<StoreInst>(I), *SJ = cast<StoreInst>(J);
623 IPtr = SI->getPointerOperand();
624 JPtr = SJ->getPointerOperand();
625 IAlignment = SI->getAlignment();
626 JAlignment = SJ->getAlignment();
627 IAddressSpace = SI->getPointerAddressSpace();
628 JAddressSpace = SJ->getPointerAddressSpace();
634 const SCEV *IPtrSCEV = SE->getSCEV(IPtr);
635 const SCEV *JPtrSCEV = SE->getSCEV(JPtr);
637 // If this is a trivial offset, then we'll get something like
638 // 1*sizeof(type). With target data, which we need anyway, this will get
639 // constant folded into a number.
640 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV);
641 if (const SCEVConstant *ConstOffSCEV =
642 dyn_cast<SCEVConstant>(OffsetSCEV)) {
643 ConstantInt *IntOff = ConstOffSCEV->getValue();
644 int64_t Offset = IntOff->getSExtValue();
646 Type *VTy = IPtr->getType()->getPointerElementType();
647 int64_t VTyTSS = (int64_t) DL->getTypeStoreSize(VTy);
649 Type *VTy2 = JPtr->getType()->getPointerElementType();
650 if (VTy != VTy2 && Offset < 0) {
651 int64_t VTy2TSS = (int64_t) DL->getTypeStoreSize(VTy2);
652 OffsetInElmts = Offset/VTy2TSS;
653 return (abs64(Offset) % VTy2TSS) == 0;
656 OffsetInElmts = Offset/VTyTSS;
657 return (abs64(Offset) % VTyTSS) == 0;
663 // Returns true if the provided CallInst represents an intrinsic that can
665 bool isVectorizableIntrinsic(CallInst* I) {
666 Function *F = I->getCalledFunction();
667 if (!F) return false;
669 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID();
670 if (!IID) return false;
675 case Intrinsic::sqrt:
676 case Intrinsic::powi:
680 case Intrinsic::log2:
681 case Intrinsic::log10:
683 case Intrinsic::exp2:
685 case Intrinsic::round:
686 case Intrinsic::copysign:
687 case Intrinsic::ceil:
688 case Intrinsic::nearbyint:
689 case Intrinsic::rint:
690 case Intrinsic::trunc:
691 case Intrinsic::floor:
692 case Intrinsic::fabs:
693 case Intrinsic::minnum:
694 case Intrinsic::maxnum:
695 return Config.VectorizeMath;
696 case Intrinsic::bswap:
697 case Intrinsic::ctpop:
698 case Intrinsic::ctlz:
699 case Intrinsic::cttz:
700 return Config.VectorizeBitManipulations;
702 case Intrinsic::fmuladd:
703 return Config.VectorizeFMA;
707 bool isPureIEChain(InsertElementInst *IE) {
708 InsertElementInst *IENext = IE;
710 if (!isa<UndefValue>(IENext->getOperand(0)) &&
711 !isa<InsertElementInst>(IENext->getOperand(0))) {
715 dyn_cast<InsertElementInst>(IENext->getOperand(0))));
721 // This function implements one vectorization iteration on the provided
722 // basic block. It returns true if the block is changed.
723 bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) {
725 BasicBlock::iterator Start = BB.getFirstInsertionPt();
727 std::vector<Value *> AllPairableInsts;
728 DenseMap<Value *, Value *> AllChosenPairs;
729 DenseSet<ValuePair> AllFixedOrderPairs;
730 DenseMap<VPPair, unsigned> AllPairConnectionTypes;
731 DenseMap<ValuePair, std::vector<ValuePair> > AllConnectedPairs,
732 AllConnectedPairDeps;
735 std::vector<Value *> PairableInsts;
736 DenseMap<Value *, std::vector<Value *> > CandidatePairs;
737 DenseSet<ValuePair> FixedOrderPairs;
738 DenseMap<ValuePair, int> CandidatePairCostSavings;
739 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs,
741 CandidatePairCostSavings,
742 PairableInsts, NonPow2Len);
743 if (PairableInsts.empty()) continue;
745 // Build the candidate pair set for faster lookups.
746 DenseSet<ValuePair> CandidatePairsSet;
747 for (DenseMap<Value *, std::vector<Value *> >::iterator I =
748 CandidatePairs.begin(), E = CandidatePairs.end(); I != E; ++I)
749 for (std::vector<Value *>::iterator J = I->second.begin(),
750 JE = I->second.end(); J != JE; ++J)
751 CandidatePairsSet.insert(ValuePair(I->first, *J));
753 // Now we have a map of all of the pairable instructions and we need to
754 // select the best possible pairing. A good pairing is one such that the
755 // users of the pair are also paired. This defines a (directed) forest
756 // over the pairs such that two pairs are connected iff the second pair
759 // Note that it only matters that both members of the second pair use some
760 // element of the first pair (to allow for splatting).
762 DenseMap<ValuePair, std::vector<ValuePair> > ConnectedPairs,
764 DenseMap<VPPair, unsigned> PairConnectionTypes;
765 computeConnectedPairs(CandidatePairs, CandidatePairsSet,
766 PairableInsts, ConnectedPairs, PairConnectionTypes);
767 if (ConnectedPairs.empty()) continue;
769 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator
770 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
772 for (std::vector<ValuePair>::iterator J = I->second.begin(),
773 JE = I->second.end(); J != JE; ++J)
774 ConnectedPairDeps[*J].push_back(I->first);
776 // Build the pairable-instruction dependency map
777 DenseSet<ValuePair> PairableInstUsers;
778 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers);
780 // There is now a graph of the connected pairs. For each variable, pick
781 // the pairing with the largest dag meeting the depth requirement on at
782 // least one branch. Then select all pairings that are part of that dag
783 // and remove them from the list of available pairings and pairable
786 DenseMap<Value *, Value *> ChosenPairs;
787 choosePairs(CandidatePairs, CandidatePairsSet,
788 CandidatePairCostSavings,
789 PairableInsts, FixedOrderPairs, PairConnectionTypes,
790 ConnectedPairs, ConnectedPairDeps,
791 PairableInstUsers, ChosenPairs);
793 if (ChosenPairs.empty()) continue;
794 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(),
795 PairableInsts.end());
796 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end());
798 // Only for the chosen pairs, propagate information on fixed-order pairs,
799 // pair connections, and their types to the data structures used by the
800 // pair fusion procedures.
801 for (DenseMap<Value *, Value *>::iterator I = ChosenPairs.begin(),
802 IE = ChosenPairs.end(); I != IE; ++I) {
803 if (FixedOrderPairs.count(*I))
804 AllFixedOrderPairs.insert(*I);
805 else if (FixedOrderPairs.count(ValuePair(I->second, I->first)))
806 AllFixedOrderPairs.insert(ValuePair(I->second, I->first));
808 for (DenseMap<Value *, Value *>::iterator J = ChosenPairs.begin();
810 DenseMap<VPPair, unsigned>::iterator K =
811 PairConnectionTypes.find(VPPair(*I, *J));
812 if (K != PairConnectionTypes.end()) {
813 AllPairConnectionTypes.insert(*K);
815 K = PairConnectionTypes.find(VPPair(*J, *I));
816 if (K != PairConnectionTypes.end())
817 AllPairConnectionTypes.insert(*K);
822 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator
823 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
825 for (std::vector<ValuePair>::iterator J = I->second.begin(),
826 JE = I->second.end(); J != JE; ++J)
827 if (AllPairConnectionTypes.count(VPPair(I->first, *J))) {
828 AllConnectedPairs[I->first].push_back(*J);
829 AllConnectedPairDeps[*J].push_back(I->first);
831 } while (ShouldContinue);
833 if (AllChosenPairs.empty()) return false;
834 NumFusedOps += AllChosenPairs.size();
836 // A set of pairs has now been selected. It is now necessary to replace the
837 // paired instructions with vector instructions. For this procedure each
838 // operand must be replaced with a vector operand. This vector is formed
839 // by using build_vector on the old operands. The replaced values are then
840 // replaced with a vector_extract on the result. Subsequent optimization
841 // passes should coalesce the build/extract combinations.
843 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs, AllFixedOrderPairs,
844 AllPairConnectionTypes,
845 AllConnectedPairs, AllConnectedPairDeps);
847 // It is important to cleanup here so that future iterations of this
848 // function have less work to do.
849 (void) SimplifyInstructionsInBlock(&BB, DL, AA->getTargetLibraryInfo());
853 // This function returns true if the provided instruction is capable of being
854 // fused into a vector instruction. This determination is based only on the
855 // type and other attributes of the instruction.
856 bool BBVectorize::isInstVectorizable(Instruction *I,
857 bool &IsSimpleLoadStore) {
858 IsSimpleLoadStore = false;
860 if (CallInst *C = dyn_cast<CallInst>(I)) {
861 if (!isVectorizableIntrinsic(C))
863 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) {
864 // Vectorize simple loads if possbile:
865 IsSimpleLoadStore = L->isSimple();
866 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
868 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) {
869 // Vectorize simple stores if possbile:
870 IsSimpleLoadStore = S->isSimple();
871 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
873 } else if (CastInst *C = dyn_cast<CastInst>(I)) {
874 // We can vectorize casts, but not casts of pointer types, etc.
875 if (!Config.VectorizeCasts)
878 Type *SrcTy = C->getSrcTy();
879 if (!SrcTy->isSingleValueType())
882 Type *DestTy = C->getDestTy();
883 if (!DestTy->isSingleValueType())
885 } else if (isa<SelectInst>(I)) {
886 if (!Config.VectorizeSelect)
888 } else if (isa<CmpInst>(I)) {
889 if (!Config.VectorizeCmp)
891 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) {
892 if (!Config.VectorizeGEP)
895 // Currently, vector GEPs exist only with one index.
896 if (G->getNumIndices() != 1)
898 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) ||
899 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) {
903 // We can't vectorize memory operations without target data
904 if (!DL && IsSimpleLoadStore)
908 getInstructionTypes(I, T1, T2);
910 // Not every type can be vectorized...
911 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) ||
912 !(VectorType::isValidElementType(T2) || T2->isVectorTy()))
915 if (T1->getScalarSizeInBits() == 1) {
916 if (!Config.VectorizeBools)
919 if (!Config.VectorizeInts && T1->isIntOrIntVectorTy())
923 if (T2->getScalarSizeInBits() == 1) {
924 if (!Config.VectorizeBools)
927 if (!Config.VectorizeInts && T2->isIntOrIntVectorTy())
931 if (!Config.VectorizeFloats
932 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy()))
935 // Don't vectorize target-specific types.
936 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy())
938 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy())
941 if ((!Config.VectorizePointers || !DL) &&
942 (T1->getScalarType()->isPointerTy() ||
943 T2->getScalarType()->isPointerTy()))
946 if (!TTI && (T1->getPrimitiveSizeInBits() >= Config.VectorBits ||
947 T2->getPrimitiveSizeInBits() >= Config.VectorBits))
953 // This function returns true if the two provided instructions are compatible
954 // (meaning that they can be fused into a vector instruction). This assumes
955 // that I has already been determined to be vectorizable and that J is not
956 // in the use dag of I.
957 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J,
958 bool IsSimpleLoadStore, bool NonPow2Len,
959 int &CostSavings, int &FixedOrder) {
960 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I <<
961 " <-> " << *J << "\n");
966 // Loads and stores can be merged if they have different alignments,
967 // but are otherwise the same.
