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/DataLayout.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<DataLayout>();
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,
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 void collectPtrInfo(std::vector<Value *> &PairableInsts,
318 DenseMap<Value *, Value *> &ChosenPairs,
319 DenseSet<Value *> &LowPtrInsts);
321 bool canMoveUsesOfIAfterJ(BasicBlock &BB,
322 std::multimap<Value *, Value *> &LoadMoveSet,
323 Instruction *I, Instruction *J);
325 void moveUsesOfIAfterJ(BasicBlock &BB,
326 std::multimap<Value *, Value *> &LoadMoveSet,
327 Instruction *&InsertionPt,
328 Instruction *I, Instruction *J);
330 void combineMetadata(Instruction *K, const Instruction *J);
332 bool vectorizeBB(BasicBlock &BB) {
333 bool changed = false;
334 // Iterate a sufficient number of times to merge types of size 1 bit,
335 // then 2 bits, then 4, etc. up to half of the target vector width of the
336 // target vector register.
339 v <= Config.VectorBits && (!Config.MaxIter || n <= Config.MaxIter);
341 DEBUG(dbgs() << "BBV: fusing loop #" << n <<
342 " for " << BB.getName() << " in " <<
343 BB.getParent()->getName() << "...\n");
344 if (vectorizePairs(BB))
350 if (changed && !Pow2LenOnly) {
352 for (; !Config.MaxIter || n <= Config.MaxIter; ++n) {
353 DEBUG(dbgs() << "BBV: fusing for non-2^n-length vectors loop #: " <<
354 n << " for " << BB.getName() << " in " <<
355 BB.getParent()->getName() << "...\n");
356 if (!vectorizePairs(BB, true)) break;
360 DEBUG(dbgs() << "BBV: done!\n");
364 virtual bool runOnBasicBlock(BasicBlock &BB) {
365 AA = &getAnalysis<AliasAnalysis>();
366 SE = &getAnalysis<ScalarEvolution>();
367 TD = getAnalysisIfAvailable<DataLayout>();
369 return vectorizeBB(BB);
372 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
373 BasicBlockPass::getAnalysisUsage(AU);
374 AU.addRequired<AliasAnalysis>();
375 AU.addRequired<ScalarEvolution>();
376 AU.addPreserved<AliasAnalysis>();
377 AU.addPreserved<ScalarEvolution>();
378 AU.setPreservesCFG();
381 static inline VectorType *getVecTypeForPair(Type *ElemTy, Type *Elem2Ty) {
382 assert(ElemTy->getScalarType() == Elem2Ty->getScalarType() &&
383 "Cannot form vector from incompatible scalar types");
384 Type *STy = ElemTy->getScalarType();
387 if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) {
388 numElem = VTy->getNumElements();
393 if (VectorType *VTy = dyn_cast<VectorType>(Elem2Ty)) {
394 numElem += VTy->getNumElements();
399 return VectorType::get(STy, numElem);
402 static inline void getInstructionTypes(Instruction *I,
403 Type *&T1, Type *&T2) {
404 if (isa<StoreInst>(I)) {
405 // For stores, it is the value type, not the pointer type that matters
406 // because the value is what will come from a vector register.
408 Value *IVal = cast<StoreInst>(I)->getValueOperand();
409 T1 = IVal->getType();
415 T2 = cast<CastInst>(I)->getSrcTy();
420 // Returns the weight associated with the provided value. A chain of
421 // candidate pairs has a length given by the sum of the weights of its
422 // members (one weight per pair; the weight of each member of the pair
423 // is assumed to be the same). This length is then compared to the
424 // chain-length threshold to determine if a given chain is significant
425 // enough to be vectorized. The length is also used in comparing
426 // candidate chains where longer chains are considered to be better.
427 // Note: when this function returns 0, the resulting instructions are
428 // not actually fused.
429 inline size_t getDepthFactor(Value *V) {
430 // InsertElement and ExtractElement have a depth factor of zero. This is
431 // for two reasons: First, they cannot be usefully fused. Second, because
432 // the pass generates a lot of these, they can confuse the simple metric
433 // used to compare the trees in the next iteration. Thus, giving them a
434 // weight of zero allows the pass to essentially ignore them in
435 // subsequent iterations when looking for vectorization opportunities
436 // while still tracking dependency chains that flow through those
438 if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V))
441 // Give a load or store half of the required depth so that load/store
442 // pairs will vectorize.
