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/Vectorize.h"
50 static cl::opt<unsigned>
51 ReqChainDepth("bb-vectorize-req-chain-depth", cl::init(6), cl::Hidden,
52 cl::desc("The required chain depth for vectorization"));
54 static cl::opt<unsigned>
55 SearchLimit("bb-vectorize-search-limit", cl::init(400), cl::Hidden,
56 cl::desc("The maximum search distance for instruction pairs"));
59 SplatBreaksChain("bb-vectorize-splat-breaks-chain", cl::init(false), cl::Hidden,
60 cl::desc("Replicating one element to a pair breaks the chain"));
62 static cl::opt<unsigned>
63 VectorBits("bb-vectorize-vector-bits", cl::init(128), cl::Hidden,
64 cl::desc("The size of the native vector registers"));
66 static cl::opt<unsigned>
67 MaxIter("bb-vectorize-max-iter", cl::init(0), cl::Hidden,
68 cl::desc("The maximum number of pairing iterations"));
70 static cl::opt<unsigned>
71 MaxInsts("bb-vectorize-max-instr-per-group", cl::init(500), cl::Hidden,
72 cl::desc("The maximum number of pairable instructions per group"));
74 static cl::opt<unsigned>
75 MaxCandPairsForCycleCheck("bb-vectorize-max-cycle-check-pairs", cl::init(200),
76 cl::Hidden, cl::desc("The maximum number of candidate pairs with which to use"
77 " a full cycle check"));
80 NoBools("bb-vectorize-no-bools", cl::init(false), cl::Hidden,
81 cl::desc("Don't try to vectorize boolean (i1) values"));
84 NoInts("bb-vectorize-no-ints", cl::init(false), cl::Hidden,
85 cl::desc("Don't try to vectorize integer values"));
88 NoFloats("bb-vectorize-no-floats", cl::init(false), cl::Hidden,
89 cl::desc("Don't try to vectorize floating-point values"));
92 NoPointers("bb-vectorize-no-pointers", cl::init(false), cl::Hidden,
93 cl::desc("Don't try to vectorize pointer values"));
96 NoCasts("bb-vectorize-no-casts", cl::init(false), cl::Hidden,
97 cl::desc("Don't try to vectorize casting (conversion) operations"));
100 NoMath("bb-vectorize-no-math", cl::init(false), cl::Hidden,
101 cl::desc("Don't try to vectorize floating-point math intrinsics"));
104 NoFMA("bb-vectorize-no-fma", cl::init(false), cl::Hidden,
105 cl::desc("Don't try to vectorize the fused-multiply-add intrinsic"));
108 NoSelect("bb-vectorize-no-select", cl::init(false), cl::Hidden,
109 cl::desc("Don't try to vectorize select instructions"));
112 NoCmp("bb-vectorize-no-cmp", cl::init(false), cl::Hidden,
113 cl::desc("Don't try to vectorize comparison instructions"));
116 NoGEP("bb-vectorize-no-gep", cl::init(false), cl::Hidden,
117 cl::desc("Don't try to vectorize getelementptr instructions"));
120 NoMemOps("bb-vectorize-no-mem-ops", cl::init(false), cl::Hidden,
121 cl::desc("Don't try to vectorize loads and stores"));
124 AlignedOnly("bb-vectorize-aligned-only", cl::init(false), cl::Hidden,
125 cl::desc("Only generate aligned loads and stores"));
128 NoMemOpBoost("bb-vectorize-no-mem-op-boost",
129 cl::init(false), cl::Hidden,
130 cl::desc("Don't boost the chain-depth contribution of loads and stores"));
133 FastDep("bb-vectorize-fast-dep", cl::init(false), cl::Hidden,
134 cl::desc("Use a fast instruction dependency analysis"));
138 DebugInstructionExamination("bb-vectorize-debug-instruction-examination",
139 cl::init(false), cl::Hidden,
140 cl::desc("When debugging is enabled, output information on the"
141 " instruction-examination process"));
143 DebugCandidateSelection("bb-vectorize-debug-candidate-selection",
144 cl::init(false), cl::Hidden,
145 cl::desc("When debugging is enabled, output information on the"
146 " candidate-selection process"));
148 DebugPairSelection("bb-vectorize-debug-pair-selection",
149 cl::init(false), cl::Hidden,
150 cl::desc("When debugging is enabled, output information on the"
151 " pair-selection process"));
153 DebugCycleCheck("bb-vectorize-debug-cycle-check",
154 cl::init(false), cl::Hidden,
155 cl::desc("When debugging is enabled, output information on the"
156 " cycle-checking process"));
159 STATISTIC(NumFusedOps, "Number of operations fused by bb-vectorize");
162 struct BBVectorize : public BasicBlockPass {
163 static char ID; // Pass identification, replacement for typeid
165 const VectorizeConfig Config;
167 BBVectorize(const VectorizeConfig &C = VectorizeConfig())
168 : BasicBlockPass(ID), Config(C) {
169 initializeBBVectorizePass(*PassRegistry::getPassRegistry());
172 BBVectorize(Pass *P, const VectorizeConfig &C)
173 : BasicBlockPass(ID), Config(C) {
174 AA = &P->getAnalysis<AliasAnalysis>();
175 SE = &P->getAnalysis<ScalarEvolution>();
176 TD = P->getAnalysisIfAvailable<TargetData>();
179 typedef std::pair<Value *, Value *> ValuePair;
180 typedef std::pair<ValuePair, size_t> ValuePairWithDepth;
181 typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair
182 typedef std::pair<std::multimap<Value *, Value *>::iterator,
183 std::multimap<Value *, Value *>::iterator> VPIteratorPair;
184 typedef std::pair<std::multimap<ValuePair, ValuePair>::iterator,
185 std::multimap<ValuePair, ValuePair>::iterator>
192 // FIXME: const correct?
194 bool vectorizePairs(BasicBlock &BB);
196 bool getCandidatePairs(BasicBlock &BB,
197 BasicBlock::iterator &Start,
198 std::multimap<Value *, Value *> &CandidatePairs,
199 std::vector<Value *> &PairableInsts);
201 void computeConnectedPairs(std::multimap<Value *, Value *> &CandidatePairs,
202 std::vector<Value *> &PairableInsts,
203 std::multimap<ValuePair, ValuePair> &ConnectedPairs);
205 void buildDepMap(BasicBlock &BB,
206 std::multimap<Value *, Value *> &CandidatePairs,
207 std::vector<Value *> &PairableInsts,
208 DenseSet<ValuePair> &PairableInstUsers);
210 void choosePairs(std::multimap<Value *, Value *> &CandidatePairs,
211 std::vector<Value *> &PairableInsts,
212 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
213 DenseSet<ValuePair> &PairableInstUsers,
214 DenseMap<Value *, Value *>& ChosenPairs);
216 void fuseChosenPairs(BasicBlock &BB,
217 std::vector<Value *> &PairableInsts,
218 DenseMap<Value *, Value *>& ChosenPairs);
220 bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore);
222 bool areInstsCompatible(Instruction *I, Instruction *J,
223 bool IsSimpleLoadStore);
225 bool trackUsesOfI(DenseSet<Value *> &Users,
226 AliasSetTracker &WriteSet, Instruction *I,
227 Instruction *J, bool UpdateUsers = true,
228 std::multimap<Value *, Value *> *LoadMoveSet = 0);
230 void computePairsConnectedTo(
231 std::multimap<Value *, Value *> &CandidatePairs,
232 std::vector<Value *> &PairableInsts,
233 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
236 bool pairsConflict(ValuePair P, ValuePair Q,
237 DenseSet<ValuePair> &PairableInstUsers,
238 std::multimap<ValuePair, ValuePair> *PairableInstUserMap = 0);
240 bool pairWillFormCycle(ValuePair P,
241 std::multimap<ValuePair, ValuePair> &PairableInstUsers,
242 DenseSet<ValuePair> &CurrentPairs);
245 std::multimap<Value *, Value *> &CandidatePairs,
246 std::vector<Value *> &PairableInsts,
247 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
248 DenseSet<ValuePair> &PairableInstUsers,
249 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
250 DenseMap<Value *, Value *> &ChosenPairs,
251 DenseMap<ValuePair, size_t> &Tree,
252 DenseSet<ValuePair> &PrunedTree, ValuePair J,
255 void buildInitialTreeFor(
256 std::multimap<Value *, Value *> &CandidatePairs,
257 std::vector<Value *> &PairableInsts,
258 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
259 DenseSet<ValuePair> &PairableInstUsers,
260 DenseMap<Value *, Value *> &ChosenPairs,
261 DenseMap<ValuePair, size_t> &Tree, ValuePair J);
263 void findBestTreeFor(
264 std::multimap<Value *, Value *> &CandidatePairs,
265 std::vector<Value *> &PairableInsts,
266 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
267 DenseSet<ValuePair> &PairableInstUsers,
268 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
269 DenseMap<Value *, Value *> &ChosenPairs,
270 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth,
271 size_t &BestEffSize, VPIteratorPair ChoiceRange,
274 Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I,
275 Instruction *J, unsigned o, bool &FlipMemInputs);
277 void fillNewShuffleMask(LLVMContext& Context, Instruction *J,
278 unsigned NumElem, unsigned MaskOffset, unsigned NumInElem,
279 unsigned IdxOffset, std::vector<Constant*> &Mask);
281 Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I,
284 Value *getReplacementInput(LLVMContext& Context, Instruction *I,
285 Instruction *J, unsigned o, bool FlipMemInputs);
287 void getReplacementInputsForPair(LLVMContext& Context, Instruction *I,
288 Instruction *J, SmallVector<Value *, 3> &ReplacedOperands,
289 bool &FlipMemInputs);
291 void replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
292 Instruction *J, Instruction *K,
293 Instruction *&InsertionPt, Instruction *&K1,
294 Instruction *&K2, bool &FlipMemInputs);
296 void collectPairLoadMoveSet(BasicBlock &BB,
297 DenseMap<Value *, Value *> &ChosenPairs,
298 std::multimap<Value *, Value *> &LoadMoveSet,
301 void collectLoadMoveSet(BasicBlock &BB,
302 std::vector<Value *> &PairableInsts,
303 DenseMap<Value *, Value *> &ChosenPairs,
304 std::multimap<Value *, Value *> &LoadMoveSet);
306 bool canMoveUsesOfIAfterJ(BasicBlock &BB,
307 std::multimap<Value *, Value *> &LoadMoveSet,
308 Instruction *I, Instruction *J);
310 void moveUsesOfIAfterJ(BasicBlock &BB,
311 std::multimap<Value *, Value *> &LoadMoveSet,
312 Instruction *&InsertionPt,
313 Instruction *I, Instruction *J);
315 void combineMetadata(Instruction *K, const Instruction *J);
317 bool vectorizeBB(BasicBlock &BB) {
318 bool changed = false;
319 // Iterate a sufficient number of times to merge types of size 1 bit,
320 // then 2 bits, then 4, etc. up to half of the target vector width of the
321 // target vector register.
