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)
752 // The powi intrinsic is special because only the first argument is
753 // vectorized, the second arguments must be equal.
754 CallInst *CI = dyn_cast<CallInst>(I);
756 if (CI && (FI = CI->getCalledFunction()) &&
757 FI->getIntrinsicID() == Intrinsic::powi) {
759 Value *A1I = CI->getArgOperand(1),
760 *A1J = cast<CallInst>(J)->getArgOperand(1);
761 const SCEV *A1ISCEV = SE->getSCEV(A1I),
762 *A1JSCEV = SE->getSCEV(A1J);
763 return (A1ISCEV == A1JSCEV);
769 // Figure out whether or not J uses I and update the users and write-set
770 // structures associated with I. Specifically, Users represents the set of
771 // instructions that depend on I. WriteSet represents the set
772 // of memory locations that are dependent on I. If UpdateUsers is true,
773 // and J uses I, then Users is updated to contain J and WriteSet is updated
774 // to contain any memory locations to which J writes. The function returns
775 // true if J uses I. By default, alias analysis is used to determine
776 // whether J reads from memory that overlaps with a location in WriteSet.
777 // If LoadMoveSet is not null, then it is a previously-computed multimap
778 // where the key is the memory-based user instruction and the value is
779 // the instruction to be compared with I. So, if LoadMoveSet is provided,
780 // then the alias analysis is not used. This is necessary because this
781 // function is called during the process of moving instructions during
782 // vectorization and the results of the alias analysis are not stable during
784 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users,
785 AliasSetTracker &WriteSet, Instruction *I,
786 Instruction *J, bool UpdateUsers,
787 std::multimap<Value *, Value *> *LoadMoveSet) {
790 // This instruction may already be marked as a user due, for example, to
791 // being a member of a selected pair.
796 for (User::op_iterator JU = J->op_begin(), JE = J->op_end();
799 if (I == V || Users.count(V)) {
804 if (!UsesI && J->mayReadFromMemory()) {
806 VPIteratorPair JPairRange = LoadMoveSet->equal_range(J);
807 UsesI = isSecondInIteratorPair<Value*>(I, JPairRange);
809 for (AliasSetTracker::iterator W = WriteSet.begin(),
810 WE = WriteSet.end(); W != WE; ++W) {
811 if (W->aliasesUnknownInst(J, *AA)) {
819 if (UsesI && UpdateUsers) {
820 if (J->mayWriteToMemory()) WriteSet.add(J);
827 // This function iterates over all instruction pairs in the provided
828 // basic block and collects all candidate pairs for vectorization.
829 bool BBVectorize::getCandidatePairs(BasicBlock &BB,
830 BasicBlock::iterator &Start,
831 std::multimap<Value *, Value *> &CandidatePairs,
832 std::vector<Value *> &PairableInsts, bool NonPow2Len) {
833 BasicBlock::iterator E = BB.end();
834 if (Start == E) return false;
836 bool ShouldContinue = false, IAfterStart = false;
837 for (BasicBlock::iterator I = Start++; I != E; ++I) {
838 if (I == Start) IAfterStart = true;
840 bool IsSimpleLoadStore;
841 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue;
843 // Look for an instruction with which to pair instruction *I...
844 DenseSet<Value *> Users;
845 AliasSetTracker WriteSet(*AA);
846 bool JAfterStart = IAfterStart;
847 BasicBlock::iterator J = llvm::next(I);
848 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) {
849 if (J == Start) JAfterStart = true;
851 // Determine if J uses I, if so, exit the loop.
852 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep);
853 if (Config.FastDep) {
854 // Note: For this heuristic to be effective, independent operations
855 // must tend to be intermixed. This is likely to be true from some
856 // kinds of grouped loop unrolling (but not the generic LLVM pass),
857 // but otherwise may require some kind of reordering pass.
859 // When using fast dependency analysis,
860 // stop searching after first use:
866 // J does not use I, and comes before the first use of I, so it can be
867 // merged with I if the instructions are compatible.
868 if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len)) continue;
870 // J is a candidate for merging with I.
871 if (!PairableInsts.size() ||
872 PairableInsts[PairableInsts.size()-1] != I) {
873 PairableInsts.push_back(I);
876 CandidatePairs.insert(ValuePair(I, J));
878 // The next call to this function must start after the last instruction
879 // selected during this invocation.
881 Start = llvm::next(J);
882 IAfterStart = JAfterStart = false;
885 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair "
886 << *I << " <-> " << *J << "\n");
888 // If we have already found too many pairs, break here and this function
889 // will be called again starting after the last instruction selected
890 // during this invocation.
891 if (PairableInsts.size() >= Config.MaxInsts) {
892 ShouldContinue = true;
901 DEBUG(dbgs() << "BBV: found " << PairableInsts.size()
902 << " instructions with candidate pairs\n");
904 return ShouldContinue;
907 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that
908 // it looks for pairs such that both members have an input which is an
909 // output of PI or PJ.
910 void BBVectorize::computePairsConnectedTo(
911 std::multimap<Value *, Value *> &CandidatePairs,
912 std::vector<Value *> &PairableInsts,
913 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
917 // For each possible pairing for this variable, look at the uses of
918 // the first value...
919 for (Value::use_iterator I = P.first->use_begin(),
920 E = P.first->use_end(); I != E; ++I) {
921 if (isa<LoadInst>(*I)) {
922 // A pair cannot be connected to a load because the load only takes one
923 // operand (the address) and it is a scalar even after vectorization.
925 } else if ((SI = dyn_cast<StoreInst>(*I)) &&
926 P.first == SI->getPointerOperand()) {
927 // Similarly, a pair cannot be connected to a store through its
932 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I);
934 // For each use of the first variable, look for uses of the second
936 for (Value::use_iterator J = P.second->use_begin(),
937 E2 = P.second->use_end(); J != E2; ++J) {
938 if ((SJ = dyn_cast<StoreInst>(*J)) &&
939 P.second == SJ->getPointerOperand())
942 VPIteratorPair JPairRange = CandidatePairs.equal_range(*J);
945 if (isSecondInIteratorPair<Value*>(*J, IPairRange))
946 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J)));
949 if (isSecondInIteratorPair<Value*>(*I, JPairRange))
950 ConnectedPairs.insert(VPPair(P, ValuePair(*J, *I)));
953 if (Config.SplatBreaksChain) continue;
954 // Look for cases where just the first value in the pair is used by
955 // both members of another pair (splatting).
956 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) {
957 if ((SJ = dyn_cast<StoreInst>(*J)) &&
958 P.first == SJ->getPointerOperand())
961 if (isSecondInIteratorPair<Value*>(*J, IPairRange))
962 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J)));
966 if (Config.SplatBreaksChain) return;
967 // Look for cases where just the second value in the pair is used by
968 // both members of another pair (splatting).
