1 //===- BBVectorize.cpp - A Basic-Block Vectorizer -------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements a basic-block vectorization pass. The algorithm was
11 // inspired by that used by the Vienna MAP Vectorizor by Franchetti and Kral,
12 // et al. It works by looking for chains of pairable operations and then
15 //===----------------------------------------------------------------------===//
17 #define BBV_NAME "bb-vectorize"
18 #define DEBUG_TYPE BBV_NAME
19 #include "llvm/Transforms/Vectorize.h"
20 #include "llvm/ADT/DenseMap.h"
21 #include "llvm/ADT/DenseSet.h"
22 #include "llvm/ADT/STLExtras.h"
23 #include "llvm/ADT/SmallSet.h"
24 #include "llvm/ADT/SmallVector.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/ADT/StringExtras.h"
27 #include "llvm/Analysis/AliasAnalysis.h"
28 #include "llvm/Analysis/AliasSetTracker.h"
29 #include "llvm/Analysis/Dominators.h"
30 #include "llvm/Analysis/ScalarEvolution.h"
31 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
32 #include "llvm/Analysis/TargetTransformInfo.h"
33 #include "llvm/Analysis/ValueTracking.h"
34 #include "llvm/IR/Constants.h"
35 #include "llvm/IR/DataLayout.h"
36 #include "llvm/IR/DerivedTypes.h"
37 #include "llvm/IR/Function.h"
38 #include "llvm/IR/Instructions.h"
39 #include "llvm/IR/IntrinsicInst.h"
40 #include "llvm/IR/Intrinsics.h"
41 #include "llvm/IR/LLVMContext.h"
42 #include "llvm/IR/Metadata.h"
43 #include "llvm/IR/Type.h"
44 #include "llvm/Pass.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/ValueHandle.h"
48 #include "llvm/Support/raw_ostream.h"
49 #include "llvm/Transforms/Utils/Local.h"
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 MaxCandPairsForCycleCheck("bb-vectorize-max-cycle-check-pairs", cl::init(200),
93 cl::Hidden, cl::desc("The maximum number of candidate pairs with which to use"
94 " a full cycle check"));
97 NoBools("bb-vectorize-no-bools", cl::init(false), cl::Hidden,
98 cl::desc("Don't try to vectorize boolean (i1) values"));
101 NoInts("bb-vectorize-no-ints", cl::init(false), cl::Hidden,
102 cl::desc("Don't try to vectorize integer values"));
105 NoFloats("bb-vectorize-no-floats", cl::init(false), cl::Hidden,
106 cl::desc("Don't try to vectorize floating-point values"));
108 // FIXME: This should default to false once pointer vector support works.
110 NoPointers("bb-vectorize-no-pointers", cl::init(/*false*/ true), cl::Hidden,
111 cl::desc("Don't try to vectorize pointer values"));
114 NoCasts("bb-vectorize-no-casts", cl::init(false), cl::Hidden,
115 cl::desc("Don't try to vectorize casting (conversion) operations"));
118 NoMath("bb-vectorize-no-math", cl::init(false), cl::Hidden,
119 cl::desc("Don't try to vectorize floating-point math intrinsics"));
122 NoFMA("bb-vectorize-no-fma", cl::init(false), cl::Hidden,
123 cl::desc("Don't try to vectorize the fused-multiply-add intrinsic"));
126 NoSelect("bb-vectorize-no-select", cl::init(false), cl::Hidden,
127 cl::desc("Don't try to vectorize select instructions"));
130 NoCmp("bb-vectorize-no-cmp", cl::init(false), cl::Hidden,
131 cl::desc("Don't try to vectorize comparison instructions"));
134 NoGEP("bb-vectorize-no-gep", cl::init(false), cl::Hidden,
135 cl::desc("Don't try to vectorize getelementptr instructions"));
138 NoMemOps("bb-vectorize-no-mem-ops", cl::init(false), cl::Hidden,
139 cl::desc("Don't try to vectorize loads and stores"));
142 AlignedOnly("bb-vectorize-aligned-only", cl::init(false), cl::Hidden,
143 cl::desc("Only generate aligned loads and stores"));
146 NoMemOpBoost("bb-vectorize-no-mem-op-boost",
147 cl::init(false), cl::Hidden,
148 cl::desc("Don't boost the chain-depth contribution of loads and stores"));
151 FastDep("bb-vectorize-fast-dep", cl::init(false), cl::Hidden,
152 cl::desc("Use a fast instruction dependency analysis"));
156 DebugInstructionExamination("bb-vectorize-debug-instruction-examination",
157 cl::init(false), cl::Hidden,
158 cl::desc("When debugging is enabled, output information on the"
159 " instruction-examination process"));
161 DebugCandidateSelection("bb-vectorize-debug-candidate-selection",
162 cl::init(false), cl::Hidden,
163 cl::desc("When debugging is enabled, output information on the"
164 " candidate-selection process"));
166 DebugPairSelection("bb-vectorize-debug-pair-selection",
167 cl::init(false), cl::Hidden,
168 cl::desc("When debugging is enabled, output information on the"
169 " pair-selection process"));
171 DebugCycleCheck("bb-vectorize-debug-cycle-check",
172 cl::init(false), cl::Hidden,
173 cl::desc("When debugging is enabled, output information on the"
174 " cycle-checking process"));
177 PrintAfterEveryPair("bb-vectorize-debug-print-after-every-pair",
178 cl::init(false), cl::Hidden,
179 cl::desc("When debugging is enabled, dump the basic block after"
180 " every pair is fused"));
183 STATISTIC(NumFusedOps, "Number of operations fused by bb-vectorize");
186 struct BBVectorize : public BasicBlockPass {
187 static char ID; // Pass identification, replacement for typeid
189 const VectorizeConfig Config;
191 BBVectorize(const VectorizeConfig &C = VectorizeConfig())
192 : BasicBlockPass(ID), Config(C) {
193 initializeBBVectorizePass(*PassRegistry::getPassRegistry());
196 BBVectorize(Pass *P, const VectorizeConfig &C)
197 : BasicBlockPass(ID), Config(C) {
198 AA = &P->getAnalysis<AliasAnalysis>();
199 DT = &P->getAnalysis<DominatorTree>();
200 SE = &P->getAnalysis<ScalarEvolution>();
201 TD = P->getAnalysisIfAvailable<DataLayout>();
202 TTI = IgnoreTargetInfo ? 0 :
203 P->getAnalysisIfAvailable<TargetTransformInfo>();
206 typedef std::pair<Value *, Value *> ValuePair;
207 typedef std::pair<ValuePair, int> ValuePairWithCost;
208 typedef std::pair<ValuePair, size_t> ValuePairWithDepth;
209 typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair
210 typedef std::pair<VPPair, unsigned> VPPairWithType;
211 typedef std::pair<std::multimap<Value *, Value *>::iterator,
212 std::multimap<Value *, Value *>::iterator> VPIteratorPair;
213 typedef std::pair<std::multimap<ValuePair, ValuePair>::iterator,
214 std::multimap<ValuePair, ValuePair>::iterator>
221 const TargetTransformInfo *TTI;
223 // FIXME: const correct?
225 bool vectorizePairs(BasicBlock &BB, bool NonPow2Len = false);
227 bool getCandidatePairs(BasicBlock &BB,
228 BasicBlock::iterator &Start,
229 std::multimap<Value *, Value *> &CandidatePairs,
230 DenseSet<ValuePair> &FixedOrderPairs,
231 DenseMap<ValuePair, int> &CandidatePairCostSavings,
232 std::vector<Value *> &PairableInsts, bool NonPow2Len);
234 // FIXME: The current implementation does not account for pairs that
235 // are connected in multiple ways. For example:
236 // C1 = A1 / A2; C2 = A2 / A1 (which may be both direct and a swap)
237 enum PairConnectionType {
238 PairConnectionDirect,
243 void computeConnectedPairs(std::multimap<Value *, Value *> &CandidatePairs,
244 std::vector<Value *> &PairableInsts,
245 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
246 DenseMap<VPPair, unsigned> &PairConnectionTypes);
248 void buildDepMap(BasicBlock &BB,
249 std::multimap<Value *, Value *> &CandidatePairs,
250 std::vector<Value *> &PairableInsts,
251 DenseSet<ValuePair> &PairableInstUsers);
253 void choosePairs(std::multimap<Value *, Value *> &CandidatePairs,
254 DenseMap<ValuePair, int> &CandidatePairCostSavings,
255 std::vector<Value *> &PairableInsts,
256 DenseSet<ValuePair> &FixedOrderPairs,
257 DenseMap<VPPair, unsigned> &PairConnectionTypes,
258 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
259 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps,
260 DenseSet<ValuePair> &PairableInstUsers,
261 DenseMap<Value *, Value *>& ChosenPairs);
263 void fuseChosenPairs(BasicBlock &BB,
264 std::vector<Value *> &PairableInsts,
265 DenseMap<Value *, Value *>& ChosenPairs,
266 DenseSet<ValuePair> &FixedOrderPairs,
267 DenseMap<VPPair, unsigned> &PairConnectionTypes,
268 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
269 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps);
272 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore);
274 bool areInstsCompatible(Instruction *I, Instruction *J,
275 bool IsSimpleLoadStore, bool NonPow2Len,
276 int &CostSavings, int &FixedOrder);
278 bool trackUsesOfI(DenseSet<Value *> &Users,
279 AliasSetTracker &WriteSet, Instruction *I,
280 Instruction *J, bool UpdateUsers = true,
281 std::multimap<Value *, Value *> *LoadMoveSet = 0);
283 void computePairsConnectedTo(
284 std::multimap<Value *, Value *> &CandidatePairs,
285 std::vector<Value *> &PairableInsts,
286 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
287 DenseMap<VPPair, unsigned> &PairConnectionTypes,
290 bool pairsConflict(ValuePair P, ValuePair Q,
291 DenseSet<ValuePair> &PairableInstUsers,
292 std::multimap<ValuePair, ValuePair> *PairableInstUserMap = 0);
294 bool pairWillFormCycle(ValuePair P,
295 std::multimap<ValuePair, ValuePair> &PairableInstUsers,
296 DenseSet<ValuePair> &CurrentPairs);
299 std::multimap<Value *, Value *> &CandidatePairs,
300 std::vector<Value *> &PairableInsts,
301 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
302 DenseSet<ValuePair> &PairableInstUsers,
303 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
304 DenseMap<Value *, Value *> &ChosenPairs,
305 DenseMap<ValuePair, size_t> &Tree,
306 DenseSet<ValuePair> &PrunedTree, ValuePair J,
309 void buildInitialTreeFor(
310 std::multimap<Value *, Value *> &CandidatePairs,
311 std::vector<Value *> &PairableInsts,
312 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
313 DenseSet<ValuePair> &PairableInstUsers,
314 DenseMap<Value *, Value *> &ChosenPairs,
315 DenseMap<ValuePair, size_t> &Tree, ValuePair J);
317 void findBestTreeFor(
318 std::multimap<Value *, Value *> &CandidatePairs,
319 DenseMap<ValuePair, int> &CandidatePairCostSavings,
320 std::vector<Value *> &PairableInsts,
321 DenseSet<ValuePair> &FixedOrderPairs,
322 DenseMap<VPPair, unsigned> &PairConnectionTypes,
323 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
324 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps,
325 DenseSet<ValuePair> &PairableInstUsers,
326 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
327 DenseMap<Value *, Value *> &ChosenPairs,
328 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth,
329 int &BestEffSize, VPIteratorPair ChoiceRange,
332 Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I,
333 Instruction *J, unsigned o);
335 void fillNewShuffleMask(LLVMContext& Context, Instruction *J,
336 unsigned MaskOffset, unsigned NumInElem,
337 unsigned NumInElem1, unsigned IdxOffset,
338 std::vector<Constant*> &Mask);
340 Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I,
343 bool expandIEChain(LLVMContext& Context, Instruction *I, Instruction *J,
344 unsigned o, Value *&LOp, unsigned numElemL,
345 Type *ArgTypeL, Type *ArgTypeR, bool IBeforeJ,
346 unsigned IdxOff = 0);
348 Value *getReplacementInput(LLVMContext& Context, Instruction *I,
349 Instruction *J, unsigned o, bool IBeforeJ);
351 void getReplacementInputsForPair(LLVMContext& Context, Instruction *I,
352 Instruction *J, SmallVector<Value *, 3> &ReplacedOperands,
355 void replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
356 Instruction *J, Instruction *K,
357 Instruction *&InsertionPt, Instruction *&K1,
360 void collectPairLoadMoveSet(BasicBlock &BB,
361 DenseMap<Value *, Value *> &ChosenPairs,
362 std::multimap<Value *, Value *> &LoadMoveSet,
365 void collectLoadMoveSet(BasicBlock &BB,
366 std::vector<Value *> &PairableInsts,
367 DenseMap<Value *, Value *> &ChosenPairs,
368 std::multimap<Value *, Value *> &LoadMoveSet);
370 bool canMoveUsesOfIAfterJ(BasicBlock &BB,
371 std::multimap<Value *, Value *> &LoadMoveSet,
372 Instruction *I, Instruction *J);
374 void moveUsesOfIAfterJ(BasicBlock &BB,
375 std::multimap<Value *, Value *> &LoadMoveSet,
376 Instruction *&InsertionPt,
377 Instruction *I, Instruction *J);
379 void combineMetadata(Instruction *K, const Instruction *J);
381 bool vectorizeBB(BasicBlock &BB) {
382 if (!DT->isReachableFromEntry(&BB)) {
383 DEBUG(dbgs() << "BBV: skipping unreachable " << BB.getName() <<
384 " in " << BB.getParent()->getName() << "\n");
388 DEBUG(if (TTI) dbgs() << "BBV: using target information\n");
390 bool changed = false;
391 // Iterate a sufficient number of times to merge types of size 1 bit,
392 // then 2 bits, then 4, etc. up to half of the target vector width of the
393 // target vector register.