322 for (unsigned v = 2, n = 1;
323 v <= Config.VectorBits && (!Config.MaxIter || n <= Config.MaxIter);
325 DEBUG(dbgs() << "BBV: fusing loop #" << n <<
326 " for " << BB.getName() << " in " <<
327 BB.getParent()->getName() << "...\n");
328 if (vectorizePairs(BB))
334 DEBUG(dbgs() << "BBV: done!\n");
338 virtual bool runOnBasicBlock(BasicBlock &BB) {
339 AA = &getAnalysis<AliasAnalysis>();
340 SE = &getAnalysis<ScalarEvolution>();
341 TD = getAnalysisIfAvailable<TargetData>();
343 return vectorizeBB(BB);
346 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
347 BasicBlockPass::getAnalysisUsage(AU);
348 AU.addRequired<AliasAnalysis>();
349 AU.addRequired<ScalarEvolution>();
350 AU.addPreserved<AliasAnalysis>();
351 AU.addPreserved<ScalarEvolution>();
352 AU.setPreservesCFG();
355 // This returns the vector type that holds a pair of the provided type.
356 // If the provided type is already a vector, then its length is doubled.
357 static inline VectorType *getVecTypeForPair(Type *ElemTy) {
358 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) {
359 unsigned numElem = VTy->getNumElements();
360 return VectorType::get(ElemTy->getScalarType(), numElem*2);
363 return VectorType::get(ElemTy, 2);
366 // Returns the weight associated with the provided value. A chain of
367 // candidate pairs has a length given by the sum of the weights of its
368 // members (one weight per pair; the weight of each member of the pair
369 // is assumed to be the same). This length is then compared to the
370 // chain-length threshold to determine if a given chain is significant
371 // enough to be vectorized. The length is also used in comparing
372 // candidate chains where longer chains are considered to be better.
373 // Note: when this function returns 0, the resulting instructions are
374 // not actually fused.
375 inline size_t getDepthFactor(Value *V) {
376 // InsertElement and ExtractElement have a depth factor of zero. This is
377 // for two reasons: First, they cannot be usefully fused. Second, because
378 // the pass generates a lot of these, they can confuse the simple metric
379 // used to compare the trees in the next iteration. Thus, giving them a
380 // weight of zero allows the pass to essentially ignore them in
381 // subsequent iterations when looking for vectorization opportunities
382 // while still tracking dependency chains that flow through those
384 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V))
387 // Give a load or store half of the required depth so that load/store
388 // pairs will vectorize.
389 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V)))
390 return Config.ReqChainDepth/2;
395 // This determines the relative offset of two loads or stores, returning
396 // true if the offset could be determined to be some constant value.
397 // For example, if OffsetInElmts == 1, then J accesses the memory directly
398 // after I; if OffsetInElmts == -1 then I accesses the memory
399 // directly after J. This function assumes that both instructions
400 // have the same type.
401 bool getPairPtrInfo(Instruction *I, Instruction *J,
402 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment,
403 int64_t &OffsetInElmts) {
405 if (isa<LoadInst>(I)) {
406 IPtr = cast<LoadInst>(I)->getPointerOperand();
407 JPtr = cast<LoadInst>(J)->getPointerOperand();
408 IAlignment = cast<LoadInst>(I)->getAlignment();
409 JAlignment = cast<LoadInst>(J)->getAlignment();
411 IPtr = cast<StoreInst>(I)->getPointerOperand();
412 JPtr = cast<StoreInst>(J)->getPointerOperand();
413 IAlignment = cast<StoreInst>(I)->getAlignment();
414 JAlignment = cast<StoreInst>(J)->getAlignment();
417 const SCEV *IPtrSCEV = SE->getSCEV(IPtr);
418 const SCEV *JPtrSCEV = SE->getSCEV(JPtr);
420 // If this is a trivial offset, then we'll get something like
421 // 1*sizeof(type). With target data, which we need anyway, this will get
422 // constant folded into a number.
423 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV);
424 if (const SCEVConstant *ConstOffSCEV =
425 dyn_cast<SCEVConstant>(OffsetSCEV)) {
426 ConstantInt *IntOff = ConstOffSCEV->getValue();
427 int64_t Offset = IntOff->getSExtValue();
429 Type *VTy = cast<PointerType>(IPtr->getType())->getElementType();
430 int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy);
432 assert(VTy == cast<PointerType>(JPtr->getType())->getElementType());
434 OffsetInElmts = Offset/VTyTSS;
435 return (abs64(Offset) % VTyTSS) == 0;
441 // Returns true if the provided CallInst represents an intrinsic that can
443 bool isVectorizableIntrinsic(CallInst* I) {
444 Function *F = I->getCalledFunction();
445 if (!F) return false;
447 unsigned IID = F->getIntrinsicID();
448 if (!IID) return false;
453 case Intrinsic::sqrt:
454 case Intrinsic::powi:
458 case Intrinsic::log2:
459 case Intrinsic::log10:
461 case Intrinsic::exp2:
463 return Config.VectorizeMath;
465 return Config.VectorizeFMA;
469 // Returns true if J is the second element in some pair referenced by
470 // some multimap pair iterator pair.
471 template <typename V>
472 bool isSecondInIteratorPair(V J, std::pair<
473 typename std::multimap<V, V>::iterator,
474 typename std::multimap<V, V>::iterator> PairRange) {
475 for (typename std::multimap<V, V>::iterator K = PairRange.first;
476 K != PairRange.second; ++K)
477 if (K->second == J) return true;
483 // This function implements one vectorization iteration on the provided
484 // basic block. It returns true if the block is changed.
485 bool BBVectorize::vectorizePairs(BasicBlock &BB) {
487 BasicBlock::iterator Start = BB.getFirstInsertionPt();
489 std::vector<Value *> AllPairableInsts;
490 DenseMap<Value *, Value *> AllChosenPairs;
493 std::vector<Value *> PairableInsts;
494 std::multimap<Value *, Value *> CandidatePairs;
495 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs,
497 if (PairableInsts.empty()) continue;
499 // Now we have a map of all of the pairable instructions and we need to
500 // select the best possible pairing. A good pairing is one such that the
501 // users of the pair are also paired. This defines a (directed) forest
502 // over the pairs such that two pairs are connected iff the second pair
505 // Note that it only matters that both members of the second pair use some
506 // element of the first pair (to allow for splatting).