968 if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment |
969 (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0)))
972 Type *IT1, *IT2, *JT1, *JT2;
973 getInstructionTypes(I, IT1, IT2);
974 getInstructionTypes(J, JT1, JT2);
975 unsigned MaxTypeBits = std::max(
976 IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(),
977 IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits());
978 if (!TTI && MaxTypeBits > Config.VectorBits)
981 // FIXME: handle addsub-type operations!
983 if (IsSimpleLoadStore) {
985 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
986 int64_t OffsetInElmts = 0;
987 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
988 IAddressSpace, JAddressSpace,
989 OffsetInElmts) && abs64(OffsetInElmts) == 1) {
990 FixedOrder = (int) OffsetInElmts;
991 unsigned BottomAlignment = IAlignment;
992 if (OffsetInElmts < 0) BottomAlignment = JAlignment;
994 Type *aTypeI = isa<StoreInst>(I) ?
995 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType();
996 Type *aTypeJ = isa<StoreInst>(J) ?
997 cast<StoreInst>(J)->getValueOperand()->getType() : J->getType();
998 Type *VType = getVecTypeForPair(aTypeI, aTypeJ);
1000 if (Config.AlignedOnly) {
1001 // An aligned load or store is possible only if the instruction
1002 // with the lower offset has an alignment suitable for the
1005 unsigned VecAlignment = DL->getPrefTypeAlignment(VType);
1006 if (BottomAlignment < VecAlignment)
1011 unsigned ICost = TTI->getMemoryOpCost(I->getOpcode(), aTypeI,
1012 IAlignment, IAddressSpace);
1013 unsigned JCost = TTI->getMemoryOpCost(J->getOpcode(), aTypeJ,
1014 JAlignment, JAddressSpace);
1015 unsigned VCost = TTI->getMemoryOpCost(I->getOpcode(), VType,
1019 ICost += TTI->getAddressComputationCost(aTypeI);
1020 JCost += TTI->getAddressComputationCost(aTypeJ);
1021 VCost += TTI->getAddressComputationCost(VType);
1023 if (VCost > ICost + JCost)
1026 // We don't want to fuse to a type that will be split, even
1027 // if the two input types will also be split and there is no other
1029 unsigned VParts = TTI->getNumberOfParts(VType);
1032 else if (!VParts && VCost == ICost + JCost)
1035 CostSavings = ICost + JCost - VCost;
1041 unsigned ICost = getInstrCost(I->getOpcode(), IT1, IT2);
1042 unsigned JCost = getInstrCost(J->getOpcode(), JT1, JT2);
1043 Type *VT1 = getVecTypeForPair(IT1, JT1),
1044 *VT2 = getVecTypeForPair(IT2, JT2);
1045 TargetTransformInfo::OperandValueKind Op1VK =
1046 TargetTransformInfo::OK_AnyValue;
1047 TargetTransformInfo::OperandValueKind Op2VK =
1048 TargetTransformInfo::OK_AnyValue;
1050 // On some targets (example X86) the cost of a vector shift may vary
1051 // depending on whether the second operand is a Uniform or
1052 // NonUniform Constant.
1053 switch (I->getOpcode()) {
1055 case Instruction::Shl:
1056 case Instruction::LShr:
1057 case Instruction::AShr:
1059 // If both I and J are scalar shifts by constant, then the
1060 // merged vector shift count would be either a constant splat value
1061 // or a non-uniform vector of constants.
1062 if (ConstantInt *CII = dyn_cast<ConstantInt>(I->getOperand(1))) {
1063 if (ConstantInt *CIJ = dyn_cast<ConstantInt>(J->getOperand(1)))
1064 Op2VK = CII == CIJ ? TargetTransformInfo::OK_UniformConstantValue :
1065 TargetTransformInfo::OK_NonUniformConstantValue;
1067 // Check for a splat of a constant or for a non uniform vector
1069 Value *IOp = I->getOperand(1);
1070 Value *JOp = J->getOperand(1);
1071 if ((isa<ConstantVector>(IOp) || isa<ConstantDataVector>(IOp)) &&
1072 (isa<ConstantVector>(JOp) || isa<ConstantDataVector>(JOp))) {
1073 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
1074 Constant *SplatValue = cast<Constant>(IOp)->getSplatValue();
1075 if (SplatValue != nullptr &&
1076 SplatValue == cast<Constant>(JOp)->getSplatValue())
1077 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
1082 // Note that this procedure is incorrect for insert and extract element
1083 // instructions (because combining these often results in a shuffle),
1084 // but this cost is ignored (because insert and extract element
1085 // instructions are assigned a zero depth factor and are not really
1086 // fused in general).
1087 unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2, Op1VK, Op2VK);
1089 if (VCost > ICost + JCost)
1092 // We don't want to fuse to a type that will be split, even
1093 // if the two input types will also be split and there is no other
1095 unsigned VParts1 = TTI->getNumberOfParts(VT1),
1096 VParts2 = TTI->getNumberOfParts(VT2);
1097 if (VParts1 > 1 || VParts2 > 1)
1099 else if ((!VParts1 || !VParts2) && VCost == ICost + JCost)
1102 CostSavings = ICost + JCost - VCost;
1105 // The powi,ctlz,cttz intrinsics are special because only the first
1106 // argument is vectorized, the second arguments must be equal.
1107 CallInst *CI = dyn_cast<CallInst>(I);
1109 if (CI && (FI = CI->getCalledFunction())) {
1110 Intrinsic::ID IID = (Intrinsic::ID) FI->getIntrinsicID();
1111 if (IID == Intrinsic::powi || IID == Intrinsic::ctlz ||
1112 IID == Intrinsic::cttz) {
1113 Value *A1I = CI->getArgOperand(1),
1114 *A1J = cast<CallInst>(J)->getArgOperand(1);
1115 const SCEV *A1ISCEV = SE->getSCEV(A1I),
1116 *A1JSCEV = SE->getSCEV(A1J);
1117 return (A1ISCEV == A1JSCEV);
1121 SmallVector<Type*, 4> Tys;
1122 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
1123 Tys.push_back(CI->getArgOperand(i)->getType());
1124 unsigned ICost = TTI->getIntrinsicInstrCost(IID, IT1, Tys);
1127 CallInst *CJ = cast<CallInst>(J);
1128 for (unsigned i = 0, ie = CJ->getNumArgOperands(); i != ie; ++i)
1129 Tys.push_back(CJ->getArgOperand(i)->getType());
1130 unsigned JCost = TTI->getIntrinsicInstrCost(IID, JT1, Tys);
1133 assert(CI->getNumArgOperands() == CJ->getNumArgOperands() &&
1134 "Intrinsic argument counts differ");
1135 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
1136 if ((IID == Intrinsic::powi || IID == Intrinsic::ctlz ||
1137 IID == Intrinsic::cttz) && i == 1)
1138 Tys.push_back(CI->getArgOperand(i)->getType());
1140 Tys.push_back(getVecTypeForPair(CI->getArgOperand(i)->getType(),
1141 CJ->getArgOperand(i)->getType()));
1144 Type *RetTy = getVecTypeForPair(IT1, JT1);
1145 unsigned VCost = TTI->getIntrinsicInstrCost(IID, RetTy, Tys);
1147 if (VCost > ICost + JCost)
1150 // We don't want to fuse to a type that will be split, even
1151 // if the two input types will also be split and there is no other
1153 unsigned RetParts = TTI->getNumberOfParts(RetTy);
1156 else if (!RetParts && VCost == ICost + JCost)
1159 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
1160 if (!Tys[i]->isVectorTy())
1163 unsigned NumParts = TTI->getNumberOfParts(Tys[i]);
1166 else if (!NumParts && VCost == ICost + JCost)
1170 CostSavings = ICost + JCost - VCost;
1177 // Figure out whether or not J uses I and update the users and write-set
1178 // structures associated with I. Specifically, Users represents the set of
1179 // instructions that depend on I. WriteSet represents the set
1180 // of memory locations that are dependent on I. If UpdateUsers is true,
1181 // and J uses I, then Users is updated to contain J and WriteSet is updated
1182 // to contain any memory locations to which J writes. The function returns
1183 // true if J uses I. By default, alias analysis is used to determine
1184 // whether J reads from memory that overlaps with a location in WriteSet.
1185 // If LoadMoveSet is not null, then it is a previously-computed map
1186 // where the key is the memory-based user instruction and the value is
1187 // the instruction to be compared with I. So, if LoadMoveSet is provided,
1188 // then the alias analysis is not used. This is necessary because this
1189 // function is called during the process of moving instructions during
1190 // vectorization and the results of the alias analysis are not stable during
1192 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users,
1193 AliasSetTracker &WriteSet, Instruction *I,
1194 Instruction *J, bool UpdateUsers,
1195 DenseSet<ValuePair> *LoadMoveSetPairs) {
1198 // This instruction may already be marked as a user due, for example, to
1199 // being a member of a selected pair.
1204 for (User::op_iterator JU = J->op_begin(), JE = J->op_end();
1207 if (I == V || Users.count(V)) {
1212 if (!UsesI && J->mayReadFromMemory()) {
1213 if (LoadMoveSetPairs) {
1214 UsesI = LoadMoveSetPairs->count(ValuePair(J, I));
1216 for (AliasSetTracker::iterator W = WriteSet.begin(),
1217 WE = WriteSet.end(); W != WE; ++W) {
1218 if (W->aliasesUnknownInst(J, *AA)) {
1226 if (UsesI && UpdateUsers) {
1227 if (J->mayWriteToMemory()) WriteSet.add(J);
1234 // This function iterates over all instruction pairs in the provided
1235 // basic block and collects all candidate pairs for vectorization.
1236 bool BBVectorize::getCandidatePairs(BasicBlock &BB,
1237 BasicBlock::iterator &Start,
1238 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1239 DenseSet<ValuePair> &FixedOrderPairs,
1240 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1241 std::vector<Value *> &PairableInsts, bool NonPow2Len) {
1242 size_t TotalPairs = 0;
1243 BasicBlock::iterator E = BB.end();
1244 if (Start == E) return false;
1246 bool ShouldContinue = false, IAfterStart = false;
1247 for (BasicBlock::iterator I = Start++; I != E; ++I) {
1248 if (I == Start) IAfterStart = true;
1250 bool IsSimpleLoadStore;
1251 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue;
1253 // Look for an instruction with which to pair instruction *I...
1254 DenseSet<Value *> Users;
1255 AliasSetTracker WriteSet(*AA);
1256 if (I->mayWriteToMemory()) WriteSet.add(I);
1258 bool JAfterStart = IAfterStart;
1259 BasicBlock::iterator J = std::next(I);
1260 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) {
1261 if (J == Start) JAfterStart = true;
1263 // Determine if J uses I, if so, exit the loop.
1264 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep);
1265 if (Config.FastDep) {
1266 // Note: For this heuristic to be effective, independent operations
1267 // must tend to be intermixed. This is likely to be true from some
1268 // kinds of grouped loop unrolling (but not the generic LLVM pass),
1269 // but otherwise may require some kind of reordering pass.
1271 // When using fast dependency analysis,
1272 // stop searching after first use:
1275 if (UsesI) continue;
1278 // J does not use I, and comes before the first use of I, so it can be
1279 // merged with I if the instructions are compatible.
1280 int CostSavings, FixedOrder;
1281 if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len,
1282 CostSavings, FixedOrder)) continue;
1284 // J is a candidate for merging with I.