396 (TTI || v <= Config.VectorBits) &&
397 (!Config.MaxIter || n <= Config.MaxIter);
399 DEBUG(dbgs() << "BBV: fusing loop #" << n <<
400 " for " << BB.getName() << " in " <<
401 BB.getParent()->getName() << "...\n");
402 if (vectorizePairs(BB))
408 if (changed && !Pow2LenOnly) {
410 for (; !Config.MaxIter || n <= Config.MaxIter; ++n) {
411 DEBUG(dbgs() << "BBV: fusing for non-2^n-length vectors loop #: " <<
412 n << " for " << BB.getName() << " in " <<
413 BB.getParent()->getName() << "...\n");
414 if (!vectorizePairs(BB, true)) break;
418 DEBUG(dbgs() << "BBV: done!\n");
422 virtual bool runOnBasicBlock(BasicBlock &BB) {
423 AA = &getAnalysis<AliasAnalysis>();
424 DT = &getAnalysis<DominatorTree>();
425 SE = &getAnalysis<ScalarEvolution>();
426 TD = getAnalysisIfAvailable<DataLayout>();
427 TTI = IgnoreTargetInfo ? 0 :
428 getAnalysisIfAvailable<TargetTransformInfo>();
430 return vectorizeBB(BB);
433 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
434 BasicBlockPass::getAnalysisUsage(AU);
435 AU.addRequired<AliasAnalysis>();
436 AU.addRequired<DominatorTree>();
437 AU.addRequired<ScalarEvolution>();
438 AU.addPreserved<AliasAnalysis>();
439 AU.addPreserved<DominatorTree>();
440 AU.addPreserved<ScalarEvolution>();
441 AU.setPreservesCFG();
444 static inline VectorType *getVecTypeForPair(Type *ElemTy, Type *Elem2Ty) {
445 assert(ElemTy->getScalarType() == Elem2Ty->getScalarType() &&
446 "Cannot form vector from incompatible scalar types");
447 Type *STy = ElemTy->getScalarType();
450 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) {
451 numElem = VTy->getNumElements();
456 if (VectorType *VTy = dyn_cast<VectorType>(Elem2Ty)) {
457 numElem += VTy->getNumElements();
462 return VectorType::get(STy, numElem);
465 static inline void getInstructionTypes(Instruction *I,
466 Type *&T1, Type *&T2) {
467 if (isa<StoreInst>(I)) {
468 // For stores, it is the value type, not the pointer type that matters
469 // because the value is what will come from a vector register.
471 Value *IVal = cast<StoreInst>(I)->getValueOperand();
472 T1 = IVal->getType();
478 T2 = cast<CastInst>(I)->getSrcTy();
482 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
483 T2 = SI->getCondition()->getType();
484 } else if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(I)) {
485 T2 = SI->getOperand(0)->getType();
486 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) {
487 T2 = CI->getOperand(0)->getType();
491 // Returns the weight associated with the provided value. A chain of
492 // candidate pairs has a length given by the sum of the weights of its
493 // members (one weight per pair; the weight of each member of the pair
494 // is assumed to be the same). This length is then compared to the
495 // chain-length threshold to determine if a given chain is significant
496 // enough to be vectorized. The length is also used in comparing
497 // candidate chains where longer chains are considered to be better.
498 // Note: when this function returns 0, the resulting instructions are
499 // not actually fused.
500 inline size_t getDepthFactor(Value *V) {
501 // InsertElement and ExtractElement have a depth factor of zero. This is
502 // for two reasons: First, they cannot be usefully fused. Second, because
503 // the pass generates a lot of these, they can confuse the simple metric
504 // used to compare the trees in the next iteration. Thus, giving them a
505 // weight of zero allows the pass to essentially ignore them in
506 // subsequent iterations when looking for vectorization opportunities
507 // while still tracking dependency chains that flow through those
509 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V))
512 // Give a load or store half of the required depth so that load/store
513 // pairs will vectorize.
514 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V)))
515 return Config.ReqChainDepth/2;
520 // Returns the cost of the provided instruction using TTI.
521 // This does not handle loads and stores.
522 unsigned getInstrCost(unsigned Opcode, Type *T1, Type *T2) {
525 case Instruction::GetElementPtr:
526 // We mark this instruction as zero-cost because scalar GEPs are usually
527 // lowered to the intruction addressing mode. At the moment we don't
528 // generate vector GEPs.
530 case Instruction::Br:
531 return TTI->getCFInstrCost(Opcode);
532 case Instruction::PHI:
534 case Instruction::Add:
535 case Instruction::FAdd:
536 case Instruction::Sub:
537 case Instruction::FSub:
538 case Instruction::Mul:
539 case Instruction::FMul:
540 case Instruction::UDiv:
541 case Instruction::SDiv:
542 case Instruction::FDiv:
543 case Instruction::URem:
544 case Instruction::SRem:
545 case Instruction::FRem:
546 case Instruction::Shl:
547 case Instruction::LShr:
548 case Instruction::AShr:
549 case Instruction::And:
550 case Instruction::Or:
551 case Instruction::Xor:
552 return TTI->getArithmeticInstrCost(Opcode, T1);
553 case Instruction::Select:
554 case Instruction::ICmp:
555 case Instruction::FCmp:
556 return TTI->getCmpSelInstrCost(Opcode, T1, T2);
557 case Instruction::ZExt:
558 case Instruction::SExt:
559 case Instruction::FPToUI:
560 case Instruction::FPToSI:
561 case Instruction::FPExt:
562 case Instruction::PtrToInt:
563 case Instruction::IntToPtr:
564 case Instruction::SIToFP:
565 case Instruction::UIToFP:
566 case Instruction::Trunc:
567 case Instruction::FPTrunc:
568 case Instruction::BitCast:
569 case Instruction::ShuffleVector:
570 return TTI->getCastInstrCost(Opcode, T1, T2);
576 // This determines the relative offset of two loads or stores, returning
577 // true if the offset could be determined to be some constant value.
578 // For example, if OffsetInElmts == 1, then J accesses the memory directly
579 // after I; if OffsetInElmts == -1 then I accesses the memory
581 bool getPairPtrInfo(Instruction *I, Instruction *J,
582 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment,
583 unsigned &IAddressSpace, unsigned &JAddressSpace,
584 int64_t &OffsetInElmts, bool ComputeOffset = true) {
586 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
587 LoadInst *LJ = cast<LoadInst>(J);
588 IPtr = LI->getPointerOperand();
589 JPtr = LJ->getPointerOperand();
590 IAlignment = LI->getAlignment();
591 JAlignment = LJ->getAlignment();
592 IAddressSpace = LI->getPointerAddressSpace();
593 JAddressSpace = LJ->getPointerAddressSpace();
595 StoreInst *SI = cast<StoreInst>(I), *SJ = cast<StoreInst>(J);
596 IPtr = SI->getPointerOperand();
597 JPtr = SJ->getPointerOperand();
598 IAlignment = SI->getAlignment();
599 JAlignment = SJ->getAlignment();
600 IAddressSpace = SI->getPointerAddressSpace();
601 JAddressSpace = SJ->getPointerAddressSpace();
607 const SCEV *IPtrSCEV = SE->getSCEV(IPtr);
608 const SCEV *JPtrSCEV = SE->getSCEV(JPtr);
610 // If this is a trivial offset, then we'll get something like
611 // 1*sizeof(type). With target data, which we need anyway, this will get
612 // constant folded into a number.
613 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV);
614 if (const SCEVConstant *ConstOffSCEV =
615 dyn_cast<SCEVConstant>(OffsetSCEV)) {
616 ConstantInt *IntOff = ConstOffSCEV->getValue();
617 int64_t Offset = IntOff->getSExtValue();
619 Type *VTy = cast<PointerType>(IPtr->getType())->getElementType();
620 int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy);
622 Type *VTy2 = cast<PointerType>(JPtr->getType())->getElementType();
623 if (VTy != VTy2 && Offset < 0) {
624 int64_t VTy2TSS = (int64_t) TD->getTypeStoreSize(VTy2);
625 OffsetInElmts = Offset/VTy2TSS;
626 return (abs64(Offset) % VTy2TSS) == 0;
629 OffsetInElmts = Offset/VTyTSS;
630 return (abs64(Offset) % VTyTSS) == 0;
636 // Returns true if the provided CallInst represents an intrinsic that can
638 bool isVectorizableIntrinsic(CallInst* I) {
639 Function *F = I->getCalledFunction();
640 if (!F) return false;
642 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID();
643 if (!IID) return false;
648 case Intrinsic::sqrt:
649 case Intrinsic::powi:
653 case Intrinsic::log2:
654 case Intrinsic::log10:
656 case Intrinsic::exp2:
658 return Config.VectorizeMath;
660 case Intrinsic::fmuladd:
661 return Config.VectorizeFMA;
665 // Returns true if J is the second element in some pair referenced by
666 // some multimap pair iterator pair.
667 template <typename V>
668 bool isSecondInIteratorPair(V J, std::pair<
669 typename std::multimap<V, V>::iterator,
670 typename std::multimap<V, V>::iterator> PairRange) {
671 for (typename std::multimap<V, V>::iterator K = PairRange.first;
672 K != PairRange.second; ++K)
673 if (K->second == J) return true;
678 bool isPureIEChain(InsertElementInst *IE) {
679 InsertElementInst *IENext = IE;
681 if (!isa<UndefValue>(IENext->getOperand(0)) &&
682 !isa<InsertElementInst>(IENext->getOperand(0))) {
686 dyn_cast<InsertElementInst>(IENext->getOperand(0))));
692 // This function implements one vectorization iteration on the provided
693 // basic block. It returns true if the block is changed.
694 bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) {
696 BasicBlock::iterator Start = BB.getFirstInsertionPt();
698 std::vector<Value *> AllPairableInsts;
699 DenseMap<Value *, Value *> AllChosenPairs;
700 DenseSet<ValuePair> AllFixedOrderPairs;
701 DenseMap<VPPair, unsigned> AllPairConnectionTypes;
702 std::multimap<ValuePair, ValuePair> AllConnectedPairs, AllConnectedPairDeps;
705 std::vector<Value *> PairableInsts;
706 std::multimap<Value *, Value *> CandidatePairs;
707 DenseSet<ValuePair> FixedOrderPairs;
708 DenseMap<ValuePair, int> CandidatePairCostSavings;
709 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs,
711 CandidatePairCostSavings,
712 PairableInsts, NonPow2Len);
713 if (PairableInsts.empty()) continue;
715 // Now we have a map of all of the pairable instructions and we need to
716 // select the best possible pairing. A good pairing is one such that the
717 // users of the pair are also paired. This defines a (directed) forest
718 // over the pairs such that two pairs are connected iff the second pair
721 // Note that it only matters that both members of the second pair use some
722 // element of the first pair (to allow for splatting).
724 std::multimap<ValuePair, ValuePair> ConnectedPairs, ConnectedPairDeps;
725 DenseMap<VPPair, unsigned> PairConnectionTypes;
726 computeConnectedPairs(CandidatePairs, PairableInsts, ConnectedPairs,
727 PairConnectionTypes);
728 if (ConnectedPairs.empty()) continue;
730 for (std::multimap<ValuePair, ValuePair>::iterator
731 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
733 ConnectedPairDeps.insert(VPPair(I->second, I->first));
736 // Build the pairable-instruction dependency map
737 DenseSet<ValuePair> PairableInstUsers;
738 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers);
740 // There is now a graph of the connected pairs. For each variable, pick
741 // the pairing with the largest tree meeting the depth requirement on at
742 // least one branch. Then select all pairings that are part of that tree
743 // and remove them from the list of available pairings and pairable
746 DenseMap<Value *, Value *> ChosenPairs;
747 choosePairs(CandidatePairs, CandidatePairCostSavings,
748 PairableInsts, FixedOrderPairs, PairConnectionTypes,
749 ConnectedPairs, ConnectedPairDeps,
750 PairableInstUsers, ChosenPairs);
752 if (ChosenPairs.empty()) continue;
753 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(),
754 PairableInsts.end());
755 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end());
757 // Only for the chosen pairs, propagate information on fixed-order pairs,
758 // pair connections, and their types to the data structures used by the
759 // pair fusion procedures.