508 std::multimap<ValuePair, ValuePair> ConnectedPairs;
509 computeConnectedPairs(CandidatePairs, PairableInsts, ConnectedPairs);
510 if (ConnectedPairs.empty()) continue;
512 // Build the pairable-instruction dependency map
513 DenseSet<ValuePair> PairableInstUsers;
514 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers);
516 // There is now a graph of the connected pairs. For each variable, pick
517 // the pairing with the largest tree meeting the depth requirement on at
518 // least one branch. Then select all pairings that are part of that tree
519 // and remove them from the list of available pairings and pairable
522 DenseMap<Value *, Value *> ChosenPairs;
523 choosePairs(CandidatePairs, PairableInsts, ConnectedPairs,
524 PairableInstUsers, ChosenPairs);
526 if (ChosenPairs.empty()) continue;
527 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(),
528 PairableInsts.end());
529 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end());
530 } while (ShouldContinue);
532 if (AllChosenPairs.empty()) return false;
533 NumFusedOps += AllChosenPairs.size();
535 // A set of pairs has now been selected. It is now necessary to replace the
536 // paired instructions with vector instructions. For this procedure each
537 // operand must be replaced with a vector operand. This vector is formed
538 // by using build_vector on the old operands. The replaced values are then
539 // replaced with a vector_extract on the result. Subsequent optimization
540 // passes should coalesce the build/extract combinations.
542 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs);
546 // This function returns true if the provided instruction is capable of being
547 // fused into a vector instruction. This determination is based only on the
548 // type and other attributes of the instruction.
549 bool BBVectorize::isInstVectorizable(Instruction *I,
550 bool &IsSimpleLoadStore) {
551 IsSimpleLoadStore = false;
553 if (CallInst *C = dyn_cast<CallInst>(I)) {
554 if (!isVectorizableIntrinsic(C))
556 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) {
557 // Vectorize simple loads if possbile:
558 IsSimpleLoadStore = L->isSimple();
559 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
561 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) {
562 // Vectorize simple stores if possbile:
563 IsSimpleLoadStore = S->isSimple();
564 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
566 } else if (CastInst *C = dyn_cast<CastInst>(I)) {
567 // We can vectorize casts, but not casts of pointer types, etc.
568 if (!Config.VectorizeCasts)
571 Type *SrcTy = C->getSrcTy();
572 if (!SrcTy->isSingleValueType())
575 Type *DestTy = C->getDestTy();
576 if (!DestTy->isSingleValueType())
578 } else if (isa<SelectInst>(I)) {
579 if (!Config.VectorizeSelect)
581 } else if (isa<CmpInst>(I)) {
582 if (!Config.VectorizeCmp)
584 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) {
585 if (!Config.VectorizeGEP)
588 // Currently, vector GEPs exist only with one index.
589 if (G->getNumIndices() != 1)
591 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) ||
592 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) {
596 // We can't vectorize memory operations without target data
597 if (TD == 0 && IsSimpleLoadStore)
601 if (isa<StoreInst>(I)) {
602 // For stores, it is the value type, not the pointer type that matters
603 // because the value is what will come from a vector register.
605 Value *IVal = cast<StoreInst>(I)->getValueOperand();
606 T1 = IVal->getType();
612 T2 = cast<CastInst>(I)->getSrcTy();
616 // Not every type can be vectorized...
617 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) ||
618 !(VectorType::isValidElementType(T2) || T2->isVectorTy()))
621 if (T1->getScalarSizeInBits() == 1 && T2->getScalarSizeInBits() == 1) {
622 if (!Config.VectorizeBools)
625 if (!Config.VectorizeInts
626 && (T1->isIntOrIntVectorTy() || T2->isIntOrIntVectorTy()))
630 if (!Config.VectorizeFloats
631 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy()))
634 // Don't vectorize target-specific types.
635 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy())
637 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy())
640 if ((!Config.VectorizePointers || TD == 0) &&
641 (T1->getScalarType()->isPointerTy() ||
642 T2->getScalarType()->isPointerTy()))
645 if (T1->getPrimitiveSizeInBits() > Config.VectorBits/2 ||
646 T2->getPrimitiveSizeInBits() > Config.VectorBits/2)
652 // This function returns true if the two provided instructions are compatible
653 // (meaning that they can be fused into a vector instruction). This assumes
654 // that I has already been determined to be vectorizable and that J is not
655 // in the use tree of I.
656 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J,
657 bool IsSimpleLoadStore) {
658 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I <<
659 " <-> " << *J << "\n");
661 // Loads and stores can be merged if they have different alignments,
662 // but are otherwise the same.
665 if ((LI = dyn_cast<LoadInst>(I)) && (LJ = dyn_cast<LoadInst>(J))) {
666 if (I->getType() != J->getType())
669 if (LI->getPointerOperand()->getType() !=
670 LJ->getPointerOperand()->getType() ||
671 LI->isVolatile() != LJ->isVolatile() ||
672 LI->getOrdering() != LJ->getOrdering() ||
673 LI->getSynchScope() != LJ->getSynchScope())
675 } else if ((SI = dyn_cast<StoreInst>(I)) && (SJ = dyn_cast<StoreInst>(J))) {
676 if (SI->getValueOperand()->getType() !=
677 SJ->getValueOperand()->getType() ||
678 SI->getPointerOperand()->getType() !=
679 SJ->getPointerOperand()->getType() ||
680 SI->isVolatile() != SJ->isVolatile() ||
681 SI->getOrdering() != SJ->getOrdering() ||
682 SI->getSynchScope() != SJ->getSynchScope())
684 } else if (!J->isSameOperationAs(I)) {
687 // FIXME: handle addsub-type operations!
689 if (IsSimpleLoadStore) {
691 unsigned IAlignment, JAlignment;
692 int64_t OffsetInElmts = 0;
693 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
694 OffsetInElmts) && abs64(OffsetInElmts) == 1) {
695 if (Config.AlignedOnly) {
696 Type *aType = isa<StoreInst>(I) ?
697 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType();
698 // An aligned load or store is possible only if the instruction
699 // with the lower offset has an alignment suitable for the
702 unsigned BottomAlignment = IAlignment;
703 if (OffsetInElmts < 0) BottomAlignment = JAlignment;
705 Type *VType = getVecTypeForPair(aType);
706 unsigned VecAlignment = TD->getPrefTypeAlignment(VType);
707 if (BottomAlignment < VecAlignment)
713 } else if (isa<ShuffleVectorInst>(I)) {
714 // Only merge two shuffles if they're both constant
715 return isa<Constant>(I->getOperand(2)) &&
716 isa<Constant>(J->getOperand(2));
717 // FIXME: We may want to vectorize non-constant shuffles also.
720 // The powi intrinsic is special because only the first argument is
721 // vectorized, the second arguments must be equal.
722 CallInst *CI = dyn_cast<CallInst>(I);
724 if (CI && (FI = CI->getCalledFunction()) &&
725 FI->getIntrinsicID() == Intrinsic::powi) {
727 Value *A1I = CI->getArgOperand(1),
728 *A1J = cast<CallInst>(J)->getArgOperand(1);
729 const SCEV *A1ISCEV = SE->getSCEV(A1I),
730 *A1JSCEV = SE->getSCEV(A1J);
731 return (A1ISCEV == A1JSCEV);
737 // Figure out whether or not J uses I and update the users and write-set
738 // structures associated with I. Specifically, Users represents the set of
739 // instructions that depend on I. WriteSet represents the set
740 // of memory locations that are dependent on I. If UpdateUsers is true,
741 // and J uses I, then Users is updated to contain J and WriteSet is updated
742 // to contain any memory locations to which J writes. The function returns
743 // true if J uses I. By default, alias analysis is used to determine
744 // whether J reads from memory that overlaps with a location in WriteSet.
745 // If LoadMoveSet is not null, then it is a previously-computed multimap
746 // where the key is the memory-based user instruction and the value is
747 // the instruction to be compared with I. So, if LoadMoveSet is provided,
748 // then the alias analysis is not used. This is necessary because this
749 // function is called during the process of moving instructions during
750 // vectorization and the results of the alias analysis are not stable during
752 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users,
753 AliasSetTracker &WriteSet, Instruction *I,
754 Instruction *J, bool UpdateUsers,
755 std::multimap<Value *, Value *> *LoadMoveSet) {
758 // This instruction may already be marked as a user due, for example, to
759 // being a member of a selected pair.
764 for (User::op_iterator JU = J->op_begin(), JE = J->op_end();
767 if (I == V || Users.count(V)) {
772 if (!UsesI && J->mayReadFromMemory()) {
774 VPIteratorPair JPairRange = LoadMoveSet->equal_range(J);
775 UsesI = isSecondInIteratorPair<Value*>(I, JPairRange);
777 for (AliasSetTracker::iterator W = WriteSet.begin(),
778 WE = WriteSet.end(); W != WE; ++W) {
779 if (W->aliasesUnknownInst(J, *AA)) {
787 if (UsesI && UpdateUsers) {
788 if (J->mayWriteToMemory()) WriteSet.add(J);
795 // This function iterates over all instruction pairs in the provided
796 // basic block and collects all candidate pairs for vectorization.