1285 if (PairableInsts.empty() ||
1286 PairableInsts[PairableInsts.size()-1] != I) {
1287 PairableInsts.push_back(I);
1290 CandidatePairs[I].push_back(J);
1293 CandidatePairCostSavings.insert(ValuePairWithCost(ValuePair(I, J),
1296 if (FixedOrder == 1)
1297 FixedOrderPairs.insert(ValuePair(I, J));
1298 else if (FixedOrder == -1)
1299 FixedOrderPairs.insert(ValuePair(J, I));
1301 // The next call to this function must start after the last instruction
1302 // selected during this invocation.
1304 Start = std::next(J);
1305 IAfterStart = JAfterStart = false;
1308 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair "
1309 << *I << " <-> " << *J << " (cost savings: " <<
1310 CostSavings << ")\n");
1312 // If we have already found too many pairs, break here and this function
1313 // will be called again starting after the last instruction selected
1314 // during this invocation.
1315 if (PairableInsts.size() >= Config.MaxInsts ||
1316 TotalPairs >= Config.MaxPairs) {
1317 ShouldContinue = true;
1326 DEBUG(dbgs() << "BBV: found " << PairableInsts.size()
1327 << " instructions with candidate pairs\n");
1329 return ShouldContinue;
1332 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that
1333 // it looks for pairs such that both members have an input which is an
1334 // output of PI or PJ.
1335 void BBVectorize::computePairsConnectedTo(
1336 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1337 DenseSet<ValuePair> &CandidatePairsSet,
1338 std::vector<Value *> &PairableInsts,
1339 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1340 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1344 // For each possible pairing for this variable, look at the uses of
1345 // the first value...
1346 for (Value::user_iterator I = P.first->user_begin(),
1347 E = P.first->user_end();
1350 if (isa<LoadInst>(UI)) {
1351 // A pair cannot be connected to a load because the load only takes one
1352 // operand (the address) and it is a scalar even after vectorization.
1354 } else if ((SI = dyn_cast<StoreInst>(UI)) &&
1355 P.first == SI->getPointerOperand()) {
1356 // Similarly, a pair cannot be connected to a store through its
1361 // For each use of the first variable, look for uses of the second
1363 for (User *UJ : P.second->users()) {
1364 if ((SJ = dyn_cast<StoreInst>(UJ)) &&
1365 P.second == SJ->getPointerOperand())
1369 if (CandidatePairsSet.count(ValuePair(UI, UJ))) {
1370 VPPair VP(P, ValuePair(UI, UJ));
1371 ConnectedPairs[VP.first].push_back(VP.second);
1372 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionDirect));
1376 if (CandidatePairsSet.count(ValuePair(UJ, UI))) {
1377 VPPair VP(P, ValuePair(UJ, UI));
1378 ConnectedPairs[VP.first].push_back(VP.second);
1379 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSwap));
1383 if (Config.SplatBreaksChain) continue;
1384 // Look for cases where just the first value in the pair is used by
1385 // both members of another pair (splatting).
1386 for (Value::user_iterator J = P.first->user_begin(); J != E; ++J) {
1388 if ((SJ = dyn_cast<StoreInst>(UJ)) &&
1389 P.first == SJ->getPointerOperand())
1392 if (CandidatePairsSet.count(ValuePair(UI, UJ))) {
1393 VPPair VP(P, ValuePair(UI, UJ));
1394 ConnectedPairs[VP.first].push_back(VP.second);
1395 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1400 if (Config.SplatBreaksChain) return;
1401 // Look for cases where just the second value in the pair is used by
1402 // both members of another pair (splatting).
1403 for (Value::user_iterator I = P.second->user_begin(),
1404 E = P.second->user_end();
1407 if (isa<LoadInst>(UI))
1409 else if ((SI = dyn_cast<StoreInst>(UI)) &&
1410 P.second == SI->getPointerOperand())
1413 for (Value::user_iterator J = P.second->user_begin(); J != E; ++J) {
1415 if ((SJ = dyn_cast<StoreInst>(UJ)) &&
1416 P.second == SJ->getPointerOperand())
1419 if (CandidatePairsSet.count(ValuePair(UI, UJ))) {
1420 VPPair VP(P, ValuePair(UI, UJ));
1421 ConnectedPairs[VP.first].push_back(VP.second);
1422 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1428 // This function figures out which pairs are connected. Two pairs are
1429 // connected if some output of the first pair forms an input to both members
1430 // of the second pair.
1431 void BBVectorize::computeConnectedPairs(
1432 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1433 DenseSet<ValuePair> &CandidatePairsSet,
1434 std::vector<Value *> &PairableInsts,
1435 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1436 DenseMap<VPPair, unsigned> &PairConnectionTypes) {
1437 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
1438 PE = PairableInsts.end(); PI != PE; ++PI) {
1439 DenseMap<Value *, std::vector<Value *> >::iterator PP =
1440 CandidatePairs.find(*PI);
1441 if (PP == CandidatePairs.end())
1444 for (std::vector<Value *>::iterator P = PP->second.begin(),
1445 E = PP->second.end(); P != E; ++P)
1446 computePairsConnectedTo(CandidatePairs, CandidatePairsSet,
1447 PairableInsts, ConnectedPairs,
1448 PairConnectionTypes, ValuePair(*PI, *P));
1451 DEBUG(size_t TotalPairs = 0;
1452 for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator I =
1453 ConnectedPairs.begin(), IE = ConnectedPairs.end(); I != IE; ++I)
1454 TotalPairs += I->second.size();
1455 dbgs() << "BBV: found " << TotalPairs
1456 << " pair connections.\n");
1459 // This function builds a set of use tuples such that <A, B> is in the set
1460 // if B is in the use dag of A. If B is in the use dag of A, then B
1461 // depends on the output of A.
1462 void BBVectorize::buildDepMap(
1464 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1465 std::vector<Value *> &PairableInsts,
1466 DenseSet<ValuePair> &PairableInstUsers) {
1467 DenseSet<Value *> IsInPair;
1468 for (DenseMap<Value *, std::vector<Value *> >::iterator C =
1469 CandidatePairs.begin(), E = CandidatePairs.end(); C != E; ++C) {
1470 IsInPair.insert(C->first);
1471 IsInPair.insert(C->second.begin(), C->second.end());
1474 // Iterate through the basic block, recording all users of each
1475 // pairable instruction.
1477 BasicBlock::iterator E = BB.end(), EL =
1478 BasicBlock::iterator(cast<Instruction>(PairableInsts.back()));
1479 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) {
1480 if (IsInPair.find(I) == IsInPair.end()) continue;
1482 DenseSet<Value *> Users;
1483 AliasSetTracker WriteSet(*AA);
1484 if (I->mayWriteToMemory()) WriteSet.add(I);
1486 for (BasicBlock::iterator J = std::next(I); J != E; ++J) {
1487 (void) trackUsesOfI(Users, WriteSet, I, J);
1493 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end();
1495 if (IsInPair.find(*U) == IsInPair.end()) continue;
1496 PairableInstUsers.insert(ValuePair(I, *U));
1504 // Returns true if an input to pair P is an output of pair Q and also an
1505 // input of pair Q is an output of pair P. If this is the case, then these
1506 // two pairs cannot be simultaneously fused.
1507 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q,
1508 DenseSet<ValuePair> &PairableInstUsers,
1509 DenseMap<ValuePair, std::vector<ValuePair> > *PairableInstUserMap,
1510 DenseSet<VPPair> *PairableInstUserPairSet) {
1511 // Two pairs are in conflict if they are mutual Users of eachother.
1512 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) ||
1513 PairableInstUsers.count(ValuePair(P.first, Q.second)) ||
1514 PairableInstUsers.count(ValuePair(P.second, Q.first)) ||
1515 PairableInstUsers.count(ValuePair(P.second, Q.second));
1516 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) ||
1517 PairableInstUsers.count(ValuePair(Q.first, P.second)) ||
1518 PairableInstUsers.count(ValuePair(Q.second, P.first)) ||
1519 PairableInstUsers.count(ValuePair(Q.second, P.second));
1520 if (PairableInstUserMap) {
1521 // FIXME: The expensive part of the cycle check is not so much the cycle
1522 // check itself but this edge insertion procedure. This needs some
1523 // profiling and probably a different data structure.
1525 if (PairableInstUserPairSet->insert(VPPair(Q, P)).second)
1526 (*PairableInstUserMap)[Q].push_back(P);
1529 if (PairableInstUserPairSet->insert(VPPair(P, Q)).second)
1530 (*PairableInstUserMap)[P].push_back(Q);
1534 return (QUsesP && PUsesQ);
1537 // This function walks the use graph of current pairs to see if, starting
1538 // from P, the walk returns to P.
1539 bool BBVectorize::pairWillFormCycle(ValuePair P,
1540 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
1541 DenseSet<ValuePair> &CurrentPairs) {
1542 DEBUG(if (DebugCycleCheck)
1543 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> "
1544 << *P.second << "\n");
1545 // A lookup table of visisted pairs is kept because the PairableInstUserMap
1546 // contains non-direct associations.
1547 DenseSet<ValuePair> Visited;
1548 SmallVector<ValuePair, 32> Q;
1549 // General depth-first post-order traversal:
1552 ValuePair QTop = Q.pop_back_val();
1553 Visited.insert(QTop);
1555 DEBUG(if (DebugCycleCheck)
1556 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> "
1557 << *QTop.second << "\n");
1558 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
1559 PairableInstUserMap.find(QTop);
1560 if (QQ == PairableInstUserMap.end())
1563 for (std::vector<ValuePair>::iterator C = QQ->second.begin(),
1564 CE = QQ->second.end(); C != CE; ++C) {
1567 << "BBV: rejected to prevent non-trivial cycle formation: "
1568 << QTop.first << " <-> " << C->second << "\n");
1572 if (CurrentPairs.count(*C) && !Visited.count(*C))
1575 } while (!Q.empty());
1580 // This function builds the initial dag of connected pairs with the
1581 // pair J at the root.
1582 void BBVectorize::buildInitialDAGFor(
1583 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1584 DenseSet<ValuePair> &CandidatePairsSet,
1585 std::vector<Value *> &PairableInsts,
1586 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1587 DenseSet<ValuePair> &PairableInstUsers,
1588 DenseMap<Value *, Value *> &ChosenPairs,
1589 DenseMap<ValuePair, size_t> &DAG, ValuePair J) {
1590 // Each of these pairs is viewed as the root node of a DAG. The DAG
1591 // is then walked (depth-first). As this happens, we keep track of
1592 // the pairs that compose the DAG and the maximum depth of the DAG.
1593 SmallVector<ValuePairWithDepth, 32> Q;
1594 // General depth-first post-order traversal:
1595 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1597 ValuePairWithDepth QTop = Q.back();
1599 // Push each child onto the queue:
1600 bool MoreChildren = false;
1601 size_t MaxChildDepth = QTop.second;
1602 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
1603 ConnectedPairs.find(QTop.first);
1604 if (QQ != ConnectedPairs.end())
1605 for (std::vector<ValuePair>::iterator k = QQ->second.begin(),
1606 ke = QQ->second.end(); k != ke; ++k) {
1607 // Make sure that this child pair is still a candidate:
1608 if (CandidatePairsSet.count(*k)) {
1609 DenseMap<ValuePair, size_t>::iterator C = DAG.find(*k);
1610 if (C == DAG.end()) {
1611 size_t d = getDepthFactor(k->first);
1612 Q.push_back(ValuePairWithDepth(*k, QTop.second+d));
1613 MoreChildren = true;
1615 MaxChildDepth = std::max(MaxChildDepth, C->second);
1620 if (!MoreChildren) {
1621 // Record the current pair as part of the DAG:
1622 DAG.insert(ValuePairWithDepth(QTop.first, MaxChildDepth));
1625 } while (!Q.empty());
1628 // Given some initial dag, prune it by removing conflicting pairs (pairs
1629 // that cannot be simultaneously chosen for vectorization).