443 if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V)))
444 return Config.ReqChainDepth/2;
449 // This determines the relative offset of two loads or stores, returning
450 // true if the offset could be determined to be some constant value.
451 // For example, if OffsetInElmts == 1, then J accesses the memory directly
452 // after I; if OffsetInElmts == -1 then I accesses the memory
454 bool getPairPtrInfo(Instruction *I, Instruction *J,
455 Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment,
456 int64_t &OffsetInElmts) {
458 if (isa<LoadInst>(I)) {
459 IPtr = cast<LoadInst>(I)->getPointerOperand();
460 JPtr = cast<LoadInst>(J)->getPointerOperand();
461 IAlignment = cast<LoadInst>(I)->getAlignment();
462 JAlignment = cast<LoadInst>(J)->getAlignment();
464 IPtr = cast<StoreInst>(I)->getPointerOperand();
465 JPtr = cast<StoreInst>(J)->getPointerOperand();
466 IAlignment = cast<StoreInst>(I)->getAlignment();
467 JAlignment = cast<StoreInst>(J)->getAlignment();
470 const SCEV *IPtrSCEV = SE->getSCEV(IPtr);
471 const SCEV *JPtrSCEV = SE->getSCEV(JPtr);
473 // If this is a trivial offset, then we'll get something like
474 // 1*sizeof(type). With target data, which we need anyway, this will get
475 // constant folded into a number.
476 const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV);
477 if (const SCEVConstant *ConstOffSCEV =
478 dyn_cast<SCEVConstant>(OffsetSCEV)) {
479 ConstantInt *IntOff = ConstOffSCEV->getValue();
480 int64_t Offset = IntOff->getSExtValue();
482 Type *VTy = cast<PointerType>(IPtr->getType())->getElementType();
483 int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy);
485 Type *VTy2 = cast<PointerType>(JPtr->getType())->getElementType();
486 if (VTy != VTy2 && Offset < 0) {
487 int64_t VTy2TSS = (int64_t) TD->getTypeStoreSize(VTy2);
488 OffsetInElmts = Offset/VTy2TSS;
489 return (abs64(Offset) % VTy2TSS) == 0;
492 OffsetInElmts = Offset/VTyTSS;
493 return (abs64(Offset) % VTyTSS) == 0;
499 // Returns true if the provided CallInst represents an intrinsic that can
501 bool isVectorizableIntrinsic(CallInst* I) {
502 Function *F = I->getCalledFunction();
503 if (!F) return false;
505 unsigned IID = F->getIntrinsicID();
506 if (!IID) return false;
511 case Intrinsic::sqrt:
512 case Intrinsic::powi:
516 case Intrinsic::log2:
517 case Intrinsic::log10:
519 case Intrinsic::exp2:
521 return Config.VectorizeMath;
523 return Config.VectorizeFMA;
527 // Returns true if J is the second element in some pair referenced by
528 // some multimap pair iterator pair.
529 template <typename V>
530 bool isSecondInIteratorPair(V J, std::pair<
531 typename std::multimap<V, V>::iterator,
532 typename std::multimap<V, V>::iterator> PairRange) {
533 for (typename std::multimap<V, V>::iterator K = PairRange.first;
534 K != PairRange.second; ++K)
535 if (K->second == J) return true;
541 // This function implements one vectorization iteration on the provided
542 // basic block. It returns true if the block is changed.
543 bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) {
545 BasicBlock::iterator Start = BB.getFirstInsertionPt();
547 std::vector<Value *> AllPairableInsts;
548 DenseMap<Value *, Value *> AllChosenPairs;
551 std::vector<Value *> PairableInsts;
552 std::multimap<Value *, Value *> CandidatePairs;
553 ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs,
554 PairableInsts, NonPow2Len);
555 if (PairableInsts.empty()) continue;
557 // Now we have a map of all of the pairable instructions and we need to
558 // select the best possible pairing. A good pairing is one such that the
559 // users of the pair are also paired. This defines a (directed) forest
560 // over the pairs such that two pairs are connected iff the second pair
563 // Note that it only matters that both members of the second pair use some
564 // element of the first pair (to allow for splatting).