797 bool BBVectorize::getCandidatePairs(BasicBlock &BB,
798 BasicBlock::iterator &Start,
799 std::multimap<Value *, Value *> &CandidatePairs,
800 std::vector<Value *> &PairableInsts) {
801 BasicBlock::iterator E = BB.end();
802 if (Start == E) return false;
804 bool ShouldContinue = false, IAfterStart = false;
805 for (BasicBlock::iterator I = Start++; I != E; ++I) {
806 if (I == Start) IAfterStart = true;
808 bool IsSimpleLoadStore;
809 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue;
811 // Look for an instruction with which to pair instruction *I...
812 DenseSet<Value *> Users;
813 AliasSetTracker WriteSet(*AA);
814 bool JAfterStart = IAfterStart;
815 BasicBlock::iterator J = llvm::next(I);
816 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) {
817 if (J == Start) JAfterStart = true;
819 // Determine if J uses I, if so, exit the loop.
820 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep);
821 if (Config.FastDep) {
822 // Note: For this heuristic to be effective, independent operations
823 // must tend to be intermixed. This is likely to be true from some
824 // kinds of grouped loop unrolling (but not the generic LLVM pass),
825 // but otherwise may require some kind of reordering pass.
827 // When using fast dependency analysis,
828 // stop searching after first use:
834 // J does not use I, and comes before the first use of I, so it can be
835 // merged with I if the instructions are compatible.
836 if (!areInstsCompatible(I, J, IsSimpleLoadStore)) continue;
838 // J is a candidate for merging with I.
839 if (!PairableInsts.size() ||
840 PairableInsts[PairableInsts.size()-1] != I) {
841 PairableInsts.push_back(I);
844 CandidatePairs.insert(ValuePair(I, J));
846 // The next call to this function must start after the last instruction
847 // selected during this invocation.
849 Start = llvm::next(J);
850 IAfterStart = JAfterStart = false;
853 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair "
854 << *I << " <-> " << *J << "\n");
856 // If we have already found too many pairs, break here and this function
857 // will be called again starting after the last instruction selected
858 // during this invocation.
859 if (PairableInsts.size() >= Config.MaxInsts) {
860 ShouldContinue = true;
869 DEBUG(dbgs() << "BBV: found " << PairableInsts.size()
870 << " instructions with candidate pairs\n");
872 return ShouldContinue;
875 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that
876 // it looks for pairs such that both members have an input which is an
877 // output of PI or PJ.
878 void BBVectorize::computePairsConnectedTo(
879 std::multimap<Value *, Value *> &CandidatePairs,
880 std::vector<Value *> &PairableInsts,
881 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
885 // For each possible pairing for this variable, look at the uses of
886 // the first value...
887 for (Value::use_iterator I = P.first->use_begin(),
888 E = P.first->use_end(); I != E; ++I) {
889 if (isa<LoadInst>(*I)) {
890 // A pair cannot be connected to a load because the load only takes one
891 // operand (the address) and it is a scalar even after vectorization.
893 } else if ((SI = dyn_cast<StoreInst>(*I)) &&
894 P.first == SI->getPointerOperand()) {
895 // Similarly, a pair cannot be connected to a store through its
900 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I);
902 // For each use of the first variable, look for uses of the second
904 for (Value::use_iterator J = P.second->use_begin(),
905 E2 = P.second->use_end(); J != E2; ++J) {
906 if ((SJ = dyn_cast<StoreInst>(*J)) &&
907 P.second == SJ->getPointerOperand())
910 VPIteratorPair JPairRange = CandidatePairs.equal_range(*J);
913 if (isSecondInIteratorPair<Value*>(*J, IPairRange))
914 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J)));
917 if (isSecondInIteratorPair<Value*>(*I, JPairRange))
918 ConnectedPairs.insert(VPPair(P, ValuePair(*J, *I)));
921 if (Config.SplatBreaksChain) continue;
922 // Look for cases where just the first value in the pair is used by
923 // both members of another pair (splatting).
924 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) {
925 if ((SJ = dyn_cast<StoreInst>(*J)) &&
926 P.first == SJ->getPointerOperand())
929 if (isSecondInIteratorPair<Value*>(*J, IPairRange))
930 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J)));
934 if (Config.SplatBreaksChain) return;
935 // Look for cases where just the second value in the pair is used by
936 // both members of another pair (splatting).
937 for (Value::use_iterator I = P.second->use_begin(),
938 E = P.second->use_end(); I != E; ++I) {
939 if (isa<LoadInst>(*I))
941 else if ((SI = dyn_cast<StoreInst>(*I)) &&
942 P.second == SI->getPointerOperand())
945 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I);
947 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) {
948 if ((SJ = dyn_cast<StoreInst>(*J)) &&
949 P.second == SJ->getPointerOperand())
952 if (isSecondInIteratorPair<Value*>(*J, IPairRange))
953 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J)));
958 // This function figures out which pairs are connected. Two pairs are
959 // connected if some output of the first pair forms an input to both members
960 // of the second pair.
961 void BBVectorize::computeConnectedPairs(
962 std::multimap<Value *, Value *> &CandidatePairs,
963 std::vector<Value *> &PairableInsts,
964 std::multimap<ValuePair, ValuePair> &ConnectedPairs) {
966 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
967 PE = PairableInsts.end(); PI != PE; ++PI) {
968 VPIteratorPair choiceRange = CandidatePairs.equal_range(*PI);
970 for (std::multimap<Value *, Value *>::iterator P = choiceRange.first;
971 P != choiceRange.second; ++P)
972 computePairsConnectedTo(CandidatePairs, PairableInsts,
976 DEBUG(dbgs() << "BBV: found " << ConnectedPairs.size()
977 << " pair connections.\n");
980 // This function builds a set of use tuples such that <A, B> is in the set
981 // if B is in the use tree of A. If B is in the use tree of A, then B
982 // depends on the output of A.
983 void BBVectorize::buildDepMap(
985 std::multimap<Value *, Value *> &CandidatePairs,
986 std::vector<Value *> &PairableInsts,
987 DenseSet<ValuePair> &PairableInstUsers) {
988 DenseSet<Value *> IsInPair;
989 for (std::multimap<Value *, Value *>::iterator C = CandidatePairs.begin(),
990 E = CandidatePairs.end(); C != E; ++C) {
991 IsInPair.insert(C->first);
992 IsInPair.insert(C->second);
995 // Iterate through the basic block, recording all Users of each
996 // pairable instruction.
998 BasicBlock::iterator E = BB.end();
999 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) {
1000 if (IsInPair.find(I) == IsInPair.end()) continue;
1002 DenseSet<Value *> Users;
1003 AliasSetTracker WriteSet(*AA);
1004 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J)
1005 (void) trackUsesOfI(Users, WriteSet, I, J);
1007 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end();
1009 PairableInstUsers.insert(ValuePair(I, *U));
1013 // Returns true if an input to pair P is an output of pair Q and also an
1014 // input of pair Q is an output of pair P. If this is the case, then these
1015 // two pairs cannot be simultaneously fused.
1016 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q,
1017 DenseSet<ValuePair> &PairableInstUsers,
1018 std::multimap<ValuePair, ValuePair> *PairableInstUserMap) {
1019 // Two pairs are in conflict if they are mutual Users of eachother.
1020 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) ||
1021 PairableInstUsers.count(ValuePair(P.first, Q.second)) ||
1022 PairableInstUsers.count(ValuePair(P.second, Q.first)) ||
1023 PairableInstUsers.count(ValuePair(P.second, Q.second));
1024 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) ||
1025 PairableInstUsers.count(ValuePair(Q.first, P.second)) ||
1026 PairableInstUsers.count(ValuePair(Q.second, P.first)) ||
1027 PairableInstUsers.count(ValuePair(Q.second, P.second));
1028 if (PairableInstUserMap) {
1029 // FIXME: The expensive part of the cycle check is not so much the cycle
1030 // check itself but this edge insertion procedure. This needs some
1031 // profiling and probably a different data structure (same is true of
1032 // most uses of std::multimap).
1034 VPPIteratorPair QPairRange = PairableInstUserMap->equal_range(Q);
1035 if (!isSecondInIteratorPair(P, QPairRange))
1036 PairableInstUserMap->insert(VPPair(Q, P));
1039 VPPIteratorPair PPairRange = PairableInstUserMap->equal_range(P);
1040 if (!isSecondInIteratorPair(Q, PPairRange))
1041 PairableInstUserMap->insert(VPPair(P, Q));
1045 return (QUsesP && PUsesQ);
1048 // This function walks the use graph of current pairs to see if, starting
1049 // from P, the walk returns to P.
1050 bool BBVectorize::pairWillFormCycle(ValuePair P,
1051 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1052 DenseSet<ValuePair> &CurrentPairs) {
1053 DEBUG(if (DebugCycleCheck)
1054 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> "
1055 << *P.second << "\n");
1056 // A lookup table of visisted pairs is kept because the PairableInstUserMap
1057 // contains non-direct associations.