1630 void BBVectorize::pruneDAGFor(
1631 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1632 std::vector<Value *> &PairableInsts,
1633 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1634 DenseSet<ValuePair> &PairableInstUsers,
1635 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
1636 DenseSet<VPPair> &PairableInstUserPairSet,
1637 DenseMap<Value *, Value *> &ChosenPairs,
1638 DenseMap<ValuePair, size_t> &DAG,
1639 DenseSet<ValuePair> &PrunedDAG, ValuePair J,
1640 bool UseCycleCheck) {
1641 SmallVector<ValuePairWithDepth, 32> Q;
1642 // General depth-first post-order traversal:
1643 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1645 ValuePairWithDepth QTop = Q.pop_back_val();
1646 PrunedDAG.insert(QTop.first);
1648 // Visit each child, pruning as necessary...
1649 SmallVector<ValuePairWithDepth, 8> BestChildren;
1650 DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
1651 ConnectedPairs.find(QTop.first);
1652 if (QQ == ConnectedPairs.end())
1655 for (std::vector<ValuePair>::iterator K = QQ->second.begin(),
1656 KE = QQ->second.end(); K != KE; ++K) {
1657 DenseMap<ValuePair, size_t>::iterator C = DAG.find(*K);
1658 if (C == DAG.end()) continue;
1660 // This child is in the DAG, now we need to make sure it is the
1661 // best of any conflicting children. There could be multiple
1662 // conflicting children, so first, determine if we're keeping
1663 // this child, then delete conflicting children as necessary.
1665 // It is also necessary to guard against pairing-induced
1666 // dependencies. Consider instructions a .. x .. y .. b
1667 // such that (a,b) are to be fused and (x,y) are to be fused
1668 // but a is an input to x and b is an output from y. This
1669 // means that y cannot be moved after b but x must be moved
1670 // after b for (a,b) to be fused. In other words, after
1671 // fusing (a,b) we have y .. a/b .. x where y is an input
1672 // to a/b and x is an output to a/b: x and y can no longer
1673 // be legally fused. To prevent this condition, we must
1674 // make sure that a child pair added to the DAG is not
1675 // both an input and output of an already-selected pair.
1677 // Pairing-induced dependencies can also form from more complicated
1678 // cycles. The pair vs. pair conflicts are easy to check, and so
1679 // that is done explicitly for "fast rejection", and because for
1680 // child vs. child conflicts, we may prefer to keep the current
1681 // pair in preference to the already-selected child.
1682 DenseSet<ValuePair> CurrentPairs;
1685 for (SmallVectorImpl<ValuePairWithDepth>::iterator C2
1686 = BestChildren.begin(), E2 = BestChildren.end();
1688 if (C2->first.first == C->first.first ||
1689 C2->first.first == C->first.second ||
1690 C2->first.second == C->first.first ||
1691 C2->first.second == C->first.second ||
1692 pairsConflict(C2->first, C->first, PairableInstUsers,
1693 UseCycleCheck ? &PairableInstUserMap : nullptr,
1694 UseCycleCheck ? &PairableInstUserPairSet
1696 if (C2->second >= C->second) {
1701 CurrentPairs.insert(C2->first);
1704 if (!CanAdd) continue;
1706 // Even worse, this child could conflict with another node already
1707 // selected for the DAG. If that is the case, ignore this child.
1708 for (DenseSet<ValuePair>::iterator T = PrunedDAG.begin(),
1709 E2 = PrunedDAG.end(); T != E2; ++T) {
1710 if (T->first == C->first.first ||
1711 T->first == C->first.second ||
1712 T->second == C->first.first ||
1713 T->second == C->first.second ||
1714 pairsConflict(*T, C->first, PairableInstUsers,
1715 UseCycleCheck ? &PairableInstUserMap : nullptr,
1716 UseCycleCheck ? &PairableInstUserPairSet
1722 CurrentPairs.insert(*T);
1724 if (!CanAdd) continue;
1726 // And check the queue too...
1727 for (SmallVectorImpl<ValuePairWithDepth>::iterator C2 = Q.begin(),
1728 E2 = Q.end(); C2 != E2; ++C2) {
1729 if (C2->first.first == C->first.first ||
1730 C2->first.first == C->first.second ||
1731 C2->first.second == C->first.first ||
1732 C2->first.second == C->first.second ||
1733 pairsConflict(C2->first, C->first, PairableInstUsers,
1734 UseCycleCheck ? &PairableInstUserMap : nullptr,
1735 UseCycleCheck ? &PairableInstUserPairSet
1741 CurrentPairs.insert(C2->first);
1743 if (!CanAdd) continue;
1745 // Last but not least, check for a conflict with any of the
1746 // already-chosen pairs.
1747 for (DenseMap<Value *, Value *>::iterator C2 =
1748 ChosenPairs.begin(), E2 = ChosenPairs.end();
1750 if (pairsConflict(*C2, C->first, PairableInstUsers,
1751 UseCycleCheck ? &PairableInstUserMap : nullptr,
1752 UseCycleCheck ? &PairableInstUserPairSet
1758 CurrentPairs.insert(*C2);
1760 if (!CanAdd) continue;
1762 // To check for non-trivial cycles formed by the addition of the
1763 // current pair we've formed a list of all relevant pairs, now use a
1764 // graph walk to check for a cycle. We start from the current pair and
1765 // walk the use dag to see if we again reach the current pair. If we
1766 // do, then the current pair is rejected.
1768 // FIXME: It may be more efficient to use a topological-ordering
1769 // algorithm to improve the cycle check. This should be investigated.
1770 if (UseCycleCheck &&
1771 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs))
1774 // This child can be added, but we may have chosen it in preference
1775 // to an already-selected child. Check for this here, and if a
1776 // conflict is found, then remove the previously-selected child
1777 // before adding this one in its place.
1778 for (SmallVectorImpl<ValuePairWithDepth>::iterator C2
1779 = BestChildren.begin(); C2 != BestChildren.end();) {
1780 if (C2->first.first == C->first.first ||
1781 C2->first.first == C->first.second ||
1782 C2->first.second == C->first.first ||
1783 C2->first.second == C->first.second ||
1784 pairsConflict(C2->first, C->first, PairableInstUsers))
1785 C2 = BestChildren.erase(C2);
1790 BestChildren.push_back(ValuePairWithDepth(C->first, C->second));
1793 for (SmallVectorImpl<ValuePairWithDepth>::iterator C
1794 = BestChildren.begin(), E2 = BestChildren.end();
1796 size_t DepthF = getDepthFactor(C->first.first);
1797 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF));
1799 } while (!Q.empty());
1802 // This function finds the best dag of mututally-compatible connected
1803 // pairs, given the choice of root pairs as an iterator range.
1804 void BBVectorize::findBestDAGFor(
1805 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1806 DenseSet<ValuePair> &CandidatePairsSet,
1807 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1808 std::vector<Value *> &PairableInsts,
1809 DenseSet<ValuePair> &FixedOrderPairs,
1810 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1811 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1812 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
1813 DenseSet<ValuePair> &PairableInstUsers,
1814 DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
1815 DenseSet<VPPair> &PairableInstUserPairSet,
1816 DenseMap<Value *, Value *> &ChosenPairs,
1817 DenseSet<ValuePair> &BestDAG, size_t &BestMaxDepth,
1818 int &BestEffSize, Value *II, std::vector<Value *>&JJ,
1819 bool UseCycleCheck) {
1820 for (std::vector<Value *>::iterator J = JJ.begin(), JE = JJ.end();
1822 ValuePair IJ(II, *J);
1823 if (!CandidatePairsSet.count(IJ))
1826 // Before going any further, make sure that this pair does not
1827 // conflict with any already-selected pairs (see comment below
1828 // near the DAG pruning for more details).
1829 DenseSet<ValuePair> ChosenPairSet;
1830 bool DoesConflict = false;
1831 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(),
1832 E = ChosenPairs.end(); C != E; ++C) {
1833 if (pairsConflict(*C, IJ, PairableInstUsers,
1834 UseCycleCheck ? &PairableInstUserMap : nullptr,
1835 UseCycleCheck ? &PairableInstUserPairSet : nullptr)) {
1836 DoesConflict = true;
1840 ChosenPairSet.insert(*C);
1842 if (DoesConflict) continue;
1844 if (UseCycleCheck &&
1845 pairWillFormCycle(IJ, PairableInstUserMap, ChosenPairSet))
1848 DenseMap<ValuePair, size_t> DAG;
1849 buildInitialDAGFor(CandidatePairs, CandidatePairsSet,
1850 PairableInsts, ConnectedPairs,
1851 PairableInstUsers, ChosenPairs, DAG, IJ);
1853 // Because we'll keep the child with the largest depth, the largest
1854 // depth is still the same in the unpruned DAG.
1855 size_t MaxDepth = DAG.lookup(IJ);
1857 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found DAG for pair {"
1858 << *IJ.first << " <-> " << *IJ.second << "} of depth " <<
1859 MaxDepth << " and size " << DAG.size() << "\n");
1861 // At this point the DAG has been constructed, but, may contain
1862 // contradictory children (meaning that different children of
1863 // some dag node may be attempting to fuse the same instruction).
1864 // So now we walk the dag again, in the case of a conflict,
1865 // keep only the child with the largest depth. To break a tie,
1866 // favor the first child.
1868 DenseSet<ValuePair> PrunedDAG;
1869 pruneDAGFor(CandidatePairs, PairableInsts, ConnectedPairs,
1870 PairableInstUsers, PairableInstUserMap,
1871 PairableInstUserPairSet,
1872 ChosenPairs, DAG, PrunedDAG, IJ, UseCycleCheck);
1876 DenseSet<Value *> PrunedDAGInstrs;
1877 for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(),
1878 E = PrunedDAG.end(); S != E; ++S) {
1879 PrunedDAGInstrs.insert(S->first);
1880 PrunedDAGInstrs.insert(S->second);
1883 // The set of pairs that have already contributed to the total cost.
1884 DenseSet<ValuePair> IncomingPairs;
1886 // If the cost model were perfect, this might not be necessary; but we
1887 // need to make sure that we don't get stuck vectorizing our own
1889 bool HasNontrivialInsts = false;
1891 // The node weights represent the cost savings associated with
1892 // fusing the pair of instructions.
1893 for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(),
1894 E = PrunedDAG.end(); S != E; ++S) {
1895 if (!isa<ShuffleVectorInst>(S->first) &&
1896 !isa<InsertElementInst>(S->first) &&
1897 !isa<ExtractElementInst>(S->first))
1898 HasNontrivialInsts = true;
1900 bool FlipOrder = false;
1902 if (getDepthFactor(S->first)) {
1903 int ESContrib = CandidatePairCostSavings.find(*S)->second;
1904 DEBUG(if (DebugPairSelection) dbgs() << "\tweight {"
1905 << *S->first << " <-> " << *S->second << "} = " <<
1907 EffSize += ESContrib;
1910 // The edge weights contribute in a negative sense: they represent
1911 // the cost of shuffles.
