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/Constants.h"
20 #include "llvm/DerivedTypes.h"
21 #include "llvm/Function.h"
22 #include "llvm/Instructions.h"
23 #include "llvm/IntrinsicInst.h"
24 #include "llvm/Intrinsics.h"
25 #include "llvm/LLVMContext.h"
26 #include "llvm/Metadata.h"
27 #include "llvm/Pass.h"
28 #include "llvm/Type.h"
29 #include "llvm/ADT/DenseMap.h"
30 #include "llvm/ADT/DenseSet.h"
31 #include "llvm/ADT/SmallVector.h"
32 #include "llvm/ADT/Statistic.h"
33 #include "llvm/ADT/STLExtras.h"
34 #include "llvm/ADT/StringExtras.h"
35 #include "llvm/Analysis/AliasAnalysis.h"
36 #include "llvm/Analysis/AliasSetTracker.h"
37 #include "llvm/Analysis/ScalarEvolution.h"
38 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
39 #include "llvm/Analysis/ValueTracking.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/raw_ostream.h"
43 #include "llvm/Support/ValueHandle.h"
44 #include "llvm/Target/TargetData.h"
45 #include "llvm/Transforms/Utils/Local.h"
46 #include "llvm/Transforms/Vectorize.h"
51 static cl::opt<unsigned>
52 ReqChainDepth("bb-vectorize-req-chain-depth", cl::init(6), cl::Hidden,
53 cl::desc("The required chain depth for vectorization"));
55 static cl::opt<unsigned>
56 SearchLimit("bb-vectorize-search-limit", cl::init(400), cl::Hidden,
57 cl::desc("The maximum search distance for instruction pairs"));
60 SplatBreaksChain("bb-vectorize-splat-breaks-chain", cl::init(false), cl::Hidden,
61 cl::desc("Replicating one element to a pair breaks the chain"));
63 static cl::opt<unsigned>
64 VectorBits("bb-vectorize-vector-bits", cl::init(128), cl::Hidden,
65 cl::desc("The size of the native vector registers"));
67 static cl::opt<unsigned>
68 MaxIter("bb-vectorize-max-iter", cl::init(0), cl::Hidden,
69 cl::desc("The maximum number of pairing iterations"));
72 Pow2LenOnly("bb-vectorize-pow2-len-only", cl::init(false), cl::Hidden,
73 cl::desc("Don't try to form non-2^n-length vectors"));
75 static cl::opt<unsigned>
76 MaxInsts("bb-vectorize-max-instr-per-group", cl::init(500), cl::Hidden,
77 cl::desc("The maximum number of pairable instructions per group"));
79 static cl::opt<unsigned>
80 MaxCandPairsForCycleCheck("bb-vectorize-max-cycle-check-pairs", cl::init(200),
81 cl::Hidden, cl::desc("The maximum number of candidate pairs with which to use"
82 " a full cycle check"));
85 NoBools("bb-vectorize-no-bools", cl::init(false), cl::Hidden,
86 cl::desc("Don't try to vectorize boolean (i1) values"));
89 NoInts("bb-vectorize-no-ints", cl::init(false), cl::Hidden,
90 cl::desc("Don't try to vectorize integer values"));
93 NoFloats("bb-vectorize-no-floats", cl::init(false), cl::Hidden,
94 cl::desc("Don't try to vectorize floating-point values"));
97 NoPointers("bb-vectorize-no-pointers", cl::init(false), cl::Hidden,
98 cl::desc("Don't try to vectorize pointer values"));
101 NoCasts("bb-vectorize-no-casts", cl::init(false), cl::Hidden,
102 cl::desc("Don't try to vectorize casting (conversion) operations"));
105 NoMath("bb-vectorize-no-math", cl::init(false), cl::Hidden,
106 cl::desc("Don't try to vectorize floating-point math intrinsics"));
109 NoFMA("bb-vectorize-no-fma", cl::init(false), cl::Hidden,
110 cl::desc("Don't try to vectorize the fused-multiply-add intrinsic"));
113 NoSelect("bb-vectorize-no-select", cl::init(false), cl::Hidden,
114 cl::desc("Don't try to vectorize select instructions"));
117 NoCmp("bb-vectorize-no-cmp", cl::init(false), cl::Hidden,
118 cl::desc("Don't try to vectorize comparison instructions"));
121 NoGEP("bb-vectorize-no-gep", cl::init(false), cl::Hidden,
122 cl::desc("Don't try to vectorize getelementptr instructions"));
125 NoMemOps("bb-vectorize-no-mem-ops", cl::init(false), cl::Hidden,
126 cl::desc("Don't try to vectorize loads and stores"));
129 AlignedOnly("bb-vectorize-aligned-only", cl::init(false), cl::Hidden,
130 cl::desc("Only generate aligned loads and stores"));
133 NoMemOpBoost("bb-vectorize-no-mem-op-boost",
134 cl::init(false), cl::Hidden,
135 cl::desc("Don't boost the chain-depth contribution of loads and stores"));
138 FastDep("bb-vectorize-fast-dep", cl::init(false), cl::Hidden,
139 cl::desc("Use a fast instruction dependency analysis"));
143 DebugInstructionExamination("bb-vectorize-debug-instruction-examination",
144 cl::init(false), cl::Hidden,
145 cl::desc("When debugging is enabled, output information on the"
146 " instruction-examination process"));
148 DebugCandidateSelection("bb-vectorize-debug-candidate-selection",
149 cl::init(false), cl::Hidden,
150 cl::desc("When debugging is enabled, output information on the"
151 " candidate-selection process"));
153 DebugPairSelection("bb-vectorize-debug-pair-selection",
154 cl::init(false), cl::Hidden,
155 cl::desc("When debugging is enabled, output information on the"
156 " pair-selection process"));
158 DebugCycleCheck("bb-vectorize-debug-cycle-check",
159 cl::init(false), cl::Hidden,
160 cl::desc("When debugging is enabled, output information on the"
161 " cycle-checking process"));
164 STATISTIC(NumFusedOps, "Number of operations fused by bb-vectorize");
167 struct BBVectorize : public BasicBlockPass {
168 static char ID; // Pass identification, replacement for typeid
170 const VectorizeConfig Config;
172 BBVectorize(const VectorizeConfig &C = VectorizeConfig())
173 : BasicBlockPass(ID), Config(C) {
174 initializeBBVectorizePass(*PassRegistry::getPassRegistry());
177 BBVectorize(Pass *P, const VectorizeConfig &C)
178 : BasicBlockPass(ID), Config(C) {
179 AA = &P->getAnalysis<AliasAnalysis>();
180 SE = &P->getAnalysis<ScalarEvolution>();
181 TD = P->getAnalysisIfAvailable<TargetData>();
184 typedef std::pair<Value *, Value *> ValuePair;
185 typedef std::pair<ValuePair, size_t> ValuePairWithDepth;
186 typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair
187 typedef std::pair<std::multimap<Value *, Value *>::iterator,
188 std::multimap<Value *, Value *>::iterator> VPIteratorPair;
189 typedef std::pair<std::multimap<ValuePair, ValuePair>::iterator,
190 std::multimap<ValuePair, ValuePair>::iterator>
197 // FIXME: const correct?
199 bool vectorizePairs(BasicBlock &BB, bool NonPow2Len = false);
201 bool getCandidatePairs(BasicBlock &BB,
202 BasicBlock::iterator &Start,
203 std::multimap<Value *, Value *> &CandidatePairs,
204 std::vector<Value *> &PairableInsts, bool NonPow2Len);
206 void computeConnectedPairs(std::multimap<Value *, Value *> &CandidatePairs,
207 std::vector<Value *> &PairableInsts,
208 std::multimap<ValuePair, ValuePair> &ConnectedPairs);
210 void buildDepMap(BasicBlock &BB,
211 std::multimap<Value *, Value *> &CandidatePairs,
212 std::vector<Value *> &PairableInsts,
213 DenseSet<ValuePair> &PairableInstUsers);
215 void choosePairs(std::multimap<Value *, Value *> &CandidatePairs,
216 std::vector<Value *> &PairableInsts,
217 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
218 DenseSet<ValuePair> &PairableInstUsers,
219 DenseMap<Value *, Value *>& ChosenPairs);
221 void fuseChosenPairs(BasicBlock &BB,
222 std::vector<Value *> &PairableInsts,
223 DenseMap<Value *, Value *>& ChosenPairs);
225 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore);
227 bool areInstsCompatible(Instruction *I, Instruction *J,
228 bool IsSimpleLoadStore, bool NonPow2Len);
230 bool trackUsesOfI(DenseSet<Value *> &Users,
231 AliasSetTracker &WriteSet, Instruction *I,
232 Instruction *J, bool UpdateUsers = true,
233 std::multimap<Value *, Value *> *LoadMoveSet = 0);
235 void computePairsConnectedTo(
236 std::multimap<Value *, Value *> &CandidatePairs,
237 std::vector<Value *> &PairableInsts,
238 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
241 bool pairsConflict(ValuePair P, ValuePair Q,
242 DenseSet<ValuePair> &PairableInstUsers,
243 std::multimap<ValuePair, ValuePair> *PairableInstUserMap = 0);
245 bool pairWillFormCycle(ValuePair P,
246 std::multimap<ValuePair, ValuePair> &PairableInstUsers,
247 DenseSet<ValuePair> &CurrentPairs);
250 std::multimap<Value *, Value *> &CandidatePairs,
251 std::vector<Value *> &PairableInsts,
252 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
253 DenseSet<ValuePair> &PairableInstUsers,
254 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
255 DenseMap<Value *, Value *> &ChosenPairs,
256 DenseMap<ValuePair, size_t> &Tree,
257 DenseSet<ValuePair> &PrunedTree, ValuePair J,
260 void buildInitialTreeFor(
261 std::multimap<Value *, Value *> &CandidatePairs,
262 std::vector<Value *> &PairableInsts,
263 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
264 DenseSet<ValuePair> &PairableInstUsers,
265 DenseMap<Value *, Value *> &ChosenPairs,
266 DenseMap<ValuePair, size_t> &Tree, ValuePair J);
268 void findBestTreeFor(
269 std::multimap<Value *, Value *> &CandidatePairs,
270 std::vector<Value *> &PairableInsts,
271 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
272 DenseSet<ValuePair> &PairableInstUsers,
273 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
274 DenseMap<Value *, Value *> &ChosenPairs,
275 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth,
276 size_t &BestEffSize, VPIteratorPair ChoiceRange,
279 Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I,
280 Instruction *J, unsigned o, bool &FlipMemInputs);
282 void fillNewShuffleMask(LLVMContext& Context, Instruction *J,
283 unsigned MaskOffset, unsigned NumInElem,
284 unsigned NumInElem1, unsigned IdxOffset,
285 std::vector<Constant*> &Mask);
287 Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I,
290 bool expandIEChain(LLVMContext& Context, Instruction *I, Instruction *J,
291 unsigned o, Value *&LOp, unsigned numElemL,
292 Type *ArgTypeL, Type *ArgTypeR,
293 unsigned IdxOff = 0);
295 Value *getReplacementInput(LLVMContext& Context, Instruction *I,
296 Instruction *J, unsigned o, bool FlipMemInputs);
298 void getReplacementInputsForPair(LLVMContext& Context, Instruction *I,
299 Instruction *J, SmallVector<Value *, 3> &ReplacedOperands,
300 bool &FlipMemInputs);
302 void replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
303 Instruction *J, Instruction *K,
304 Instruction *&InsertionPt, Instruction *&K1,
305 Instruction *&K2, bool &FlipMemInputs);
307 void collectPairLoadMoveSet(BasicBlock &BB,
308 DenseMap<Value *, Value *> &ChosenPairs,
309 std::multimap<Value *, Value *> &LoadMoveSet,
312 void collectLoadMoveSet(BasicBlock &BB,
313 std::vector<Value *> &PairableInsts,
314 DenseMap<Value *, Value *> &ChosenPairs,
315 std::multimap<Value *, Value *> &LoadMoveSet);
317 bool canMoveUsesOfIAfterJ(BasicBlock &BB,
318 std::multimap<Value *, Value *> &LoadMoveSet,
319 Instruction *I, Instruction *J);
321 void moveUsesOfIAfterJ(BasicBlock &BB,
322 std::multimap<Value *, Value *> &LoadMoveSet,
323 Instruction *&InsertionPt,
324 Instruction *I, Instruction *J);
326 void combineMetadata(Instruction *K, const Instruction *J);
328 bool vectorizeBB(BasicBlock &BB) {
329 bool changed = false;
330 // Iterate a sufficient number of times to merge types of size 1 bit,
331 // then 2 bits, then 4, etc. up to half of the target vector width of the
332 // target vector register.