969 for (Value::use_iterator I = P.second->use_begin(),
970 E = P.second->use_end(); I != E; ++I) {
971 if (isa<LoadInst>(*I))
973 else if ((SI = dyn_cast<StoreInst>(*I)) &&
974 P.second == SI->getPointerOperand())
977 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I);
979 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) {
980 if ((SJ = dyn_cast<StoreInst>(*J)) &&
981 P.second == SJ->getPointerOperand())
984 if (isSecondInIteratorPair<Value*>(*J, IPairRange))
985 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J)));
990 // This function figures out which pairs are connected. Two pairs are
991 // connected if some output of the first pair forms an input to both members
992 // of the second pair.
993 void BBVectorize::computeConnectedPairs(
994 std::multimap<Value *, Value *> &CandidatePairs,
995 std::vector<Value *> &PairableInsts,
996 std::multimap<ValuePair, ValuePair> &ConnectedPairs) {
998 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
999 PE = PairableInsts.end(); PI != PE; ++PI) {
1000 VPIteratorPair choiceRange = CandidatePairs.equal_range(*PI);
1002 for (std::multimap<Value *, Value *>::iterator P = choiceRange.first;
1003 P != choiceRange.second; ++P)
1004 computePairsConnectedTo(CandidatePairs, PairableInsts,
1005 ConnectedPairs, *P);
1008 DEBUG(dbgs() << "BBV: found " << ConnectedPairs.size()
1009 << " pair connections.\n");
1012 // This function builds a set of use tuples such that <A, B> is in the set
1013 // if B is in the use tree of A. If B is in the use tree of A, then B
1014 // depends on the output of A.
1015 void BBVectorize::buildDepMap(
1017 std::multimap<Value *, Value *> &CandidatePairs,
1018 std::vector<Value *> &PairableInsts,
1019 DenseSet<ValuePair> &PairableInstUsers) {
1020 DenseSet<Value *> IsInPair;
1021 for (std::multimap<Value *, Value *>::iterator C = CandidatePairs.begin(),
1022 E = CandidatePairs.end(); C != E; ++C) {
1023 IsInPair.insert(C->first);
1024 IsInPair.insert(C->second);
1027 // Iterate through the basic block, recording all Users of each
1028 // pairable instruction.
1030 BasicBlock::iterator E = BB.end();
1031 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) {
1032 if (IsInPair.find(I) == IsInPair.end()) continue;
1034 DenseSet<Value *> Users;
1035 AliasSetTracker WriteSet(*AA);
1036 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J)
1037 (void) trackUsesOfI(Users, WriteSet, I, J);
1039 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end();
1041 PairableInstUsers.insert(ValuePair(I, *U));
1045 // Returns true if an input to pair P is an output of pair Q and also an
1046 // input of pair Q is an output of pair P. If this is the case, then these
1047 // two pairs cannot be simultaneously fused.
1048 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q,
1049 DenseSet<ValuePair> &PairableInstUsers,
1050 std::multimap<ValuePair, ValuePair> *PairableInstUserMap) {
1051 // Two pairs are in conflict if they are mutual Users of eachother.
1052 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) ||
1053 PairableInstUsers.count(ValuePair(P.first, Q.second)) ||
1054 PairableInstUsers.count(ValuePair(P.second, Q.first)) ||
1055 PairableInstUsers.count(ValuePair(P.second, Q.second));
1056 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) ||
1057 PairableInstUsers.count(ValuePair(Q.first, P.second)) ||
1058 PairableInstUsers.count(ValuePair(Q.second, P.first)) ||
1059 PairableInstUsers.count(ValuePair(Q.second, P.second));
1060 if (PairableInstUserMap) {
1061 // FIXME: The expensive part of the cycle check is not so much the cycle
1062 // check itself but this edge insertion procedure. This needs some
1063 // profiling and probably a different data structure (same is true of
1064 // most uses of std::multimap).
1066 VPPIteratorPair QPairRange = PairableInstUserMap->equal_range(Q);
1067 if (!isSecondInIteratorPair(P, QPairRange))
1068 PairableInstUserMap->insert(VPPair(Q, P));
1071 VPPIteratorPair PPairRange = PairableInstUserMap->equal_range(P);
1072 if (!isSecondInIteratorPair(Q, PPairRange))
1073 PairableInstUserMap->insert(VPPair(P, Q));
1077 return (QUsesP && PUsesQ);
1080 // This function walks the use graph of current pairs to see if, starting
1081 // from P, the walk returns to P.
1082 bool BBVectorize::pairWillFormCycle(ValuePair P,
1083 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1084 DenseSet<ValuePair> &CurrentPairs) {
1085 DEBUG(if (DebugCycleCheck)
1086 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> "
1087 << *P.second << "\n");
1088 // A lookup table of visisted pairs is kept because the PairableInstUserMap
1089 // contains non-direct associations.
1090 DenseSet<ValuePair> Visited;
1091 SmallVector<ValuePair, 32> Q;
1092 // General depth-first post-order traversal:
1095 ValuePair QTop = Q.pop_back_val();
1096 Visited.insert(QTop);
1098 DEBUG(if (DebugCycleCheck)
1099 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> "
1100 << *QTop.second << "\n");
1101 VPPIteratorPair QPairRange = PairableInstUserMap.equal_range(QTop);
1102 for (std::multimap<ValuePair, ValuePair>::iterator C = QPairRange.first;
1103 C != QPairRange.second; ++C) {
1104 if (C->second == P) {
1106 << "BBV: rejected to prevent non-trivial cycle formation: "
1107 << *C->first.first << " <-> " << *C->first.second << "\n");
1111 if (CurrentPairs.count(C->second) && !Visited.count(C->second))
1112 Q.push_back(C->second);
1114 } while (!Q.empty());
1119 // This function builds the initial tree of connected pairs with the
1120 // pair J at the root.
1121 void BBVectorize::buildInitialTreeFor(
1122 std::multimap<Value *, Value *> &CandidatePairs,
1123 std::vector<Value *> &PairableInsts,
1124 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1125 DenseSet<ValuePair> &PairableInstUsers,
1126 DenseMap<Value *, Value *> &ChosenPairs,
1127 DenseMap<ValuePair, size_t> &Tree, ValuePair J) {
1128 // Each of these pairs is viewed as the root node of a Tree. The Tree
1129 // is then walked (depth-first). As this happens, we keep track of
1130 // the pairs that compose the Tree and the maximum depth of the Tree.