1912 DenseMap<ValuePair, std::vector<ValuePair> >::iterator SS =
1913 ConnectedPairDeps.find(*S);
1914 if (SS != ConnectedPairDeps.end()) {
1915 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
1916 for (std::vector<ValuePair>::iterator T = SS->second.begin(),
1917 TE = SS->second.end(); T != TE; ++T) {
1919 if (!PrunedDAG.count(Q.second))
1921 DenseMap<VPPair, unsigned>::iterator R =
1922 PairConnectionTypes.find(VPPair(Q.second, Q.first));
1923 assert(R != PairConnectionTypes.end() &&
1924 "Cannot find pair connection type");
1925 if (R->second == PairConnectionDirect)
1927 else if (R->second == PairConnectionSwap)
1931 // If there are more swaps than direct connections, then
1932 // the pair order will be flipped during fusion. So the real
1933 // number of swaps is the minimum number.
1934 FlipOrder = !FixedOrderPairs.count(*S) &&
1935 ((NumDepsSwap > NumDepsDirect) ||
1936 FixedOrderPairs.count(ValuePair(S->second, S->first)));
1938 for (std::vector<ValuePair>::iterator T = SS->second.begin(),
1939 TE = SS->second.end(); T != TE; ++T) {
1941 if (!PrunedDAG.count(Q.second))
1943 DenseMap<VPPair, unsigned>::iterator R =
1944 PairConnectionTypes.find(VPPair(Q.second, Q.first));
1945 assert(R != PairConnectionTypes.end() &&
1946 "Cannot find pair connection type");
1947 Type *Ty1 = Q.second.first->getType(),
1948 *Ty2 = Q.second.second->getType();
1949 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1950 if ((R->second == PairConnectionDirect && FlipOrder) ||
1951 (R->second == PairConnectionSwap && !FlipOrder) ||
1952 R->second == PairConnectionSplat) {
1953 int ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1956 if (VTy->getVectorNumElements() == 2) {
1957 if (R->second == PairConnectionSplat)
1958 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1959 TargetTransformInfo::SK_Broadcast, VTy));
1961 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1962 TargetTransformInfo::SK_Reverse, VTy));
1965 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1966 *Q.second.first << " <-> " << *Q.second.second <<
1968 *S->first << " <-> " << *S->second << "} = " <<
1970 EffSize -= ESContrib;
1975 // Compute the cost of outgoing edges. We assume that edges outgoing
1976 // to shuffles, inserts or extracts can be merged, and so contribute
1977 // no additional cost.
1978 if (!S->first->getType()->isVoidTy()) {
1979 Type *Ty1 = S->first->getType(),
1980 *Ty2 = S->second->getType();
1981 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1983 bool NeedsExtraction = false;
1984 for (User *U : S->first->users()) {
1985 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(U)) {
1986 // Shuffle can be folded if it has no other input
1987 if (isa<UndefValue>(SI->getOperand(1)))
1990 if (isa<ExtractElementInst>(U))
1992 if (PrunedDAGInstrs.count(U))
1994 NeedsExtraction = true;
1998 if (NeedsExtraction) {
2000 if (Ty1->isVectorTy()) {
2001 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2003 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
2004 TargetTransformInfo::SK_ExtractSubvector, VTy, 0, Ty1));
2006 ESContrib = (int) TTI->getVectorInstrCost(
2007 Instruction::ExtractElement, VTy, 0);
2009 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
2010 *S->first << "} = " << ESContrib << "\n");
2011 EffSize -= ESContrib;
2014 NeedsExtraction = false;
2015 for (User *U : S->second->users()) {
2016 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(U)) {
2017 // Shuffle can be folded if it has no other input
2018 if (isa<UndefValue>(SI->getOperand(1)))
2021 if (isa<ExtractElementInst>(U))
2023 if (PrunedDAGInstrs.count(U))
2025 NeedsExtraction = true;
2029 if (NeedsExtraction) {
2031 if (Ty2->isVectorTy()) {
2032 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2034 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
2035 TargetTransformInfo::SK_ExtractSubvector, VTy,
2036 Ty1->isVectorTy() ? Ty1->getVectorNumElements() : 1, Ty2));
2038 ESContrib = (int) TTI->getVectorInstrCost(
2039 Instruction::ExtractElement, VTy, 1);
2040 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
2041 *S->second << "} = " << ESContrib << "\n");
2042 EffSize -= ESContrib;
2046 // Compute the cost of incoming edges.
2047 if (!isa<LoadInst>(S->first) && !isa<StoreInst>(S->first)) {
2048 Instruction *S1 = cast<Instruction>(S->first),
2049 *S2 = cast<Instruction>(S->second);
2050 for (unsigned o = 0; o < S1->getNumOperands(); ++o) {
2051 Value *O1 = S1->getOperand(o), *O2 = S2->getOperand(o);
2053 // Combining constants into vector constants (or small vector
2054 // constants into larger ones are assumed free).
2055 if (isa<Constant>(O1) && isa<Constant>(O2))
2061 ValuePair VP = ValuePair(O1, O2);
2062 ValuePair VPR = ValuePair(O2, O1);
2064 // Internal edges are not handled here.
2065 if (PrunedDAG.count(VP) || PrunedDAG.count(VPR))
2068 Type *Ty1 = O1->getType(),
2069 *Ty2 = O2->getType();
2070 Type *VTy = getVecTypeForPair(Ty1, Ty2);
2072 // Combining vector operations of the same type is also assumed
2073 // folded with other operations.
2075 // If both are insert elements, then both can be widened.
2076 InsertElementInst *IEO1 = dyn_cast<InsertElementInst>(O1),
2077 *IEO2 = dyn_cast<InsertElementInst>(O2);
2078 if (IEO1 && IEO2 && isPureIEChain(IEO1) && isPureIEChain(IEO2))
2080 // If both are extract elements, and both have the same input
2081 // type, then they can be replaced with a shuffle
2082 ExtractElementInst *EIO1 = dyn_cast<ExtractElementInst>(O1),
2083 *EIO2 = dyn_cast<ExtractElementInst>(O2);
2085 EIO1->getOperand(0)->getType() ==
2086 EIO2->getOperand(0)->getType())
2088 // If both are a shuffle with equal operand types and only two
2089 // unqiue operands, then they can be replaced with a single
2091 ShuffleVectorInst *SIO1 = dyn_cast<ShuffleVectorInst>(O1),
2092 *SIO2 = dyn_cast<ShuffleVectorInst>(O2);
2094 SIO1->getOperand(0)->getType() ==
2095 SIO2->getOperand(0)->getType()) {
2096 SmallSet<Value *, 4> SIOps;
2097 SIOps.insert(SIO1->getOperand(0));
2098 SIOps.insert(SIO1->getOperand(1));
2099 SIOps.insert(SIO2->getOperand(0));
2100 SIOps.insert(SIO2->getOperand(1));
2101 if (SIOps.size() <= 2)
2107 // This pair has already been formed.
2108 if (IncomingPairs.count(VP)) {
2110 } else if (IncomingPairs.count(VPR)) {
2111 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2114 if (VTy->getVectorNumElements() == 2)
2115 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
2116 TargetTransformInfo::SK_Reverse, VTy));
2117 } else if (!Ty1->isVectorTy() && !Ty2->isVectorTy()) {
2118 ESContrib = (int) TTI->getVectorInstrCost(
2119 Instruction::InsertElement, VTy, 0);
2120 ESContrib += (int) TTI->getVectorInstrCost(
2121 Instruction::InsertElement, VTy, 1);
2122 } else if (!Ty1->isVectorTy()) {
2123 // O1 needs to be inserted into a vector of size O2, and then
2124 // both need to be shuffled together.
2125 ESContrib = (int) TTI->getVectorInstrCost(
2126 Instruction::InsertElement, Ty2, 0);
2127 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2129 } else if (!Ty2->isVectorTy()) {
2130 // O2 needs to be inserted into a vector of size O1, and then
2131 // both need to be shuffled together.
2132 ESContrib = (int) TTI->getVectorInstrCost(
2133 Instruction::InsertElement, Ty1, 0);
2134 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2137 Type *TyBig = Ty1, *TySmall = Ty2;
2138 if (Ty2->getVectorNumElements() > Ty1->getVectorNumElements())
2139 std::swap(TyBig, TySmall);
2141 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2143 if (TyBig != TySmall)
2144 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2148 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {"
2149 << *O1 << " <-> " << *O2 << "} = " <<
2151 EffSize -= ESContrib;
2152 IncomingPairs.insert(VP);
2157 if (!HasNontrivialInsts) {
2158 DEBUG(if (DebugPairSelection) dbgs() <<
2159 "\tNo non-trivial instructions in DAG;"
2160 " override to zero effective size\n");
2164 for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(),
2165 E = PrunedDAG.end(); S != E; ++S)
2166 EffSize += (int) getDepthFactor(S->first);
2169 DEBUG(if (DebugPairSelection)
2170 dbgs() << "BBV: found pruned DAG for pair {"
2171 << *IJ.first << " <-> " << *IJ.second << "} of depth " <<
2172 MaxDepth << " and size " << PrunedDAG.size() <<
2173 " (effective size: " << EffSize << ")\n");
2174 if (((TTI && !UseChainDepthWithTI) ||
2175 MaxDepth >= Config.ReqChainDepth) &&
2176 EffSize > 0 && EffSize > BestEffSize) {
2177 BestMaxDepth = MaxDepth;
2178 BestEffSize = EffSize;
2179 BestDAG = PrunedDAG;
2184 // Given the list of candidate pairs, this function selects those
2185 // that will be fused into vector instructions.
2186 void BBVectorize::choosePairs(
2187 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
2188 DenseSet<ValuePair> &CandidatePairsSet,
2189 DenseMap<ValuePair, int> &CandidatePairCostSavings,
2190 std::vector<Value *> &PairableInsts,
2191 DenseSet<ValuePair> &FixedOrderPairs,
2192 DenseMap<VPPair, unsigned> &PairConnectionTypes,
2193 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
2194 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
2195 DenseSet<ValuePair> &PairableInstUsers,
2196 DenseMap<Value *, Value *>& ChosenPairs) {
2197 bool UseCycleCheck =
2198 CandidatePairsSet.size() <= Config.MaxCandPairsForCycleCheck;
2200 DenseMap<Value *, std::vector<Value *> > CandidatePairs2;
2201 for (DenseSet<ValuePair>::iterator I = CandidatePairsSet.begin(),
2202 E = CandidatePairsSet.end(); I != E; ++I) {
2203 std::vector<Value *> &JJ = CandidatePairs2[I->second];
2204 if (JJ.empty()) JJ.reserve(32);
2205 JJ.push_back(I->first);
2208 DenseMap<ValuePair, std::vector<ValuePair> > PairableInstUserMap;
2209 DenseSet<VPPair> PairableInstUserPairSet;
2210 for (std::vector<Value *>::iterator I = PairableInsts.begin(),
2211 E = PairableInsts.end(); I != E; ++I) {
2212 // The number of possible pairings for this variable:
2213 size_t NumChoices = CandidatePairs.lookup(*I).size();
2214 if (!NumChoices) continue;
2216 std::vector<Value *> &JJ = CandidatePairs[*I];
2218 // The best pair to choose and its dag:
2219 size_t BestMaxDepth = 0;
2220 int BestEffSize = 0;
2221 DenseSet<ValuePair> BestDAG;
2222 findBestDAGFor(CandidatePairs, CandidatePairsSet,
2223 CandidatePairCostSavings,
2224 PairableInsts, FixedOrderPairs, PairConnectionTypes,
2225 ConnectedPairs, ConnectedPairDeps,
2226 PairableInstUsers, PairableInstUserMap,
2227 PairableInstUserPairSet, ChosenPairs,
2228 BestDAG, BestMaxDepth, BestEffSize, *I, JJ,
2231 if (BestDAG.empty())
2234 // A dag has been chosen (or not) at this point. If no dag was
2235 // chosen, then this instruction, I, cannot be paired (and is no longer
2238 DEBUG(dbgs() << "BBV: selected pairs in the best DAG for: "
2239 << *cast<Instruction>(*I) << "\n");
2241 for (DenseSet<ValuePair>::iterator S = BestDAG.begin(),
2242 SE2 = BestDAG.end(); S != SE2; ++S) {
2243 // Insert the members of this dag into the list of chosen pairs.