760 for (DenseMap<Value *, Value *>::iterator I = ChosenPairs.begin(),
761 IE = ChosenPairs.end(); I != IE; ++I) {
762 if (FixedOrderPairs.count(*I))
763 AllFixedOrderPairs.insert(*I);
764 else if (FixedOrderPairs.count(ValuePair(I->second, I->first)))
765 AllFixedOrderPairs.insert(ValuePair(I->second, I->first));
767 for (DenseMap<Value *, Value *>::iterator J = ChosenPairs.begin();
769 DenseMap<VPPair, unsigned>::iterator K =
770 PairConnectionTypes.find(VPPair(*I, *J));
771 if (K != PairConnectionTypes.end()) {
772 AllPairConnectionTypes.insert(*K);
774 K = PairConnectionTypes.find(VPPair(*J, *I));
775 if (K != PairConnectionTypes.end())
776 AllPairConnectionTypes.insert(*K);
781 for (std::multimap<ValuePair, ValuePair>::iterator
782 I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
784 if (AllPairConnectionTypes.count(*I)) {
785 AllConnectedPairs.insert(*I);
786 AllConnectedPairDeps.insert(VPPair(I->second, I->first));
789 } while (ShouldContinue);
791 if (AllChosenPairs.empty()) return false;
792 NumFusedOps += AllChosenPairs.size();
794 // A set of pairs has now been selected. It is now necessary to replace the
795 // paired instructions with vector instructions. For this procedure each
796 // operand must be replaced with a vector operand. This vector is formed
797 // by using build_vector on the old operands. The replaced values are then
798 // replaced with a vector_extract on the result. Subsequent optimization
799 // passes should coalesce the build/extract combinations.
801 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs, AllFixedOrderPairs,
802 AllPairConnectionTypes,
803 AllConnectedPairs, AllConnectedPairDeps);
805 // It is important to cleanup here so that future iterations of this
806 // function have less work to do.
807 (void) SimplifyInstructionsInBlock(&BB, TD, AA->getTargetLibraryInfo());
811 // This function returns true if the provided instruction is capable of being
812 // fused into a vector instruction. This determination is based only on the
813 // type and other attributes of the instruction.
814 bool BBVectorize::isInstVectorizable(Instruction *I,
815 bool &IsSimpleLoadStore) {
816 IsSimpleLoadStore = false;
818 if (CallInst *C = dyn_cast<CallInst>(I)) {
819 if (!isVectorizableIntrinsic(C))
821 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) {
822 // Vectorize simple loads if possbile:
823 IsSimpleLoadStore = L->isSimple();
824 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
826 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) {
827 // Vectorize simple stores if possbile:
828 IsSimpleLoadStore = S->isSimple();
829 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
831 } else if (CastInst *C = dyn_cast<CastInst>(I)) {
832 // We can vectorize casts, but not casts of pointer types, etc.
833 if (!Config.VectorizeCasts)
836 Type *SrcTy = C->getSrcTy();
837 if (!SrcTy->isSingleValueType())
840 Type *DestTy = C->getDestTy();
841 if (!DestTy->isSingleValueType())
843 } else if (isa<SelectInst>(I)) {
844 if (!Config.VectorizeSelect)
846 } else if (isa<CmpInst>(I)) {
847 if (!Config.VectorizeCmp)
849 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) {
850 if (!Config.VectorizeGEP)
853 // Currently, vector GEPs exist only with one index.
854 if (G->getNumIndices() != 1)
856 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) ||
857 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) {
861 // We can't vectorize memory operations without target data
862 if (TD == 0 && IsSimpleLoadStore)
866 getInstructionTypes(I, T1, T2);
868 // Not every type can be vectorized...
869 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) ||
870 !(VectorType::isValidElementType(T2) || T2->isVectorTy()))
873 if (T1->getScalarSizeInBits() == 1) {
874 if (!Config.VectorizeBools)
877 if (!Config.VectorizeInts && T1->isIntOrIntVectorTy())
881 if (T2->getScalarSizeInBits() == 1) {
882 if (!Config.VectorizeBools)
885 if (!Config.VectorizeInts && T2->isIntOrIntVectorTy())
889 if (!Config.VectorizeFloats
890 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy()))
893 // Don't vectorize target-specific types.
894 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy())
896 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy())
899 if ((!Config.VectorizePointers || TD == 0) &&
900 (T1->getScalarType()->isPointerTy() ||
901 T2->getScalarType()->isPointerTy()))
904 if (!TTI && (T1->getPrimitiveSizeInBits() >= Config.VectorBits ||
905 T2->getPrimitiveSizeInBits() >= Config.VectorBits))
911 // This function returns true if the two provided instructions are compatible
912 // (meaning that they can be fused into a vector instruction). This assumes
913 // that I has already been determined to be vectorizable and that J is not
914 // in the use tree of I.
915 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J,
916 bool IsSimpleLoadStore, bool NonPow2Len,
917 int &CostSavings, int &FixedOrder) {
918 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I <<
919 " <-> " << *J << "\n");
924 // Loads and stores can be merged if they have different alignments,
925 // but are otherwise the same.
926 if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment |
927 (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0)))
930 Type *IT1, *IT2, *JT1, *JT2;
931 getInstructionTypes(I, IT1, IT2);
932 getInstructionTypes(J, JT1, JT2);
933 unsigned MaxTypeBits = std::max(
934 IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(),
935 IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits());
936 if (!TTI && MaxTypeBits > Config.VectorBits)
939 // FIXME: handle addsub-type operations!
941 if (IsSimpleLoadStore) {
943 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
944 int64_t OffsetInElmts = 0;
945 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
946 IAddressSpace, JAddressSpace,
947 OffsetInElmts) && abs64(OffsetInElmts) == 1) {
948 FixedOrder = (int) OffsetInElmts;
949 unsigned BottomAlignment = IAlignment;
950 if (OffsetInElmts < 0) BottomAlignment = JAlignment;
952 Type *aTypeI = isa<StoreInst>(I) ?
953 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType();
954 Type *aTypeJ = isa<StoreInst>(J) ?
955 cast<StoreInst>(J)->getValueOperand()->getType() : J->getType();
956 Type *VType = getVecTypeForPair(aTypeI, aTypeJ);
958 if (Config.AlignedOnly) {
959 // An aligned load or store is possible only if the instruction
960 // with the lower offset has an alignment suitable for the
963 unsigned VecAlignment = TD->getPrefTypeAlignment(VType);
964 if (BottomAlignment < VecAlignment)
969 unsigned ICost = TTI->getMemoryOpCost(I->getOpcode(), aTypeI,
970 IAlignment, IAddressSpace);
971 unsigned JCost = TTI->getMemoryOpCost(J->getOpcode(), aTypeJ,
972 JAlignment, JAddressSpace);
973 unsigned VCost = TTI->getMemoryOpCost(I->getOpcode(), VType,
976 if (VCost > ICost + JCost)
979 // We don't want to fuse to a type that will be split, even
980 // if the two input types will also be split and there is no other
982 unsigned VParts = TTI->getNumberOfParts(VType);
985 else if (!VParts && VCost == ICost + JCost)
988 CostSavings = ICost + JCost - VCost;
994 unsigned ICost = getInstrCost(I->getOpcode(), IT1, IT2);
995 unsigned JCost = getInstrCost(J->getOpcode(), JT1, JT2);
996 Type *VT1 = getVecTypeForPair(IT1, JT1),
997 *VT2 = getVecTypeForPair(IT2, JT2);
998 unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2);
1000 if (VCost > ICost + JCost)
1003 // We don't want to fuse to a type that will be split, even
1004 // if the two input types will also be split and there is no other
1006 unsigned VParts1 = TTI->getNumberOfParts(VT1),
1007 VParts2 = TTI->getNumberOfParts(VT2);
1008 if (VParts1 > 1 || VParts2 > 1)
1010 else if ((!VParts1 || !VParts2) && VCost == ICost + JCost)
1013 CostSavings = ICost + JCost - VCost;
1016 // The powi intrinsic is special because only the first argument is
1017 // vectorized, the second arguments must be equal.
1018 CallInst *CI = dyn_cast<CallInst>(I);
1020 if (CI && (FI = CI->getCalledFunction())) {
1021 Intrinsic::ID IID = (Intrinsic::ID) FI->getIntrinsicID();
1022 if (IID == Intrinsic::powi) {
1023 Value *A1I = CI->getArgOperand(1),
1024 *A1J = cast<CallInst>(J)->getArgOperand(1);
1025 const SCEV *A1ISCEV = SE->getSCEV(A1I),
1026 *A1JSCEV = SE->getSCEV(A1J);
1027 return (A1ISCEV == A1JSCEV);
1031 SmallVector<Type*, 4> Tys;
1032 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
1033 Tys.push_back(CI->getArgOperand(i)->getType());
1034 unsigned ICost = TTI->getIntrinsicInstrCost(IID, IT1, Tys);
1037 CallInst *CJ = cast<CallInst>(J);
1038 for (unsigned i = 0, ie = CJ->getNumArgOperands(); i != ie; ++i)
1039 Tys.push_back(CJ->getArgOperand(i)->getType());
1040 unsigned JCost = TTI->getIntrinsicInstrCost(IID, JT1, Tys);
1043 assert(CI->getNumArgOperands() == CJ->getNumArgOperands() &&
1044 "Intrinsic argument counts differ");
1045 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
1046 if (IID == Intrinsic::powi && i == 1)
1047 Tys.push_back(CI->getArgOperand(i)->getType());
1049 Tys.push_back(getVecTypeForPair(CI->getArgOperand(i)->getType(),
1050 CJ->getArgOperand(i)->getType()));
1053 Type *RetTy = getVecTypeForPair(IT1, JT1);
1054 unsigned VCost = TTI->getIntrinsicInstrCost(IID, RetTy, Tys);
1056 if (VCost > ICost + JCost)
1059 // We don't want to fuse to a type that will be split, even
1060 // if the two input types will also be split and there is no other
1062 unsigned RetParts = TTI->getNumberOfParts(RetTy);
1065 else if (!RetParts && VCost == ICost + JCost)
1068 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
1069 if (!Tys[i]->isVectorTy())
1072 unsigned NumParts = TTI->getNumberOfParts(Tys[i]);
1075 else if (!NumParts && VCost == ICost + JCost)
1079 CostSavings = ICost + JCost - VCost;
1086 // Figure out whether or not J uses I and update the users and write-set
1087 // structures associated with I. Specifically, Users represents the set of
1088 // instructions that depend on I. WriteSet represents the set
1089 // of memory locations that are dependent on I. If UpdateUsers is true,
1090 // and J uses I, then Users is updated to contain J and WriteSet is updated
1091 // to contain any memory locations to which J writes. The function returns
1092 // true if J uses I. By default, alias analysis is used to determine
1093 // whether J reads from memory that overlaps with a location in WriteSet.
1094 // If LoadMoveSet is not null, then it is a previously-computed multimap
1095 // where the key is the memory-based user instruction and the value is
1096 // the instruction to be compared with I. So, if LoadMoveSet is provided,
1097 // then the alias analysis is not used. This is necessary because this
1098 // function is called during the process of moving instructions during
1099 // vectorization and the results of the alias analysis are not stable during
1101 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users,
1102 AliasSetTracker &WriteSet, Instruction *I,
1103 Instruction *J, bool UpdateUsers,
1104 std::multimap<Value *, Value *> *LoadMoveSet) {
1107 // This instruction may already be marked as a user due, for example, to
1108 // being a member of a selected pair.
1113 for (User::op_iterator JU = J->op_begin(), JE = J->op_end();
1116 if (I == V || Users.count(V)) {
1121 if (!UsesI && J->mayReadFromMemory()) {
1123 VPIteratorPair JPairRange = LoadMoveSet->equal_range(J);
1124 UsesI = isSecondInIteratorPair<Value*>(I, JPairRange);
1126 for (AliasSetTracker::iterator W = WriteSet.begin(),
1127 WE = WriteSet.end(); W != WE; ++W) {
1128 if (W->aliasesUnknownInst(J, *AA)) {
1136 if (UsesI && UpdateUsers) {
1137 if (J->mayWriteToMemory()) WriteSet.add(J);
1144 // This function iterates over all instruction pairs in the provided
1145 // basic block and collects all candidate pairs for vectorization.
1146 bool BBVectorize::getCandidatePairs(BasicBlock &BB,
1147 BasicBlock::iterator &Start,
1148 std::multimap<Value *, Value *> &CandidatePairs,
1149 DenseSet<ValuePair> &FixedOrderPairs,
1150 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1151 std::vector<Value *> &PairableInsts, bool NonPow2Len) {
1152 BasicBlock::iterator E = BB.end();
1153 if (Start == E) return false;
1155 bool ShouldContinue = false, IAfterStart = false;
1156 for (BasicBlock::iterator I = Start++; I != E; ++I) {
1157 if (I == Start) IAfterStart = true;
1159 bool IsSimpleLoadStore;
1160 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue;
1162 // Look for an instruction with which to pair instruction *I...