1058 DenseSet<ValuePair> Visited;
1059 SmallVector<ValuePair, 32> Q;
1060 // General depth-first post-order traversal:
1063 ValuePair QTop = Q.pop_back_val();
1064 Visited.insert(QTop);
1066 DEBUG(if (DebugCycleCheck)
1067 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> "
1068 << *QTop.second << "\n");
1069 VPPIteratorPair QPairRange = PairableInstUserMap.equal_range(QTop);
1070 for (std::multimap<ValuePair, ValuePair>::iterator C = QPairRange.first;
1071 C != QPairRange.second; ++C) {
1072 if (C->second == P) {
1074 << "BBV: rejected to prevent non-trivial cycle formation: "
1075 << *C->first.first << " <-> " << *C->first.second << "\n");
1079 if (CurrentPairs.count(C->second) && !Visited.count(C->second))
1080 Q.push_back(C->second);
1082 } while (!Q.empty());
1087 // This function builds the initial tree of connected pairs with the
1088 // pair J at the root.
1089 void BBVectorize::buildInitialTreeFor(
1090 std::multimap<Value *, Value *> &CandidatePairs,
1091 std::vector<Value *> &PairableInsts,
1092 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1093 DenseSet<ValuePair> &PairableInstUsers,
1094 DenseMap<Value *, Value *> &ChosenPairs,
1095 DenseMap<ValuePair, size_t> &Tree, ValuePair J) {
1096 // Each of these pairs is viewed as the root node of a Tree. The Tree
1097 // is then walked (depth-first). As this happens, we keep track of
1098 // the pairs that compose the Tree and the maximum depth of the Tree.
1099 SmallVector<ValuePairWithDepth, 32> Q;
1100 // General depth-first post-order traversal:
1101 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1103 ValuePairWithDepth QTop = Q.back();
1105 // Push each child onto the queue:
1106 bool MoreChildren = false;
1107 size_t MaxChildDepth = QTop.second;
1108 VPPIteratorPair qtRange = ConnectedPairs.equal_range(QTop.first);
1109 for (std::multimap<ValuePair, ValuePair>::iterator k = qtRange.first;
1110 k != qtRange.second; ++k) {
1111 // Make sure that this child pair is still a candidate:
1112 bool IsStillCand = false;
1113 VPIteratorPair checkRange =
1114 CandidatePairs.equal_range(k->second.first);
1115 for (std::multimap<Value *, Value *>::iterator m = checkRange.first;
1116 m != checkRange.second; ++m) {
1117 if (m->second == k->second.second) {
1124 DenseMap<ValuePair, size_t>::iterator C = Tree.find(k->second);
1125 if (C == Tree.end()) {
1126 size_t d = getDepthFactor(k->second.first);
1127 Q.push_back(ValuePairWithDepth(k->second, QTop.second+d));
1128 MoreChildren = true;
1130 MaxChildDepth = std::max(MaxChildDepth, C->second);
1135 if (!MoreChildren) {
1136 // Record the current pair as part of the Tree:
1137 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth));
1140 } while (!Q.empty());
1143 // Given some initial tree, prune it by removing conflicting pairs (pairs
1144 // that cannot be simultaneously chosen for vectorization).
1145 void BBVectorize::pruneTreeFor(
1146 std::multimap<Value *, Value *> &CandidatePairs,
1147 std::vector<Value *> &PairableInsts,
1148 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1149 DenseSet<ValuePair> &PairableInstUsers,
1150 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1151 DenseMap<Value *, Value *> &ChosenPairs,
1152 DenseMap<ValuePair, size_t> &Tree,
1153 DenseSet<ValuePair> &PrunedTree, ValuePair J,
1154 bool UseCycleCheck) {
1155 SmallVector<ValuePairWithDepth, 32> Q;
1156 // General depth-first post-order traversal:
1157 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1159 ValuePairWithDepth QTop = Q.pop_back_val();
1160 PrunedTree.insert(QTop.first);
1162 // Visit each child, pruning as necessary...
1163 DenseMap<ValuePair, size_t> BestChildren;
1164 VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first);
1165 for (std::multimap<ValuePair, ValuePair>::iterator K = QTopRange.first;
1166 K != QTopRange.second; ++K) {
1167 DenseMap<ValuePair, size_t>::iterator C = Tree.find(K->second);
1168 if (C == Tree.end()) continue;
1170 // This child is in the Tree, now we need to make sure it is the
1171 // best of any conflicting children. There could be multiple
1172 // conflicting children, so first, determine if we're keeping
1173 // this child, then delete conflicting children as necessary.
1175 // It is also necessary to guard against pairing-induced
1176 // dependencies. Consider instructions a .. x .. y .. b
1177 // such that (a,b) are to be fused and (x,y) are to be fused
1178 // but a is an input to x and b is an output from y. This
1179 // means that y cannot be moved after b but x must be moved
1180 // after b for (a,b) to be fused. In other words, after
1181 // fusing (a,b) we have y .. a/b .. x where y is an input
1182 // to a/b and x is an output to a/b: x and y can no longer
1183 // be legally fused. To prevent this condition, we must
1184 // make sure that a child pair added to the Tree is not
1185 // both an input and output of an already-selected pair.
1187 // Pairing-induced dependencies can also form from more complicated
1188 // cycles. The pair vs. pair conflicts are easy to check, and so
1189 // that is done explicitly for "fast rejection", and because for
1190 // child vs. child conflicts, we may prefer to keep the current
1191 // pair in preference to the already-selected child.
1192 DenseSet<ValuePair> CurrentPairs;
1195 for (DenseMap<ValuePair, size_t>::iterator C2
1196 = BestChildren.begin(), E2 = BestChildren.end();
1198 if (C2->first.first == C->first.first ||
1199 C2->first.first == C->first.second ||
1200 C2->first.second == C->first.first ||
1201 C2->first.second == C->first.second ||
1202 pairsConflict(C2->first, C->first, PairableInstUsers,
1203 UseCycleCheck ? &PairableInstUserMap : 0)) {
1204 if (C2->second >= C->second) {
1209 CurrentPairs.insert(C2->first);
1212 if (!CanAdd) continue;
1214 // Even worse, this child could conflict with another node already
1215 // selected for the Tree. If that is the case, ignore this child.
1216 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(),
1217 E2 = PrunedTree.end(); T != E2; ++T) {
1218 if (T->first == C->first.first ||
1219 T->first == C->first.second ||
1220 T->second == C->first.first ||
1221 T->second == C->first.second ||
1222 pairsConflict(*T, C->first, PairableInstUsers,
1223 UseCycleCheck ? &PairableInstUserMap : 0)) {
1228 CurrentPairs.insert(*T);
1230 if (!CanAdd) continue;
1232 // And check the queue too...
1233 for (SmallVector<ValuePairWithDepth, 32>::iterator C2 = Q.begin(),
1234 E2 = Q.end(); C2 != E2; ++C2) {
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)) {
1245 CurrentPairs.insert(C2->first);
1247 if (!CanAdd) continue;
1249 // Last but not least, check for a conflict with any of the
1250 // already-chosen pairs.
1251 for (DenseMap<Value *, Value *>::iterator C2 =
1252 ChosenPairs.begin(), E2 = ChosenPairs.end();
1254 if (pairsConflict(*C2, C->first, PairableInstUsers,
1255 UseCycleCheck ? &PairableInstUserMap : 0)) {
1260 CurrentPairs.insert(*C2);
1262 if (!CanAdd) continue;
1264 // To check for non-trivial cycles formed by the addition of the
1265 // current pair we've formed a list of all relevant pairs, now use a
1266 // graph walk to check for a cycle. We start from the current pair and
1267 // walk the use tree to see if we again reach the current pair. If we
1268 // do, then the current pair is rejected.
1270 // FIXME: It may be more efficient to use a topological-ordering
1271 // algorithm to improve the cycle check. This should be investigated.
1272 if (UseCycleCheck &&
1273 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs))
1276 // This child can be added, but we may have chosen it in preference
1277 // to an already-selected child. Check for this here, and if a
1278 // conflict is found, then remove the previously-selected child
1279 // before adding this one in its place.
1280 for (DenseMap<ValuePair, size_t>::iterator C2
1281 = BestChildren.begin(); C2 != BestChildren.end();) {
1282 if (C2->first.first == C->first.first ||
1283 C2->first.first == C->first.second ||
1284 C2->first.second == C->first.first ||
1285 C2->first.second == C->first.second ||
1286 pairsConflict(C2->first, C->first, PairableInstUsers))
1287 BestChildren.erase(C2++);
1292 BestChildren.insert(ValuePairWithDepth(C->first, C->second));
1295 for (DenseMap<ValuePair, size_t>::iterator C
1296 = BestChildren.begin(), E2 = BestChildren.end();
1298 size_t DepthF = getDepthFactor(C->first.first);
1299 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF));
1301 } while (!Q.empty());
1304 // This function finds the best tree of mututally-compatible connected
1305 // pairs, given the choice of root pairs as an iterator range.