335 v <= Config.VectorBits && (!Config.MaxIter || n <= Config.MaxIter);
337 DEBUG(dbgs() << "BBV: fusing loop #" << n <<
338 " for " << BB.getName() << " in " <<
339 BB.getParent()->getName() << "...\n");
340 if (vectorizePairs(BB))
346 if (changed && !Pow2LenOnly) {
348 for (; !Config.MaxIter || n <= Config.MaxIter; ++n) {
349 DEBUG(dbgs() << "BBV: fusing for non-2^n-length vectors loop #: " <<
350 n << " for " << BB.getName() << " in " <<
351 BB.getParent()->getName() << "...\n");
352 if (!vectorizePairs(BB, true)) break;
356 DEBUG(dbgs() << "BBV: done!\n");
360 virtual bool runOnBasicBlock(BasicBlock &BB) {
361 AA = &getAnalysis<AliasAnalysis>();
362 SE = &getAnalysis<ScalarEvolution>();
363 TD = getAnalysisIfAvailable<TargetData>();
365 return vectorizeBB(BB);
368 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
369 BasicBlockPass::getAnalysisUsage(AU);
370 AU.addRequired<AliasAnalysis>();
371 AU.addRequired<ScalarEvolution>();
372 AU.addPreserved<AliasAnalysis>();
373 AU.addPreserved<ScalarEvolution>();
374 AU.setPreservesCFG();
377 static inline VectorType *getVecTypeForPair(Type *ElemTy, Type *Elem2Ty) {
378 assert(ElemTy->getScalarType() == Elem2Ty->getScalarType() &&
379 "Cannot form vector from incompatible scalar types");
380 Type *STy = ElemTy->getScalarType();
383 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) {
384 numElem = VTy->getNumElements();
389 if (VectorType *VTy = dyn_cast<VectorType>(Elem2Ty)) {
390 numElem += VTy->getNumElements();
395 return VectorType::get(STy, numElem);
398 static inline void getInstructionTypes(Instruction *I,
399 Type *&T1, Type *&T2) {
400 if (isa<StoreInst>(I)) {
401 // For stores, it is the value type, not the pointer type that matters
402 // because the value is what will come from a vector register.
404 Value *IVal = cast<StoreInst>(I)->getValueOperand();
405 T1 = IVal->getType();
411 T2 = cast<CastInst>(I)->getSrcTy();
416 // Returns the weight associated with the provided value. A chain of
417 // candidate pairs has a length given by the sum of the weights of its
418 // members (one weight per pair; the weight of each member of the pair
419 // is assumed to be the same). This length is then compared to the
420 // chain-length threshold to determine if a given chain is significant
421 // enough to be vectorized. The length is also used in comparing
422 // candidate chains where longer chains are considered to be better.
423 // Note: when this function returns 0, the resulting instructions are
424 // not actually fused.
425 inline size_t getDepthFactor(Value *V) {
426 // InsertElement and ExtractElement have a depth factor of zero. This is
427 // for two reasons: First, they cannot be usefully fused. Second, because
428 // the pass generates a lot of these, they can confuse the simple metric
429 // used to compare the trees in the next iteration. Thus, giving them a
430 // weight of zero allows the pass to essentially ignore them in
431 // subsequent iterations when looking for vectorization opportunities
432 // while still tracking dependency chains that flow through those
434 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V))
437 // Give a load or store half of the required depth so that load/store
438 // pairs will vectorize.
439 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V)))
440 return Config.ReqChainDepth/2;
445 // This determines the relative offset of two loads or stores, returning
446 // true if the offset could be determined to be some constant value.
447 // For example, if OffsetInElmts == 1, then J accesses the memory directly
448 // after I; if OffsetInElmts == -1 then I accesses the memory
450 bool getPairPtrInfo(Instruction *I, Instruction *J,
451 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment,
452 int64_t &OffsetInElmts) {
454 if (isa<LoadInst>(I)) {
455 IPtr = cast<LoadInst>(I)->getPointerOperand();
456 JPtr = cast<LoadInst>(J)->getPointerOperand();
457 IAlignment = cast<LoadInst>(I)->getAlignment();
458 JAlignment = cast<LoadInst>(J)->getAlignment();
460 IPtr = cast<StoreInst>(I)->getPointerOperand();
461 JPtr = cast<StoreInst>(J)->getPointerOperand();
462 IAlignment = cast<StoreInst>(I)->getAlignment();
463 JAlignment = cast<StoreInst>(J)->getAlignment();
466 const SCEV *IPtrSCEV = SE->getSCEV(IPtr);
467 const SCEV *JPtrSCEV = SE->getSCEV(JPtr);
469 // If this is a trivial offset, then we'll get something like
470 // 1*sizeof(type). With target data, which we need anyway, this will get
471 // constant folded into a number.
472 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV);
473 if (const SCEVConstant *ConstOffSCEV =
474 dyn_cast<SCEVConstant>(OffsetSCEV)) {
475 ConstantInt *IntOff = ConstOffSCEV->getValue();
476 int64_t Offset = IntOff->getSExtValue();
478 Type *VTy = cast<PointerType>(IPtr->getType())->getElementType();
479 int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy);
481 Type *VTy2 = cast<PointerType>(JPtr->getType())->getElementType();
482 if (VTy != VTy2 && Offset < 0) {
483 int64_t VTy2TSS = (int64_t) TD->getTypeStoreSize(VTy2);
484 OffsetInElmts = Offset/VTy2TSS;
485 return (abs64(Offset) % VTy2TSS) == 0;
488 OffsetInElmts = Offset/VTyTSS;
489 return (abs64(Offset) % VTyTSS) == 0;
495 // Returns true if the provided CallInst represents an intrinsic that can
497 bool isVectorizableIntrinsic(CallInst* I) {
498 Function *F = I->getCalledFunction();
499 if (!F) return false;
501 unsigned IID = F->getIntrinsicID();
502 if (!IID) return false;
507 case Intrinsic::sqrt:
508 case Intrinsic::powi:
512 case Intrinsic::log2:
513 case Intrinsic::log10:
515 case Intrinsic::exp2:
517 return Config.VectorizeMath;
519 return Config.VectorizeFMA;
523 // Returns true if J is the second element in some pair referenced by
524 // some multimap pair iterator pair.
525 template <typename V>
526 bool isSecondInIteratorPair(V J, std::pair<
527 typename std::multimap<V, V>::iterator,
528 typename std::multimap<V, V>::iterator> PairRange) {
529 for (typename std::multimap<V, V>::iterator K = PairRange.first;
530 K != PairRange.second; ++K)
531 if (K->second == J) return true;
537 // This function implements one vectorization iteration on the provided
538 // basic block. It returns true if the block is changed.
539 bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) {
541 BasicBlock::iterator Start = BB.getFirstInsertionPt();
543 std::vector<Value *> AllPairableInsts;
544 DenseMap<Value *, Value *> AllChosenPairs;
547 std::vector<Value *> PairableInsts;
548 std::multimap<Value *, Value *> CandidatePairs;
549 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs,
550 PairableInsts, NonPow2Len);
551 if (PairableInsts.empty()) continue;
553 // Now we have a map of all of the pairable instructions and we need to
554 // select the best possible pairing. A good pairing is one such that the
555 // users of the pair are also paired. This defines a (directed) forest
556 // over the pairs such that two pairs are connected iff the second pair
559 // Note that it only matters that both members of the second pair use some
560 // element of the first pair (to allow for splatting).