1163 DenseSet<Value *> Users;
1164 AliasSetTracker WriteSet(*AA);
1165 bool JAfterStart = IAfterStart;
1166 BasicBlock::iterator J = llvm::next(I);
1167 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) {
1168 if (J == Start) JAfterStart = true;
1170 // Determine if J uses I, if so, exit the loop.
1171 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep);
1172 if (Config.FastDep) {
1173 // Note: For this heuristic to be effective, independent operations
1174 // must tend to be intermixed. This is likely to be true from some
1175 // kinds of grouped loop unrolling (but not the generic LLVM pass),
1176 // but otherwise may require some kind of reordering pass.
1178 // When using fast dependency analysis,
1179 // stop searching after first use:
1182 if (UsesI) continue;
1185 // J does not use I, and comes before the first use of I, so it can be
1186 // merged with I if the instructions are compatible.
1187 int CostSavings, FixedOrder;
1188 if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len,
1189 CostSavings, FixedOrder)) continue;
1191 // J is a candidate for merging with I.
1192 if (!PairableInsts.size() ||
1193 PairableInsts[PairableInsts.size()-1] != I) {
1194 PairableInsts.push_back(I);
1197 CandidatePairs.insert(ValuePair(I, J));
1199 CandidatePairCostSavings.insert(ValuePairWithCost(ValuePair(I, J),
1202 if (FixedOrder == 1)
1203 FixedOrderPairs.insert(ValuePair(I, J));
1204 else if (FixedOrder == -1)
1205 FixedOrderPairs.insert(ValuePair(J, I));
1207 // The next call to this function must start after the last instruction
1208 // selected during this invocation.
1210 Start = llvm::next(J);
1211 IAfterStart = JAfterStart = false;
1214 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair "
1215 << *I << " <-> " << *J << " (cost savings: " <<
1216 CostSavings << ")\n");
1218 // If we have already found too many pairs, break here and this function
1219 // will be called again starting after the last instruction selected
1220 // during this invocation.
1221 if (PairableInsts.size() >= Config.MaxInsts) {
1222 ShouldContinue = true;
1231 DEBUG(dbgs() << "BBV: found " << PairableInsts.size()
1232 << " instructions with candidate pairs\n");
1234 return ShouldContinue;
1237 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that
1238 // it looks for pairs such that both members have an input which is an
1239 // output of PI or PJ.
1240 void BBVectorize::computePairsConnectedTo(
1241 std::multimap<Value *, Value *> &CandidatePairs,
1242 std::vector<Value *> &PairableInsts,
1243 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1244 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1248 // For each possible pairing for this variable, look at the uses of
1249 // the first value...
1250 for (Value::use_iterator I = P.first->use_begin(),
1251 E = P.first->use_end(); I != E; ++I) {
1252 if (isa<LoadInst>(*I)) {
1253 // A pair cannot be connected to a load because the load only takes one
1254 // operand (the address) and it is a scalar even after vectorization.
1256 } else if ((SI = dyn_cast<StoreInst>(*I)) &&
1257 P.first == SI->getPointerOperand()) {
1258 // Similarly, a pair cannot be connected to a store through its
1263 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I);
1265 // For each use of the first variable, look for uses of the second
1267 for (Value::use_iterator J = P.second->use_begin(),
1268 E2 = P.second->use_end(); J != E2; ++J) {
1269 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1270 P.second == SJ->getPointerOperand())
1273 VPIteratorPair JPairRange = CandidatePairs.equal_range(*J);
1276 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) {
1277 VPPair VP(P, ValuePair(*I, *J));
1278 ConnectedPairs.insert(VP);
1279 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionDirect));
1283 if (isSecondInIteratorPair<Value*>(*I, JPairRange)) {
1284 VPPair VP(P, ValuePair(*J, *I));
1285 ConnectedPairs.insert(VP);
1286 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSwap));
1290 if (Config.SplatBreaksChain) continue;
1291 // Look for cases where just the first value in the pair is used by
1292 // both members of another pair (splatting).
1293 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) {
1294 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1295 P.first == SJ->getPointerOperand())
1298 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) {
1299 VPPair VP(P, ValuePair(*I, *J));
1300 ConnectedPairs.insert(VP);
1301 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1306 if (Config.SplatBreaksChain) return;
1307 // Look for cases where just the second value in the pair is used by
1308 // both members of another pair (splatting).
1309 for (Value::use_iterator I = P.second->use_begin(),
1310 E = P.second->use_end(); I != E; ++I) {
1311 if (isa<LoadInst>(*I))
1313 else if ((SI = dyn_cast<StoreInst>(*I)) &&
1314 P.second == SI->getPointerOperand())
1317 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I);
1319 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) {
1320 if ((SJ = dyn_cast<StoreInst>(*J)) &&
1321 P.second == SJ->getPointerOperand())
1324 if (isSecondInIteratorPair<Value*>(*J, IPairRange)) {
1325 VPPair VP(P, ValuePair(*I, *J));
1326 ConnectedPairs.insert(VP);
1327 PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1333 // This function figures out which pairs are connected. Two pairs are
1334 // connected if some output of the first pair forms an input to both members
1335 // of the second pair.
1336 void BBVectorize::computeConnectedPairs(
1337 std::multimap<Value *, Value *> &CandidatePairs,
1338 std::vector<Value *> &PairableInsts,
1339 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1340 DenseMap<VPPair, unsigned> &PairConnectionTypes) {
1342 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
1343 PE = PairableInsts.end(); PI != PE; ++PI) {
1344 VPIteratorPair choiceRange = CandidatePairs.equal_range(*PI);
1346 for (std::multimap<Value *, Value *>::iterator P = choiceRange.first;
1347 P != choiceRange.second; ++P)
1348 computePairsConnectedTo(CandidatePairs, PairableInsts,
1349 ConnectedPairs, PairConnectionTypes, *P);
1352 DEBUG(dbgs() << "BBV: found " << ConnectedPairs.size()
1353 << " pair connections.\n");
1356 // This function builds a set of use tuples such that <A, B> is in the set
1357 // if B is in the use tree of A. If B is in the use tree of A, then B
1358 // depends on the output of A.
1359 void BBVectorize::buildDepMap(
1361 std::multimap<Value *, Value *> &CandidatePairs,
1362 std::vector<Value *> &PairableInsts,
1363 DenseSet<ValuePair> &PairableInstUsers) {
1364 DenseSet<Value *> IsInPair;
1365 for (std::multimap<Value *, Value *>::iterator C = CandidatePairs.begin(),
1366 E = CandidatePairs.end(); C != E; ++C) {
1367 IsInPair.insert(C->first);
1368 IsInPair.insert(C->second);
1371 // Iterate through the basic block, recording all Users of each
1372 // pairable instruction.
1374 BasicBlock::iterator E = BB.end();
1375 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) {
1376 if (IsInPair.find(I) == IsInPair.end()) continue;
1378 DenseSet<Value *> Users;
1379 AliasSetTracker WriteSet(*AA);
1380 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J)
1381 (void) trackUsesOfI(Users, WriteSet, I, J);
1383 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end();
1385 PairableInstUsers.insert(ValuePair(I, *U));
1389 // Returns true if an input to pair P is an output of pair Q and also an
1390 // input of pair Q is an output of pair P. If this is the case, then these
1391 // two pairs cannot be simultaneously fused.
1392 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q,
1393 DenseSet<ValuePair> &PairableInstUsers,
1394 std::multimap<ValuePair, ValuePair> *PairableInstUserMap) {
1395 // Two pairs are in conflict if they are mutual Users of eachother.
1396 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) ||
1397 PairableInstUsers.count(ValuePair(P.first, Q.second)) ||
1398 PairableInstUsers.count(ValuePair(P.second, Q.first)) ||
1399 PairableInstUsers.count(ValuePair(P.second, Q.second));
1400 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) ||
1401 PairableInstUsers.count(ValuePair(Q.first, P.second)) ||
1402 PairableInstUsers.count(ValuePair(Q.second, P.first)) ||
1403 PairableInstUsers.count(ValuePair(Q.second, P.second));
1404 if (PairableInstUserMap) {
1405 // FIXME: The expensive part of the cycle check is not so much the cycle
1406 // check itself but this edge insertion procedure. This needs some
1407 // profiling and probably a different data structure (same is true of
1408 // most uses of std::multimap).
1410 VPPIteratorPair QPairRange = PairableInstUserMap->equal_range(Q);
1411 if (!isSecondInIteratorPair(P, QPairRange))
1412 PairableInstUserMap->insert(VPPair(Q, P));
1415 VPPIteratorPair PPairRange = PairableInstUserMap->equal_range(P);
1416 if (!isSecondInIteratorPair(Q, PPairRange))
1417 PairableInstUserMap->insert(VPPair(P, Q));
1421 return (QUsesP && PUsesQ);
1424 // This function walks the use graph of current pairs to see if, starting
1425 // from P, the walk returns to P.
1426 bool BBVectorize::pairWillFormCycle(ValuePair P,
1427 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1428 DenseSet<ValuePair> &CurrentPairs) {
1429 DEBUG(if (DebugCycleCheck)
1430 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> "
1431 << *P.second << "\n");
1432 // A lookup table of visisted pairs is kept because the PairableInstUserMap
1433 // contains non-direct associations.
1434 DenseSet<ValuePair> Visited;
1435 SmallVector<ValuePair, 32> Q;
1436 // General depth-first post-order traversal:
1439 ValuePair QTop = Q.pop_back_val();
1440 Visited.insert(QTop);
1442 DEBUG(if (DebugCycleCheck)
1443 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> "
1444 << *QTop.second << "\n");
1445 VPPIteratorPair QPairRange = PairableInstUserMap.equal_range(QTop);
1446 for (std::multimap<ValuePair, ValuePair>::iterator C = QPairRange.first;
1447 C != QPairRange.second; ++C) {
1448 if (C->second == P) {
1450 << "BBV: rejected to prevent non-trivial cycle formation: "
1451 << *C->first.first << " <-> " << *C->first.second << "\n");
1455 if (CurrentPairs.count(C->second) && !Visited.count(C->second))
1456 Q.push_back(C->second);
1458 } while (!Q.empty());
1463 // This function builds the initial tree of connected pairs with the
1464 // pair J at the root.
1465 void BBVectorize::buildInitialTreeFor(
1466 std::multimap<Value *, Value *> &CandidatePairs,
1467 std::vector<Value *> &PairableInsts,
1468 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1469 DenseSet<ValuePair> &PairableInstUsers,
1470 DenseMap<Value *, Value *> &ChosenPairs,
1471 DenseMap<ValuePair, size_t> &Tree, ValuePair J) {
1472 // Each of these pairs is viewed as the root node of a Tree. The Tree
1473 // is then walked (depth-first). As this happens, we keep track of
1474 // the pairs that compose the Tree and the maximum depth of the Tree.
1475 SmallVector<ValuePairWithDepth, 32> Q;
1476 // General depth-first post-order traversal:
1477 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1479 ValuePairWithDepth QTop = Q.back();
1481 // Push each child onto the queue:
1482 bool MoreChildren = false;
1483 size_t MaxChildDepth = QTop.second;
1484 VPPIteratorPair qtRange = ConnectedPairs.equal_range(QTop.first);
1485 for (std::multimap<ValuePair, ValuePair>::iterator k = qtRange.first;
1486 k != qtRange.second; ++k) {
1487 // Make sure that this child pair is still a candidate:
1488 bool IsStillCand = false;
1489 VPIteratorPair checkRange =
1490 CandidatePairs.equal_range(k->second.first);
1491 for (std::multimap<Value *, Value *>::iterator m = checkRange.first;
1492 m != checkRange.second; ++m) {
1493 if (m->second == k->second.second) {
1500 DenseMap<ValuePair, size_t>::iterator C = Tree.find(k->second);
1501 if (C == Tree.end()) {
1502 size_t d = getDepthFactor(k->second.first);
1503 Q.push_back(ValuePairWithDepth(k->second, QTop.second+d));
1504 MoreChildren = true;
1506 MaxChildDepth = std::max(MaxChildDepth, C->second);
1511 if (!MoreChildren) {
1512 // Record the current pair as part of the Tree:
1513 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth));
1516 } while (!Q.empty());
1519 // Given some initial tree, prune it by removing conflicting pairs (pairs
1520 // that cannot be simultaneously chosen for vectorization).
1521 void BBVectorize::pruneTreeFor(
1522 std::multimap<Value *, Value *> &CandidatePairs,
1523 std::vector<Value *> &PairableInsts,
1524 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1525 DenseSet<ValuePair> &PairableInstUsers,
1526 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1527 DenseMap<Value *, Value *> &ChosenPairs,
1528 DenseMap<ValuePair, size_t> &Tree,
1529 DenseSet<ValuePair> &PrunedTree, ValuePair J,
1530 bool UseCycleCheck) {
1531 SmallVector<ValuePairWithDepth, 32> Q;
1532 // General depth-first post-order traversal:
1533 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1535 ValuePairWithDepth QTop = Q.pop_back_val();
1536 PrunedTree.insert(QTop.first);
1538 // Visit each child, pruning as necessary...