2244 ChosenPairs.insert(ValuePair(S->first, S->second));
2245 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " <<
2246 *S->second << "\n");
2248 // Remove all candidate pairs that have values in the chosen dag.
2249 std::vector<Value *> &KK = CandidatePairs[S->first];
2250 for (std::vector<Value *>::iterator K = KK.begin(), KE = KK.end();
2252 if (*K == S->second)
2255 CandidatePairsSet.erase(ValuePair(S->first, *K));
2258 std::vector<Value *> &LL = CandidatePairs2[S->second];
2259 for (std::vector<Value *>::iterator L = LL.begin(), LE = LL.end();
2264 CandidatePairsSet.erase(ValuePair(*L, S->second));
2267 std::vector<Value *> &MM = CandidatePairs[S->second];
2268 for (std::vector<Value *>::iterator M = MM.begin(), ME = MM.end();
2270 assert(*M != S->first && "Flipped pair in candidate list?");
2271 CandidatePairsSet.erase(ValuePair(S->second, *M));
2274 std::vector<Value *> &NN = CandidatePairs2[S->first];
2275 for (std::vector<Value *>::iterator N = NN.begin(), NE = NN.end();
2277 assert(*N != S->second && "Flipped pair in candidate list?");
2278 CandidatePairsSet.erase(ValuePair(*N, S->first));
2283 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n");
2286 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o,
2291 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) +
2292 (n > 0 ? "." + utostr(n) : "")).str();
2295 // Returns the value that is to be used as the pointer input to the vector
2296 // instruction that fuses I with J.
2297 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context,
2298 Instruction *I, Instruction *J, unsigned o) {
2300 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
2301 int64_t OffsetInElmts;
2303 // Note: the analysis might fail here, that is why the pair order has
2304 // been precomputed (OffsetInElmts must be unused here).
2305 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
2306 IAddressSpace, JAddressSpace,
2307 OffsetInElmts, false);
2309 // The pointer value is taken to be the one with the lowest offset.
2312 Type *ArgTypeI = IPtr->getType()->getPointerElementType();
2313 Type *ArgTypeJ = JPtr->getType()->getPointerElementType();
2314 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2316 = PointerType::get(VArgType,
2317 IPtr->getType()->getPointerAddressSpace());
2318 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o),
2319 /* insert before */ I);
2322 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J,
2323 unsigned MaskOffset, unsigned NumInElem,
2324 unsigned NumInElem1, unsigned IdxOffset,
2325 std::vector<Constant*> &Mask) {
2326 unsigned NumElem1 = J->getType()->getVectorNumElements();
2327 for (unsigned v = 0; v < NumElem1; ++v) {
2328 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v);
2330 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context));
2332 unsigned mm = m + (int) IdxOffset;
2333 if (m >= (int) NumInElem1)
2334 mm += (int) NumInElem;
2336 Mask[v+MaskOffset] =
2337 ConstantInt::get(Type::getInt32Ty(Context), mm);
2342 // Returns the value that is to be used as the vector-shuffle mask to the
2343 // vector instruction that fuses I with J.
2344 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context,
2345 Instruction *I, Instruction *J) {
2346 // This is the shuffle mask. We need to append the second
2347 // mask to the first, and the numbers need to be adjusted.
2349 Type *ArgTypeI = I->getType();
2350 Type *ArgTypeJ = J->getType();
2351 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2353 unsigned NumElemI = ArgTypeI->getVectorNumElements();
2355 // Get the total number of elements in the fused vector type.
2356 // By definition, this must equal the number of elements in
2358 unsigned NumElem = VArgType->getVectorNumElements();
2359 std::vector<Constant*> Mask(NumElem);
2361 Type *OpTypeI = I->getOperand(0)->getType();
2362 unsigned NumInElemI = OpTypeI->getVectorNumElements();
2363 Type *OpTypeJ = J->getOperand(0)->getType();
2364 unsigned NumInElemJ = OpTypeJ->getVectorNumElements();
2366 // The fused vector will be:
2367 // -----------------------------------------------------
2368 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ |
2369 // -----------------------------------------------------
2370 // from which we'll extract NumElem total elements (where the first NumElemI
2371 // of them come from the mask in I and the remainder come from the mask
2374 // For the mask from the first pair...
2375 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI,
2378 // For the mask from the second pair...
2379 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ,
2382 return ConstantVector::get(Mask);
2385 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I,
2386 Instruction *J, unsigned o, Value *&LOp,
2388 Type *ArgTypeL, Type *ArgTypeH,
2389 bool IBeforeJ, unsigned IdxOff) {
2390 bool ExpandedIEChain = false;
2391 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) {
2392 // If we have a pure insertelement chain, then this can be rewritten
2393 // into a chain that directly builds the larger type.
2394 if (isPureIEChain(LIE)) {
2395 SmallVector<Value *, 8> VectElemts(numElemL,
2396 UndefValue::get(ArgTypeL->getScalarType()));
2397 InsertElementInst *LIENext = LIE;
2400 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue();
2401 VectElemts[Idx] = LIENext->getOperand(1);
2403 dyn_cast<InsertElementInst>(LIENext->getOperand(0))));
2406 Value *LIEPrev = UndefValue::get(ArgTypeH);
2407 for (unsigned i = 0; i < numElemL; ++i) {
2408 if (isa<UndefValue>(VectElemts[i])) continue;
2409 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i],
2410 ConstantInt::get(Type::getInt32Ty(Context),
2412 getReplacementName(IBeforeJ ? I : J,
2414 LIENext->insertBefore(IBeforeJ ? J : I);
2418 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH);
2419 ExpandedIEChain = true;
2423 return ExpandedIEChain;
2426 static unsigned getNumScalarElements(Type *Ty) {
2427 if (VectorType *VecTy = dyn_cast<VectorType>(Ty))
2428 return VecTy->getNumElements();
2432 // Returns the value to be used as the specified operand of the vector
2433 // instruction that fuses I with J.
2434 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I,
2435 Instruction *J, unsigned o, bool IBeforeJ) {
2436 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2437 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1);
2439 // Compute the fused vector type for this operand
2440 Type *ArgTypeI = I->getOperand(o)->getType();
2441 Type *ArgTypeJ = J->getOperand(o)->getType();
2442 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2444 Instruction *L = I, *H = J;
2445 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ;
2447 unsigned numElemL = getNumScalarElements(ArgTypeL);
2448 unsigned numElemH = getNumScalarElements(ArgTypeH);
2450 Value *LOp = L->getOperand(o);
2451 Value *HOp = H->getOperand(o);
2452 unsigned numElem = VArgType->getNumElements();
2454 // First, we check if we can reuse the "original" vector outputs (if these
2455 // exist). We might need a shuffle.
2456 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp);
2457 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp);
2458 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp);
2459 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp);
2461 // FIXME: If we're fusing shuffle instructions, then we can't apply this
2462 // optimization. The input vectors to the shuffle might be a different
2463 // length from the shuffle outputs. Unfortunately, the replacement
2464 // shuffle mask has already been formed, and the mask entries are sensitive
2465 // to the sizes of the inputs.
2466 bool IsSizeChangeShuffle =
2467 isa<ShuffleVectorInst>(L) &&
2468 (LOp->getType() != L->getType() || HOp->getType() != H->getType());
2470 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) {
2471 // We can have at most two unique vector inputs.
2472 bool CanUseInputs = true;
2473 Value *I1, *I2 = nullptr;
2475 I1 = LEE->getOperand(0);
2477 I1 = LSV->getOperand(0);
2478 I2 = LSV->getOperand(1);
2479 if (I2 == I1 || isa<UndefValue>(I2))
2484 Value *I3 = HEE->getOperand(0);
2485 if (!I2 && I3 != I1)
2487 else if (I3 != I1 && I3 != I2)
2488 CanUseInputs = false;
2490 Value *I3 = HSV->getOperand(0);
2491 if (!I2 && I3 != I1)
2493 else if (I3 != I1 && I3 != I2)
2494 CanUseInputs = false;
2497 Value *I4 = HSV->getOperand(1);
2498 if (!isa<UndefValue>(I4)) {
2499 if (!I2 && I4 != I1)
2501 else if (I4 != I1 && I4 != I2)
2502 CanUseInputs = false;
2509 cast<Instruction>(LOp)->getOperand(0)->getType()
2510 ->getVectorNumElements();
2513 cast<Instruction>(HOp)->getOperand(0)->getType()
2514 ->getVectorNumElements();
2516 // We have one or two input vectors. We need to map each index of the
2517 // operands to the index of the original vector.
2518 SmallVector<std::pair<int, int>, 8> II(numElem);
2519 for (unsigned i = 0; i < numElemL; ++i) {
2523 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue();
2524 INum = LEE->getOperand(0) == I1 ? 0 : 1;
2526 Idx = LSV->getMaskValue(i);
2527 if (Idx < (int) LOpElem) {
2528 INum = LSV->getOperand(0) == I1 ? 0 : 1;
2531 INum = LSV->getOperand(1) == I1 ? 0 : 1;
2535 II[i] = std::pair<int, int>(Idx, INum);
2537 for (unsigned i = 0; i < numElemH; ++i) {
2541 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue();
2542 INum = HEE->getOperand(0) == I1 ? 0 : 1;
2544 Idx = HSV->getMaskValue(i);
2545 if (Idx < (int) HOpElem) {
2546 INum = HSV->getOperand(0) == I1 ? 0 : 1;
2549 INum = HSV->getOperand(1) == I1 ? 0 : 1;
2553 II[i + numElemL] = std::pair<int, int>(Idx, INum);
2556 // We now have an array which tells us from which index of which
2557 // input vector each element of the operand comes.
2558 VectorType *I1T = cast<VectorType>(I1->getType());
2559 unsigned I1Elem = I1T->getNumElements();
2562 // In this case there is only one underlying vector input. Check for
2563 // the trivial case where we can use the input directly.
2564 if (I1Elem == numElem) {
2565 bool ElemInOrder = true;
2566 for (unsigned i = 0; i < numElem; ++i) {
2567 if (II[i].first != (int) i && II[i].first != -1) {
2568 ElemInOrder = false;
2577 // A shuffle is needed.