1131 SmallVector<ValuePairWithDepth, 32> Q;
1132 // General depth-first post-order traversal:
1133 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1135 ValuePairWithDepth QTop = Q.back();
1137 // Push each child onto the queue:
1138 bool MoreChildren = false;
1139 size_t MaxChildDepth = QTop.second;
1140 VPPIteratorPair qtRange = ConnectedPairs.equal_range(QTop.first);
1141 for (std::multimap<ValuePair, ValuePair>::iterator k = qtRange.first;
1142 k != qtRange.second; ++k) {
1143 // Make sure that this child pair is still a candidate:
1144 bool IsStillCand = false;
1145 VPIteratorPair checkRange =
1146 CandidatePairs.equal_range(k->second.first);
1147 for (std::multimap<Value *, Value *>::iterator m = checkRange.first;
1148 m != checkRange.second; ++m) {
1149 if (m->second == k->second.second) {
1156 DenseMap<ValuePair, size_t>::iterator C = Tree.find(k->second);
1157 if (C == Tree.end()) {
1158 size_t d = getDepthFactor(k->second.first);
1159 Q.push_back(ValuePairWithDepth(k->second, QTop.second+d));
1160 MoreChildren = true;
1162 MaxChildDepth = std::max(MaxChildDepth, C->second);
1167 if (!MoreChildren) {
1168 // Record the current pair as part of the Tree:
1169 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth));
1172 } while (!Q.empty());
1175 // Given some initial tree, prune it by removing conflicting pairs (pairs
1176 // that cannot be simultaneously chosen for vectorization).
1177 void BBVectorize::pruneTreeFor(
1178 std::multimap<Value *, Value *> &CandidatePairs,
1179 std::vector<Value *> &PairableInsts,
1180 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1181 DenseSet<ValuePair> &PairableInstUsers,
1182 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1183 DenseMap<Value *, Value *> &ChosenPairs,
1184 DenseMap<ValuePair, size_t> &Tree,
1185 DenseSet<ValuePair> &PrunedTree, ValuePair J,
1186 bool UseCycleCheck) {
1187 SmallVector<ValuePairWithDepth, 32> Q;
1188 // General depth-first post-order traversal:
1189 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1191 ValuePairWithDepth QTop = Q.pop_back_val();
1192 PrunedTree.insert(QTop.first);
1194 // Visit each child, pruning as necessary...
1195 DenseMap<ValuePair, size_t> BestChildren;
1196 VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first);
1197 for (std::multimap<ValuePair, ValuePair>::iterator K = QTopRange.first;
1198 K != QTopRange.second; ++K) {
1199 DenseMap<ValuePair, size_t>::iterator C = Tree.find(K->second);
1200 if (C == Tree.end()) continue;
1202 // This child is in the Tree, now we need to make sure it is the
1203 // best of any conflicting children. There could be multiple
1204 // conflicting children, so first, determine if we're keeping
1205 // this child, then delete conflicting children as necessary.
1207 // It is also necessary to guard against pairing-induced
1208 // dependencies. Consider instructions a .. x .. y .. b
1209 // such that (a,b) are to be fused and (x,y) are to be fused
1210 // but a is an input to x and b is an output from y. This
1211 // means that y cannot be moved after b but x must be moved
1212 // after b for (a,b) to be fused. In other words, after
1213 // fusing (a,b) we have y .. a/b .. x where y is an input
1214 // to a/b and x is an output to a/b: x and y can no longer
1215 // be legally fused. To prevent this condition, we must
1216 // make sure that a child pair added to the Tree is not
1217 // both an input and output of an already-selected pair.
1219 // Pairing-induced dependencies can also form from more complicated
1220 // cycles. The pair vs. pair conflicts are easy to check, and so
1221 // that is done explicitly for "fast rejection", and because for
1222 // child vs. child conflicts, we may prefer to keep the current
1223 // pair in preference to the already-selected child.
1224 DenseSet<ValuePair> CurrentPairs;
1227 for (DenseMap<ValuePair, size_t>::iterator C2
1228 = BestChildren.begin(), E2 = BestChildren.end();
1230 if (C2->first.first == C->first.first ||
1231 C2->first.first == C->first.second ||
1232 C2->first.second == C->first.first ||
1233 C2->first.second == C->first.second ||
1234 pairsConflict(C2->first, C->first, PairableInstUsers,
1235 UseCycleCheck ? &PairableInstUserMap : 0)) {
1236 if (C2->second >= C->second) {
1241 CurrentPairs.insert(C2->first);
1244 if (!CanAdd) continue;
1246 // Even worse, this child could conflict with another node already
1247 // selected for the Tree. If that is the case, ignore this child.
1248 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(),
1249 E2 = PrunedTree.end(); T != E2; ++T) {
1250 if (T->first == C->first.first ||
1251 T->first == C->first.second ||
1252 T->second == C->first.first ||
1253 T->second == C->first.second ||
1254 pairsConflict(*T, C->first, PairableInstUsers,
1255 UseCycleCheck ? &PairableInstUserMap : 0)) {
1260 CurrentPairs.insert(*T);
1262 if (!CanAdd) continue;
1264 // And check the queue too...
1265 for (SmallVector<ValuePairWithDepth, 32>::iterator C2 = Q.begin(),
1266 E2 = Q.end(); C2 != E2; ++C2) {
1267 if (C2->first.first == C->first.first ||
1268 C2->first.first == C->first.second ||
1269 C2->first.second == C->first.first ||
1270 C2->first.second == C->first.second ||
1271 pairsConflict(C2->first, C->first, PairableInstUsers,
1272 UseCycleCheck ? &PairableInstUserMap : 0)) {
1277 CurrentPairs.insert(C2->first);
1279 if (!CanAdd) continue;
1281 // Last but not least, check for a conflict with any of the
1282 // already-chosen pairs.
1283 for (DenseMap<Value *, Value *>::iterator C2 =
1284 ChosenPairs.begin(), E2 = ChosenPairs.end();
1286 if (pairsConflict(*C2, C->first, PairableInstUsers,
1287 UseCycleCheck ? &PairableInstUserMap : 0)) {
1292 CurrentPairs.insert(*C2);
1294 if (!CanAdd) continue;
1296 // To check for non-trivial cycles formed by the addition of the
1297 // current pair we've formed a list of all relevant pairs, now use a
1298 // graph walk to check for a cycle. We start from the current pair and
1299 // walk the use tree to see if we again reach the current pair. If we
1300 // do, then the current pair is rejected.
1302 // FIXME: It may be more efficient to use a topological-ordering
1303 // algorithm to improve the cycle check. This should be investigated.
1304 if (UseCycleCheck &&
1305 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs))
1308 // This child can be added, but we may have chosen it in preference
1309 // to an already-selected child. Check for this here, and if a
1310 // conflict is found, then remove the previously-selected child
1311 // before adding this one in its place.
1312 for (DenseMap<ValuePair, size_t>::iterator C2
1313 = BestChildren.begin(); C2 != BestChildren.end();) {
1314 if (C2->first.first == C->first.first ||
1315 C2->first.first == C->first.second ||
1316 C2->first.second == C->first.first ||
1317 C2->first.second == C->first.second ||
1318 pairsConflict(C2->first, C->first, PairableInstUsers))
1319 BestChildren.erase(C2++);
1324 BestChildren.insert(ValuePairWithDepth(C->first, C->second));
1327 for (DenseMap<ValuePair, size_t>::iterator C
1328 = BestChildren.begin(), E2 = BestChildren.end();
1330 size_t DepthF = getDepthFactor(C->first.first);
1331 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF));
1333 } while (!Q.empty());
1336 // This function finds the best tree of mututally-compatible connected
1337 // pairs, given the choice of root pairs as an iterator range.