1539 SmallVector<ValuePairWithDepth, 8> BestChildren;
1540 VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first);
1541 for (std::multimap<ValuePair, ValuePair>::iterator K = QTopRange.first;
1542 K != QTopRange.second; ++K) {
1543 DenseMap<ValuePair, size_t>::iterator C = Tree.find(K->second);
1544 if (C == Tree.end()) continue;
1546 // This child is in the Tree, now we need to make sure it is the
1547 // best of any conflicting children. There could be multiple
1548 // conflicting children, so first, determine if we're keeping
1549 // this child, then delete conflicting children as necessary.
1551 // It is also necessary to guard against pairing-induced
1552 // dependencies. Consider instructions a .. x .. y .. b
1553 // such that (a,b) are to be fused and (x,y) are to be fused
1554 // but a is an input to x and b is an output from y. This
1555 // means that y cannot be moved after b but x must be moved
1556 // after b for (a,b) to be fused. In other words, after
1557 // fusing (a,b) we have y .. a/b .. x where y is an input
1558 // to a/b and x is an output to a/b: x and y can no longer
1559 // be legally fused. To prevent this condition, we must
1560 // make sure that a child pair added to the Tree is not
1561 // both an input and output of an already-selected pair.
1563 // Pairing-induced dependencies can also form from more complicated
1564 // cycles. The pair vs. pair conflicts are easy to check, and so
1565 // that is done explicitly for "fast rejection", and because for
1566 // child vs. child conflicts, we may prefer to keep the current
1567 // pair in preference to the already-selected child.
1568 DenseSet<ValuePair> CurrentPairs;
1571 for (SmallVector<ValuePairWithDepth, 8>::iterator C2
1572 = BestChildren.begin(), E2 = BestChildren.end();
1574 if (C2->first.first == C->first.first ||
1575 C2->first.first == C->first.second ||
1576 C2->first.second == C->first.first ||
1577 C2->first.second == C->first.second ||
1578 pairsConflict(C2->first, C->first, PairableInstUsers,
1579 UseCycleCheck ? &PairableInstUserMap : 0)) {
1580 if (C2->second >= C->second) {
1585 CurrentPairs.insert(C2->first);
1588 if (!CanAdd) continue;
1590 // Even worse, this child could conflict with another node already
1591 // selected for the Tree. If that is the case, ignore this child.
1592 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(),
1593 E2 = PrunedTree.end(); T != E2; ++T) {
1594 if (T->first == C->first.first ||
1595 T->first == C->first.second ||
1596 T->second == C->first.first ||
1597 T->second == C->first.second ||
1598 pairsConflict(*T, C->first, PairableInstUsers,
1599 UseCycleCheck ? &PairableInstUserMap : 0)) {
1604 CurrentPairs.insert(*T);
1606 if (!CanAdd) continue;
1608 // And check the queue too...
1609 for (SmallVector<ValuePairWithDepth, 32>::iterator C2 = Q.begin(),
1610 E2 = Q.end(); C2 != E2; ++C2) {
1611 if (C2->first.first == C->first.first ||
1612 C2->first.first == C->first.second ||
1613 C2->first.second == C->first.first ||
1614 C2->first.second == C->first.second ||
1615 pairsConflict(C2->first, C->first, PairableInstUsers,
1616 UseCycleCheck ? &PairableInstUserMap : 0)) {
1621 CurrentPairs.insert(C2->first);
1623 if (!CanAdd) continue;
1625 // Last but not least, check for a conflict with any of the
1626 // already-chosen pairs.
1627 for (DenseMap<Value *, Value *>::iterator C2 =
1628 ChosenPairs.begin(), E2 = ChosenPairs.end();
1630 if (pairsConflict(*C2, C->first, PairableInstUsers,
1631 UseCycleCheck ? &PairableInstUserMap : 0)) {
1636 CurrentPairs.insert(*C2);
1638 if (!CanAdd) continue;
1640 // To check for non-trivial cycles formed by the addition of the
1641 // current pair we've formed a list of all relevant pairs, now use a
1642 // graph walk to check for a cycle. We start from the current pair and
1643 // walk the use tree to see if we again reach the current pair. If we
1644 // do, then the current pair is rejected.
1646 // FIXME: It may be more efficient to use a topological-ordering
1647 // algorithm to improve the cycle check. This should be investigated.
1648 if (UseCycleCheck &&
1649 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs))
1652 // This child can be added, but we may have chosen it in preference
1653 // to an already-selected child. Check for this here, and if a
1654 // conflict is found, then remove the previously-selected child
1655 // before adding this one in its place.
1656 for (SmallVector<ValuePairWithDepth, 8>::iterator C2
1657 = BestChildren.begin(); C2 != BestChildren.end();) {
1658 if (C2->first.first == C->first.first ||
1659 C2->first.first == C->first.second ||
1660 C2->first.second == C->first.first ||
1661 C2->first.second == C->first.second ||
1662 pairsConflict(C2->first, C->first, PairableInstUsers))
1663 C2 = BestChildren.erase(C2);
1668 BestChildren.push_back(ValuePairWithDepth(C->first, C->second));
1671 for (SmallVector<ValuePairWithDepth, 8>::iterator C
1672 = BestChildren.begin(), E2 = BestChildren.end();
1674 size_t DepthF = getDepthFactor(C->first.first);
1675 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF));
1677 } while (!Q.empty());
1680 // This function finds the best tree of mututally-compatible connected
1681 // pairs, given the choice of root pairs as an iterator range.
1682 void BBVectorize::findBestTreeFor(
1683 std::multimap<Value *, Value *> &CandidatePairs,
1684 DenseMap<ValuePair, int> &CandidatePairCostSavings,
1685 std::vector<Value *> &PairableInsts,
1686 DenseSet<ValuePair> &FixedOrderPairs,
1687 DenseMap<VPPair, unsigned> &PairConnectionTypes,
1688 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1689 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps,
1690 DenseSet<ValuePair> &PairableInstUsers,
1691 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1692 DenseMap<Value *, Value *> &ChosenPairs,
1693 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth,
1694 int &BestEffSize, VPIteratorPair ChoiceRange,
1695 bool UseCycleCheck) {
1696 for (std::multimap<Value *, Value *>::iterator J = ChoiceRange.first;
1697 J != ChoiceRange.second; ++J) {
1699 // Before going any further, make sure that this pair does not
1700 // conflict with any already-selected pairs (see comment below
1701 // near the Tree pruning for more details).
1702 DenseSet<ValuePair> ChosenPairSet;
1703 bool DoesConflict = false;
1704 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(),
1705 E = ChosenPairs.end(); C != E; ++C) {
1706 if (pairsConflict(*C, *J, PairableInstUsers,
1707 UseCycleCheck ? &PairableInstUserMap : 0)) {
1708 DoesConflict = true;
1712 ChosenPairSet.insert(*C);
1714 if (DoesConflict) continue;
1716 if (UseCycleCheck &&
1717 pairWillFormCycle(*J, PairableInstUserMap, ChosenPairSet))
1720 DenseMap<ValuePair, size_t> Tree;
1721 buildInitialTreeFor(CandidatePairs, PairableInsts, ConnectedPairs,
1722 PairableInstUsers, ChosenPairs, Tree, *J);
1724 // Because we'll keep the child with the largest depth, the largest
1725 // depth is still the same in the unpruned Tree.
1726 size_t MaxDepth = Tree.lookup(*J);
1728 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found Tree for pair {"
1729 << *J->first << " <-> " << *J->second << "} of depth " <<
1730 MaxDepth << " and size " << Tree.size() << "\n");
1732 // At this point the Tree has been constructed, but, may contain
1733 // contradictory children (meaning that different children of
1734 // some tree node may be attempting to fuse the same instruction).
1735 // So now we walk the tree again, in the case of a conflict,
1736 // keep only the child with the largest depth. To break a tie,
1737 // favor the first child.
1739 DenseSet<ValuePair> PrunedTree;
1740 pruneTreeFor(CandidatePairs, PairableInsts, ConnectedPairs,
1741 PairableInstUsers, PairableInstUserMap, ChosenPairs, Tree,
1742 PrunedTree, *J, UseCycleCheck);
1746 DenseSet<Value *> PrunedTreeInstrs;
1747 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
1748 E = PrunedTree.end(); S != E; ++S) {
1749 PrunedTreeInstrs.insert(S->first);
1750 PrunedTreeInstrs.insert(S->second);
1753 // The set of pairs that have already contributed to the total cost.
1754 DenseSet<ValuePair> IncomingPairs;
1756 // If the cost model were perfect, this might not be necessary; but we
1757 // need to make sure that we don't get stuck vectorizing our own
1759 bool HasNontrivialInsts = false;
1761 // The node weights represent the cost savings associated with
1762 // fusing the pair of instructions.
1763 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
1764 E = PrunedTree.end(); S != E; ++S) {
1765 if (!isa<ShuffleVectorInst>(S->first) &&
1766 !isa<InsertElementInst>(S->first) &&
1767 !isa<ExtractElementInst>(S->first))
1768 HasNontrivialInsts = true;
1770 bool FlipOrder = false;
1772 if (getDepthFactor(S->first)) {
1773 int ESContrib = CandidatePairCostSavings.find(*S)->second;
1774 DEBUG(if (DebugPairSelection) dbgs() << "\tweight {"
1775 << *S->first << " <-> " << *S->second << "} = " <<
1777 EffSize += ESContrib;
1780 // The edge weights contribute in a negative sense: they represent
1781 // the cost of shuffles.
1782 VPPIteratorPair IP = ConnectedPairDeps.equal_range(*S);
1783 if (IP.first != ConnectedPairDeps.end()) {
1784 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
1785 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first;
1786 Q != IP.second; ++Q) {
1787 if (!PrunedTree.count(Q->second))
1789 DenseMap<VPPair, unsigned>::iterator R =
1790 PairConnectionTypes.find(VPPair(Q->second, Q->first));
1791 assert(R != PairConnectionTypes.end() &&
1792 "Cannot find pair connection type");
1793 if (R->second == PairConnectionDirect)
1795 else if (R->second == PairConnectionSwap)
1799 // If there are more swaps than direct connections, then
1800 // the pair order will be flipped during fusion. So the real
1801 // number of swaps is the minimum number.
1802 FlipOrder = !FixedOrderPairs.count(*S) &&
1803 ((NumDepsSwap > NumDepsDirect) ||
1804 FixedOrderPairs.count(ValuePair(S->second, S->first)));
1806 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first;
1807 Q != IP.second; ++Q) {
1808 if (!PrunedTree.count(Q->second))
1810 DenseMap<VPPair, unsigned>::iterator R =
1811 PairConnectionTypes.find(VPPair(Q->second, Q->first));
1812 assert(R != PairConnectionTypes.end() &&
1813 "Cannot find pair connection type");
1814 Type *Ty1 = Q->second.first->getType(),
1815 *Ty2 = Q->second.second->getType();
1816 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1817 if ((R->second == PairConnectionDirect && FlipOrder) ||
1818 (R->second == PairConnectionSwap && !FlipOrder) ||
1819 R->second == PairConnectionSplat) {
1820 int ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1822 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1823 *Q->second.first << " <-> " << *Q->second.second <<
1825 *S->first << " <-> " << *S->second << "} = " <<
1827 EffSize -= ESContrib;
1832 // Compute the cost of outgoing edges. We assume that edges outgoing
1833 // to shuffles, inserts or extracts can be merged, and so contribute
1834 // no additional cost.
1835 if (!S->first->getType()->isVoidTy()) {
1836 Type *Ty1 = S->first->getType(),
1837 *Ty2 = S->second->getType();
1838 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1840 bool NeedsExtraction = false;
1841 for (Value::use_iterator I = S->first->use_begin(),
1842 IE = S->first->use_end(); I != IE; ++I) {
1843 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) {
1844 // Shuffle can be folded if it has no other input
1845 if (isa<UndefValue>(SI->getOperand(1)))
1848 if (isa<ExtractElementInst>(*I))
1850 if (PrunedTreeInstrs.count(*I))
1852 NeedsExtraction = true;
1856 if (NeedsExtraction) {
1858 if (Ty1->isVectorTy())
1859 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1862 ESContrib = (int) TTI->getVectorInstrCost(
1863 Instruction::ExtractElement, VTy, 0);
1865 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1866 *S->first << "} = " << ESContrib << "\n");
1867 EffSize -= ESContrib;
1870 NeedsExtraction = false;
1871 for (Value::use_iterator I = S->second->use_begin(),
1872 IE = S->second->use_end(); I != IE; ++I) {
1873 if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) {
1874 // Shuffle can be folded if it has no other input
1875 if (isa<UndefValue>(SI->getOperand(1)))
1878 if (isa<ExtractElementInst>(*I))
1880 if (PrunedTreeInstrs.count(*I))
1882 NeedsExtraction = true;
1886 if (NeedsExtraction) {
1888 if (Ty2->isVectorTy())
1889 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1892 ESContrib = (int) TTI->getVectorInstrCost(
1893 Instruction::ExtractElement, VTy, 1);
1894 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1895 *S->second << "} = " << ESContrib << "\n");
1896 EffSize -= ESContrib;
1900 // Compute the cost of incoming edges.