566 std::multimap<ValuePair, ValuePair> ConnectedPairs;
567 computeConnectedPairs(CandidatePairs, PairableInsts, ConnectedPairs);
568 if (ConnectedPairs.empty()) continue;
570 // Build the pairable-instruction dependency map
571 DenseSet<ValuePair> PairableInstUsers;
572 buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers);
574 // There is now a graph of the connected pairs. For each variable, pick
575 // the pairing with the largest tree meeting the depth requirement on at
576 // least one branch. Then select all pairings that are part of that tree
577 // and remove them from the list of available pairings and pairable
580 DenseMap<Value *, Value *> ChosenPairs;
581 choosePairs(CandidatePairs, PairableInsts, ConnectedPairs,
582 PairableInstUsers, ChosenPairs);
584 if (ChosenPairs.empty()) continue;
585 AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(),
586 PairableInsts.end());
587 AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end());
588 } while (ShouldContinue);
590 if (AllChosenPairs.empty()) return false;
591 NumFusedOps += AllChosenPairs.size();
593 // A set of pairs has now been selected. It is now necessary to replace the
594 // paired instructions with vector instructions. For this procedure each
595 // operand must be replaced with a vector operand. This vector is formed
596 // by using build_vector on the old operands. The replaced values are then
597 // replaced with a vector_extract on the result. Subsequent optimization
598 // passes should coalesce the build/extract combinations.
600 fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs);
602 // It is important to cleanup here so that future iterations of this
603 // function have less work to do.
604 (void) SimplifyInstructionsInBlock(&BB, TD, AA->getTargetLibraryInfo());
608 // This function returns true if the provided instruction is capable of being
609 // fused into a vector instruction. This determination is based only on the
610 // type and other attributes of the instruction.
611 bool BBVectorize::isInstVectorizable(Instruction *I,
612 bool &IsSimpleLoadStore) {
613 IsSimpleLoadStore = false;
615 if (CallInst *C = dyn_cast<CallInst>(I)) {
616 if (!isVectorizableIntrinsic(C))
618 } else if (LoadInst *L = dyn_cast<LoadInst>(I)) {
619 // Vectorize simple loads if possbile:
620 IsSimpleLoadStore = L->isSimple();
621 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
623 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) {
624 // Vectorize simple stores if possbile:
625 IsSimpleLoadStore = S->isSimple();
626 if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
628 } else if (CastInst *C = dyn_cast<CastInst>(I)) {
629 // We can vectorize casts, but not casts of pointer types, etc.
630 if (!Config.VectorizeCasts)
633 Type *SrcTy = C->getSrcTy();
634 if (!SrcTy->isSingleValueType())
637 Type *DestTy = C->getDestTy();
638 if (!DestTy->isSingleValueType())
640 } else if (isa<SelectInst>(I)) {
641 if (!Config.VectorizeSelect)
643 } else if (isa<CmpInst>(I)) {
644 if (!Config.VectorizeCmp)
646 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) {
647 if (!Config.VectorizeGEP)
650 // Currently, vector GEPs exist only with one index.
651 if (G->getNumIndices() != 1)
653 } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) ||
654 isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) {
658 // We can't vectorize memory operations without target data
659 if (TD == 0 && IsSimpleLoadStore)
663 getInstructionTypes(I, T1, T2);
665 // Not every type can be vectorized...
666 if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) ||
667 !(VectorType::isValidElementType(T2) || T2->isVectorTy()))
670 if (T1->getScalarSizeInBits() == 1 && T2->getScalarSizeInBits() == 1) {
671 if (!Config.VectorizeBools)
674 if (!Config.VectorizeInts
675 && (T1->isIntOrIntVectorTy() || T2->isIntOrIntVectorTy()))
679 if (!Config.VectorizeFloats
680 && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy()))
683 // Don't vectorize target-specific types.
684 if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy())
686 if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy())
689 if ((!Config.VectorizePointers || TD == 0) &&
690 (T1->getScalarType()->isPointerTy() ||
691 T2->getScalarType()->isPointerTy()))
694 if (T1->getPrimitiveSizeInBits() >= Config.VectorBits ||
695 T2->getPrimitiveSizeInBits() >= Config.VectorBits)
701 // This function returns true if the two provided instructions are compatible
702 // (meaning that they can be fused into a vector instruction). This assumes
703 // that I has already been determined to be vectorizable and that J is not
704 // in the use tree of I.
705 bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J,
706 bool IsSimpleLoadStore, bool NonPow2Len) {
707 DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I <<
708 " <-> " << *J << "\n");
710 // Loads and stores can be merged if they have different alignments,
711 // but are otherwise the same.