1338 void BBVectorize::findBestTreeFor(
1339 std::multimap<Value *, Value *> &CandidatePairs,
1340 std::vector<Value *> &PairableInsts,
1341 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1342 DenseSet<ValuePair> &PairableInstUsers,
1343 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1344 DenseMap<Value *, Value *> &ChosenPairs,
1345 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth,
1346 size_t &BestEffSize, VPIteratorPair ChoiceRange,
1347 bool UseCycleCheck) {
1348 for (std::multimap<Value *, Value *>::iterator J = ChoiceRange.first;
1349 J != ChoiceRange.second; ++J) {
1351 // Before going any further, make sure that this pair does not
1352 // conflict with any already-selected pairs (see comment below
1353 // near the Tree pruning for more details).
1354 DenseSet<ValuePair> ChosenPairSet;
1355 bool DoesConflict = false;
1356 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(),
1357 E = ChosenPairs.end(); C != E; ++C) {
1358 if (pairsConflict(*C, *J, PairableInstUsers,
1359 UseCycleCheck ? &PairableInstUserMap : 0)) {
1360 DoesConflict = true;
1364 ChosenPairSet.insert(*C);
1366 if (DoesConflict) continue;
1368 if (UseCycleCheck &&
1369 pairWillFormCycle(*J, PairableInstUserMap, ChosenPairSet))
1372 DenseMap<ValuePair, size_t> Tree;
1373 buildInitialTreeFor(CandidatePairs, PairableInsts, ConnectedPairs,
1374 PairableInstUsers, ChosenPairs, Tree, *J);
1376 // Because we'll keep the child with the largest depth, the largest
1377 // depth is still the same in the unpruned Tree.
1378 size_t MaxDepth = Tree.lookup(*J);
1380 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found Tree for pair {"
1381 << *J->first << " <-> " << *J->second << "} of depth " <<
1382 MaxDepth << " and size " << Tree.size() << "\n");
1384 // At this point the Tree has been constructed, but, may contain
1385 // contradictory children (meaning that different children of
1386 // some tree node may be attempting to fuse the same instruction).
1387 // So now we walk the tree again, in the case of a conflict,
1388 // keep only the child with the largest depth. To break a tie,
1389 // favor the first child.
1391 DenseSet<ValuePair> PrunedTree;
1392 pruneTreeFor(CandidatePairs, PairableInsts, ConnectedPairs,
1393 PairableInstUsers, PairableInstUserMap, ChosenPairs, Tree,
1394 PrunedTree, *J, UseCycleCheck);
1397 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
1398 E = PrunedTree.end(); S != E; ++S)
1399 EffSize += getDepthFactor(S->first);
1401 DEBUG(if (DebugPairSelection)
1402 dbgs() << "BBV: found pruned Tree for pair {"
1403 << *J->first << " <-> " << *J->second << "} of depth " <<
1404 MaxDepth << " and size " << PrunedTree.size() <<
1405 " (effective size: " << EffSize << ")\n");
1406 if (MaxDepth >= Config.ReqChainDepth && EffSize > BestEffSize) {
1407 BestMaxDepth = MaxDepth;
1408 BestEffSize = EffSize;
1409 BestTree = PrunedTree;
1414 // Given the list of candidate pairs, this function selects those
1415 // that will be fused into vector instructions.
1416 void BBVectorize::choosePairs(
1417 std::multimap<Value *, Value *> &CandidatePairs,
1418 std::vector<Value *> &PairableInsts,
1419 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1420 DenseSet<ValuePair> &PairableInstUsers,
1421 DenseMap<Value *, Value *>& ChosenPairs) {
1422 bool UseCycleCheck =
1423 CandidatePairs.size() <= Config.MaxCandPairsForCycleCheck;
1424 std::multimap<ValuePair, ValuePair> PairableInstUserMap;
1425 for (std::vector<Value *>::iterator I = PairableInsts.begin(),
1426 E = PairableInsts.end(); I != E; ++I) {
1427 // The number of possible pairings for this variable:
1428 size_t NumChoices = CandidatePairs.count(*I);
1429 if (!NumChoices) continue;
1431 VPIteratorPair ChoiceRange = CandidatePairs.equal_range(*I);
1433 // The best pair to choose and its tree:
1434 size_t BestMaxDepth = 0, BestEffSize = 0;
1435 DenseSet<ValuePair> BestTree;
1436 findBestTreeFor(CandidatePairs, PairableInsts, ConnectedPairs,
1437 PairableInstUsers, PairableInstUserMap, ChosenPairs,
1438 BestTree, BestMaxDepth, BestEffSize, ChoiceRange,
1441 // A tree has been chosen (or not) at this point. If no tree was
1442 // chosen, then this instruction, I, cannot be paired (and is no longer
1445 DEBUG(if (BestTree.size() > 0)
1446 dbgs() << "BBV: selected pairs in the best tree for: "
1447 << *cast<Instruction>(*I) << "\n");
1449 for (DenseSet<ValuePair>::iterator S = BestTree.begin(),
1450 SE2 = BestTree.end(); S != SE2; ++S) {
1451 // Insert the members of this tree into the list of chosen pairs.
1452 ChosenPairs.insert(ValuePair(S->first, S->second));
1453 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " <<
1454 *S->second << "\n");
1456 // Remove all candidate pairs that have values in the chosen tree.
1457 for (std::multimap<Value *, Value *>::iterator K =
1458 CandidatePairs.begin(); K != CandidatePairs.end();) {
1459 if (K->first == S->first || K->second == S->first ||
1460 K->second == S->second || K->first == S->second) {
1461 // Don't remove the actual pair chosen so that it can be used
1462 // in subsequent tree selections.
1463 if (!(K->first == S->first && K->second == S->second))
1464 CandidatePairs.erase(K++);
1474 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n");
1477 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o,
1482 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) +
1483 (n > 0 ? "." + utostr(n) : "")).str();
1486 // Returns the value that is to be used as the pointer input to the vector
1487 // instruction that fuses I with J.
1488 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context,
1489 Instruction *I, Instruction *J, unsigned o,
1490 bool &FlipMemInputs) {
1492 unsigned IAlignment, JAlignment;
1493 int64_t OffsetInElmts;
1494 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
1497 // The pointer value is taken to be the one with the lowest offset.