1901 if (!isa<LoadInst>(S->first) && !isa<StoreInst>(S->first)) {
1902 Instruction *S1 = cast<Instruction>(S->first),
1903 *S2 = cast<Instruction>(S->second);
1904 for (unsigned o = 0; o < S1->getNumOperands(); ++o) {
1905 Value *O1 = S1->getOperand(o), *O2 = S2->getOperand(o);
1907 // Combining constants into vector constants (or small vector
1908 // constants into larger ones are assumed free).
1909 if (isa<Constant>(O1) && isa<Constant>(O2))
1915 ValuePair VP = ValuePair(O1, O2);
1916 ValuePair VPR = ValuePair(O2, O1);
1918 // Internal edges are not handled here.
1919 if (PrunedTree.count(VP) || PrunedTree.count(VPR))
1922 Type *Ty1 = O1->getType(),
1923 *Ty2 = O2->getType();
1924 Type *VTy = getVecTypeForPair(Ty1, Ty2);
1926 // Combining vector operations of the same type is also assumed
1927 // folded with other operations.
1929 // If both are insert elements, then both can be widened.
1930 InsertElementInst *IEO1 = dyn_cast<InsertElementInst>(O1),
1931 *IEO2 = dyn_cast<InsertElementInst>(O2);
1932 if (IEO1 && IEO2 && isPureIEChain(IEO1) && isPureIEChain(IEO2))
1934 // If both are extract elements, and both have the same input
1935 // type, then they can be replaced with a shuffle
1936 ExtractElementInst *EIO1 = dyn_cast<ExtractElementInst>(O1),
1937 *EIO2 = dyn_cast<ExtractElementInst>(O2);
1939 EIO1->getOperand(0)->getType() ==
1940 EIO2->getOperand(0)->getType())
1942 // If both are a shuffle with equal operand types and only two
1943 // unqiue operands, then they can be replaced with a single
1945 ShuffleVectorInst *SIO1 = dyn_cast<ShuffleVectorInst>(O1),
1946 *SIO2 = dyn_cast<ShuffleVectorInst>(O2);
1948 SIO1->getOperand(0)->getType() ==
1949 SIO2->getOperand(0)->getType()) {
1950 SmallSet<Value *, 4> SIOps;
1951 SIOps.insert(SIO1->getOperand(0));
1952 SIOps.insert(SIO1->getOperand(1));
1953 SIOps.insert(SIO2->getOperand(0));
1954 SIOps.insert(SIO2->getOperand(1));
1955 if (SIOps.size() <= 2)
1961 // This pair has already been formed.
1962 if (IncomingPairs.count(VP)) {
1964 } else if (IncomingPairs.count(VPR)) {
1965 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1967 } else if (!Ty1->isVectorTy() && !Ty2->isVectorTy()) {
1968 ESContrib = (int) TTI->getVectorInstrCost(
1969 Instruction::InsertElement, VTy, 0);
1970 ESContrib += (int) TTI->getVectorInstrCost(
1971 Instruction::InsertElement, VTy, 1);
1972 } else if (!Ty1->isVectorTy()) {
1973 // O1 needs to be inserted into a vector of size O2, and then
1974 // both need to be shuffled together.
1975 ESContrib = (int) TTI->getVectorInstrCost(
1976 Instruction::InsertElement, Ty2, 0);
1977 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
1979 } else if (!Ty2->isVectorTy()) {
1980 // O2 needs to be inserted into a vector of size O1, and then
1981 // both need to be shuffled together.
1982 ESContrib = (int) TTI->getVectorInstrCost(
1983 Instruction::InsertElement, Ty1, 0);
1984 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
1987 Type *TyBig = Ty1, *TySmall = Ty2;
1988 if (Ty2->getVectorNumElements() > Ty1->getVectorNumElements())
1989 std::swap(TyBig, TySmall);
1991 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1993 if (TyBig != TySmall)
1994 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
1998 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {"
1999 << *O1 << " <-> " << *O2 << "} = " <<
2001 EffSize -= ESContrib;
2002 IncomingPairs.insert(VP);
2007 if (!HasNontrivialInsts) {
2008 DEBUG(if (DebugPairSelection) dbgs() <<
2009 "\tNo non-trivial instructions in tree;"
2010 " override to zero effective size\n");
2014 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
2015 E = PrunedTree.end(); S != E; ++S)
2016 EffSize += (int) getDepthFactor(S->first);
2019 DEBUG(if (DebugPairSelection)
2020 dbgs() << "BBV: found pruned Tree for pair {"
2021 << *J->first << " <-> " << *J->second << "} of depth " <<
2022 MaxDepth << " and size " << PrunedTree.size() <<
2023 " (effective size: " << EffSize << ")\n");
2024 if (((TTI && !UseChainDepthWithTI) ||
2025 MaxDepth >= Config.ReqChainDepth) &&
2026 EffSize > 0 && EffSize > BestEffSize) {
2027 BestMaxDepth = MaxDepth;
2028 BestEffSize = EffSize;
2029 BestTree = PrunedTree;
2034 // Given the list of candidate pairs, this function selects those
2035 // that will be fused into vector instructions.
2036 void BBVectorize::choosePairs(
2037 std::multimap<Value *, Value *> &CandidatePairs,
2038 DenseMap<ValuePair, int> &CandidatePairCostSavings,
2039 std::vector<Value *> &PairableInsts,
2040 DenseSet<ValuePair> &FixedOrderPairs,
2041 DenseMap<VPPair, unsigned> &PairConnectionTypes,
2042 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
2043 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps,
2044 DenseSet<ValuePair> &PairableInstUsers,
2045 DenseMap<Value *, Value *>& ChosenPairs) {
2046 bool UseCycleCheck =
2047 CandidatePairs.size() <= Config.MaxCandPairsForCycleCheck;
2048 std::multimap<ValuePair, ValuePair> PairableInstUserMap;
2049 for (std::vector<Value *>::iterator I = PairableInsts.begin(),
2050 E = PairableInsts.end(); I != E; ++I) {
2051 // The number of possible pairings for this variable:
2052 size_t NumChoices = CandidatePairs.count(*I);
2053 if (!NumChoices) continue;
2055 VPIteratorPair ChoiceRange = CandidatePairs.equal_range(*I);
2057 // The best pair to choose and its tree:
2058 size_t BestMaxDepth = 0;
2059 int BestEffSize = 0;
2060 DenseSet<ValuePair> BestTree;
2061 findBestTreeFor(CandidatePairs, CandidatePairCostSavings,
2062 PairableInsts, FixedOrderPairs, PairConnectionTypes,
2063 ConnectedPairs, ConnectedPairDeps,
2064 PairableInstUsers, PairableInstUserMap, ChosenPairs,
2065 BestTree, BestMaxDepth, BestEffSize, ChoiceRange,
2068 // A tree has been chosen (or not) at this point. If no tree was
2069 // chosen, then this instruction, I, cannot be paired (and is no longer
2072 DEBUG(if (BestTree.size() > 0)
2073 dbgs() << "BBV: selected pairs in the best tree for: "
2074 << *cast<Instruction>(*I) << "\n");
2076 for (DenseSet<ValuePair>::iterator S = BestTree.begin(),
2077 SE2 = BestTree.end(); S != SE2; ++S) {
2078 // Insert the members of this tree into the list of chosen pairs.
2079 ChosenPairs.insert(ValuePair(S->first, S->second));
2080 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " <<
2081 *S->second << "\n");
2083 // Remove all candidate pairs that have values in the chosen tree.
2084 for (std::multimap<Value *, Value *>::iterator K =
2085 CandidatePairs.begin(); K != CandidatePairs.end();) {
2086 if (K->first == S->first || K->second == S->first ||
2087 K->second == S->second || K->first == S->second) {
2088 // Don't remove the actual pair chosen so that it can be used
2089 // in subsequent tree selections.
2090 if (!(K->first == S->first && K->second == S->second))
2091 CandidatePairs.erase(K++);
2101 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n");
2104 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o,
2109 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) +
2110 (n > 0 ? "." + utostr(n) : "")).str();
2113 // Returns the value that is to be used as the pointer input to the vector
2114 // instruction that fuses I with J.
2115 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context,
2116 Instruction *I, Instruction *J, unsigned o) {
2118 unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
2119 int64_t OffsetInElmts;
2121 // Note: the analysis might fail here, that is why the pair order has
2122 // been precomputed (OffsetInElmts must be unused here).
2123 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
2124 IAddressSpace, JAddressSpace,
2125 OffsetInElmts, false);
2127 // The pointer value is taken to be the one with the lowest offset.
2130 Type *ArgTypeI = cast<PointerType>(IPtr->getType())->getElementType();
2131 Type *ArgTypeJ = cast<PointerType>(JPtr->getType())->getElementType();
2132 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2133 Type *VArgPtrType = PointerType::get(VArgType,
2134 cast<PointerType>(IPtr->getType())->getAddressSpace());
2135 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o),
2136 /* insert before */ I);
2139 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J,
2140 unsigned MaskOffset, unsigned NumInElem,
2141 unsigned NumInElem1, unsigned IdxOffset,
2142 std::vector<Constant*> &Mask) {
2143 unsigned NumElem1 = cast<VectorType>(J->getType())->getNumElements();
2144 for (unsigned v = 0; v < NumElem1; ++v) {
2145 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v);
2147 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context));
2149 unsigned mm = m + (int) IdxOffset;
2150 if (m >= (int) NumInElem1)
2151 mm += (int) NumInElem;
2153 Mask[v+MaskOffset] =
2154 ConstantInt::get(Type::getInt32Ty(Context), mm);
2159 // Returns the value that is to be used as the vector-shuffle mask to the
2160 // vector instruction that fuses I with J.
2161 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context,
2162 Instruction *I, Instruction *J) {
2163 // This is the shuffle mask. We need to append the second
2164 // mask to the first, and the numbers need to be adjusted.
2166 Type *ArgTypeI = I->getType();
2167 Type *ArgTypeJ = J->getType();
2168 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2170 unsigned NumElemI = cast<VectorType>(ArgTypeI)->getNumElements();
2172 // Get the total number of elements in the fused vector type.
2173 // By definition, this must equal the number of elements in
2175 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements();
2176 std::vector<Constant*> Mask(NumElem);
2178 Type *OpTypeI = I->getOperand(0)->getType();
2179 unsigned NumInElemI = cast<VectorType>(OpTypeI)->getNumElements();
2180 Type *OpTypeJ = J->getOperand(0)->getType();
2181 unsigned NumInElemJ = cast<VectorType>(OpTypeJ)->getNumElements();
2183 // The fused vector will be:
2184 // -----------------------------------------------------
2185 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ |
2186 // -----------------------------------------------------
2187 // from which we'll extract NumElem total elements (where the first NumElemI
2188 // of them come from the mask in I and the remainder come from the mask
2191 // For the mask from the first pair...
2192 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI,
2195 // For the mask from the second pair...
2196 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ,
2199 return ConstantVector::get(Mask);
2202 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I,
2203 Instruction *J, unsigned o, Value *&LOp,
2205 Type *ArgTypeL, Type *ArgTypeH,
2206 bool IBeforeJ, unsigned IdxOff) {
2207 bool ExpandedIEChain = false;
2208 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) {
2209 // If we have a pure insertelement chain, then this can be rewritten
2210 // into a chain that directly builds the larger type.
2211 if (isPureIEChain(LIE)) {
2212 SmallVector<Value *, 8> VectElemts(numElemL,
2213 UndefValue::get(ArgTypeL->getScalarType()));
2214 InsertElementInst *LIENext = LIE;
2217 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue();
2218 VectElemts[Idx] = LIENext->getOperand(1);
2220 dyn_cast<InsertElementInst>(LIENext->getOperand(0))));
2223 Value *LIEPrev = UndefValue::get(ArgTypeH);
2224 for (unsigned i = 0; i < numElemL; ++i) {
2225 if (isa<UndefValue>(VectElemts[i])) continue;
2226 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i],
2227 ConstantInt::get(Type::getInt32Ty(Context),
2229 getReplacementName(IBeforeJ ? I : J,
2231 LIENext->insertBefore(IBeforeJ ? J : I);
2235 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH);
2236 ExpandedIEChain = true;
2240 return ExpandedIEChain;
2243 // Returns the value to be used as the specified operand of the vector
2244 // instruction that fuses I with J.