562 std::multimap<ValuePair, ValuePair> ConnectedPairs;
563 computeConnectedPairs(CandidatePairs, PairableInsts, ConnectedPairs);
564 if (ConnectedPairs.empty()) continue;
566 // Build the pairable-instruction dependency map
567 DenseSet<ValuePair> PairableInstUsers;
568 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers);
570 // There is now a graph of the connected pairs. For each variable, pick
571 // the pairing with the largest tree meeting the depth requirement on at
572 // least one branch. Then select all pairings that are part of that tree
573 // and remove them from the list of available pairings and pairable
576 DenseMap<Value *, Value *> ChosenPairs;
577 choosePairs(CandidatePairs, PairableInsts, ConnectedPairs,
578 PairableInstUsers, ChosenPairs);
580 if (ChosenPairs.empty()) continue;
581 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(),
582 PairableInsts.end());
583 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end());
584 } while (ShouldContinue);
586 if (AllChosenPairs.empty()) return false;
587 NumFusedOps += AllChosenPairs.size();
589 // A set of pairs has now been selected. It is now necessary to replace the
590 // paired instructions with vector instructions. For this procedure each
591 // operand must be replaced with a vector operand. This vector is formed
592 // by using build_vector on the old operands. The replaced values are then
593 // replaced with a vector_extract on the result. Subsequent optimization
594 // passes should coalesce the build/extract combinations.
596 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs);
598 // It is important to cleanup here so that future iterations of this
599 // function have less work to do.
600 (void) SimplifyInstructionsInBlock(&BB, TD);
604 // This function returns true if the provided instruction is capable of being
605 // fused into a vector instruction. This determination is based only on the
606 // type and other attributes of the instruction.
607 bool BBVectorize::isInstVectorizable(Instruction *I,
608 bool &IsSimpleLoadStore) {
609 IsSimpleLoadStore = false;
611 if (CallInst *C = dyn_cast<CallInst>(I)) {
612 if (!isVectorizableIntrinsic(C))
614 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) {
615 // Vectorize simple loads if possbile:
616 IsSimpleLoadStore = L->isSimple();
617 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
619 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) {
620 // Vectorize simple stores if possbile:
621 IsSimpleLoadStore = S->isSimple();
622 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
624 } else if (CastInst *C = dyn_cast<CastInst>(I)) {
625 // We can vectorize casts, but not casts of pointer types, etc.
626 if (!Config.VectorizeCasts)
629 Type *SrcTy = C->getSrcTy();
630 if (!SrcTy->isSingleValueType())
633 Type *DestTy = C->getDestTy();
634 if (!DestTy->isSingleValueType())
636 } else if (isa<SelectInst>(I)) {
637 if (!Config.VectorizeSelect)
639 } else if (isa<CmpInst>(I)) {
640 if (!Config.VectorizeCmp)
642 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) {
643 if (!Config.VectorizeGEP)
646 // Currently, vector GEPs exist only with one index.
647 if (G->getNumIndices() != 1)
649 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) ||
650 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) {
654 // We can't vectorize memory operations without target data
655 if (TD == 0 && IsSimpleLoadStore)
659 getInstructionTypes(I, T1, T2);
661 // Not every type can be vectorized...
662 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) ||
663 !(VectorType::isValidElementType(T2) || T2->isVectorTy()))
666 if (T1->getScalarSizeInBits() == 1 && T2->getScalarSizeInBits() == 1) {
667 if (!Config.VectorizeBools)
670 if (!Config.VectorizeInts
671 && (T1->isIntOrIntVectorTy() || T2->isIntOrIntVectorTy()))
675 if (!Config.VectorizeFloats
676 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy()))
679 // Don't vectorize target-specific types.
680 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy())
682 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy())
685 if ((!Config.VectorizePointers || TD == 0) &&
686 (T1->getScalarType()->isPointerTy() ||
687 T2->getScalarType()->isPointerTy()))
690 if (T1->getPrimitiveSizeInBits() >= Config.VectorBits ||
691 T2->getPrimitiveSizeInBits() >= Config.VectorBits)
697 // This function returns true if the two provided instructions are compatible
698 // (meaning that they can be fused into a vector instruction). This assumes
699 // that I has already been determined to be vectorizable and that J is not
700 // in the use tree of I.
701 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J,
702 bool IsSimpleLoadStore, bool NonPow2Len) {
703 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I <<
704 " <-> " << *J << "\n");
706 // Loads and stores can be merged if they have different alignments,
707 // but are otherwise the same.
708 if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment |
709 (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0)))
712 Type *IT1, *IT2, *JT1, *JT2;
713 getInstructionTypes(I, IT1, IT2);
714 getInstructionTypes(J, JT1, JT2);
715 unsigned MaxTypeBits = std::max(
716 IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(),
717 IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits());
718 if (MaxTypeBits > Config.VectorBits)
721 // FIXME: handle addsub-type operations!
723 if (IsSimpleLoadStore) {
725 unsigned IAlignment, JAlignment;
726 int64_t OffsetInElmts = 0;
727 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
728 OffsetInElmts) && abs64(OffsetInElmts) == 1) {
729 if (Config.AlignedOnly) {
730 Type *aTypeI = isa<StoreInst>(I) ?
731 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType();
732 Type *aTypeJ = isa<StoreInst>(J) ?
733 cast<StoreInst>(J)->getValueOperand()->getType() : J->getType();
735 // An aligned load or store is possible only if the instruction
736 // with the lower offset has an alignment suitable for the
739 unsigned BottomAlignment = IAlignment;
740 if (OffsetInElmts < 0) BottomAlignment = JAlignment;
742 Type *VType = getVecTypeForPair(aTypeI, aTypeJ);
743 unsigned VecAlignment = TD->getPrefTypeAlignment(VType);
744 if (BottomAlignment < VecAlignment)
750 } else if (isa<ShuffleVectorInst>(I)) {
751 // Only merge two shuffles if they're both constant
752 return isa<Constant>(I->getOperand(2)) &&
753 isa<Constant>(J->getOperand(2));
754 // FIXME: We may want to vectorize non-constant shuffles also.
757 // The powi intrinsic is special because only the first argument is
758 // vectorized, the second arguments must be equal.
759 CallInst *CI = dyn_cast<CallInst>(I);
761 if (CI && (FI = CI->getCalledFunction()) &&
762 FI->getIntrinsicID() == Intrinsic::powi) {
764 Value *A1I = CI->getArgOperand(1),
765 *A1J = cast<CallInst>(J)->getArgOperand(1);
766 const SCEV *A1ISCEV = SE->getSCEV(A1I),
767 *A1JSCEV = SE->getSCEV(A1J);
768 return (A1ISCEV == A1JSCEV);
774 // Figure out whether or not J uses I and update the users and write-set
775 // structures associated with I. Specifically, Users represents the set of
776 // instructions that depend on I. WriteSet represents the set
777 // of memory locations that are dependent on I. If UpdateUsers is true,
778 // and J uses I, then Users is updated to contain J and WriteSet is updated
779 // to contain any memory locations to which J writes. The function returns
780 // true if J uses I. By default, alias analysis is used to determine
781 // whether J reads from memory that overlaps with a location in WriteSet.
782 // If LoadMoveSet is not null, then it is a previously-computed multimap
783 // where the key is the memory-based user instruction and the value is
784 // the instruction to be compared with I. So, if LoadMoveSet is provided,
785 // then the alias analysis is not used. This is necessary because this
786 // function is called during the process of moving instructions during
787 // vectorization and the results of the alias analysis are not stable during
789 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users,
790 AliasSetTracker &WriteSet, Instruction *I,
791 Instruction *J, bool UpdateUsers,
792 std::multimap<Value *, Value *> *LoadMoveSet) {
795 // This instruction may already be marked as a user due, for example, to
796 // being a member of a selected pair.
801 for (User::op_iterator JU = J->op_begin(), JE = J->op_end();
804 if (I == V || Users.count(V)) {
809 if (!UsesI && J->mayReadFromMemory()) {
811 VPIteratorPair JPairRange = LoadMoveSet->equal_range(J);
812 UsesI = isSecondInIteratorPair<Value*>(I, JPairRange);
814 for (AliasSetTracker::iterator W = WriteSet.begin(),
815 WE = WriteSet.end(); W != WE; ++W) {
816 if (W->aliasesUnknownInst(J, *AA)) {
824 if (UsesI && UpdateUsers) {
825 if (J->mayWriteToMemory()) WriteSet.add(J);
832 // This function iterates over all instruction pairs in the provided
833 // basic block and collects all candidate pairs for vectorization.
834 bool BBVectorize::getCandidatePairs(BasicBlock &BB,
835 BasicBlock::iterator &Start,
836 std::multimap<Value *, Value *> &CandidatePairs,
837 std::vector<Value *> &PairableInsts, bool NonPow2Len) {
838 BasicBlock::iterator E = BB.end();
839 if (Start == E) return false;
841 bool ShouldContinue = false, IAfterStart = false;
842 for (BasicBlock::iterator I = Start++; I != E; ++I) {
843 if (I == Start) IAfterStart = true;
845 bool IsSimpleLoadStore;
846 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue;
848 // Look for an instruction with which to pair instruction *I...
849 DenseSet<Value *> Users;
850 AliasSetTracker WriteSet(*AA);
851 bool JAfterStart = IAfterStart;
852 BasicBlock::iterator J = llvm::next(I);
853 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) {
854 if (J == Start) JAfterStart = true;
856 // Determine if J uses I, if so, exit the loop.
857 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep);
858 if (Config.FastDep) {
859 // Note: For this heuristic to be effective, independent operations
860 // must tend to be intermixed. This is likely to be true from some
861 // kinds of grouped loop unrolling (but not the generic LLVM pass),
862 // but otherwise may require some kind of reordering pass.
864 // When using fast dependency analysis,
865 // stop searching after first use:
871 // J does not use I, and comes before the first use of I, so it can be
872 // merged with I if the instructions are compatible.
873 if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len)) continue;
875 // J is a candidate for merging with I.
876 if (!PairableInsts.size() ||
877 PairableInsts[PairableInsts.size()-1] != I) {
878 PairableInsts.push_back(I);
881 CandidatePairs.insert(ValuePair(I, J));
883 // The next call to this function must start after the last instruction
884 // selected during this invocation.