1499 if (OffsetInElmts > 0) {
1502 FlipMemInputs = true;
1506 Type *ArgTypeI = cast<PointerType>(IPtr->getType())->getElementType();
1507 Type *ArgTypeJ = cast<PointerType>(JPtr->getType())->getElementType();
1508 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
1509 Type *VArgPtrType = PointerType::get(VArgType,
1510 cast<PointerType>(IPtr->getType())->getAddressSpace());
1511 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o),
1512 /* insert before */ FlipMemInputs ? J : I);
1515 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J,
1516 unsigned MaskOffset, unsigned NumInElem,
1517 unsigned NumInElem1, unsigned IdxOffset,
1518 std::vector<Constant*> &Mask) {
1519 unsigned NumElem1 = cast<VectorType>(J->getType())->getNumElements();
1520 for (unsigned v = 0; v < NumElem1; ++v) {
1521 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v);
1523 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context));
1525 unsigned mm = m + (int) IdxOffset;
1526 if (m >= (int) NumInElem1)
1527 mm += (int) NumInElem;
1529 Mask[v+MaskOffset] =
1530 ConstantInt::get(Type::getInt32Ty(Context), mm);
1535 // Returns the value that is to be used as the vector-shuffle mask to the
1536 // vector instruction that fuses I with J.
1537 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context,
1538 Instruction *I, Instruction *J) {
1539 // This is the shuffle mask. We need to append the second
1540 // mask to the first, and the numbers need to be adjusted.
1542 Type *ArgTypeI = I->getType();
1543 Type *ArgTypeJ = J->getType();
1544 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
1546 unsigned NumElemI = cast<VectorType>(ArgTypeI)->getNumElements();
1548 // Get the total number of elements in the fused vector type.
1549 // By definition, this must equal the number of elements in
1551 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements();
1552 std::vector<Constant*> Mask(NumElem);
1554 Type *OpTypeI = I->getOperand(0)->getType();
1555 unsigned NumInElemI = cast<VectorType>(OpTypeI)->getNumElements();
1556 Type *OpTypeJ = J->getOperand(0)->getType();
1557 unsigned NumInElemJ = cast<VectorType>(OpTypeJ)->getNumElements();
1559 // The fused vector will be:
1560 // -----------------------------------------------------
1561 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ |
1562 // -----------------------------------------------------
1563 // from which we'll extract NumElem total elements (where the first NumElemI
1564 // of them come from the mask in I and the remainder come from the mask
1567 // For the mask from the first pair...
1568 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI,
1571 // For the mask from the second pair...
1572 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ,
1575 return ConstantVector::get(Mask);
1578 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I,
1579 Instruction *J, unsigned o, Value *&LOp,
1581 Type *ArgTypeL, Type *ArgTypeH,
1583 bool ExpandedIEChain = false;
1584 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) {
1585 // If we have a pure insertelement chain, then this can be rewritten
1586 // into a chain that directly builds the larger type.
1587 bool PureChain = true;
1588 InsertElementInst *LIENext = LIE;
1590 if (!isa<UndefValue>(LIENext->getOperand(0)) &&
1591 !isa<InsertElementInst>(LIENext->getOperand(0))) {
1596 dyn_cast<InsertElementInst>(LIENext->getOperand(0))));
1599 SmallVector<Value *, 8> VectElemts(numElemL,
1600 UndefValue::get(ArgTypeL->getScalarType()));
1601 InsertElementInst *LIENext = LIE;
1604 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue();
1605 VectElemts[Idx] = LIENext->getOperand(1);
1607 dyn_cast<InsertElementInst>(LIENext->getOperand(0))));
1610 Value *LIEPrev = UndefValue::get(ArgTypeH);
1611 for (unsigned i = 0; i < numElemL; ++i) {
1612 if (isa<UndefValue>(VectElemts[i])) continue;
1613 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i],
1614 ConstantInt::get(Type::getInt32Ty(Context),
1616 getReplacementName(I, true, o, i+1));
1617 LIENext->insertBefore(J);
1621 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH);
1622 ExpandedIEChain = true;
1626 return ExpandedIEChain;
1629 // Returns the value to be used as the specified operand of the vector
1630 // instruction that fuses I with J.
1631 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I,
1632 Instruction *J, unsigned o, bool FlipMemInputs) {
1633 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
1634 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1);
1636 // Compute the fused vector type for this operand
1637 Type *ArgTypeI = I->getOperand(o)->getType();
1638 Type *ArgTypeJ = J->getOperand(o)->getType();
1639 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
1641 Instruction *L = I, *H = J;
1642 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ;
1643 if (FlipMemInputs) {
1646 ArgTypeL = ArgTypeJ;
1647 ArgTypeH = ArgTypeI;
1651 if (ArgTypeL->isVectorTy())
1652 numElemL = cast<VectorType>(ArgTypeL)->getNumElements();
1657 if (ArgTypeH->isVectorTy())
1658 numElemH = cast<VectorType>(ArgTypeH)->getNumElements();
1662 Value *LOp = L->getOperand(o);
1663 Value *HOp = H->getOperand(o);
1664 unsigned numElem = VArgType->getNumElements();
1666 // First, we check if we can reuse the "original" vector outputs (if these
1667 // exist). We might need a shuffle.
1668 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp);
1669 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp);
1670 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp);
1671 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp);
1673 // FIXME: If we're fusing shuffle instructions, then we can't apply this
1674 // optimization. The input vectors to the shuffle might be a different
1675 // length from the shuffle outputs. Unfortunately, the replacement
1676 // shuffle mask has already been formed, and the mask entries are sensitive
1677 // to the sizes of the inputs.
1678 bool IsSizeChangeShuffle =
1679 isa<ShuffleVectorInst>(L) &&
1680 (LOp->getType() != L->getType() || HOp->getType() != H->getType());
1682 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) {
1683 // We can have at most two unique vector inputs.
1684 bool CanUseInputs = true;
1687 I1 = LEE->getOperand(0);
1689 I1 = LSV->getOperand(0);
1690 I2 = LSV->getOperand(1);
1691 if (I2 == I1 || isa<UndefValue>(I2))
1696 Value *I3 = HEE->getOperand(0);
1697 if (!I2 && I3 != I1)
1699 else if (I3 != I1 && I3 != I2)
1700 CanUseInputs = false;
1702 Value *I3 = HSV->getOperand(0);
1703 if (!I2 && I3 != I1)
1705 else if (I3 != I1 && I3 != I2)
1706 CanUseInputs = false;
1709 Value *I4 = HSV->getOperand(1);
1710 if (!isa<UndefValue>(I4)) {
1711 if (!I2 && I4 != I1)
1713 else if (I4 != I1 && I4 != I2)
1714 CanUseInputs = false;
1721 cast<VectorType>(cast<Instruction>(LOp)->getOperand(0)->getType())
1724 cast<VectorType>(cast<Instruction>(HOp)->getOperand(0)->getType())
1727 // We have one or two input vectors. We need to map each index of the
1728 // operands to the index of the original vector.