2245 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I,
2246 Instruction *J, unsigned o, bool IBeforeJ) {
2247 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2248 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1);
2250 // Compute the fused vector type for this operand
2251 Type *ArgTypeI = I->getOperand(o)->getType();
2252 Type *ArgTypeJ = J->getOperand(o)->getType();
2253 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2255 Instruction *L = I, *H = J;
2256 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ;
2259 if (ArgTypeL->isVectorTy())
2260 numElemL = cast<VectorType>(ArgTypeL)->getNumElements();
2265 if (ArgTypeH->isVectorTy())
2266 numElemH = cast<VectorType>(ArgTypeH)->getNumElements();
2270 Value *LOp = L->getOperand(o);
2271 Value *HOp = H->getOperand(o);
2272 unsigned numElem = VArgType->getNumElements();
2274 // First, we check if we can reuse the "original" vector outputs (if these
2275 // exist). We might need a shuffle.
2276 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp);
2277 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp);
2278 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp);
2279 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp);
2281 // FIXME: If we're fusing shuffle instructions, then we can't apply this
2282 // optimization. The input vectors to the shuffle might be a different
2283 // length from the shuffle outputs. Unfortunately, the replacement
2284 // shuffle mask has already been formed, and the mask entries are sensitive
2285 // to the sizes of the inputs.
2286 bool IsSizeChangeShuffle =
2287 isa<ShuffleVectorInst>(L) &&
2288 (LOp->getType() != L->getType() || HOp->getType() != H->getType());
2290 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) {
2291 // We can have at most two unique vector inputs.
2292 bool CanUseInputs = true;
2295 I1 = LEE->getOperand(0);
2297 I1 = LSV->getOperand(0);
2298 I2 = LSV->getOperand(1);
2299 if (I2 == I1 || isa<UndefValue>(I2))
2304 Value *I3 = HEE->getOperand(0);
2305 if (!I2 && I3 != I1)
2307 else if (I3 != I1 && I3 != I2)
2308 CanUseInputs = false;
2310 Value *I3 = HSV->getOperand(0);
2311 if (!I2 && I3 != I1)
2313 else if (I3 != I1 && I3 != I2)
2314 CanUseInputs = false;
2317 Value *I4 = HSV->getOperand(1);
2318 if (!isa<UndefValue>(I4)) {
2319 if (!I2 && I4 != I1)
2321 else if (I4 != I1 && I4 != I2)
2322 CanUseInputs = false;
2329 cast<VectorType>(cast<Instruction>(LOp)->getOperand(0)->getType())
2332 cast<VectorType>(cast<Instruction>(HOp)->getOperand(0)->getType())
2335 // We have one or two input vectors. We need to map each index of the
2336 // operands to the index of the original vector.
2337 SmallVector<std::pair<int, int>, 8> II(numElem);
2338 for (unsigned i = 0; i < numElemL; ++i) {
2342 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue();
2343 INum = LEE->getOperand(0) == I1 ? 0 : 1;
2345 Idx = LSV->getMaskValue(i);
2346 if (Idx < (int) LOpElem) {
2347 INum = LSV->getOperand(0) == I1 ? 0 : 1;
2350 INum = LSV->getOperand(1) == I1 ? 0 : 1;
2354 II[i] = std::pair<int, int>(Idx, INum);
2356 for (unsigned i = 0; i < numElemH; ++i) {
2360 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue();
2361 INum = HEE->getOperand(0) == I1 ? 0 : 1;
2363 Idx = HSV->getMaskValue(i);
2364 if (Idx < (int) HOpElem) {
2365 INum = HSV->getOperand(0) == I1 ? 0 : 1;
2368 INum = HSV->getOperand(1) == I1 ? 0 : 1;
2372 II[i + numElemL] = std::pair<int, int>(Idx, INum);
2375 // We now have an array which tells us from which index of which
2376 // input vector each element of the operand comes.
2377 VectorType *I1T = cast<VectorType>(I1->getType());
2378 unsigned I1Elem = I1T->getNumElements();
2381 // In this case there is only one underlying vector input. Check for
2382 // the trivial case where we can use the input directly.
2383 if (I1Elem == numElem) {
2384 bool ElemInOrder = true;
2385 for (unsigned i = 0; i < numElem; ++i) {
2386 if (II[i].first != (int) i && II[i].first != -1) {
2387 ElemInOrder = false;
2396 // A shuffle is needed.
2397 std::vector<Constant *> Mask(numElem);
2398 for (unsigned i = 0; i < numElem; ++i) {
2399 int Idx = II[i].first;
2401 Mask[i] = UndefValue::get(Type::getInt32Ty(Context));
2403 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2407 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2408 ConstantVector::get(Mask),
2409 getReplacementName(IBeforeJ ? I : J,
2411 S->insertBefore(IBeforeJ ? J : I);
2415 VectorType *I2T = cast<VectorType>(I2->getType());
2416 unsigned I2Elem = I2T->getNumElements();
2418 // This input comes from two distinct vectors. The first step is to
2419 // make sure that both vectors are the same length. If not, the
2420 // smaller one will need to grow before they can be shuffled together.
2421 if (I1Elem < I2Elem) {
2422 std::vector<Constant *> Mask(I2Elem);
2424 for (; v < I1Elem; ++v)
2425 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2426 for (; v < I2Elem; ++v)
2427 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2429 Instruction *NewI1 =
2430 new ShuffleVectorInst(I1, UndefValue::get(I1T),
2431 ConstantVector::get(Mask),
2432 getReplacementName(IBeforeJ ? I : J,
2434 NewI1->insertBefore(IBeforeJ ? J : I);
2438 } else if (I1Elem > I2Elem) {
2439 std::vector<Constant *> Mask(I1Elem);
2441 for (; v < I2Elem; ++v)
2442 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2443 for (; v < I1Elem; ++v)
2444 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2446 Instruction *NewI2 =
2447 new ShuffleVectorInst(I2, UndefValue::get(I2T),
2448 ConstantVector::get(Mask),
2449 getReplacementName(IBeforeJ ? I : J,
2451 NewI2->insertBefore(IBeforeJ ? J : I);
2457 // Now that both I1 and I2 are the same length we can shuffle them
2458 // together (and use the result).
2459 std::vector<Constant *> Mask(numElem);
2460 for (unsigned v = 0; v < numElem; ++v) {
2461 if (II[v].first == -1) {
2462 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2464 int Idx = II[v].first + II[v].second * I1Elem;
2465 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2469 Instruction *NewOp =
2470 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask),
2471 getReplacementName(IBeforeJ ? I : J, true, o));
2472 NewOp->insertBefore(IBeforeJ ? J : I);
2477 Type *ArgType = ArgTypeL;
2478 if (numElemL < numElemH) {
2479 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH,
2480 ArgTypeL, VArgType, IBeforeJ, 1)) {
2481 // This is another short-circuit case: we're combining a scalar into
2482 // a vector that is formed by an IE chain. We've just expanded the IE
2483 // chain, now insert the scalar and we're done.
2485 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0,
2486 getReplacementName(IBeforeJ ? I : J, true, o));
2487 S->insertBefore(IBeforeJ ? J : I);
2489 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL,
2490 ArgTypeH, IBeforeJ)) {
2491 // The two vector inputs to the shuffle must be the same length,
2492 // so extend the smaller vector to be the same length as the larger one.
2496 std::vector<Constant *> Mask(numElemH);
2498 for (; v < numElemL; ++v)
2499 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2500 for (; v < numElemH; ++v)
2501 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2503 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL),
2504 ConstantVector::get(Mask),
2505 getReplacementName(IBeforeJ ? I : J,
2508 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0,
2509 getReplacementName(IBeforeJ ? I : J,
2513 NLOp->insertBefore(IBeforeJ ? J : I);
2518 } else if (numElemL > numElemH) {
2519 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL,
2520 ArgTypeH, VArgType, IBeforeJ)) {
2522 InsertElementInst::Create(LOp, HOp,
2523 ConstantInt::get(Type::getInt32Ty(Context),
2525 getReplacementName(IBeforeJ ? I : J,
2527 S->insertBefore(IBeforeJ ? J : I);
2529 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH,
2530 ArgTypeL, IBeforeJ)) {
2533 std::vector<Constant *> Mask(numElemL);
2535 for (; v < numElemH; ++v)
2536 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2537 for (; v < numElemL; ++v)
2538 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2540 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH),
2541 ConstantVector::get(Mask),
2542 getReplacementName(IBeforeJ ? I : J,
2545 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0,
2546 getReplacementName(IBeforeJ ? I : J,
2550 NHOp->insertBefore(IBeforeJ ? J : I);
2555 if (ArgType->isVectorTy()) {
2556 unsigned numElem = cast<VectorType>(VArgType)->getNumElements();
2557 std::vector<Constant*> Mask(numElem);
2558 for (unsigned v = 0; v < numElem; ++v) {
2560 // If the low vector was expanded, we need to skip the extra
2561 // undefined entries.
2562 if (v >= numElemL && numElemH > numElemL)
2563 Idx += (numElemH - numElemL);
2564 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2567 Instruction *BV = new ShuffleVectorInst(LOp, HOp,
2568 ConstantVector::get(Mask),
2569 getReplacementName(IBeforeJ ? I : J, true, o));
2570 BV->insertBefore(IBeforeJ ? J : I);
2574 Instruction *BV1 = InsertElementInst::Create(
2575 UndefValue::get(VArgType), LOp, CV0,
2576 getReplacementName(IBeforeJ ? I : J,
2578 BV1->insertBefore(IBeforeJ ? J : I);
2579 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1,
2580 getReplacementName(IBeforeJ ? I : J,
2582 BV2->insertBefore(IBeforeJ ? J : I);
2586 // This function creates an array of values that will be used as the inputs
2587 // to the vector instruction that fuses I with J.
2588 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context,
2589 Instruction *I, Instruction *J,
2590 SmallVector<Value *, 3> &ReplacedOperands,
2592 unsigned NumOperands = I->getNumOperands();
2594 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) {
2595 // Iterate backward so that we look at the store pointer
2596 // first and know whether or not we need to flip the inputs.
2598 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) {
2599 // This is the pointer for a load/store instruction.
2600 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o);
2602 } else if (isa<CallInst>(I)) {
2603 Function *F = cast<CallInst>(I)->getCalledFunction();
2604 Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID();
2605 if (o == NumOperands-1) {
2606 BasicBlock &BB = *I->getParent();
2608 Module *M = BB.getParent()->getParent();
2609 Type *ArgTypeI = I->getType();
2610 Type *ArgTypeJ = J->getType();
2611 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2613 ReplacedOperands[o] = Intrinsic::getDeclaration(M, IID, VArgType);
2615 } else if (IID == Intrinsic::powi && o == 1) {
2616 // The second argument of powi is a single integer and we've already
2617 // checked that both arguments are equal. As a result, we just keep
2618 // I's second argument.
2619 ReplacedOperands[o] = I->getOperand(o);
2622 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) {
2623 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J);
2627 ReplacedOperands[o] = getReplacementInput(Context, I, J, o, IBeforeJ);
2631 // This function creates two values that represent the outputs of the
2632 // original I and J instructions. These are generally vector shuffles
2633 // or extracts. In many cases, these will end up being unused and, thus,
2634 // eliminated by later passes.
2635 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
2636 Instruction *J, Instruction *K,
2637 Instruction *&InsertionPt,
2638 Instruction *&K1, Instruction *&K2) {
2639 if (isa<StoreInst>(I)) {
2640 AA->replaceWithNewValue(I, K);
2641 AA->replaceWithNewValue(J, K);
2643 Type *IType = I->getType();
2644 Type *JType = J->getType();
2646 VectorType *VType = getVecTypeForPair(IType, JType);
2647 unsigned numElem = VType->getNumElements();
2649 unsigned numElemI, numElemJ;
2650 if (IType->isVectorTy())
2651 numElemI = cast<VectorType>(IType)->getNumElements();
2655 if (JType->isVectorTy())
2656 numElemJ = cast<VectorType>(JType)->getNumElements();
2660 if (IType->isVectorTy()) {
2661 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI);
2662 for (unsigned v = 0; v < numElemI; ++v) {
2663 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2664 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v);
2667 K1 = new ShuffleVectorInst(K, UndefValue::get(VType),
2668 ConstantVector::get( Mask1),
2669 getReplacementName(K, false, 1));
2671 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2672 K1 = ExtractElementInst::Create(K, CV0,
2673 getReplacementName(K, false, 1));
2676 if (JType->isVectorTy()) {
2677 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ);
2678 for (unsigned v = 0; v < numElemJ; ++v) {
2679 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2680 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v);
2683 K2 = new ShuffleVectorInst(K, UndefValue::get(VType),
2684 ConstantVector::get( Mask2),
2685 getReplacementName(K, false, 2));
2687 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1);
2688 K2 = ExtractElementInst::Create(K, CV1,
2689 getReplacementName(K, false, 2));
2693 K2->insertAfter(K1);
2698 // Move all uses of the function I (including pairing-induced uses) after J.
2699 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB,
2700 std::multimap<Value *, Value *> &LoadMoveSet,
2701 Instruction *I, Instruction *J) {
2702 // Skip to the first instruction past I.