2578 std::vector<Constant *> Mask(numElem);
2579 for (unsigned i = 0; i < numElem; ++i) {
2580 int Idx = II[i].first;
2582 Mask[i] = UndefValue::get(Type::getInt32Ty(Context));
2584 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2588 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2589 ConstantVector::get(Mask),
2590 getReplacementName(IBeforeJ ? I : J,
2592 S->insertBefore(IBeforeJ ? J : I);
2596 VectorType *I2T = cast<VectorType>(I2->getType());
2597 unsigned I2Elem = I2T->getNumElements();
2599 // This input comes from two distinct vectors. The first step is to
2600 // make sure that both vectors are the same length. If not, the
2601 // smaller one will need to grow before they can be shuffled together.
2602 if (I1Elem < I2Elem) {
2603 std::vector<Constant *> Mask(I2Elem);
2605 for (; v < I1Elem; ++v)
2606 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2607 for (; v < I2Elem; ++v)
2608 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2610 Instruction *NewI1 =
2611 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2612 ConstantVector::get(Mask),
2613 getReplacementName(IBeforeJ ? I : J,
2615 NewI1->insertBefore(IBeforeJ ? J : I);
2618 } else if (I1Elem > I2Elem) {
2619 std::vector<Constant *> Mask(I1Elem);
2621 for (; v < I2Elem; ++v)
2622 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2623 for (; v < I1Elem; ++v)
2624 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2626 Instruction *NewI2 =
2627 new ShuffleVectorInst(I2, UndefValue::get(I2T),
2628 ConstantVector::get(Mask),
2629 getReplacementName(IBeforeJ ? I : J,
2631 NewI2->insertBefore(IBeforeJ ? J : I);
2635 // Now that both I1 and I2 are the same length we can shuffle them
2636 // together (and use the result).
2637 std::vector<Constant *> Mask(numElem);
2638 for (unsigned v = 0; v < numElem; ++v) {
2639 if (II[v].first == -1) {
2640 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2642 int Idx = II[v].first + II[v].second * I1Elem;
2643 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2647 Instruction *NewOp =
2648 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask),
2649 getReplacementName(IBeforeJ ? I : J, true, o));
2650 NewOp->insertBefore(IBeforeJ ? J : I);
2655 Type *ArgType = ArgTypeL;
2656 if (numElemL < numElemH) {
2657 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH,
2658 ArgTypeL, VArgType, IBeforeJ, 1)) {
2659 // This is another short-circuit case: we're combining a scalar into
2660 // a vector that is formed by an IE chain. We've just expanded the IE
2661 // chain, now insert the scalar and we're done.
2663 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0,
2664 getReplacementName(IBeforeJ ? I : J, true, o));
2665 S->insertBefore(IBeforeJ ? J : I);
2667 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL,
2668 ArgTypeH, IBeforeJ)) {
2669 // The two vector inputs to the shuffle must be the same length,
2670 // so extend the smaller vector to be the same length as the larger one.
2674 std::vector<Constant *> Mask(numElemH);
2676 for (; v < numElemL; ++v)
2677 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2678 for (; v < numElemH; ++v)
2679 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2681 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL),
2682 ConstantVector::get(Mask),
2683 getReplacementName(IBeforeJ ? I : J,
2686 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0,
2687 getReplacementName(IBeforeJ ? I : J,
2691 NLOp->insertBefore(IBeforeJ ? J : I);
2696 } else if (numElemL > numElemH) {
2697 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL,
2698 ArgTypeH, VArgType, IBeforeJ)) {
2700 InsertElementInst::Create(LOp, HOp,
2701 ConstantInt::get(Type::getInt32Ty(Context),
2703 getReplacementName(IBeforeJ ? I : J,
2705 S->insertBefore(IBeforeJ ? J : I);
2707 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH,
2708 ArgTypeL, IBeforeJ)) {
2711 std::vector<Constant *> Mask(numElemL);
2713 for (; v < numElemH; ++v)
2714 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2715 for (; v < numElemL; ++v)
2716 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2718 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH),
2719 ConstantVector::get(Mask),
2720 getReplacementName(IBeforeJ ? I : J,
2723 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0,
2724 getReplacementName(IBeforeJ ? I : J,
2728 NHOp->insertBefore(IBeforeJ ? J : I);
2733 if (ArgType->isVectorTy()) {
2734 unsigned numElem = VArgType->getVectorNumElements();
2735 std::vector<Constant*> Mask(numElem);
2736 for (unsigned v = 0; v < numElem; ++v) {
2738 // If the low vector was expanded, we need to skip the extra
2739 // undefined entries.
2740 if (v >= numElemL && numElemH > numElemL)
2741 Idx += (numElemH - numElemL);
2742 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2745 Instruction *BV = new ShuffleVectorInst(LOp, HOp,
2746 ConstantVector::get(Mask),
2747 getReplacementName(IBeforeJ ? I : J, true, o));
2748 BV->insertBefore(IBeforeJ ? J : I);
2752 Instruction *BV1 = InsertElementInst::Create(
2753 UndefValue::get(VArgType), LOp, CV0,
2754 getReplacementName(IBeforeJ ? I : J,
2756 BV1->insertBefore(IBeforeJ ? J : I);
2757 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1,
2758 getReplacementName(IBeforeJ ? I : J,
2760 BV2->insertBefore(IBeforeJ ? J : I);
2764 // This function creates an array of values that will be used as the inputs
2765 // to the vector instruction that fuses I with J.
2766 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context,
2767 Instruction *I, Instruction *J,
2768 SmallVectorImpl<Value *> &ReplacedOperands,
2770 unsigned NumOperands = I->getNumOperands();
2772 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) {
2773 // Iterate backward so that we look at the store pointer
2774 // first and know whether or not we need to flip the inputs.
2776 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) {
2777 // This is the pointer for a load/store instruction.
2778 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o);
2780 } else if (isa<CallInst>(I)) {
2781 Function *F = cast<CallInst>(I)->getCalledFunction();
2782 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID();
2783 if (o == NumOperands-1) {
2784 BasicBlock &BB = *I->getParent();
2786 Module *M = BB.getParent()->getParent();
2787 Type *ArgTypeI = I->getType();
2788 Type *ArgTypeJ = J->getType();
2789 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2791 ReplacedOperands[o] = Intrinsic::getDeclaration(M, IID, VArgType);
2793 } else if ((IID == Intrinsic::powi || IID == Intrinsic::ctlz ||
2794 IID == Intrinsic::cttz) && o == 1) {
2795 // The second argument of powi/ctlz/cttz is a single integer/constant
2796 // and we've already checked that both arguments are equal.
2797 // As a result, we just keep I's second argument.
2798 ReplacedOperands[o] = I->getOperand(o);
2801 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) {
2802 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J);
2806 ReplacedOperands[o] = getReplacementInput(Context, I, J, o, IBeforeJ);
2810 // This function creates two values that represent the outputs of the
2811 // original I and J instructions. These are generally vector shuffles
2812 // or extracts. In many cases, these will end up being unused and, thus,
2813 // eliminated by later passes.
2814 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
2815 Instruction *J, Instruction *K,
2816 Instruction *&InsertionPt,
2817 Instruction *&K1, Instruction *&K2) {
2818 if (isa<StoreInst>(I)) {
2819 AA->replaceWithNewValue(I, K);
2820 AA->replaceWithNewValue(J, K);
2822 Type *IType = I->getType();
2823 Type *JType = J->getType();
2825 VectorType *VType = getVecTypeForPair(IType, JType);
2826 unsigned numElem = VType->getNumElements();
2828 unsigned numElemI = getNumScalarElements(IType);
2829 unsigned numElemJ = getNumScalarElements(JType);
2831 if (IType->isVectorTy()) {
2832 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI);
2833 for (unsigned v = 0; v < numElemI; ++v) {
2834 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2835 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v);
2838 K1 = new ShuffleVectorInst(K, UndefValue::get(VType),
2839 ConstantVector::get( Mask1),
2840 getReplacementName(K, false, 1));
2842 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2843 K1 = ExtractElementInst::Create(K, CV0,
2844 getReplacementName(K, false, 1));
2847 if (JType->isVectorTy()) {
2848 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ);
2849 for (unsigned v = 0; v < numElemJ; ++v) {
2850 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2851 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v);
2854 K2 = new ShuffleVectorInst(K, UndefValue::get(VType),
2855 ConstantVector::get( Mask2),
2856 getReplacementName(K, false, 2));
2858 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1);
2859 K2 = ExtractElementInst::Create(K, CV1,
2860 getReplacementName(K, false, 2));
2864 K2->insertAfter(K1);
2869 // Move all uses of the function I (including pairing-induced uses) after J.
2870 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB,
2871 DenseSet<ValuePair> &LoadMoveSetPairs,
2872 Instruction *I, Instruction *J) {
2873 // Skip to the first instruction past I.
2874 BasicBlock::iterator L = std::next(BasicBlock::iterator(I));
2876 DenseSet<Value *> Users;
2877 AliasSetTracker WriteSet(*AA);
2878 if (I->mayWriteToMemory()) WriteSet.add(I);
2880 for (; cast<Instruction>(L) != J; ++L)
2881 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs);
2883 assert(cast<Instruction>(L) == J &&
2884 "Tracking has not proceeded far enough to check for dependencies");
2885 // If J is now in the use set of I, then trackUsesOfI will return true
2886 // and we have a dependency cycle (and the fusing operation must abort).
2887 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSetPairs);
2890 // Move all uses of the function I (including pairing-induced uses) after J.
2891 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB,
2892 DenseSet<ValuePair> &LoadMoveSetPairs,
2893 Instruction *&InsertionPt,
2894 Instruction *I, Instruction *J) {
2895 // Skip to the first instruction past I.
2896 BasicBlock::iterator L = std::next(BasicBlock::iterator(I));
2898 DenseSet<Value *> Users;
2899 AliasSetTracker WriteSet(*AA);
2900 if (I->mayWriteToMemory()) WriteSet.add(I);
2902 for (; cast<Instruction>(L) != J;) {
2903 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs)) {
2904 // Move this instruction
2905 Instruction *InstToMove = L; ++L;
2907 DEBUG(dbgs() << "BBV: moving: " << *InstToMove <<
2908 " to after " << *InsertionPt << "\n");
2909 InstToMove->removeFromParent();
2910 InstToMove->insertAfter(InsertionPt);
2911 InsertionPt = InstToMove;
2918 // Collect all load instruction that are in the move set of a given first
2919 // pair member. These loads depend on the first instruction, I, and so need
2920 // to be moved after J (the second instruction) when the pair is fused.
2921 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB,
2922 DenseMap<Value *, Value *> &ChosenPairs,
2923 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
2924 DenseSet<ValuePair> &LoadMoveSetPairs,
2926 // Skip to the first instruction past I.
2927 BasicBlock::iterator L = std::next(BasicBlock::iterator(I));
2929 DenseSet<Value *> Users;
2930 AliasSetTracker WriteSet(*AA);
2931 if (I->mayWriteToMemory()) WriteSet.add(I);
2933 // Note: We cannot end the loop when we reach J because J could be moved
2934 // farther down the use chain by another instruction pairing. Also, J
2935 // could be before I if this is an inverted input.
2936 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) {
2937 if (trackUsesOfI(Users, WriteSet, I, L)) {
2938 if (L->mayReadFromMemory()) {
2939 LoadMoveSet[L].push_back(I);
2940 LoadMoveSetPairs.insert(ValuePair(L, I));
2946 // In cases where both load/stores and the computation of their pointers
2947 // are chosen for vectorization, we can end up in a situation where the
2948 // aliasing analysis starts returning different query results as the
2949 // process of fusing instruction pairs continues. Because the algorithm
2950 // relies on finding the same use dags here as were found earlier, we'll
2951 // need to precompute the necessary aliasing information here and then
2952 // manually update it during the fusion process.