886 Start = llvm::next(J);
887 IAfterStart = JAfterStart = false;
890 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair "
891 << *I << " <-> " << *J << "\n");
893 // If we have already found too many pairs, break here and this function
894 // will be called again starting after the last instruction selected
895 // during this invocation.
896 if (PairableInsts.size() >= Config.MaxInsts) {
897 ShouldContinue = true;
906 DEBUG(dbgs() << "BBV: found " << PairableInsts.size()
907 << " instructions with candidate pairs\n");
909 return ShouldContinue;
912 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that
913 // it looks for pairs such that both members have an input which is an
914 // output of PI or PJ.
915 void BBVectorize::computePairsConnectedTo(
916 std::multimap<Value *, Value *> &CandidatePairs,
917 std::vector<Value *> &PairableInsts,
918 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
922 // For each possible pairing for this variable, look at the uses of
923 // the first value...
924 for (Value::use_iterator I = P.first->use_begin(),
925 E = P.first->use_end(); I != E; ++I) {
926 if (isa<LoadInst>(*I)) {
927 // A pair cannot be connected to a load because the load only takes one
928 // operand (the address) and it is a scalar even after vectorization.
930 } else if ((SI = dyn_cast<StoreInst>(*I)) &&
931 P.first == SI->getPointerOperand()) {
932 // Similarly, a pair cannot be connected to a store through its
937 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I);
939 // For each use of the first variable, look for uses of the second
941 for (Value::use_iterator J = P.second->use_begin(),
942 E2 = P.second->use_end(); J != E2; ++J) {
943 if ((SJ = dyn_cast<StoreInst>(*J)) &&
944 P.second == SJ->getPointerOperand())
947 VPIteratorPair JPairRange = CandidatePairs.equal_range(*J);
950 if (isSecondInIteratorPair<Value*>(*J, IPairRange))
951 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J)));
954 if (isSecondInIteratorPair<Value*>(*I, JPairRange))
955 ConnectedPairs.insert(VPPair(P, ValuePair(*J, *I)));
958 if (Config.SplatBreaksChain) continue;
959 // Look for cases where just the first value in the pair is used by
960 // both members of another pair (splatting).
961 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) {
962 if ((SJ = dyn_cast<StoreInst>(*J)) &&
963 P.first == SJ->getPointerOperand())
966 if (isSecondInIteratorPair<Value*>(*J, IPairRange))
967 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J)));
971 if (Config.SplatBreaksChain) return;
972 // Look for cases where just the second value in the pair is used by
973 // both members of another pair (splatting).
974 for (Value::use_iterator I = P.second->use_begin(),
975 E = P.second->use_end(); I != E; ++I) {
976 if (isa<LoadInst>(*I))
978 else if ((SI = dyn_cast<StoreInst>(*I)) &&
979 P.second == SI->getPointerOperand())
982 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I);
984 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) {
985 if ((SJ = dyn_cast<StoreInst>(*J)) &&
986 P.second == SJ->getPointerOperand())
989 if (isSecondInIteratorPair<Value*>(*J, IPairRange))
990 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J)));
995 // This function figures out which pairs are connected. Two pairs are
996 // connected if some output of the first pair forms an input to both members
997 // of the second pair.
998 void BBVectorize::computeConnectedPairs(
999 std::multimap<Value *, Value *> &CandidatePairs,
1000 std::vector<Value *> &PairableInsts,
1001 std::multimap<ValuePair, ValuePair> &ConnectedPairs) {
1003 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
1004 PE = PairableInsts.end(); PI != PE; ++PI) {
1005 VPIteratorPair choiceRange = CandidatePairs.equal_range(*PI);
1007 for (std::multimap<Value *, Value *>::iterator P = choiceRange.first;
1008 P != choiceRange.second; ++P)
1009 computePairsConnectedTo(CandidatePairs, PairableInsts,
1010 ConnectedPairs, *P);
1013 DEBUG(dbgs() << "BBV: found " << ConnectedPairs.size()
1014 << " pair connections.\n");
1017 // This function builds a set of use tuples such that <A, B> is in the set
1018 // if B is in the use tree of A. If B is in the use tree of A, then B
1019 // depends on the output of A.
1020 void BBVectorize::buildDepMap(
1022 std::multimap<Value *, Value *> &CandidatePairs,
1023 std::vector<Value *> &PairableInsts,
1024 DenseSet<ValuePair> &PairableInstUsers) {
1025 DenseSet<Value *> IsInPair;
1026 for (std::multimap<Value *, Value *>::iterator C = CandidatePairs.begin(),
1027 E = CandidatePairs.end(); C != E; ++C) {
1028 IsInPair.insert(C->first);
1029 IsInPair.insert(C->second);
1032 // Iterate through the basic block, recording all Users of each
1033 // pairable instruction.
1035 BasicBlock::iterator E = BB.end();
1036 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) {
1037 if (IsInPair.find(I) == IsInPair.end()) continue;
1039 DenseSet<Value *> Users;
1040 AliasSetTracker WriteSet(*AA);
1041 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J)
1042 (void) trackUsesOfI(Users, WriteSet, I, J);
1044 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end();
1046 PairableInstUsers.insert(ValuePair(I, *U));
1050 // Returns true if an input to pair P is an output of pair Q and also an
1051 // input of pair Q is an output of pair P. If this is the case, then these
1052 // two pairs cannot be simultaneously fused.
1053 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q,
1054 DenseSet<ValuePair> &PairableInstUsers,
1055 std::multimap<ValuePair, ValuePair> *PairableInstUserMap) {
1056 // Two pairs are in conflict if they are mutual Users of eachother.
1057 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) ||
1058 PairableInstUsers.count(ValuePair(P.first, Q.second)) ||
1059 PairableInstUsers.count(ValuePair(P.second, Q.first)) ||
1060 PairableInstUsers.count(ValuePair(P.second, Q.second));
1061 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) ||
1062 PairableInstUsers.count(ValuePair(Q.first, P.second)) ||
1063 PairableInstUsers.count(ValuePair(Q.second, P.first)) ||
1064 PairableInstUsers.count(ValuePair(Q.second, P.second));
1065 if (PairableInstUserMap) {
1066 // FIXME: The expensive part of the cycle check is not so much the cycle
1067 // check itself but this edge insertion procedure. This needs some
1068 // profiling and probably a different data structure (same is true of
1069 // most uses of std::multimap).
1071 VPPIteratorPair QPairRange = PairableInstUserMap->equal_range(Q);
1072 if (!isSecondInIteratorPair(P, QPairRange))
1073 PairableInstUserMap->insert(VPPair(Q, P));
1076 VPPIteratorPair PPairRange = PairableInstUserMap->equal_range(P);
1077 if (!isSecondInIteratorPair(Q, PPairRange))
1078 PairableInstUserMap->insert(VPPair(P, Q));
1082 return (QUsesP && PUsesQ);
1085 // This function walks the use graph of current pairs to see if, starting
1086 // from P, the walk returns to P.
1087 bool BBVectorize::pairWillFormCycle(ValuePair P,
1088 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1089 DenseSet<ValuePair> &CurrentPairs) {
1090 DEBUG(if (DebugCycleCheck)
1091 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> "
1092 << *P.second << "\n");
1093 // A lookup table of visisted pairs is kept because the PairableInstUserMap
1094 // contains non-direct associations.
1095 DenseSet<ValuePair> Visited;
1096 SmallVector<ValuePair, 32> Q;
1097 // General depth-first post-order traversal:
1100 ValuePair QTop = Q.pop_back_val();
1101 Visited.insert(QTop);
1103 DEBUG(if (DebugCycleCheck)
1104 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> "
1105 << *QTop.second << "\n");
1106 VPPIteratorPair QPairRange = PairableInstUserMap.equal_range(QTop);
1107 for (std::multimap<ValuePair, ValuePair>::iterator C = QPairRange.first;
1108 C != QPairRange.second; ++C) {
1109 if (C->second == P) {
1111 << "BBV: rejected to prevent non-trivial cycle formation: "
1112 << *C->first.first << " <-> " << *C->first.second << "\n");
1116 if (CurrentPairs.count(C->second) && !Visited.count(C->second))
1117 Q.push_back(C->second);
1119 } while (!Q.empty());
1124 // This function builds the initial tree of connected pairs with the
1125 // pair J at the root.
1126 void BBVectorize::buildInitialTreeFor(
1127 std::multimap<Value *, Value *> &CandidatePairs,
1128 std::vector<Value *> &PairableInsts,
1129 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1130 DenseSet<ValuePair> &PairableInstUsers,
1131 DenseMap<Value *, Value *> &ChosenPairs,
1132 DenseMap<ValuePair, size_t> &Tree, ValuePair J) {
1133 // Each of these pairs is viewed as the root node of a Tree. The Tree
1134 // is then walked (depth-first). As this happens, we keep track of
1135 // the pairs that compose the Tree and the maximum depth of the Tree.
1136 SmallVector<ValuePairWithDepth, 32> Q;
1137 // General depth-first post-order traversal:
1138 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1140 ValuePairWithDepth QTop = Q.back();
1142 // Push each child onto the queue:
1143 bool MoreChildren = false;
1144 size_t MaxChildDepth = QTop.second;
1145 VPPIteratorPair qtRange = ConnectedPairs.equal_range(QTop.first);
1146 for (std::multimap<ValuePair, ValuePair>::iterator k = qtRange.first;
1147 k != qtRange.second; ++k) {
1148 // Make sure that this child pair is still a candidate:
1149 bool IsStillCand = false;
1150 VPIteratorPair checkRange =
1151 CandidatePairs.equal_range(k->second.first);
1152 for (std::multimap<Value *, Value *>::iterator m = checkRange.first;
1153 m != checkRange.second; ++m) {
1154 if (m->second == k->second.second) {
1161 DenseMap<ValuePair, size_t>::iterator C = Tree.find(k->second);
1162 if (C == Tree.end()) {
1163 size_t d = getDepthFactor(k->second.first);
1164 Q.push_back(ValuePairWithDepth(k->second, QTop.second+d));
1165 MoreChildren = true;
1167 MaxChildDepth = std::max(MaxChildDepth, C->second);
1172 if (!MoreChildren) {
1173 // Record the current pair as part of the Tree:
1174 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth));
1177 } while (!Q.empty());
1180 // Given some initial tree, prune it by removing conflicting pairs (pairs
1181 // that cannot be simultaneously chosen for vectorization).