1729 SmallVector<std::pair<int, int>, 8> II(numElem);
1730 for (unsigned i = 0; i < numElemL; ++i) {
1734 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue();
1735 INum = LEE->getOperand(0) == I1 ? 0 : 1;
1737 Idx = LSV->getMaskValue(i);
1738 if (Idx < (int) LOpElem) {
1739 INum = LSV->getOperand(0) == I1 ? 0 : 1;
1742 INum = LSV->getOperand(1) == I1 ? 0 : 1;
1746 II[i] = std::pair<int, int>(Idx, INum);
1748 for (unsigned i = 0; i < numElemH; ++i) {
1752 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue();
1753 INum = HEE->getOperand(0) == I1 ? 0 : 1;
1755 Idx = HSV->getMaskValue(i);
1756 if (Idx < (int) HOpElem) {
1757 INum = HSV->getOperand(0) == I1 ? 0 : 1;
1760 INum = HSV->getOperand(1) == I1 ? 0 : 1;
1764 II[i + numElemL] = std::pair<int, int>(Idx, INum);
1767 // We now have an array which tells us from which index of which
1768 // input vector each element of the operand comes.
1769 VectorType *I1T = cast<VectorType>(I1->getType());
1770 unsigned I1Elem = I1T->getNumElements();
1773 // In this case there is only one underlying vector input. Check for
1774 // the trivial case where we can use the input directly.
1775 if (I1Elem == numElem) {
1776 bool ElemInOrder = true;
1777 for (unsigned i = 0; i < numElem; ++i) {
1778 if (II[i].first != (int) i && II[i].first != -1) {
1779 ElemInOrder = false;
1788 // A shuffle is needed.
1789 std::vector<Constant *> Mask(numElem);
1790 for (unsigned i = 0; i < numElem; ++i) {
1791 int Idx = II[i].first;
1793 Mask[i] = UndefValue::get(Type::getInt32Ty(Context));
1795 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
1799 new ShuffleVectorInst(I1, UndefValue::get(I1T),
1800 ConstantVector::get(Mask),
1801 getReplacementName(I, true, o));
1806 VectorType *I2T = cast<VectorType>(I2->getType());
1807 unsigned I2Elem = I2T->getNumElements();
1809 // This input comes from two distinct vectors. The first step is to
1810 // make sure that both vectors are the same length. If not, the
1811 // smaller one will need to grow before they can be shuffled together.
1812 if (I1Elem < I2Elem) {
1813 std::vector<Constant *> Mask(I2Elem);
1815 for (; v < I1Elem; ++v)
1816 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
1817 for (; v < I2Elem; ++v)
1818 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
1820 Instruction *NewI1 =
1821 new ShuffleVectorInst(I1, UndefValue::get(I1T),
1822 ConstantVector::get(Mask),
1823 getReplacementName(I, true, o, 1));
1824 NewI1->insertBefore(J);
1828 } else if (I1Elem > I2Elem) {
1829 std::vector<Constant *> Mask(I1Elem);
1831 for (; v < I2Elem; ++v)
1832 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
1833 for (; v < I1Elem; ++v)
1834 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
1836 Instruction *NewI2 =
1837 new ShuffleVectorInst(I2, UndefValue::get(I2T),
1838 ConstantVector::get(Mask),
1839 getReplacementName(I, true, o, 1));
1840 NewI2->insertBefore(J);
1846 // Now that both I1 and I2 are the same length we can shuffle them
1847 // together (and use the result).
1848 std::vector<Constant *> Mask(numElem);
1849 for (unsigned v = 0; v < numElem; ++v) {
1850 if (II[v].first == -1) {
1851 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
1853 int Idx = II[v].first + II[v].second * I1Elem;
1854 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
1858 Instruction *NewOp =
1859 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask),
1860 getReplacementName(I, true, o));
1861 NewOp->insertBefore(J);
1866 Type *ArgType = ArgTypeL;
1867 if (numElemL < numElemH) {
1868 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH,
1869 ArgTypeL, VArgType, 1)) {
1870 // This is another short-circuit case: we're combining a scalar into
1871 // a vector that is formed by an IE chain. We've just expanded the IE
1872 // chain, now insert the scalar and we're done.
1874 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0,
1875 getReplacementName(I, true, o));
1878 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL,
1880 // The two vector inputs to the shuffle must be the same length,
1881 // so extend the smaller vector to be the same length as the larger one.
1885 std::vector<Constant *> Mask(numElemH);
1887 for (; v < numElemL; ++v)
1888 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
1889 for (; v < numElemH; ++v)
1890 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
1892 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL),
1893 ConstantVector::get(Mask),
1894 getReplacementName(I, true, o, 1));
1896 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0,
1897 getReplacementName(I, true, o, 1));
1900 NLOp->insertBefore(J);
1905 } else if (numElemL > numElemH) {
1906 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL,
1907 ArgTypeH, VArgType)) {
1909 InsertElementInst::Create(LOp, HOp,
1910 ConstantInt::get(Type::getInt32Ty(Context),
1912 getReplacementName(I, true, o));
1915 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH,
1919 std::vector<Constant *> Mask(numElemL);
1921 for (; v < numElemH; ++v)
1922 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
1923 for (; v < numElemL; ++v)
1924 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
1926 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH),
1927 ConstantVector::get(Mask),
1928 getReplacementName(I, true, o, 1));
1930 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0,
1931 getReplacementName(I, true, o, 1));
1934 NHOp->insertBefore(J);
1939 if (ArgType->isVectorTy()) {
1940 unsigned numElem = cast<VectorType>(VArgType)->getNumElements();
1941 std::vector<Constant*> Mask(numElem);
1942 for (unsigned v = 0; v < numElem; ++v) {
1944 // If the low vector was expanded, we need to skip the extra
1945 // undefined entries.
1946 if (v >= numElemL && numElemH > numElemL)
1947 Idx += (numElemH - numElemL);
1948 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
1951 Instruction *BV = new ShuffleVectorInst(LOp, HOp,
1952 ConstantVector::get(Mask),
1953 getReplacementName(I, true, o));
1954 BV->insertBefore(J);
1958 Instruction *BV1 = InsertElementInst::Create(
1959 UndefValue::get(VArgType), LOp, CV0,
1960 getReplacementName(I, true, o, 1));
1961 BV1->insertBefore(I);
1962 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1,
1963 getReplacementName(I, true, o, 2));
1964 BV2->insertBefore(J);
1968 // This function creates an array of values that will be used as the inputs
1969 // to the vector instruction that fuses I with J.
1970 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context,
1971 Instruction *I, Instruction *J,
1972 SmallVector<Value *, 3> &ReplacedOperands,
1973 bool &FlipMemInputs) {
1974 FlipMemInputs = false;
1975 unsigned NumOperands = I->getNumOperands();
1977 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) {
1978 // Iterate backward so that we look at the store pointer
1979 // first and know whether or not we need to flip the inputs.
1981 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) {
1982 // This is the pointer for a load/store instruction.
1983 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o,
1986 } else if (isa<CallInst>(I)) {
1987 Function *F = cast<CallInst>(I)->getCalledFunction();
1988 unsigned IID = F->getIntrinsicID();
1989 if (o == NumOperands-1) {
1990 BasicBlock &BB = *I->getParent();
1992 Module *M = BB.getParent()->getParent();
1993 Type *ArgTypeI = I->getType();
1994 Type *ArgTypeJ = J->getType();
1995 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
1997 ReplacedOperands[o] = Intrinsic::getDeclaration(M,
1998 (Intrinsic::ID) IID, VArgType);
2000 } else if (IID == Intrinsic::powi && o == 1) {
2001 // The second argument of powi is a single integer and we've already
2002 // checked that both arguments are equal. As a result, we just keep
2003 // I's second argument.