1306 void BBVectorize::findBestTreeFor(
1307 std::multimap<Value *, Value *> &CandidatePairs,
1308 std::vector<Value *> &PairableInsts,
1309 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1310 DenseSet<ValuePair> &PairableInstUsers,
1311 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1312 DenseMap<Value *, Value *> &ChosenPairs,
1313 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth,
1314 size_t &BestEffSize, VPIteratorPair ChoiceRange,
1315 bool UseCycleCheck) {
1316 for (std::multimap<Value *, Value *>::iterator J = ChoiceRange.first;
1317 J != ChoiceRange.second; ++J) {
1319 // Before going any further, make sure that this pair does not
1320 // conflict with any already-selected pairs (see comment below
1321 // near the Tree pruning for more details).
1322 DenseSet<ValuePair> ChosenPairSet;
1323 bool DoesConflict = false;
1324 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(),
1325 E = ChosenPairs.end(); C != E; ++C) {
1326 if (pairsConflict(*C, *J, PairableInstUsers,
1327 UseCycleCheck ? &PairableInstUserMap : 0)) {
1328 DoesConflict = true;
1332 ChosenPairSet.insert(*C);
1334 if (DoesConflict) continue;
1336 if (UseCycleCheck &&
1337 pairWillFormCycle(*J, PairableInstUserMap, ChosenPairSet))
1340 DenseMap<ValuePair, size_t> Tree;
1341 buildInitialTreeFor(CandidatePairs, PairableInsts, ConnectedPairs,
1342 PairableInstUsers, ChosenPairs, Tree, *J);
1344 // Because we'll keep the child with the largest depth, the largest
1345 // depth is still the same in the unpruned Tree.
1346 size_t MaxDepth = Tree.lookup(*J);
1348 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found Tree for pair {"
1349 << *J->first << " <-> " << *J->second << "} of depth " <<
1350 MaxDepth << " and size " << Tree.size() << "\n");
1352 // At this point the Tree has been constructed, but, may contain
1353 // contradictory children (meaning that different children of
1354 // some tree node may be attempting to fuse the same instruction).
1355 // So now we walk the tree again, in the case of a conflict,
1356 // keep only the child with the largest depth. To break a tie,
1357 // favor the first child.
1359 DenseSet<ValuePair> PrunedTree;
1360 pruneTreeFor(CandidatePairs, PairableInsts, ConnectedPairs,
1361 PairableInstUsers, PairableInstUserMap, ChosenPairs, Tree,
1362 PrunedTree, *J, UseCycleCheck);
1365 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
1366 E = PrunedTree.end(); S != E; ++S)
1367 EffSize += getDepthFactor(S->first);
1369 DEBUG(if (DebugPairSelection)
1370 dbgs() << "BBV: found pruned Tree for pair {"
1371 << *J->first << " <-> " << *J->second << "} of depth " <<
1372 MaxDepth << " and size " << PrunedTree.size() <<
1373 " (effective size: " << EffSize << ")\n");
1374 if (MaxDepth >= Config.ReqChainDepth && EffSize > BestEffSize) {
1375 BestMaxDepth = MaxDepth;
1376 BestEffSize = EffSize;
1377 BestTree = PrunedTree;
1382 // Given the list of candidate pairs, this function selects those
1383 // that will be fused into vector instructions.
1384 void BBVectorize::choosePairs(
1385 std::multimap<Value *, Value *> &CandidatePairs,
1386 std::vector<Value *> &PairableInsts,
1387 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1388 DenseSet<ValuePair> &PairableInstUsers,
1389 DenseMap<Value *, Value *>& ChosenPairs) {
1390 bool UseCycleCheck =
1391 CandidatePairs.size() <= Config.MaxCandPairsForCycleCheck;
1392 std::multimap<ValuePair, ValuePair> PairableInstUserMap;
1393 for (std::vector<Value *>::iterator I = PairableInsts.begin(),
1394 E = PairableInsts.end(); I != E; ++I) {
1395 // The number of possible pairings for this variable:
1396 size_t NumChoices = CandidatePairs.count(*I);
1397 if (!NumChoices) continue;
1399 VPIteratorPair ChoiceRange = CandidatePairs.equal_range(*I);
1401 // The best pair to choose and its tree:
1402 size_t BestMaxDepth = 0, BestEffSize = 0;
1403 DenseSet<ValuePair> BestTree;
1404 findBestTreeFor(CandidatePairs, PairableInsts, ConnectedPairs,
1405 PairableInstUsers, PairableInstUserMap, ChosenPairs,
1406 BestTree, BestMaxDepth, BestEffSize, ChoiceRange,
1409 // A tree has been chosen (or not) at this point. If no tree was
1410 // chosen, then this instruction, I, cannot be paired (and is no longer
1413 DEBUG(if (BestTree.size() > 0)
1414 dbgs() << "BBV: selected pairs in the best tree for: "
1415 << *cast<Instruction>(*I) << "\n");
1417 for (DenseSet<ValuePair>::iterator S = BestTree.begin(),
1418 SE2 = BestTree.end(); S != SE2; ++S) {
1419 // Insert the members of this tree into the list of chosen pairs.
1420 ChosenPairs.insert(ValuePair(S->first, S->second));
1421 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " <<
1422 *S->second << "\n");
1424 // Remove all candidate pairs that have values in the chosen tree.
1425 for (std::multimap<Value *, Value *>::iterator K =
1426 CandidatePairs.begin(); K != CandidatePairs.end();) {
1427 if (K->first == S->first || K->second == S->first ||
1428 K->second == S->second || K->first == S->second) {
1429 // Don't remove the actual pair chosen so that it can be used
1430 // in subsequent tree selections.
1431 if (!(K->first == S->first && K->second == S->second))
1432 CandidatePairs.erase(K++);
1442 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n");
1445 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o,
1450 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) +
1451 (n > 0 ? "." + utostr(n) : "")).str();
1454 // Returns the value that is to be used as the pointer input to the vector
1455 // instruction that fuses I with J.
1456 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context,
1457 Instruction *I, Instruction *J, unsigned o,
1458 bool &FlipMemInputs) {
1460 unsigned IAlignment, JAlignment;
1461 int64_t OffsetInElmts;
1462 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
1465 // The pointer value is taken to be the one with the lowest offset.
1467 if (OffsetInElmts > 0) {
1470 FlipMemInputs = true;
1474 Type *ArgType = cast<PointerType>(IPtr->getType())->getElementType();
1475 Type *VArgType = getVecTypeForPair(ArgType);
1476 Type *VArgPtrType = PointerType::get(VArgType,
1477 cast<PointerType>(IPtr->getType())->getAddressSpace());
1478 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o),
1479 /* insert before */ FlipMemInputs ? J : I);
1482 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J,
1483 unsigned NumElem, unsigned MaskOffset, unsigned NumInElem,
1484 unsigned IdxOffset, std::vector<Constant*> &Mask) {
1485 for (unsigned v = 0; v < NumElem/2; ++v) {
1486 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v);
1488 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context));
1490 unsigned mm = m + (int) IdxOffset;
1491 if (m >= (int) NumInElem)
1492 mm += (int) NumInElem;
1494 Mask[v+MaskOffset] =
1495 ConstantInt::get(Type::getInt32Ty(Context), mm);
1500 // Returns the value that is to be used as the vector-shuffle mask to the
1501 // vector instruction that fuses I with J.
1502 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context,
1503 Instruction *I, Instruction *J) {
1504 // This is the shuffle mask. We need to append the second
1505 // mask to the first, and the numbers need to be adjusted.
1507 Type *ArgType = I->getType();
1508 Type *VArgType = getVecTypeForPair(ArgType);
1510 // Get the total number of elements in the fused vector type.
1511 // By definition, this must equal the number of elements in
1513 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements();
1514 std::vector<Constant*> Mask(NumElem);
1516 Type *OpType = I->getOperand(0)->getType();
1517 unsigned NumInElem = cast<VectorType>(OpType)->getNumElements();
1519 // For the mask from the first pair...
1520 fillNewShuffleMask(Context, I, NumElem, 0, NumInElem, 0, Mask);
1522 // For the mask from the second pair...
1523 fillNewShuffleMask(Context, J, NumElem, NumElem/2, NumInElem, NumInElem,
1526 return ConstantVector::get(Mask);
1529 // Returns the value to be used as the specified operand of the vector
1530 // instruction that fuses I with J.