1182 void BBVectorize::pruneTreeFor(
1183 std::multimap<Value *, Value *> &CandidatePairs,
1184 std::vector<Value *> &PairableInsts,
1185 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1186 DenseSet<ValuePair> &PairableInstUsers,
1187 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1188 DenseMap<Value *, Value *> &ChosenPairs,
1189 DenseMap<ValuePair, size_t> &Tree,
1190 DenseSet<ValuePair> &PrunedTree, ValuePair J,
1191 bool UseCycleCheck) {
1192 SmallVector<ValuePairWithDepth, 32> Q;
1193 // General depth-first post-order traversal:
1194 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1196 ValuePairWithDepth QTop = Q.pop_back_val();
1197 PrunedTree.insert(QTop.first);
1199 // Visit each child, pruning as necessary...
1200 DenseMap<ValuePair, size_t> BestChildren;
1201 VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first);
1202 for (std::multimap<ValuePair, ValuePair>::iterator K = QTopRange.first;
1203 K != QTopRange.second; ++K) {
1204 DenseMap<ValuePair, size_t>::iterator C = Tree.find(K->second);
1205 if (C == Tree.end()) continue;
1207 // This child is in the Tree, now we need to make sure it is the
1208 // best of any conflicting children. There could be multiple
1209 // conflicting children, so first, determine if we're keeping
1210 // this child, then delete conflicting children as necessary.
1212 // It is also necessary to guard against pairing-induced
1213 // dependencies. Consider instructions a .. x .. y .. b
1214 // such that (a,b) are to be fused and (x,y) are to be fused
1215 // but a is an input to x and b is an output from y. This
1216 // means that y cannot be moved after b but x must be moved
1217 // after b for (a,b) to be fused. In other words, after
1218 // fusing (a,b) we have y .. a/b .. x where y is an input
1219 // to a/b and x is an output to a/b: x and y can no longer
1220 // be legally fused. To prevent this condition, we must
1221 // make sure that a child pair added to the Tree is not
1222 // both an input and output of an already-selected pair.
1224 // Pairing-induced dependencies can also form from more complicated
1225 // cycles. The pair vs. pair conflicts are easy to check, and so
1226 // that is done explicitly for "fast rejection", and because for
1227 // child vs. child conflicts, we may prefer to keep the current
1228 // pair in preference to the already-selected child.
1229 DenseSet<ValuePair> CurrentPairs;
1232 for (DenseMap<ValuePair, size_t>::iterator C2
1233 = BestChildren.begin(), E2 = BestChildren.end();
1235 if (C2->first.first == C->first.first ||
1236 C2->first.first == C->first.second ||
1237 C2->first.second == C->first.first ||
1238 C2->first.second == C->first.second ||
1239 pairsConflict(C2->first, C->first, PairableInstUsers,
1240 UseCycleCheck ? &PairableInstUserMap : 0)) {
1241 if (C2->second >= C->second) {
1246 CurrentPairs.insert(C2->first);
1249 if (!CanAdd) continue;
1251 // Even worse, this child could conflict with another node already
1252 // selected for the Tree. If that is the case, ignore this child.
1253 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(),
1254 E2 = PrunedTree.end(); T != E2; ++T) {
1255 if (T->first == C->first.first ||
1256 T->first == C->first.second ||
1257 T->second == C->first.first ||
1258 T->second == C->first.second ||
1259 pairsConflict(*T, C->first, PairableInstUsers,
1260 UseCycleCheck ? &PairableInstUserMap : 0)) {
1265 CurrentPairs.insert(*T);
1267 if (!CanAdd) continue;
1269 // And check the queue too...
1270 for (SmallVector<ValuePairWithDepth, 32>::iterator C2 = Q.begin(),
1271 E2 = Q.end(); C2 != E2; ++C2) {
1272 if (C2->first.first == C->first.first ||
1273 C2->first.first == C->first.second ||
1274 C2->first.second == C->first.first ||
1275 C2->first.second == C->first.second ||
1276 pairsConflict(C2->first, C->first, PairableInstUsers,
1277 UseCycleCheck ? &PairableInstUserMap : 0)) {
1282 CurrentPairs.insert(C2->first);
1284 if (!CanAdd) continue;
1286 // Last but not least, check for a conflict with any of the
1287 // already-chosen pairs.
1288 for (DenseMap<Value *, Value *>::iterator C2 =
1289 ChosenPairs.begin(), E2 = ChosenPairs.end();
1291 if (pairsConflict(*C2, C->first, PairableInstUsers,
1292 UseCycleCheck ? &PairableInstUserMap : 0)) {
1297 CurrentPairs.insert(*C2);
1299 if (!CanAdd) continue;
1301 // To check for non-trivial cycles formed by the addition of the
1302 // current pair we've formed a list of all relevant pairs, now use a
1303 // graph walk to check for a cycle. We start from the current pair and
1304 // walk the use tree to see if we again reach the current pair. If we
1305 // do, then the current pair is rejected.
1307 // FIXME: It may be more efficient to use a topological-ordering
1308 // algorithm to improve the cycle check. This should be investigated.
1309 if (UseCycleCheck &&
1310 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs))
1313 // This child can be added, but we may have chosen it in preference
1314 // to an already-selected child. Check for this here, and if a
1315 // conflict is found, then remove the previously-selected child
1316 // before adding this one in its place.
1317 for (DenseMap<ValuePair, size_t>::iterator C2
1318 = BestChildren.begin(); C2 != BestChildren.end();) {
1319 if (C2->first.first == C->first.first ||
1320 C2->first.first == C->first.second ||
1321 C2->first.second == C->first.first ||
1322 C2->first.second == C->first.second ||
1323 pairsConflict(C2->first, C->first, PairableInstUsers))
1324 BestChildren.erase(C2++);
1329 BestChildren.insert(ValuePairWithDepth(C->first, C->second));
1332 for (DenseMap<ValuePair, size_t>::iterator C
1333 = BestChildren.begin(), E2 = BestChildren.end();
1335 size_t DepthF = getDepthFactor(C->first.first);
1336 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF));
1338 } while (!Q.empty());
1341 // This function finds the best tree of mututally-compatible connected
1342 // pairs, given the choice of root pairs as an iterator range.
1343 void BBVectorize::findBestTreeFor(
1344 std::multimap<Value *, Value *> &CandidatePairs,
1345 std::vector<Value *> &PairableInsts,
1346 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1347 DenseSet<ValuePair> &PairableInstUsers,
1348 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1349 DenseMap<Value *, Value *> &ChosenPairs,
1350 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth,
1351 size_t &BestEffSize, VPIteratorPair ChoiceRange,
1352 bool UseCycleCheck) {
1353 for (std::multimap<Value *, Value *>::iterator J = ChoiceRange.first;
1354 J != ChoiceRange.second; ++J) {
1356 // Before going any further, make sure that this pair does not
1357 // conflict with any already-selected pairs (see comment below
1358 // near the Tree pruning for more details).
1359 DenseSet<ValuePair> ChosenPairSet;
1360 bool DoesConflict = false;
1361 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(),
1362 E = ChosenPairs.end(); C != E; ++C) {
1363 if (pairsConflict(*C, *J, PairableInstUsers,
1364 UseCycleCheck ? &PairableInstUserMap : 0)) {
1365 DoesConflict = true;
1369 ChosenPairSet.insert(*C);
1371 if (DoesConflict) continue;
1373 if (UseCycleCheck &&
1374 pairWillFormCycle(*J, PairableInstUserMap, ChosenPairSet))
1377 DenseMap<ValuePair, size_t> Tree;
1378 buildInitialTreeFor(CandidatePairs, PairableInsts, ConnectedPairs,
1379 PairableInstUsers, ChosenPairs, Tree, *J);
1381 // Because we'll keep the child with the largest depth, the largest
1382 // depth is still the same in the unpruned Tree.
1383 size_t MaxDepth = Tree.lookup(*J);
1385 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found Tree for pair {"
1386 << *J->first << " <-> " << *J->second << "} of depth " <<
1387 MaxDepth << " and size " << Tree.size() << "\n");
1389 // At this point the Tree has been constructed, but, may contain
1390 // contradictory children (meaning that different children of
1391 // some tree node may be attempting to fuse the same instruction).
1392 // So now we walk the tree again, in the case of a conflict,
1393 // keep only the child with the largest depth. To break a tie,
1394 // favor the first child.
1396 DenseSet<ValuePair> PrunedTree;
1397 pruneTreeFor(CandidatePairs, PairableInsts, ConnectedPairs,
1398 PairableInstUsers, PairableInstUserMap, ChosenPairs, Tree,
1399 PrunedTree, *J, UseCycleCheck);
1402 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
1403 E = PrunedTree.end(); S != E; ++S)
1404 EffSize += getDepthFactor(S->first);
1406 DEBUG(if (DebugPairSelection)
1407 dbgs() << "BBV: found pruned Tree for pair {"
1408 << *J->first << " <-> " << *J->second << "} of depth " <<
1409 MaxDepth << " and size " << PrunedTree.size() <<
1410 " (effective size: " << EffSize << ")\n");
1411 if (MaxDepth >= Config.ReqChainDepth && EffSize > BestEffSize) {
1412 BestMaxDepth = MaxDepth;
1413 BestEffSize = EffSize;
1414 BestTree = PrunedTree;
1419 // Given the list of candidate pairs, this function selects those
1420 // that will be fused into vector instructions.