712 if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment |
713 (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0)))
716 Type *IT1, *IT2, *JT1, *JT2;
717 getInstructionTypes(I, IT1, IT2);
718 getInstructionTypes(J, JT1, JT2);
719 unsigned MaxTypeBits = std::max(
720 IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(),
721 IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits());
722 if (MaxTypeBits > Config.VectorBits)
725 // FIXME: handle addsub-type operations!
727 if (IsSimpleLoadStore) {
729 unsigned IAlignment, JAlignment;
730 int64_t OffsetInElmts = 0;
731 if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
732 OffsetInElmts) && abs64(OffsetInElmts) == 1) {
733 if (Config.AlignedOnly) {
734 Type *aTypeI = isa<StoreInst>(I) ?
735 cast<StoreInst>(I)->getValueOperand()->getType() : I->getType();
736 Type *aTypeJ = isa<StoreInst>(J) ?
737 cast<StoreInst>(J)->getValueOperand()->getType() : J->getType();
739 // An aligned load or store is possible only if the instruction
740 // with the lower offset has an alignment suitable for the
743 unsigned BottomAlignment = IAlignment;
744 if (OffsetInElmts < 0) BottomAlignment = JAlignment;
746 Type *VType = getVecTypeForPair(aTypeI, aTypeJ);
747 unsigned VecAlignment = TD->getPrefTypeAlignment(VType);
748 if (BottomAlignment < VecAlignment)
756 // The powi intrinsic is special because only the first argument is
757 // vectorized, the second arguments must be equal.
758 CallInst *CI = dyn_cast<CallInst>(I);
760 if (CI && (FI = CI->getCalledFunction()) &&
761 FI->getIntrinsicID() == Intrinsic::powi) {
763 Value *A1I = CI->getArgOperand(1),
764 *A1J = cast<CallInst>(J)->getArgOperand(1);
765 const SCEV *A1ISCEV = SE->getSCEV(A1I),
766 *A1JSCEV = SE->getSCEV(A1J);
767 return (A1ISCEV == A1JSCEV);
773 // Figure out whether or not J uses I and update the users and write-set
774 // structures associated with I. Specifically, Users represents the set of
775 // instructions that depend on I. WriteSet represents the set
776 // of memory locations that are dependent on I. If UpdateUsers is true,
777 // and J uses I, then Users is updated to contain J and WriteSet is updated
778 // to contain any memory locations to which J writes. The function returns
779 // true if J uses I. By default, alias analysis is used to determine
780 // whether J reads from memory that overlaps with a location in WriteSet.
781 // If LoadMoveSet is not null, then it is a previously-computed multimap
782 // where the key is the memory-based user instruction and the value is
783 // the instruction to be compared with I. So, if LoadMoveSet is provided,
784 // then the alias analysis is not used. This is necessary because this
785 // function is called during the process of moving instructions during
786 // vectorization and the results of the alias analysis are not stable during
788 bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users,
789 AliasSetTracker &WriteSet, Instruction *I,
790 Instruction *J, bool UpdateUsers,
791 std::multimap<Value *, Value *> *LoadMoveSet) {
794 // This instruction may already be marked as a user due, for example, to
795 // being a member of a selected pair.
800 for (User::op_iterator JU = J->op_begin(), JE = J->op_end();
803 if (I == V || Users.count(V)) {
808 if (!UsesI && J->mayReadFromMemory()) {
810 VPIteratorPair JPairRange = LoadMoveSet->equal_range(J);
811 UsesI = isSecondInIteratorPair<Value*>(I, JPairRange);
813 for (AliasSetTracker::iterator W = WriteSet.begin(),
814 WE = WriteSet.end(); W != WE; ++W) {
815 if (W->aliasesUnknownInst(J, *AA)) {
823 if (UsesI && UpdateUsers) {
824 if (J->mayWriteToMemory()) WriteSet.add(J);
831 // This function iterates over all instruction pairs in the provided
832 // basic block and collects all candidate pairs for vectorization.
833 bool BBVectorize::getCandidatePairs(BasicBlock &BB,
834 BasicBlock::iterator &Start,
835 std::multimap<Value *, Value *> &CandidatePairs,
836 std::vector<Value *> &PairableInsts, bool NonPow2Len) {
837 BasicBlock::iterator E = BB.end();
838 if (Start == E) return false;
840 bool ShouldContinue = false, IAfterStart = false;
841 for (BasicBlock::iterator I = Start++; I != E; ++I) {
842 if (I == Start) IAfterStart = true;
844 bool IsSimpleLoadStore;
845 if (!isInstVectorizable(I, IsSimpleLoadStore)) continue;
847 // Look for an instruction with which to pair instruction *I...