2004 ReplacedOperands[o] = I->getOperand(o);
2007 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) {
2008 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J);
2012 ReplacedOperands[o] =
2013 getReplacementInput(Context, I, J, o, FlipMemInputs);
2017 // This function creates two values that represent the outputs of the
2018 // original I and J instructions. These are generally vector shuffles
2019 // or extracts. In many cases, these will end up being unused and, thus,
2020 // eliminated by later passes.
2021 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
2022 Instruction *J, Instruction *K,
2023 Instruction *&InsertionPt,
2024 Instruction *&K1, Instruction *&K2,
2025 bool &FlipMemInputs) {
2026 if (isa<StoreInst>(I)) {
2027 AA->replaceWithNewValue(I, K);
2028 AA->replaceWithNewValue(J, K);
2030 Type *IType = I->getType();
2031 Type *JType = J->getType();
2033 VectorType *VType = getVecTypeForPair(IType, JType);
2034 unsigned numElem = VType->getNumElements();
2036 unsigned numElemI, numElemJ;
2037 if (IType->isVectorTy())
2038 numElemI = cast<VectorType>(IType)->getNumElements();
2042 if (JType->isVectorTy())
2043 numElemJ = cast<VectorType>(JType)->getNumElements();
2047 if (IType->isVectorTy()) {
2048 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI);
2049 for (unsigned v = 0; v < numElemI; ++v) {
2050 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2051 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v);
2054 K1 = new ShuffleVectorInst(K, UndefValue::get(VType),
2055 ConstantVector::get(
2056 FlipMemInputs ? Mask2 : Mask1),
2057 getReplacementName(K, false, 1));
2059 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2060 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1);
2061 K1 = ExtractElementInst::Create(K, FlipMemInputs ? CV1 : CV0,
2062 getReplacementName(K, false, 1));
2065 if (JType->isVectorTy()) {
2066 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ);
2067 for (unsigned v = 0; v < numElemJ; ++v) {
2068 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2069 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v);
2072 K2 = new ShuffleVectorInst(K, UndefValue::get(VType),
2073 ConstantVector::get(
2074 FlipMemInputs ? Mask1 : Mask2),
2075 getReplacementName(K, false, 2));
2077 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2078 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1);
2079 K2 = ExtractElementInst::Create(K, FlipMemInputs ? CV0 : CV1,
2080 getReplacementName(K, false, 2));
2084 K2->insertAfter(K1);
2089 // Move all uses of the function I (including pairing-induced uses) after J.
2090 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB,
2091 std::multimap<Value *, Value *> &LoadMoveSet,
2092 Instruction *I, Instruction *J) {
2093 // Skip to the first instruction past I.
2094 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2096 DenseSet<Value *> Users;
2097 AliasSetTracker WriteSet(*AA);
2098 for (; cast<Instruction>(L) != J; ++L)
2099 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet);
2101 assert(cast<Instruction>(L) == J &&
2102 "Tracking has not proceeded far enough to check for dependencies");
2103 // If J is now in the use set of I, then trackUsesOfI will return true
2104 // and we have a dependency cycle (and the fusing operation must abort).
2105 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSet);
2108 // Move all uses of the function I (including pairing-induced uses) after J.
2109 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB,
2110 std::multimap<Value *, Value *> &LoadMoveSet,
2111 Instruction *&InsertionPt,
2112 Instruction *I, Instruction *J) {
2113 // Skip to the first instruction past I.
2114 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2116 DenseSet<Value *> Users;
2117 AliasSetTracker WriteSet(*AA);
2118 for (; cast<Instruction>(L) != J;) {
2119 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet)) {
2120 // Move this instruction
2121 Instruction *InstToMove = L; ++L;
2123 DEBUG(dbgs() << "BBV: moving: " << *InstToMove <<
2124 " to after " << *InsertionPt << "\n");
2125 InstToMove->removeFromParent();
2126 InstToMove->insertAfter(InsertionPt);
2127 InsertionPt = InstToMove;
2134 // Collect all load instruction that are in the move set of a given first
2135 // pair member. These loads depend on the first instruction, I, and so need
2136 // to be moved after J (the second instruction) when the pair is fused.
2137 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB,
2138 DenseMap<Value *, Value *> &ChosenPairs,
2139 std::multimap<Value *, Value *> &LoadMoveSet,
2141 // Skip to the first instruction past I.
2142 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2144 DenseSet<Value *> Users;
2145 AliasSetTracker WriteSet(*AA);
2147 // Note: We cannot end the loop when we reach J because J could be moved
2148 // farther down the use chain by another instruction pairing. Also, J
2149 // could be before I if this is an inverted input.
2150 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) {
2151 if (trackUsesOfI(Users, WriteSet, I, L)) {
2152 if (L->mayReadFromMemory())
2153 LoadMoveSet.insert(ValuePair(L, I));
2158 // In cases where both load/stores and the computation of their pointers
2159 // are chosen for vectorization, we can end up in a situation where the
2160 // aliasing analysis starts returning different query results as the
2161 // process of fusing instruction pairs continues. Because the algorithm
2162 // relies on finding the same use trees here as were found earlier, we'll
2163 // need to precompute the necessary aliasing information here and then
2164 // manually update it during the fusion process.
2165 void BBVectorize::collectLoadMoveSet(BasicBlock &BB,
2166 std::vector<Value *> &PairableInsts,
2167 DenseMap<Value *, Value *> &ChosenPairs,
2168 std::multimap<Value *, Value *> &LoadMoveSet) {
2169 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
2170 PIE = PairableInsts.end(); PI != PIE; ++PI) {
2171 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI);
2172 if (P == ChosenPairs.end()) continue;
2174 Instruction *I = cast<Instruction>(P->first);
2175 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, I);
2179 // When the first instruction in each pair is cloned, it will inherit its
2180 // parent's metadata. This metadata must be combined with that of the other
2181 // instruction in a safe way.
2182 void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) {
2183 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata;
2184 K->getAllMetadataOtherThanDebugLoc(Metadata);
2185 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) {
2186 unsigned Kind = Metadata[i].first;
2187 MDNode *JMD = J->getMetadata(Kind);
2188 MDNode *KMD = Metadata[i].second;
2192 K->setMetadata(Kind, 0); // Remove unknown metadata
2194 case LLVMContext::MD_tbaa:
2195 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD));
2197 case LLVMContext::MD_fpmath:
2198 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD));
2204 // This function fuses the chosen instruction pairs into vector instructions,
2205 // taking care preserve any needed scalar outputs and, then, it reorders the
2206 // remaining instructions as needed (users of the first member of the pair
2207 // need to be moved to after the location of the second member of the pair
2208 // because the vector instruction is inserted in the location of the pair's
2210 void BBVectorize::fuseChosenPairs(BasicBlock &BB,
2211 std::vector<Value *> &PairableInsts,
2212 DenseMap<Value *, Value *> &ChosenPairs) {
2213 LLVMContext& Context = BB.getContext();
2215 // During the vectorization process, the order of the pairs to be fused
2216 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs
2217 // list. After a pair is fused, the flipped pair is removed from the list.