1421 void BBVectorize::choosePairs(
1422 std::multimap<Value *, Value *> &CandidatePairs,
1423 std::vector<Value *> &PairableInsts,
1424 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1425 DenseSet<ValuePair> &PairableInstUsers,
1426 DenseMap<Value *, Value *>& ChosenPairs) {
1427 bool UseCycleCheck =
1428 CandidatePairs.size() <= Config.MaxCandPairsForCycleCheck;
1429 std::multimap<ValuePair, ValuePair> PairableInstUserMap;
1430 for (std::vector<Value *>::iterator I = PairableInsts.begin(),
1431 E = PairableInsts.end(); I != E; ++I) {
1432 // The number of possible pairings for this variable:
1433 size_t NumChoices = CandidatePairs.count(*I);
1434 if (!NumChoices) continue;
1436 VPIteratorPair ChoiceRange = CandidatePairs.equal_range(*I);
1438 // The best pair to choose and its tree:
1439 size_t BestMaxDepth = 0, BestEffSize = 0;
1440 DenseSet<ValuePair> BestTree;
1441 findBestTreeFor(CandidatePairs, PairableInsts, ConnectedPairs,
1442 PairableInstUsers, PairableInstUserMap, ChosenPairs,
1443 BestTree, BestMaxDepth, BestEffSize, ChoiceRange,
1446 // A tree has been chosen (or not) at this point. If no tree was
1447 // chosen, then this instruction, I, cannot be paired (and is no longer
1450 DEBUG(if (BestTree.size() > 0)
1451 dbgs() << "BBV: selected pairs in the best tree for: "
1452 << *cast<Instruction>(*I) << "\n");
1454 for (DenseSet<ValuePair>::iterator S = BestTree.begin(),
1455 SE2 = BestTree.end(); S != SE2; ++S) {
1456 // Insert the members of this tree into the list of chosen pairs.
1457 ChosenPairs.insert(ValuePair(S->first, S->second));
1458 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " <<
1459 *S->second << "\n");
1461 // Remove all candidate pairs that have values in the chosen tree.
1462 for (std::multimap<Value *, Value *>::iterator K =
1463 CandidatePairs.begin(); K != CandidatePairs.end();) {
1464 if (K->first == S->first || K->second == S->first ||
1465 K->second == S->second || K->first == S->second) {
1466 // Don't remove the actual pair chosen so that it can be used
1467 // in subsequent tree selections.
1468 if (!(K->first == S->first && K->second == S->second))
1469 CandidatePairs.erase(K++);
1479 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n");
1482 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o,
1487 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) +
1488 (n > 0 ? "." + utostr(n) : "")).str();
1491 // Returns the value that is to be used as the pointer input to the vector
1492 // instruction that fuses I with J.
1493 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context,
1494 Instruction *I, Instruction *J, unsigned o,
1495 bool &FlipMemInputs) {
1497 unsigned IAlignment, JAlignment;
1498 int64_t OffsetInElmts;
1499 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
1502 // The pointer value is taken to be the one with the lowest offset.
1504 if (OffsetInElmts > 0) {
1507 FlipMemInputs = true;
1511 Type *ArgTypeI = cast<PointerType>(IPtr->getType())->getElementType();
1512 Type *ArgTypeJ = cast<PointerType>(JPtr->getType())->getElementType();
1513 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
1514 Type *VArgPtrType = PointerType::get(VArgType,
1515 cast<PointerType>(IPtr->getType())->getAddressSpace());
1516 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o),
1517 /* insert before */ FlipMemInputs ? J : I);
1520 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J,
1521 unsigned MaskOffset, unsigned NumInElem,
1522 unsigned NumInElem1, unsigned IdxOffset,
1523 std::vector<Constant*> &Mask) {
1524 unsigned NumElem1 = cast<VectorType>(J->getType())->getNumElements();
1525 for (unsigned v = 0; v < NumElem1; ++v) {
1526 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v);
1528 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context));
1530 unsigned mm = m + (int) IdxOffset;
1531 if (m >= (int) NumInElem1)
1532 mm += (int) NumInElem;
1534 Mask[v+MaskOffset] =
1535 ConstantInt::get(Type::getInt32Ty(Context), mm);
1540 // Returns the value that is to be used as the vector-shuffle mask to the
1541 // vector instruction that fuses I with J.
1542 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context,
1543 Instruction *I, Instruction *J) {
1544 // This is the shuffle mask. We need to append the second
1545 // mask to the first, and the numbers need to be adjusted.
1547 Type *ArgTypeI = I->getType();
1548 Type *ArgTypeJ = J->getType();
1549 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
1551 unsigned NumElemI = cast<VectorType>(ArgTypeI)->getNumElements();
1553 // Get the total number of elements in the fused vector type.
1554 // By definition, this must equal the number of elements in
1556 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements();
1557 std::vector<Constant*> Mask(NumElem);
1559 Type *OpTypeI = I->getOperand(0)->getType();
1560 unsigned NumInElemI = cast<VectorType>(OpTypeI)->getNumElements();
1561 Type *OpTypeJ = J->getOperand(0)->getType();
1562 unsigned NumInElemJ = cast<VectorType>(OpTypeJ)->getNumElements();
1564 // The fused vector will be:
1565 // -----------------------------------------------------
1566 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ |
1567 // -----------------------------------------------------
1568 // from which we'll extract NumElem total elements (where the first NumElemI
1569 // of them come from the mask in I and the remainder come from the mask
1572 // For the mask from the first pair...
1573 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI,
1576 // For the mask from the second pair...
1577 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ,
1580 return ConstantVector::get(Mask);
1583 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I,
1584 Instruction *J, unsigned o, Value *&LOp,
1586 Type *ArgTypeL, Type *ArgTypeH,
1588 bool ExpandedIEChain = false;
1589 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) {
1590 // If we have a pure insertelement chain, then this can be rewritten
1591 // into a chain that directly builds the larger type.
1592 bool PureChain = true;
1593 InsertElementInst *LIENext = LIE;
1595 if (!isa<UndefValue>(LIENext->getOperand(0)) &&
1596 !isa<InsertElementInst>(LIENext->getOperand(0))) {
1601 dyn_cast<InsertElementInst>(LIENext->getOperand(0))));
1604 SmallVector<Value *, 8> VectElemts(numElemL,
1605 UndefValue::get(ArgTypeL->getScalarType()));
1606 InsertElementInst *LIENext = LIE;
1609 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue();
1610 VectElemts[Idx] = LIENext->getOperand(1);
1612 dyn_cast<InsertElementInst>(LIENext->getOperand(0))));
1615 Value *LIEPrev = UndefValue::get(ArgTypeH);
1616 for (unsigned i = 0; i < numElemL; ++i) {
1617 if (isa<UndefValue>(VectElemts[i])) continue;
1618 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i],
1619 ConstantInt::get(Type::getInt32Ty(Context),
1621 getReplacementName(I, true, o, i+1));
1622 LIENext->insertBefore(J);
1626 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH);
1627 ExpandedIEChain = true;
1631 return ExpandedIEChain;
1634 // Returns the value to be used as the specified operand of the vector
1635 // instruction that fuses I with J.
1636 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I,
1637 Instruction *J, unsigned o, bool FlipMemInputs) {
1638 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
1639 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1);
1641 // Compute the fused vector type for this operand
1642 Type *ArgTypeI = I->getOperand(o)->getType();
1643 Type *ArgTypeJ = J->getOperand(o)->getType();
1644 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
1646 Instruction *L = I, *H = J;
1647 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ;
1648 if (FlipMemInputs) {
1651 ArgTypeL = ArgTypeJ;
1652 ArgTypeH = ArgTypeI;
1656 if (ArgTypeL->isVectorTy())
1657 numElemL = cast<VectorType>(ArgTypeL)->getNumElements();
1662 if (ArgTypeH->isVectorTy())
1663 numElemH = cast<VectorType>(ArgTypeH)->getNumElements();
1667 Value *LOp = L->getOperand(o);
1668 Value *HOp = H->getOperand(o);
1669 unsigned numElem = VArgType->getNumElements();
1671 // First, we check if we can reuse the "original" vector outputs (if these
1672 // exist). We might need a shuffle.
1673 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp);
1674 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp);
1675 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp);
1676 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp);
1678 // FIXME: If we're fusing shuffle instructions, then we can't apply this
1679 // optimization. The input vectors to the shuffle might be a different
1680 // length from the shuffle outputs. Unfortunately, the replacement
1681 // shuffle mask has already been formed, and the mask entries are sensitive
1682 // to the sizes of the inputs.
1683 bool IsSizeChangeShuffle =
1684 isa<ShuffleVectorInst>(L) &&
1685 (LOp->getType() != L->getType() || HOp->getType() != H->getType());
1687 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) {
1688 // We can have at most two unique vector inputs.
1689 bool CanUseInputs = true;
1692 I1 = LEE->getOperand(0);
1694 I1 = LSV->getOperand(0);
1695 I2 = LSV->getOperand(1);
1696 if (I2 == I1 || isa<UndefValue>(I2))
1701 Value *I3 = HEE->getOperand(0);
1702 if (!I2 && I3 != I1)
1704 else if (I3 != I1 && I3 != I2)
1705 CanUseInputs = false;
1707 Value *I3 = HSV->getOperand(0);
1708 if (!I2 && I3 != I1)
1710 else if (I3 != I1 && I3 != I2)
1711 CanUseInputs = false;
1714 Value *I4 = HSV->getOperand(1);
1715 if (!isa<UndefValue>(I4)) {
1716 if (!I2 && I4 != I1)
1718 else if (I4 != I1 && I4 != I2)
1719 CanUseInputs = false;
1726 cast<VectorType>(cast<Instruction>(LOp)->getOperand(0)->getType())
1729 cast<VectorType>(cast<Instruction>(HOp)->getOperand(0)->getType())
1732 // We have one or two input vectors. We need to map each index of the
1733 // operands to the index of the original vector.
1734 SmallVector<std::pair<int, int>, 8> II(numElem);
1735 for (unsigned i = 0; i < numElemL; ++i) {
1739 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue();
1740 INum = LEE->getOperand(0) == I1 ? 0 : 1;
1742 Idx = LSV->getMaskValue(i);
1743 if (Idx < (int) LOpElem) {
1744 INum = LSV->getOperand(0) == I1 ? 0 : 1;
1747 INum = LSV->getOperand(1) == I1 ? 0 : 1;
1751 II[i] = std::pair<int, int>(Idx, INum);
1753 for (unsigned i = 0; i < numElemH; ++i) {
1757 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue();
1758 INum = HEE->getOperand(0) == I1 ? 0 : 1;
1760 Idx = HSV->getMaskValue(i);
1761 if (Idx < (int) HOpElem) {
1762 INum = HSV->getOperand(0) == I1 ? 0 : 1;
1765 INum = HSV->getOperand(1) == I1 ? 0 : 1;
1769 II[i + numElemL] = std::pair<int, int>(Idx, INum);
1772 // We now have an array which tells us from which index of which
1773 // input vector each element of the operand comes.