1531 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I,
1532 Instruction *J, unsigned o, bool FlipMemInputs) {
1533 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
1534 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1);
1536 // Compute the fused vector type for this operand
1537 Type *ArgType = I->getOperand(o)->getType();
1538 VectorType *VArgType = getVecTypeForPair(ArgType);
1540 Instruction *L = I, *H = J;
1541 if (FlipMemInputs) {
1546 if (ArgType->isVectorTy()) {
1547 unsigned numElem = cast<VectorType>(VArgType)->getNumElements();
1548 std::vector<Constant*> Mask(numElem);
1549 for (unsigned v = 0; v < numElem; ++v)
1550 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
1552 Instruction *BV = new ShuffleVectorInst(L->getOperand(o),
1554 ConstantVector::get(Mask),
1555 getReplacementName(I, true, o));
1556 BV->insertBefore(J);
1560 // If these two inputs are the output of another vector instruction,
1561 // then we should use that output directly. It might be necessary to
1562 // permute it first. [When pairings are fused recursively, you can
1563 // end up with cases where a large vector is decomposed into scalars
1564 // using extractelement instructions, then built into size-2
1565 // vectors using insertelement and the into larger vectors using
1566 // shuffles. InstCombine does not simplify all of these cases well,
1567 // and so we make sure that shuffles are generated here when possible.
1568 ExtractElementInst *LEE
1569 = dyn_cast<ExtractElementInst>(L->getOperand(o));
1570 ExtractElementInst *HEE
1571 = dyn_cast<ExtractElementInst>(H->getOperand(o));
1574 LEE->getOperand(0)->getType() == HEE->getOperand(0)->getType()) {
1575 VectorType *EEType = cast<VectorType>(LEE->getOperand(0)->getType());
1576 unsigned LowIndx = cast<ConstantInt>(LEE->getOperand(1))->getZExtValue();
1577 unsigned HighIndx = cast<ConstantInt>(HEE->getOperand(1))->getZExtValue();
1578 if (LEE->getOperand(0) == HEE->getOperand(0)) {
1579 if (LowIndx == 0 && HighIndx == 1)
1580 return LEE->getOperand(0);
1582 std::vector<Constant*> Mask(2);
1583 Mask[0] = ConstantInt::get(Type::getInt32Ty(Context), LowIndx);
1584 Mask[1] = ConstantInt::get(Type::getInt32Ty(Context), HighIndx);
1586 Instruction *BV = new ShuffleVectorInst(LEE->getOperand(0),
1587 UndefValue::get(EEType),
1588 ConstantVector::get(Mask),
1589 getReplacementName(I, true, o));
1590 BV->insertBefore(J);
1594 std::vector<Constant*> Mask(2);
1595 HighIndx += EEType->getNumElements();
1596 Mask[0] = ConstantInt::get(Type::getInt32Ty(Context), LowIndx);
1597 Mask[1] = ConstantInt::get(Type::getInt32Ty(Context), HighIndx);
1599 Instruction *BV = new ShuffleVectorInst(LEE->getOperand(0),
1601 ConstantVector::get(Mask),
1602 getReplacementName(I, true, o));
1603 BV->insertBefore(J);
1607 Instruction *BV1 = InsertElementInst::Create(
1608 UndefValue::get(VArgType),
1609 L->getOperand(o), CV0,
1610 getReplacementName(I, true, o, 1));
1611 BV1->insertBefore(I);
1612 Instruction *BV2 = InsertElementInst::Create(BV1, H->getOperand(o),
1614 getReplacementName(I, true, o, 2));
1615 BV2->insertBefore(J);
1619 // This function creates an array of values that will be used as the inputs
1620 // to the vector instruction that fuses I with J.
1621 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context,
1622 Instruction *I, Instruction *J,
1623 SmallVector<Value *, 3> &ReplacedOperands,
1624 bool &FlipMemInputs) {
1625 FlipMemInputs = false;
1626 unsigned NumOperands = I->getNumOperands();
1628 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) {
1629 // Iterate backward so that we look at the store pointer
1630 // first and know whether or not we need to flip the inputs.
1632 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) {
1633 // This is the pointer for a load/store instruction.
1634 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o,
1637 } else if (isa<CallInst>(I)) {
1638 Function *F = cast<CallInst>(I)->getCalledFunction();
1639 unsigned IID = F->getIntrinsicID();
1640 if (o == NumOperands-1) {
1641 BasicBlock &BB = *I->getParent();
1643 Module *M = BB.getParent()->getParent();
1644 Type *ArgType = I->getType();
1645 Type *VArgType = getVecTypeForPair(ArgType);
1647 // FIXME: is it safe to do this here?
1648 ReplacedOperands[o] = Intrinsic::getDeclaration(M,
1649 (Intrinsic::ID) IID, VArgType);
1651 } else if (IID == Intrinsic::powi && o == 1) {
1652 // The second argument of powi is a single integer and we've already
1653 // checked that both arguments are equal. As a result, we just keep
1654 // I's second argument.
1655 ReplacedOperands[o] = I->getOperand(o);
1658 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) {
1659 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J);
1663 ReplacedOperands[o] =
1664 getReplacementInput(Context, I, J, o, FlipMemInputs);
1668 // This function creates two values that represent the outputs of the
1669 // original I and J instructions. These are generally vector shuffles
1670 // or extracts. In many cases, these will end up being unused and, thus,
1671 // eliminated by later passes.
1672 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
1673 Instruction *J, Instruction *K,
1674 Instruction *&InsertionPt,
1675 Instruction *&K1, Instruction *&K2,
1676 bool &FlipMemInputs) {
1677 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
1678 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1);
1680 if (isa<StoreInst>(I)) {
1681 AA->replaceWithNewValue(I, K);
1682 AA->replaceWithNewValue(J, K);
1684 Type *IType = I->getType();
1685 Type *VType = getVecTypeForPair(IType);
1687 if (IType->isVectorTy()) {
1688 unsigned numElem = cast<VectorType>(IType)->getNumElements();
1689 std::vector<Constant*> Mask1(numElem), Mask2(numElem);
1690 for (unsigned v = 0; v < numElem; ++v) {
1691 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
1692 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElem+v);
1695 K1 = new ShuffleVectorInst(K, UndefValue::get(VType),
1696 ConstantVector::get(
1697 FlipMemInputs ? Mask2 : Mask1),
1698 getReplacementName(K, false, 1));
1699 K2 = new ShuffleVectorInst(K, UndefValue::get(VType),
1700 ConstantVector::get(
1701 FlipMemInputs ? Mask1 : Mask2),
1702 getReplacementName(K, false, 2));
1704 K1 = ExtractElementInst::Create(K, FlipMemInputs ? CV1 : CV0,
1705 getReplacementName(K, false, 1));
1706 K2 = ExtractElementInst::Create(K, FlipMemInputs ? CV0 : CV1,
1707 getReplacementName(K, false, 2));
1711 K2->insertAfter(K1);
1716 // Move all uses of the function I (including pairing-induced uses) after J.
1717 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB,
1718 std::multimap<Value *, Value *> &LoadMoveSet,
1719 Instruction *I, Instruction *J) {
1720 // Skip to the first instruction past I.
1721 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
1723 DenseSet<Value *> Users;
1724 AliasSetTracker WriteSet(*AA);
1725 for (; cast<Instruction>(L) != J; ++L)
1726 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet);
1728 assert(cast<Instruction>(L) == J &&
1729 "Tracking has not proceeded far enough to check for dependencies");
1730 // If J is now in the use set of I, then trackUsesOfI will return true
1731 // and we have a dependency cycle (and the fusing operation must abort).
1732 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSet);
1735 // Move all uses of the function I (including pairing-induced uses) after J.
1736 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB,
1737 std::multimap<Value *, Value *> &LoadMoveSet,
1738 Instruction *&InsertionPt,
1739 Instruction *I, Instruction *J) {
1740 // Skip to the first instruction past I.
1741 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
1743 DenseSet<Value *> Users;
1744 AliasSetTracker WriteSet(*AA);
1745 for (; cast<Instruction>(L) != J;) {
1746 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet)) {
1747 // Move this instruction
1748 Instruction *InstToMove = L; ++L;
1750 DEBUG(dbgs() << "BBV: moving: " << *InstToMove <<
1751 " to after " << *InsertionPt << "\n");
1752 InstToMove->removeFromParent();
1753 InstToMove->insertAfter(InsertionPt);
1754 InsertionPt = InstToMove;
1761 // Collect all load instruction that are in the move set of a given first
1762 // pair member. These loads depend on the first instruction, I, and so need
1763 // to be moved after J (the second instruction) when the pair is fused.
1764 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB,
1765 DenseMap<Value *, Value *> &ChosenPairs,
1766 std::multimap<Value *, Value *> &LoadMoveSet,
1768 // Skip to the first instruction past I.
1769 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
1771 DenseSet<Value *> Users;
1772 AliasSetTracker WriteSet(*AA);
1774 // Note: We cannot end the loop when we reach J because J could be moved
1775 // farther down the use chain by another instruction pairing. Also, J
1776 // could be before I if this is an inverted input.