2218 std::vector<ValuePair> FlippedPairs;
2219 FlippedPairs.reserve(ChosenPairs.size());
2220 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(),
2221 E = ChosenPairs.end(); P != E; ++P)
2222 FlippedPairs.push_back(ValuePair(P->second, P->first));
2223 for (std::vector<ValuePair>::iterator P = FlippedPairs.begin(),
2224 E = FlippedPairs.end(); P != E; ++P)
2225 ChosenPairs.insert(*P);
2227 std::multimap<Value *, Value *> LoadMoveSet;
2228 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, LoadMoveSet);
2230 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n");
2232 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) {
2233 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI);
2234 if (P == ChosenPairs.end()) {
2239 if (getDepthFactor(P->first) == 0) {
2240 // These instructions are not really fused, but are tracked as though
2241 // they are. Any case in which it would be interesting to fuse them
2242 // will be taken care of by InstCombine.
2248 Instruction *I = cast<Instruction>(P->first),
2249 *J = cast<Instruction>(P->second);
2251 DEBUG(dbgs() << "BBV: fusing: " << *I <<
2252 " <-> " << *J << "\n");
2254 // Remove the pair and flipped pair from the list.
2255 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second);
2256 assert(FP != ChosenPairs.end() && "Flipped pair not found in list");
2257 ChosenPairs.erase(FP);
2258 ChosenPairs.erase(P);
2260 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSet, I, J)) {
2261 DEBUG(dbgs() << "BBV: fusion of: " << *I <<
2263 " aborted because of non-trivial dependency cycle\n");
2270 unsigned NumOperands = I->getNumOperands();
2271 SmallVector<Value *, 3> ReplacedOperands(NumOperands);
2272 getReplacementInputsForPair(Context, I, J, ReplacedOperands,
2275 // Make a copy of the original operation, change its type to the vector
2276 // type and replace its operands with the vector operands.
2277 Instruction *K = I->clone();
2278 if (I->hasName()) K->takeName(I);
2280 if (!isa<StoreInst>(K))
2281 K->mutateType(getVecTypeForPair(I->getType(), J->getType()));
2283 combineMetadata(K, J);
2285 for (unsigned o = 0; o < NumOperands; ++o)
2286 K->setOperand(o, ReplacedOperands[o]);
2288 // If we've flipped the memory inputs, make sure that we take the correct
2290 if (FlipMemInputs) {
2291 if (isa<StoreInst>(K))
2292 cast<StoreInst>(K)->setAlignment(cast<StoreInst>(J)->getAlignment());
2294 cast<LoadInst>(K)->setAlignment(cast<LoadInst>(J)->getAlignment());
2299 // Instruction insertion point:
2300 Instruction *InsertionPt = K;
2301 Instruction *K1 = 0, *K2 = 0;
2302 replaceOutputsOfPair(Context, I, J, K, InsertionPt, K1, K2,
2305 // The use tree of the first original instruction must be moved to after
2306 // the location of the second instruction. The entire use tree of the
2307 // first instruction is disjoint from the input tree of the second
2308 // (by definition), and so commutes with it.
2310 moveUsesOfIAfterJ(BB, LoadMoveSet, InsertionPt, I, J);
2312 if (!isa<StoreInst>(I)) {
2313 I->replaceAllUsesWith(K1);
2314 J->replaceAllUsesWith(K2);
2315 AA->replaceWithNewValue(I, K1);
2316 AA->replaceWithNewValue(J, K2);
2319 // Instructions that may read from memory may be in the load move set.
2320 // Once an instruction is fused, we no longer need its move set, and so
2321 // the values of the map never need to be updated. However, when a load
2322 // is fused, we need to merge the entries from both instructions in the
2323 // pair in case those instructions were in the move set of some other
2324 // yet-to-be-fused pair. The loads in question are the keys of the map.
2325 if (I->mayReadFromMemory()) {
2326 std::vector<ValuePair> NewSetMembers;
2327 VPIteratorPair IPairRange = LoadMoveSet.equal_range(I);
2328 VPIteratorPair JPairRange = LoadMoveSet.equal_range(J);
2329 for (std::multimap<Value *, Value *>::iterator N = IPairRange.first;
2330 N != IPairRange.second; ++N)
2331 NewSetMembers.push_back(ValuePair(K, N->second));
2332 for (std::multimap<Value *, Value *>::iterator N = JPairRange.first;
2333 N != JPairRange.second; ++N)
2334 NewSetMembers.push_back(ValuePair(K, N->second));
2335 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(),
2336 AE = NewSetMembers.end(); A != AE; ++A)
2337 LoadMoveSet.insert(*A);
2340 // Before removing I, set the iterator to the next instruction.
2341 PI = llvm::next(BasicBlock::iterator(I));
2342 if (cast<Instruction>(PI) == J)
2347 I->eraseFromParent();
2348 J->eraseFromParent();
2351 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n");
2355 char BBVectorize::ID = 0;
2356 static const char bb_vectorize_name[] = "Basic-Block Vectorization";
2357 INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
2358 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
2359 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
2360 INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
2362 BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) {
2363 return new BBVectorize(C);
2367 llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) {
2368 BBVectorize BBVectorizer(P, C);
2369 return BBVectorizer.vectorizeBB(BB);
2372 //===----------------------------------------------------------------------===//
2373 VectorizeConfig::VectorizeConfig() {
2374 VectorBits = ::VectorBits;
2375 VectorizeBools = !::NoBools;
2376 VectorizeInts = !::NoInts;
2377 VectorizeFloats = !::NoFloats;
2378 VectorizePointers = !::NoPointers;
2379 VectorizeCasts = !::NoCasts;
2380 VectorizeMath = !::NoMath;
2381 VectorizeFMA = !::NoFMA;
2382 VectorizeSelect = !::NoSelect;
2383 VectorizeCmp = !::NoCmp;
2384 VectorizeGEP = !::NoGEP;
2385 VectorizeMemOps = !::NoMemOps;
2386 AlignedOnly = ::AlignedOnly;
2387 ReqChainDepth= ::ReqChainDepth;
2388 SearchLimit = ::SearchLimit;
2389 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck;
2390 SplatBreaksChain = ::SplatBreaksChain;
2391 MaxInsts = ::MaxInsts;
2392 MaxIter = ::MaxIter;
2393 Pow2LenOnly = ::Pow2LenOnly;
2394 NoMemOpBoost = ::NoMemOpBoost;
2395 FastDep = ::FastDep;