1774 VectorType *I1T = cast<VectorType>(I1->getType());
1775 unsigned I1Elem = I1T->getNumElements();
1778 // In this case there is only one underlying vector input. Check for
1779 // the trivial case where we can use the input directly.
1780 if (I1Elem == numElem) {
1781 bool ElemInOrder = true;
1782 for (unsigned i = 0; i < numElem; ++i) {
1783 if (II[i].first != (int) i && II[i].first != -1) {
1784 ElemInOrder = false;
1793 // A shuffle is needed.
1794 std::vector<Constant *> Mask(numElem);
1795 for (unsigned i = 0; i < numElem; ++i) {
1796 int Idx = II[i].first;
1798 Mask[i] = UndefValue::get(Type::getInt32Ty(Context));
1800 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
1804 new ShuffleVectorInst(I1, UndefValue::get(I1T),
1805 ConstantVector::get(Mask),
1806 getReplacementName(I, true, o));
1811 VectorType *I2T = cast<VectorType>(I2->getType());
1812 unsigned I2Elem = I2T->getNumElements();
1814 // This input comes from two distinct vectors. The first step is to
1815 // make sure that both vectors are the same length. If not, the
1816 // smaller one will need to grow before they can be shuffled together.
1817 if (I1Elem < I2Elem) {
1818 std::vector<Constant *> Mask(I2Elem);
1820 for (; v < I1Elem; ++v)
1821 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
1822 for (; v < I2Elem; ++v)
1823 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
1825 Instruction *NewI1 =
1826 new ShuffleVectorInst(I1, UndefValue::get(I1T),
1827 ConstantVector::get(Mask),
1828 getReplacementName(I, true, o, 1));
1829 NewI1->insertBefore(J);
1833 } else if (I1Elem > I2Elem) {
1834 std::vector<Constant *> Mask(I1Elem);
1836 for (; v < I2Elem; ++v)
1837 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
1838 for (; v < I1Elem; ++v)
1839 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
1841 Instruction *NewI2 =
1842 new ShuffleVectorInst(I2, UndefValue::get(I2T),
1843 ConstantVector::get(Mask),
1844 getReplacementName(I, true, o, 1));
1845 NewI2->insertBefore(J);
1851 // Now that both I1 and I2 are the same length we can shuffle them
1852 // together (and use the result).
1853 std::vector<Constant *> Mask(numElem);
1854 for (unsigned v = 0; v < numElem; ++v) {
1855 if (II[v].first == -1) {
1856 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
1858 int Idx = II[v].first + II[v].second * I1Elem;
1859 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
1863 Instruction *NewOp =
1864 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask),
1865 getReplacementName(I, true, o));
1866 NewOp->insertBefore(J);
1871 Type *ArgType = ArgTypeL;
1872 if (numElemL < numElemH) {
1873 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH,
1874 ArgTypeL, VArgType, 1)) {
1875 // This is another short-circuit case: we're combining a scalar into
1876 // a vector that is formed by an IE chain. We've just expanded the IE
1877 // chain, now insert the scalar and we're done.
1879 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0,
1880 getReplacementName(I, true, o));
1883 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL,
1885 // The two vector inputs to the shuffle must be the same length,
1886 // so extend the smaller vector to be the same length as the larger one.
1890 std::vector<Constant *> Mask(numElemH);
1892 for (; v < numElemL; ++v)
1893 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
1894 for (; v < numElemH; ++v)
1895 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
1897 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL),
1898 ConstantVector::get(Mask),
1899 getReplacementName(I, true, o, 1));
1901 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0,
1902 getReplacementName(I, true, o, 1));
1905 NLOp->insertBefore(J);
1910 } else if (numElemL > numElemH) {
1911 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL,
1912 ArgTypeH, VArgType)) {
1914 InsertElementInst::Create(LOp, HOp,
1915 ConstantInt::get(Type::getInt32Ty(Context),
1917 getReplacementName(I, true, o));
1920 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH,
1924 std::vector<Constant *> Mask(numElemL);
1926 for (; v < numElemH; ++v)
1927 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
1928 for (; v < numElemL; ++v)
1929 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
1931 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH),
1932 ConstantVector::get(Mask),
1933 getReplacementName(I, true, o, 1));
1935 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0,
1936 getReplacementName(I, true, o, 1));
1939 NHOp->insertBefore(J);
1944 if (ArgType->isVectorTy()) {
1945 unsigned numElem = cast<VectorType>(VArgType)->getNumElements();
1946 std::vector<Constant*> Mask(numElem);
1947 for (unsigned v = 0; v < numElem; ++v) {
1949 // If the low vector was expanded, we need to skip the extra
1950 // undefined entries.
1951 if (v >= numElemL && numElemH > numElemL)
1952 Idx += (numElemH - numElemL);
1953 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
1956 Instruction *BV = new ShuffleVectorInst(LOp, HOp,
1957 ConstantVector::get(Mask),
1958 getReplacementName(I, true, o));
1959 BV->insertBefore(J);
1963 Instruction *BV1 = InsertElementInst::Create(
1964 UndefValue::get(VArgType), LOp, CV0,
1965 getReplacementName(I, true, o, 1));
1966 BV1->insertBefore(I);
1967 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1,
1968 getReplacementName(I, true, o, 2));
1969 BV2->insertBefore(J);
1973 // This function creates an array of values that will be used as the inputs
1974 // to the vector instruction that fuses I with J.
1975 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context,
1976 Instruction *I, Instruction *J,
1977 SmallVector<Value *, 3> &ReplacedOperands,
1978 bool &FlipMemInputs) {
1979 FlipMemInputs = false;
1980 unsigned NumOperands = I->getNumOperands();
1982 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) {
1983 // Iterate backward so that we look at the store pointer
1984 // first and know whether or not we need to flip the inputs.
1986 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) {
1987 // This is the pointer for a load/store instruction.
1988 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o,
1991 } else if (isa<CallInst>(I)) {
1992 Function *F = cast<CallInst>(I)->getCalledFunction();
1993 unsigned IID = F->getIntrinsicID();
1994 if (o == NumOperands-1) {
1995 BasicBlock &BB = *I->getParent();
1997 Module *M = BB.getParent()->getParent();
1998 Type *ArgTypeI = I->getType();
1999 Type *ArgTypeJ = J->getType();
2000 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2002 ReplacedOperands[o] = Intrinsic::getDeclaration(M,
2003 (Intrinsic::ID) IID, VArgType);
2005 } else if (IID == Intrinsic::powi && o == 1) {
2006 // The second argument of powi is a single integer and we've already
2007 // checked that both arguments are equal. As a result, we just keep
2008 // I's second argument.
2009 ReplacedOperands[o] = I->getOperand(o);
2012 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) {
2013 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J);
2017 ReplacedOperands[o] =
2018 getReplacementInput(Context, I, J, o, FlipMemInputs);
2022 // This function creates two values that represent the outputs of the
2023 // original I and J instructions. These are generally vector shuffles
2024 // or extracts. In many cases, these will end up being unused and, thus,
2025 // eliminated by later passes.
2026 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
2027 Instruction *J, Instruction *K,
2028 Instruction *&InsertionPt,
2029 Instruction *&K1, Instruction *&K2,
2030 bool &FlipMemInputs) {
2031 if (isa<StoreInst>(I)) {
2032 AA->replaceWithNewValue(I, K);
2033 AA->replaceWithNewValue(J, K);
2035 Type *IType = I->getType();
2036 Type *JType = J->getType();
2038 VectorType *VType = getVecTypeForPair(IType, JType);
2039 unsigned numElem = VType->getNumElements();
2041 unsigned numElemI, numElemJ;
2042 if (IType->isVectorTy())
2043 numElemI = cast<VectorType>(IType)->getNumElements();
2047 if (JType->isVectorTy())
2048 numElemJ = cast<VectorType>(JType)->getNumElements();
2052 if (IType->isVectorTy()) {
2053 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI);
2054 for (unsigned v = 0; v < numElemI; ++v) {
2055 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2056 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v);
2059 K1 = new ShuffleVectorInst(K, UndefValue::get(VType),
2060 ConstantVector::get(
2061 FlipMemInputs ? Mask2 : Mask1),
2062 getReplacementName(K, false, 1));
2064 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2065 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1);
2066 K1 = ExtractElementInst::Create(K, FlipMemInputs ? CV1 : CV0,
2067 getReplacementName(K, false, 1));
2070 if (JType->isVectorTy()) {
2071 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ);
2072 for (unsigned v = 0; v < numElemJ; ++v) {
2073 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2074 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v);
2077 K2 = new ShuffleVectorInst(K, UndefValue::get(VType),
2078 ConstantVector::get(
2079 FlipMemInputs ? Mask1 : Mask2),
2080 getReplacementName(K, false, 2));
2082 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2083 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1);
2084 K2 = ExtractElementInst::Create(K, FlipMemInputs ? CV0 : CV1,
2085 getReplacementName(K, false, 2));
2089 K2->insertAfter(K1);
2094 // Move all uses of the function I (including pairing-induced uses) after J.
2095 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB,
2096 std::multimap<Value *, Value *> &LoadMoveSet,
2097 Instruction *I, Instruction *J) {
2098 // Skip to the first instruction past I.
2099 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2101 DenseSet<Value *> Users;
2102 AliasSetTracker WriteSet(*AA);
2103 for (; cast<Instruction>(L) != J; ++L)
2104 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet);
2106 assert(cast<Instruction>(L) == J &&
2107 "Tracking has not proceeded far enough to check for dependencies");
2108 // If J is now in the use set of I, then trackUsesOfI will return true
2109 // and we have a dependency cycle (and the fusing operation must abort).
2110 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSet);
2113 // Move all uses of the function I (including pairing-induced uses) after J.
2114 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB,
2115 std::multimap<Value *, Value *> &LoadMoveSet,
2116 Instruction *&InsertionPt,
2117 Instruction *I, Instruction *J) {
2118 // Skip to the first instruction past I.