1777 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) {
1778 if (trackUsesOfI(Users, WriteSet, I, L)) {
1779 if (L->mayReadFromMemory())
1780 LoadMoveSet.insert(ValuePair(L, I));
1785 // In cases where both load/stores and the computation of their pointers
1786 // are chosen for vectorization, we can end up in a situation where the
1787 // aliasing analysis starts returning different query results as the
1788 // process of fusing instruction pairs continues. Because the algorithm
1789 // relies on finding the same use trees here as were found earlier, we'll
1790 // need to precompute the necessary aliasing information here and then
1791 // manually update it during the fusion process.
1792 void BBVectorize::collectLoadMoveSet(BasicBlock &BB,
1793 std::vector<Value *> &PairableInsts,
1794 DenseMap<Value *, Value *> &ChosenPairs,
1795 std::multimap<Value *, Value *> &LoadMoveSet) {
1796 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
1797 PIE = PairableInsts.end(); PI != PIE; ++PI) {
1798 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI);
1799 if (P == ChosenPairs.end()) continue;
1801 Instruction *I = cast<Instruction>(P->first);
1802 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, I);
1806 // When the first instruction in each pair is cloned, it will inherit its
1807 // parent's metadata. This metadata must be combined with that of the other
1808 // instruction in a safe way.
1809 void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) {
1810 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata;
1811 K->getAllMetadataOtherThanDebugLoc(Metadata);
1812 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) {
1813 unsigned Kind = Metadata[i].first;
1814 MDNode *JMD = J->getMetadata(Kind);
1815 MDNode *KMD = Metadata[i].second;
1819 K->setMetadata(Kind, 0); // Remove unknown metadata
1821 case LLVMContext::MD_tbaa:
1822 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD));
1824 case LLVMContext::MD_fpmath:
1825 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD));
1831 // This function fuses the chosen instruction pairs into vector instructions,
1832 // taking care preserve any needed scalar outputs and, then, it reorders the
1833 // remaining instructions as needed (users of the first member of the pair
1834 // need to be moved to after the location of the second member of the pair
1835 // because the vector instruction is inserted in the location of the pair's
1837 void BBVectorize::fuseChosenPairs(BasicBlock &BB,
1838 std::vector<Value *> &PairableInsts,
1839 DenseMap<Value *, Value *> &ChosenPairs) {
1840 LLVMContext& Context = BB.getContext();
1842 // During the vectorization process, the order of the pairs to be fused
1843 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs
1844 // list. After a pair is fused, the flipped pair is removed from the list.
1845 std::vector<ValuePair> FlippedPairs;
1846 FlippedPairs.reserve(ChosenPairs.size());
1847 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(),
1848 E = ChosenPairs.end(); P != E; ++P)
1849 FlippedPairs.push_back(ValuePair(P->second, P->first));
1850 for (std::vector<ValuePair>::iterator P = FlippedPairs.begin(),
1851 E = FlippedPairs.end(); P != E; ++P)
1852 ChosenPairs.insert(*P);
1854 std::multimap<Value *, Value *> LoadMoveSet;
1855 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, LoadMoveSet);
1857 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n");
1859 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) {
1860 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI);
1861 if (P == ChosenPairs.end()) {
1866 if (getDepthFactor(P->first) == 0) {
1867 // These instructions are not really fused, but are tracked as though
1868 // they are. Any case in which it would be interesting to fuse them
1869 // will be taken care of by InstCombine.
1875 Instruction *I = cast<Instruction>(P->first),
1876 *J = cast<Instruction>(P->second);
1878 DEBUG(dbgs() << "BBV: fusing: " << *I <<
1879 " <-> " << *J << "\n");
1881 // Remove the pair and flipped pair from the list.
1882 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second);
1883 assert(FP != ChosenPairs.end() && "Flipped pair not found in list");
1884 ChosenPairs.erase(FP);
1885 ChosenPairs.erase(P);
1887 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSet, I, J)) {
1888 DEBUG(dbgs() << "BBV: fusion of: " << *I <<
1890 " aborted because of non-trivial dependency cycle\n");
1897 unsigned NumOperands = I->getNumOperands();
1898 SmallVector<Value *, 3> ReplacedOperands(NumOperands);
1899 getReplacementInputsForPair(Context, I, J, ReplacedOperands,
1902 // Make a copy of the original operation, change its type to the vector
1903 // type and replace its operands with the vector operands.
1904 Instruction *K = I->clone();
1905 if (I->hasName()) K->takeName(I);
1907 if (!isa<StoreInst>(K))
1908 K->mutateType(getVecTypeForPair(I->getType()));
1910 combineMetadata(K, J);
1912 for (unsigned o = 0; o < NumOperands; ++o)
1913 K->setOperand(o, ReplacedOperands[o]);
1915 // If we've flipped the memory inputs, make sure that we take the correct
1917 if (FlipMemInputs) {
1918 if (isa<StoreInst>(K))
1919 cast<StoreInst>(K)->setAlignment(cast<StoreInst>(J)->getAlignment());
1921 cast<LoadInst>(K)->setAlignment(cast<LoadInst>(J)->getAlignment());
1926 // Instruction insertion point:
1927 Instruction *InsertionPt = K;
1928 Instruction *K1 = 0, *K2 = 0;
1929 replaceOutputsOfPair(Context, I, J, K, InsertionPt, K1, K2,
1932 // The use tree of the first original instruction must be moved to after
1933 // the location of the second instruction. The entire use tree of the
1934 // first instruction is disjoint from the input tree of the second
1935 // (by definition), and so commutes with it.
1937 moveUsesOfIAfterJ(BB, LoadMoveSet, InsertionPt, I, J);
1939 if (!isa<StoreInst>(I)) {
1940 I->replaceAllUsesWith(K1);
1941 J->replaceAllUsesWith(K2);
1942 AA->replaceWithNewValue(I, K1);
1943 AA->replaceWithNewValue(J, K2);
1946 // Instructions that may read from memory may be in the load move set.
1947 // Once an instruction is fused, we no longer need its move set, and so
1948 // the values of the map never need to be updated. However, when a load
1949 // is fused, we need to merge the entries from both instructions in the
1950 // pair in case those instructions were in the move set of some other
1951 // yet-to-be-fused pair. The loads in question are the keys of the map.
1952 if (I->mayReadFromMemory()) {
1953 std::vector<ValuePair> NewSetMembers;
1954 VPIteratorPair IPairRange = LoadMoveSet.equal_range(I);
1955 VPIteratorPair JPairRange = LoadMoveSet.equal_range(J);
1956 for (std::multimap<Value *, Value *>::iterator N = IPairRange.first;
1957 N != IPairRange.second; ++N)
1958 NewSetMembers.push_back(ValuePair(K, N->second));
1959 for (std::multimap<Value *, Value *>::iterator N = JPairRange.first;
1960 N != JPairRange.second; ++N)
1961 NewSetMembers.push_back(ValuePair(K, N->second));
1962 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(),
1963 AE = NewSetMembers.end(); A != AE; ++A)
1964 LoadMoveSet.insert(*A);
1967 // Before removing I, set the iterator to the next instruction.
1968 PI = llvm::next(BasicBlock::iterator(I));
1969 if (cast<Instruction>(PI) == J)
1974 I->eraseFromParent();
1975 J->eraseFromParent();
1978 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n");
1982 char BBVectorize::ID = 0;
1983 static const char bb_vectorize_name[] = "Basic-Block Vectorization";
1984 INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
1985 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
1986 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
1987 INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
1989 BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) {
1990 return new BBVectorize(C);
1994 llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) {
1995 BBVectorize BBVectorizer(P, C);
1996 return BBVectorizer.vectorizeBB(BB);
1999 //===----------------------------------------------------------------------===//
2000 VectorizeConfig::VectorizeConfig() {
2001 VectorBits = ::VectorBits;
2002 VectorizeBools = !::NoBools;
2003 VectorizeInts = !::NoInts;
2004 VectorizeFloats = !::NoFloats;
2005 VectorizePointers = !::NoPointers;
2006 VectorizeCasts = !::NoCasts;
2007 VectorizeMath = !::NoMath;
2008 VectorizeFMA = !::NoFMA;
2009 VectorizeSelect = !::NoSelect;
2010 VectorizeCmp = !::NoCmp;
2011 VectorizeGEP = !::NoGEP;
2012 VectorizeMemOps = !::NoMemOps;
2013 AlignedOnly = ::AlignedOnly;
2014 ReqChainDepth= ::ReqChainDepth;
2015 SearchLimit = ::SearchLimit;
2016 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck;
2017 SplatBreaksChain = ::SplatBreaksChain;
2018 MaxInsts = ::MaxInsts;
2019 MaxIter = ::MaxIter;
2020 NoMemOpBoost = ::NoMemOpBoost;
2021 FastDep = ::FastDep;