2703 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2705 DenseSet<Value *> Users;
2706 AliasSetTracker WriteSet(*AA);
2707 for (; cast<Instruction>(L) != J; ++L)
2708 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet);
2710 assert(cast<Instruction>(L) == J &&
2711 "Tracking has not proceeded far enough to check for dependencies");
2712 // If J is now in the use set of I, then trackUsesOfI will return true
2713 // and we have a dependency cycle (and the fusing operation must abort).
2714 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSet);
2717 // Move all uses of the function I (including pairing-induced uses) after J.
2718 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB,
2719 std::multimap<Value *, Value *> &LoadMoveSet,
2720 Instruction *&InsertionPt,
2721 Instruction *I, Instruction *J) {
2722 // Skip to the first instruction past I.
2723 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2725 DenseSet<Value *> Users;
2726 AliasSetTracker WriteSet(*AA);
2727 for (; cast<Instruction>(L) != J;) {
2728 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet)) {
2729 // Move this instruction
2730 Instruction *InstToMove = L; ++L;
2732 DEBUG(dbgs() << "BBV: moving: " << *InstToMove <<
2733 " to after " << *InsertionPt << "\n");
2734 InstToMove->removeFromParent();
2735 InstToMove->insertAfter(InsertionPt);
2736 InsertionPt = InstToMove;
2743 // Collect all load instruction that are in the move set of a given first
2744 // pair member. These loads depend on the first instruction, I, and so need
2745 // to be moved after J (the second instruction) when the pair is fused.
2746 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB,
2747 DenseMap<Value *, Value *> &ChosenPairs,
2748 std::multimap<Value *, Value *> &LoadMoveSet,
2750 // Skip to the first instruction past I.
2751 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2753 DenseSet<Value *> Users;
2754 AliasSetTracker WriteSet(*AA);
2756 // Note: We cannot end the loop when we reach J because J could be moved
2757 // farther down the use chain by another instruction pairing. Also, J
2758 // could be before I if this is an inverted input.
2759 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) {
2760 if (trackUsesOfI(Users, WriteSet, I, L)) {
2761 if (L->mayReadFromMemory())
2762 LoadMoveSet.insert(ValuePair(L, I));
2767 // In cases where both load/stores and the computation of their pointers
2768 // are chosen for vectorization, we can end up in a situation where the
2769 // aliasing analysis starts returning different query results as the
2770 // process of fusing instruction pairs continues. Because the algorithm
2771 // relies on finding the same use trees here as were found earlier, we'll
2772 // need to precompute the necessary aliasing information here and then
2773 // manually update it during the fusion process.
2774 void BBVectorize::collectLoadMoveSet(BasicBlock &BB,
2775 std::vector<Value *> &PairableInsts,
2776 DenseMap<Value *, Value *> &ChosenPairs,
2777 std::multimap<Value *, Value *> &LoadMoveSet) {
2778 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
2779 PIE = PairableInsts.end(); PI != PIE; ++PI) {
2780 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI);
2781 if (P == ChosenPairs.end()) continue;
2783 Instruction *I = cast<Instruction>(P->first);
2784 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, I);
2788 // When the first instruction in each pair is cloned, it will inherit its
2789 // parent's metadata. This metadata must be combined with that of the other
2790 // instruction in a safe way.
2791 void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) {
2792 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata;
2793 K->getAllMetadataOtherThanDebugLoc(Metadata);
2794 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) {
2795 unsigned Kind = Metadata[i].first;
2796 MDNode *JMD = J->getMetadata(Kind);
2797 MDNode *KMD = Metadata[i].second;
2801 K->setMetadata(Kind, 0); // Remove unknown metadata
2803 case LLVMContext::MD_tbaa:
2804 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD));
2806 case LLVMContext::MD_fpmath:
2807 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD));
2813 // This function fuses the chosen instruction pairs into vector instructions,
2814 // taking care preserve any needed scalar outputs and, then, it reorders the
2815 // remaining instructions as needed (users of the first member of the pair
2816 // need to be moved to after the location of the second member of the pair
2817 // because the vector instruction is inserted in the location of the pair's
2819 void BBVectorize::fuseChosenPairs(BasicBlock &BB,
2820 std::vector<Value *> &PairableInsts,
2821 DenseMap<Value *, Value *> &ChosenPairs,
2822 DenseSet<ValuePair> &FixedOrderPairs,
2823 DenseMap<VPPair, unsigned> &PairConnectionTypes,
2824 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
2825 std::multimap<ValuePair, ValuePair> &ConnectedPairDeps) {
2826 LLVMContext& Context = BB.getContext();
2828 // During the vectorization process, the order of the pairs to be fused
2829 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs
2830 // list. After a pair is fused, the flipped pair is removed from the list.
2831 DenseSet<ValuePair> FlippedPairs;
2832 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(),
2833 E = ChosenPairs.end(); P != E; ++P)
2834 FlippedPairs.insert(ValuePair(P->second, P->first));
2835 for (DenseSet<ValuePair>::iterator P = FlippedPairs.begin(),
2836 E = FlippedPairs.end(); P != E; ++P)
2837 ChosenPairs.insert(*P);
2839 std::multimap<Value *, Value *> LoadMoveSet;
2840 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, LoadMoveSet);
2842 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n");
2844 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) {
2845 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI);
2846 if (P == ChosenPairs.end()) {
2851 if (getDepthFactor(P->first) == 0) {
2852 // These instructions are not really fused, but are tracked as though
2853 // they are. Any case in which it would be interesting to fuse them
2854 // will be taken care of by InstCombine.
2860 Instruction *I = cast<Instruction>(P->first),
2861 *J = cast<Instruction>(P->second);
2863 DEBUG(dbgs() << "BBV: fusing: " << *I <<
2864 " <-> " << *J << "\n");
2866 // Remove the pair and flipped pair from the list.
2867 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second);
2868 assert(FP != ChosenPairs.end() && "Flipped pair not found in list");
2869 ChosenPairs.erase(FP);
2870 ChosenPairs.erase(P);
2872 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSet, I, J)) {
2873 DEBUG(dbgs() << "BBV: fusion of: " << *I <<
2875 " aborted because of non-trivial dependency cycle\n");
2881 // If the pair must have the other order, then flip it.
2882 bool FlipPairOrder = FixedOrderPairs.count(ValuePair(J, I));
2883 if (!FlipPairOrder && !FixedOrderPairs.count(ValuePair(I, J))) {
2884 // This pair does not have a fixed order, and so we might want to
2885 // flip it if that will yield fewer shuffles. We count the number
2886 // of dependencies connected via swaps, and those directly connected,
2887 // and flip the order if the number of swaps is greater.
2888 bool OrigOrder = true;
2889 VPPIteratorPair IP = ConnectedPairDeps.equal_range(ValuePair(I, J));
2890 if (IP.first == ConnectedPairDeps.end()) {
2891 IP = ConnectedPairDeps.equal_range(ValuePair(J, I));
2895 if (IP.first != ConnectedPairDeps.end()) {
2896 unsigned NumDepsDirect = 0, NumDepsSwap = 0;
2897 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first;
2898 Q != IP.second; ++Q) {
2899 DenseMap<VPPair, unsigned>::iterator R =
2900 PairConnectionTypes.find(VPPair(Q->second, Q->first));
2901 assert(R != PairConnectionTypes.end() &&
2902 "Cannot find pair connection type");
2903 if (R->second == PairConnectionDirect)
2905 else if (R->second == PairConnectionSwap)
2910 std::swap(NumDepsDirect, NumDepsSwap);
2912 if (NumDepsSwap > NumDepsDirect) {
2913 FlipPairOrder = true;
2914 DEBUG(dbgs() << "BBV: reordering pair: " << *I <<
2915 " <-> " << *J << "\n");
2920 Instruction *L = I, *H = J;
2924 // If the pair being fused uses the opposite order from that in the pair
2925 // connection map, then we need to flip the types.
2926 VPPIteratorPair IP = ConnectedPairs.equal_range(ValuePair(H, L));
2927 for (std::multimap<ValuePair, ValuePair>::iterator Q = IP.first;
2928 Q != IP.second; ++Q) {
2929 DenseMap<VPPair, unsigned>::iterator R = PairConnectionTypes.find(*Q);
2930 assert(R != PairConnectionTypes.end() &&
2931 "Cannot find pair connection type");
2932 if (R->second == PairConnectionDirect)
2933 R->second = PairConnectionSwap;
2934 else if (R->second == PairConnectionSwap)
2935 R->second = PairConnectionDirect;
2938 bool LBeforeH = !FlipPairOrder;
2939 unsigned NumOperands = I->getNumOperands();
2940 SmallVector<Value *, 3> ReplacedOperands(NumOperands);
2941 getReplacementInputsForPair(Context, L, H, ReplacedOperands,
2944 // Make a copy of the original operation, change its type to the vector
2945 // type and replace its operands with the vector operands.
2946 Instruction *K = L->clone();
2949 else if (H->hasName())
2952 if (!isa<StoreInst>(K))
2953 K->mutateType(getVecTypeForPair(L->getType(), H->getType()));
2955 combineMetadata(K, H);
2956 K->intersectOptionalDataWith(H);
2958 for (unsigned o = 0; o < NumOperands; ++o)
2959 K->setOperand(o, ReplacedOperands[o]);
2963 // Instruction insertion point:
2964 Instruction *InsertionPt = K;
2965 Instruction *K1 = 0, *K2 = 0;
2966 replaceOutputsOfPair(Context, L, H, K, InsertionPt, K1, K2);
2968 // The use tree of the first original instruction must be moved to after
2969 // the location of the second instruction. The entire use tree of the
2970 // first instruction is disjoint from the input tree of the second
2971 // (by definition), and so commutes with it.
2973 moveUsesOfIAfterJ(BB, LoadMoveSet, InsertionPt, I, J);
2975 if (!isa<StoreInst>(I)) {
2976 L->replaceAllUsesWith(K1);
2977 H->replaceAllUsesWith(K2);
2978 AA->replaceWithNewValue(L, K1);
2979 AA->replaceWithNewValue(H, K2);
2982 // Instructions that may read from memory may be in the load move set.
2983 // Once an instruction is fused, we no longer need its move set, and so
2984 // the values of the map never need to be updated. However, when a load
2985 // is fused, we need to merge the entries from both instructions in the
2986 // pair in case those instructions were in the move set of some other
2987 // yet-to-be-fused pair. The loads in question are the keys of the map.
2988 if (I->mayReadFromMemory()) {
2989 std::vector<ValuePair> NewSetMembers;
2990 VPIteratorPair IPairRange = LoadMoveSet.equal_range(I);
2991 VPIteratorPair JPairRange = LoadMoveSet.equal_range(J);
2992 for (std::multimap<Value *, Value *>::iterator N = IPairRange.first;
2993 N != IPairRange.second; ++N)
2994 NewSetMembers.push_back(ValuePair(K, N->second));
2995 for (std::multimap<Value *, Value *>::iterator N = JPairRange.first;
2996 N != JPairRange.second; ++N)
2997 NewSetMembers.push_back(ValuePair(K, N->second));
2998 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(),
2999 AE = NewSetMembers.end(); A != AE; ++A)
3000 LoadMoveSet.insert(*A);
3003 // Before removing I, set the iterator to the next instruction.
3004 PI = llvm::next(BasicBlock::iterator(I));
3005 if (cast<Instruction>(PI) == J)
3010 I->eraseFromParent();
3011 J->eraseFromParent();
3013 DEBUG(if (PrintAfterEveryPair) dbgs() << "BBV: block is now: \n" <<
3017 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n");
3021 char BBVectorize::ID = 0;
3022 static const char bb_vectorize_name[] = "Basic-Block Vectorization";
3023 INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
3024 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
3025 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
3026 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3027 INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
3029 BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) {
3030 return new BBVectorize(C);
3034 llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) {
3035 BBVectorize BBVectorizer(P, C);
3036 return BBVectorizer.vectorizeBB(BB);
3039 //===----------------------------------------------------------------------===//
3040 VectorizeConfig::VectorizeConfig() {
3041 VectorBits = ::VectorBits;
3042 VectorizeBools = !::NoBools;
3043 VectorizeInts = !::NoInts;
3044 VectorizeFloats = !::NoFloats;
3045 VectorizePointers = !::NoPointers;
3046 VectorizeCasts = !::NoCasts;
3047 VectorizeMath = !::NoMath;
3048 VectorizeFMA = !::NoFMA;
3049 VectorizeSelect = !::NoSelect;
3050 VectorizeCmp = !::NoCmp;
3051 VectorizeGEP = !::NoGEP;
3052 VectorizeMemOps = !::NoMemOps;
3053 AlignedOnly = ::AlignedOnly;
3054 ReqChainDepth= ::ReqChainDepth;
3055 SearchLimit = ::SearchLimit;
3056 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck;
3057 SplatBreaksChain = ::SplatBreaksChain;
3058 MaxInsts = ::MaxInsts;
3059 MaxIter = ::MaxIter;
3060 Pow2LenOnly = ::Pow2LenOnly;
3061 NoMemOpBoost = ::NoMemOpBoost;
3062 FastDep = ::FastDep;