2953 void BBVectorize::collectLoadMoveSet(BasicBlock &BB,
2954 std::vector<Value *> &PairableInsts,
2955 DenseMap<Value *, Value *> &ChosenPairs,
2956 DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
2957 DenseSet<ValuePair> &LoadMoveSetPairs) {
2958 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
2959 PIE = PairableInsts.end(); PI != PIE; ++PI) {
2960 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI);
2961 if (P == ChosenPairs.end()) continue;
2963 Instruction *I = cast<Instruction>(P->first);
2964 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet,
2965 LoadMoveSetPairs, I);
2969 // This function fuses the chosen instruction pairs into vector instructions,
2970 // taking care preserve any needed scalar outputs and, then, it reorders the
2971 // remaining instructions as needed (users of the first member of the pair
2972 // need to be moved to after the location of the second member of the pair
2973 // because the vector instruction is inserted in the location of the pair's
2975 void BBVectorize::fuseChosenPairs(BasicBlock &BB,
2976 std::vector<Value *> &PairableInsts,
2977 DenseMap<Value *, Value *> &ChosenPairs,
2978 DenseSet<ValuePair> &FixedOrderPairs,
2979 DenseMap<VPPair, unsigned> &PairConnectionTypes,
2980 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
2981 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps) {
2982 LLVMContext& Context = BB.getContext();
2984 // During the vectorization process, the order of the pairs to be fused
2985 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs
2986 // list. After a pair is fused, the flipped pair is removed from the list.
2987 DenseSet<ValuePair> FlippedPairs;
2988 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(),
2989 E = ChosenPairs.end(); P != E; ++P)
2990 FlippedPairs.insert(ValuePair(P->second, P->first));
2991 for (DenseSet<ValuePair>::iterator P = FlippedPairs.begin(),
2992 E = FlippedPairs.end(); P != E; ++P)
2993 ChosenPairs.insert(*P);
2995 DenseMap<Value *, std::vector<Value *> > LoadMoveSet;
2996 DenseSet<ValuePair> LoadMoveSetPairs;
2997 collectLoadMoveSet(BB, PairableInsts, ChosenPairs,
2998 LoadMoveSet, LoadMoveSetPairs);
3000 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n");
3002 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) {
3003 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI);
3004 if (P == ChosenPairs.end()) {
3009 if (getDepthFactor(P->first) == 0) {
3010 // These instructions are not really fused, but are tracked as though
3011 // they are. Any case in which it would be interesting to fuse them
3012 // will be taken care of by InstCombine.
3018 Instruction *I = cast<Instruction>(P->first),
3019 *J = cast<Instruction>(P->second);
3021 DEBUG(dbgs() << "BBV: fusing: " << *I <<
3022 " <-> " << *J << "\n");
3024 // Remove the pair and flipped pair from the list.
3025 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second);
3026 assert(FP != ChosenPairs.end() && "Flipped pair not found in list");
3027 ChosenPairs.erase(FP);
3028 ChosenPairs.erase(P);
3030 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSetPairs, I, J)) {
3031 DEBUG(dbgs() << "BBV: fusion of: " << *I <<
3033 " aborted because of non-trivial dependency cycle\n");
3039 // If the pair must have the other order, then flip it.
3040 bool FlipPairOrder = FixedOrderPairs.count(ValuePair(J, I));
3041 if (!FlipPairOrder && !FixedOrderPairs.count(ValuePair(I, J))) {
3042 // This pair does not have a fixed order, and so we might want to
3043 // flip it if that will yield fewer shuffles. We count the number
3044 // of dependencies connected via swaps, and those directly connected,
3045 // and flip the order if the number of swaps is greater.
3046 bool OrigOrder = true;
3047 DenseMap<ValuePair, std::vector<ValuePair> >::iterator IJ =
3048 ConnectedPairDeps.find(ValuePair(I, J));
3049 if (IJ == ConnectedPairDeps.end()) {
3050 IJ = ConnectedPairDeps.find(ValuePair(J, I));
3054 if (IJ != ConnectedPairDeps.end()) {
3055 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
3056 for (std::vector<ValuePair>::iterator T = IJ->second.begin(),
3057 TE = IJ->second.end(); T != TE; ++T) {
3058 VPPair Q(IJ->first, *T);
3059 DenseMap<VPPair, unsigned>::iterator R =
3060 PairConnectionTypes.find(VPPair(Q.second, Q.first));
3061 assert(R != PairConnectionTypes.end() &&
3062 "Cannot find pair connection type");
3063 if (R->second == PairConnectionDirect)
3065 else if (R->second == PairConnectionSwap)
3070 std::swap(NumDepsDirect, NumDepsSwap);
3072 if (NumDepsSwap > NumDepsDirect) {
3073 FlipPairOrder = true;
3074 DEBUG(dbgs() << "BBV: reordering pair: " << *I <<
3075 " <-> " << *J << "\n");
3080 Instruction *L = I, *H = J;
3084 // If the pair being fused uses the opposite order from that in the pair
3085 // connection map, then we need to flip the types.
3086 DenseMap<ValuePair, std::vector<ValuePair> >::iterator HL =
3087 ConnectedPairs.find(ValuePair(H, L));
3088 if (HL != ConnectedPairs.end())
3089 for (std::vector<ValuePair>::iterator T = HL->second.begin(),
3090 TE = HL->second.end(); T != TE; ++T) {
3091 VPPair Q(HL->first, *T);
3092 DenseMap<VPPair, unsigned>::iterator R = PairConnectionTypes.find(Q);
3093 assert(R != PairConnectionTypes.end() &&
3094 "Cannot find pair connection type");
3095 if (R->second == PairConnectionDirect)
3096 R->second = PairConnectionSwap;
3097 else if (R->second == PairConnectionSwap)
3098 R->second = PairConnectionDirect;
3101 bool LBeforeH = !FlipPairOrder;
3102 unsigned NumOperands = I->getNumOperands();
3103 SmallVector<Value *, 3> ReplacedOperands(NumOperands);
3104 getReplacementInputsForPair(Context, L, H, ReplacedOperands,
3107 // Make a copy of the original operation, change its type to the vector
3108 // type and replace its operands with the vector operands.
3109 Instruction *K = L->clone();
3112 else if (H->hasName())
3115 if (!isa<StoreInst>(K))
3116 K->mutateType(getVecTypeForPair(L->getType(), H->getType()));
3118 unsigned KnownIDs[] = {
3119 LLVMContext::MD_tbaa,
3120 LLVMContext::MD_alias_scope,
3121 LLVMContext::MD_noalias,
3122 LLVMContext::MD_fpmath
3124 combineMetadata(K, H, KnownIDs);
3125 K->intersectOptionalDataWith(H);
3127 for (unsigned o = 0; o < NumOperands; ++o)
3128 K->setOperand(o, ReplacedOperands[o]);
3132 // Instruction insertion point:
3133 Instruction *InsertionPt = K;
3134 Instruction *K1 = nullptr, *K2 = nullptr;
3135 replaceOutputsOfPair(Context, L, H, K, InsertionPt, K1, K2);
3137 // The use dag of the first original instruction must be moved to after
3138 // the location of the second instruction. The entire use dag of the
3139 // first instruction is disjoint from the input dag of the second
3140 // (by definition), and so commutes with it.
3142 moveUsesOfIAfterJ(BB, LoadMoveSetPairs, InsertionPt, I, J);
3144 if (!isa<StoreInst>(I)) {
3145 L->replaceAllUsesWith(K1);
3146 H->replaceAllUsesWith(K2);
3147 AA->replaceWithNewValue(L, K1);
3148 AA->replaceWithNewValue(H, K2);
3151 // Instructions that may read from memory may be in the load move set.
3152 // Once an instruction is fused, we no longer need its move set, and so
3153 // the values of the map never need to be updated. However, when a load
3154 // is fused, we need to merge the entries from both instructions in the
3155 // pair in case those instructions were in the move set of some other
3156 // yet-to-be-fused pair. The loads in question are the keys of the map.
3157 if (I->mayReadFromMemory()) {
3158 std::vector<ValuePair> NewSetMembers;
3159 DenseMap<Value *, std::vector<Value *> >::iterator II =
3160 LoadMoveSet.find(I);
3161 if (II != LoadMoveSet.end())
3162 for (std::vector<Value *>::iterator N = II->second.begin(),
3163 NE = II->second.end(); N != NE; ++N)
3164 NewSetMembers.push_back(ValuePair(K, *N));
3165 DenseMap<Value *, std::vector<Value *> >::iterator JJ =
3166 LoadMoveSet.find(J);
3167 if (JJ != LoadMoveSet.end())
3168 for (std::vector<Value *>::iterator N = JJ->second.begin(),
3169 NE = JJ->second.end(); N != NE; ++N)
3170 NewSetMembers.push_back(ValuePair(K, *N));
3171 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(),
3172 AE = NewSetMembers.end(); A != AE; ++A) {
3173 LoadMoveSet[A->first].push_back(A->second);
3174 LoadMoveSetPairs.insert(*A);
3178 // Before removing I, set the iterator to the next instruction.
3179 PI = std::next(BasicBlock::iterator(I));
3180 if (cast<Instruction>(PI) == J)
3185 I->eraseFromParent();
3186 J->eraseFromParent();
3188 DEBUG(if (PrintAfterEveryPair) dbgs() << "BBV: block is now: \n" <<
3192 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n");
3196 char BBVectorize::ID = 0;
3197 static const char bb_vectorize_name[] = "Basic-Block Vectorization";
3198 INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
3199 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
3200 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
3201 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
3202 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3203 INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
3205 BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) {
3206 return new BBVectorize(C);
3210 llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) {
3211 BBVectorize BBVectorizer(P, *BB.getParent(), C);
3212 return BBVectorizer.vectorizeBB(BB);
3215 //===----------------------------------------------------------------------===//
3216 VectorizeConfig::VectorizeConfig() {
3217 VectorBits = ::VectorBits;
3218 VectorizeBools = !::NoBools;
3219 VectorizeInts = !::NoInts;
3220 VectorizeFloats = !::NoFloats;
3221 VectorizePointers = !::NoPointers;
3222 VectorizeCasts = !::NoCasts;
3223 VectorizeMath = !::NoMath;
3224 VectorizeBitManipulations = !::NoBitManipulation;
3225 VectorizeFMA = !::NoFMA;
3226 VectorizeSelect = !::NoSelect;
3227 VectorizeCmp = !::NoCmp;
3228 VectorizeGEP = !::NoGEP;
3229 VectorizeMemOps = !::NoMemOps;
3230 AlignedOnly = ::AlignedOnly;
3231 ReqChainDepth= ::ReqChainDepth;
3232 SearchLimit = ::SearchLimit;
3233 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck;
3234 SplatBreaksChain = ::SplatBreaksChain;
3235 MaxInsts = ::MaxInsts;
3236 MaxPairs = ::MaxPairs;
3237 MaxIter = ::MaxIter;
3238 Pow2LenOnly = ::Pow2LenOnly;
3239 NoMemOpBoost = ::NoMemOpBoost;
3240 FastDep = ::FastDep;