2119 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2121 DenseSet<Value *> Users;
2122 AliasSetTracker WriteSet(*AA);
2123 for (; cast<Instruction>(L) != J;) {
2124 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet)) {
2125 // Move this instruction
2126 Instruction *InstToMove = L; ++L;
2128 DEBUG(dbgs() << "BBV: moving: " << *InstToMove <<
2129 " to after " << *InsertionPt << "\n");
2130 InstToMove->removeFromParent();
2131 InstToMove->insertAfter(InsertionPt);
2132 InsertionPt = InstToMove;
2139 // Collect all load instruction that are in the move set of a given first
2140 // pair member. These loads depend on the first instruction, I, and so need
2141 // to be moved after J (the second instruction) when the pair is fused.
2142 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB,
2143 DenseMap<Value *, Value *> &ChosenPairs,
2144 std::multimap<Value *, Value *> &LoadMoveSet,
2146 // Skip to the first instruction past I.
2147 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2149 DenseSet<Value *> Users;
2150 AliasSetTracker WriteSet(*AA);
2152 // Note: We cannot end the loop when we reach J because J could be moved
2153 // farther down the use chain by another instruction pairing. Also, J
2154 // could be before I if this is an inverted input.
2155 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) {
2156 if (trackUsesOfI(Users, WriteSet, I, L)) {
2157 if (L->mayReadFromMemory())
2158 LoadMoveSet.insert(ValuePair(L, I));
2163 // In cases where both load/stores and the computation of their pointers
2164 // are chosen for vectorization, we can end up in a situation where the
2165 // aliasing analysis starts returning different query results as the
2166 // process of fusing instruction pairs continues. Because the algorithm
2167 // relies on finding the same use trees here as were found earlier, we'll
2168 // need to precompute the necessary aliasing information here and then
2169 // manually update it during the fusion process.
2170 void BBVectorize::collectLoadMoveSet(BasicBlock &BB,
2171 std::vector<Value *> &PairableInsts,
2172 DenseMap<Value *, Value *> &ChosenPairs,
2173 std::multimap<Value *, Value *> &LoadMoveSet) {
2174 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
2175 PIE = PairableInsts.end(); PI != PIE; ++PI) {
2176 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI);
2177 if (P == ChosenPairs.end()) continue;
2179 Instruction *I = cast<Instruction>(P->first);
2180 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, I);
2184 // When the first instruction in each pair is cloned, it will inherit its
2185 // parent's metadata. This metadata must be combined with that of the other
2186 // instruction in a safe way.
2187 void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) {
2188 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata;
2189 K->getAllMetadataOtherThanDebugLoc(Metadata);
2190 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) {
2191 unsigned Kind = Metadata[i].first;
2192 MDNode *JMD = J->getMetadata(Kind);
2193 MDNode *KMD = Metadata[i].second;
2197 K->setMetadata(Kind, 0); // Remove unknown metadata
2199 case LLVMContext::MD_tbaa:
2200 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD));
2202 case LLVMContext::MD_fpmath:
2203 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD));
2209 // This function fuses the chosen instruction pairs into vector instructions,
2210 // taking care preserve any needed scalar outputs and, then, it reorders the
2211 // remaining instructions as needed (users of the first member of the pair
2212 // need to be moved to after the location of the second member of the pair
2213 // because the vector instruction is inserted in the location of the pair's
2215 void BBVectorize::fuseChosenPairs(BasicBlock &BB,
2216 std::vector<Value *> &PairableInsts,
2217 DenseMap<Value *, Value *> &ChosenPairs) {
2218 LLVMContext& Context = BB.getContext();
2220 // During the vectorization process, the order of the pairs to be fused
2221 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs
2222 // list. After a pair is fused, the flipped pair is removed from the list.
2223 std::vector<ValuePair> FlippedPairs;
2224 FlippedPairs.reserve(ChosenPairs.size());
2225 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(),
2226 E = ChosenPairs.end(); P != E; ++P)
2227 FlippedPairs.push_back(ValuePair(P->second, P->first));
2228 for (std::vector<ValuePair>::iterator P = FlippedPairs.begin(),
2229 E = FlippedPairs.end(); P != E; ++P)
2230 ChosenPairs.insert(*P);
2232 std::multimap<Value *, Value *> LoadMoveSet;
2233 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, LoadMoveSet);
2235 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n");
2237 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) {
2238 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI);
2239 if (P == ChosenPairs.end()) {
2244 if (getDepthFactor(P->first) == 0) {
2245 // These instructions are not really fused, but are tracked as though
2246 // they are. Any case in which it would be interesting to fuse them
2247 // will be taken care of by InstCombine.
2253 Instruction *I = cast<Instruction>(P->first),
2254 *J = cast<Instruction>(P->second);
2256 DEBUG(dbgs() << "BBV: fusing: " << *I <<
2257 " <-> " << *J << "\n");
2259 // Remove the pair and flipped pair from the list.
2260 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second);
2261 assert(FP != ChosenPairs.end() && "Flipped pair not found in list");
2262 ChosenPairs.erase(FP);
2263 ChosenPairs.erase(P);
2265 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSet, I, J)) {
2266 DEBUG(dbgs() << "BBV: fusion of: " << *I <<
2268 " aborted because of non-trivial dependency cycle\n");
2275 unsigned NumOperands = I->getNumOperands();
2276 SmallVector<Value *, 3> ReplacedOperands(NumOperands);
2277 getReplacementInputsForPair(Context, I, J, ReplacedOperands,
2280 // Make a copy of the original operation, change its type to the vector
2281 // type and replace its operands with the vector operands.
2282 Instruction *K = I->clone();
2283 if (I->hasName()) K->takeName(I);
2285 if (!isa<StoreInst>(K))
2286 K->mutateType(getVecTypeForPair(I->getType(), J->getType()));
2288 combineMetadata(K, J);
2290 for (unsigned o = 0; o < NumOperands; ++o)
2291 K->setOperand(o, ReplacedOperands[o]);
2293 // If we've flipped the memory inputs, make sure that we take the correct
2295 if (FlipMemInputs) {
2296 if (isa<StoreInst>(K))
2297 cast<StoreInst>(K)->setAlignment(cast<StoreInst>(J)->getAlignment());
2299 cast<LoadInst>(K)->setAlignment(cast<LoadInst>(J)->getAlignment());
2304 // Instruction insertion point:
2305 Instruction *InsertionPt = K;
2306 Instruction *K1 = 0, *K2 = 0;
2307 replaceOutputsOfPair(Context, I, J, K, InsertionPt, K1, K2,
2310 // The use tree of the first original instruction must be moved to after
2311 // the location of the second instruction. The entire use tree of the
2312 // first instruction is disjoint from the input tree of the second
2313 // (by definition), and so commutes with it.
2315 moveUsesOfIAfterJ(BB, LoadMoveSet, InsertionPt, I, J);
2317 if (!isa<StoreInst>(I)) {
2318 I->replaceAllUsesWith(K1);
2319 J->replaceAllUsesWith(K2);
2320 AA->replaceWithNewValue(I, K1);
2321 AA->replaceWithNewValue(J, K2);
2324 // Instructions that may read from memory may be in the load move set.
2325 // Once an instruction is fused, we no longer need its move set, and so
2326 // the values of the map never need to be updated. However, when a load
2327 // is fused, we need to merge the entries from both instructions in the
2328 // pair in case those instructions were in the move set of some other
2329 // yet-to-be-fused pair. The loads in question are the keys of the map.
2330 if (I->mayReadFromMemory()) {
2331 std::vector<ValuePair> NewSetMembers;
2332 VPIteratorPair IPairRange = LoadMoveSet.equal_range(I);
2333 VPIteratorPair JPairRange = LoadMoveSet.equal_range(J);
2334 for (std::multimap<Value *, Value *>::iterator N = IPairRange.first;
2335 N != IPairRange.second; ++N)
2336 NewSetMembers.push_back(ValuePair(K, N->second));
2337 for (std::multimap<Value *, Value *>::iterator N = JPairRange.first;
2338 N != JPairRange.second; ++N)
2339 NewSetMembers.push_back(ValuePair(K, N->second));
2340 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(),
2341 AE = NewSetMembers.end(); A != AE; ++A)
2342 LoadMoveSet.insert(*A);
2345 // Before removing I, set the iterator to the next instruction.
2346 PI = llvm::next(BasicBlock::iterator(I));
2347 if (cast<Instruction>(PI) == J)
2352 I->eraseFromParent();
2353 J->eraseFromParent();
2356 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n");
2360 char BBVectorize::ID = 0;
2361 static const char bb_vectorize_name[] = "Basic-Block Vectorization";
2362 INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
2363 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
2364 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
2365 INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
2367 BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) {
2368 return new BBVectorize(C);
2372 llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) {
2373 BBVectorize BBVectorizer(P, C);
2374 return BBVectorizer.vectorizeBB(BB);
2377 //===----------------------------------------------------------------------===//
2378 VectorizeConfig::VectorizeConfig() {
2379 VectorBits = ::VectorBits;
2380 VectorizeBools = !::NoBools;
2381 VectorizeInts = !::NoInts;
2382 VectorizeFloats = !::NoFloats;
2383 VectorizePointers = !::NoPointers;
2384 VectorizeCasts = !::NoCasts;
2385 VectorizeMath = !::NoMath;
2386 VectorizeFMA = !::NoFMA;
2387 VectorizeSelect = !::NoSelect;
2388 VectorizeCmp = !::NoCmp;
2389 VectorizeGEP = !::NoGEP;
2390 VectorizeMemOps = !::NoMemOps;
2391 AlignedOnly = ::AlignedOnly;
2392 ReqChainDepth= ::ReqChainDepth;
2393 SearchLimit = ::SearchLimit;
2394 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck;
2395 SplatBreaksChain = ::SplatBreaksChain;
2396 MaxInsts = ::MaxInsts;
2397 MaxIter = ::MaxIter;
2398 Pow2LenOnly = ::Pow2LenOnly;
2399 NoMemOpBoost = ::NoMemOpBoost;
2400 FastDep = ::FastDep;