848 DenseSet<Value *> Users;
849 AliasSetTracker WriteSet(*AA);
850 bool JAfterStart = IAfterStart;
851 BasicBlock::iterator J = llvm::next(I);
852 for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) {
853 if (J == Start) JAfterStart = true;
855 // Determine if J uses I, if so, exit the loop.
856 bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep);
857 if (Config.FastDep) {
858 // Note: For this heuristic to be effective, independent operations
859 // must tend to be intermixed. This is likely to be true from some
860 // kinds of grouped loop unrolling (but not the generic LLVM pass),
861 // but otherwise may require some kind of reordering pass.
863 // When using fast dependency analysis,
864 // stop searching after first use:
870 // J does not use I, and comes before the first use of I, so it can be
871 // merged with I if the instructions are compatible.
872 if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len)) continue;
874 // J is a candidate for merging with I.
875 if (!PairableInsts.size() ||
876 PairableInsts[PairableInsts.size()-1] != I) {
877 PairableInsts.push_back(I);
880 CandidatePairs.insert(ValuePair(I, J));
882 // The next call to this function must start after the last instruction
883 // selected during this invocation.
885 Start = llvm::next(J);
886 IAfterStart = JAfterStart = false;
889 DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair "
890 << *I << " <-> " << *J << "\n");
892 // If we have already found too many pairs, break here and this function
893 // will be called again starting after the last instruction selected
894 // during this invocation.
895 if (PairableInsts.size() >= Config.MaxInsts) {
896 ShouldContinue = true;
905 DEBUG(dbgs() << "BBV: found " << PairableInsts.size()
906 << " instructions with candidate pairs\n");
908 return ShouldContinue;
911 // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that
912 // it looks for pairs such that both members have an input which is an
913 // output of PI or PJ.
914 void BBVectorize::computePairsConnectedTo(
915 std::multimap<Value *, Value *> &CandidatePairs,
916 std::vector<Value *> &PairableInsts,
917 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
921 // For each possible pairing for this variable, look at the uses of
922 // the first value...
923 for (Value::use_iterator I = P.first->use_begin(),
924 E = P.first->use_end(); I != E; ++I) {
925 if (isa<LoadInst>(*I)) {
926 // A pair cannot be connected to a load because the load only takes one
927 // operand (the address) and it is a scalar even after vectorization.
929 } else if ((SI = dyn_cast<StoreInst>(*I)) &&
930 P.first == SI->getPointerOperand()) {
931 // Similarly, a pair cannot be connected to a store through its
936 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I);
938 // For each use of the first variable, look for uses of the second
940 for (Value::use_iterator J = P.second->use_begin(),
941 E2 = P.second->use_end(); J != E2; ++J) {
942 if ((SJ = dyn_cast<StoreInst>(*J)) &&
943 P.second == SJ->getPointerOperand())
946 VPIteratorPair JPairRange = CandidatePairs.equal_range(*J);
949 if (isSecondInIteratorPair<Value*>(*J, IPairRange))
950 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J)));
953 if (isSecondInIteratorPair<Value*>(*I, JPairRange))
954 ConnectedPairs.insert(VPPair(P, ValuePair(*J, *I)));
957 if (Config.SplatBreaksChain) continue;
958 // Look for cases where just the first value in the pair is used by
959 // both members of another pair (splatting).
960 for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) {
961 if ((SJ = dyn_cast<StoreInst>(*J)) &&
962 P.first == SJ->getPointerOperand())
965 if (isSecondInIteratorPair<Value*>(*J, IPairRange))
966 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J)));
970 if (Config.SplatBreaksChain) return;
971 // Look for cases where just the second value in the pair is used by
972 // both members of another pair (splatting).
973 for (Value::use_iterator I = P.second->use_begin(),
974 E = P.second->use_end(); I != E; ++I) {
975 if (isa<LoadInst>(*I))
977 else if ((SI = dyn_cast<StoreInst>(*I)) &&
978 P.second == SI->getPointerOperand())
981 VPIteratorPair IPairRange = CandidatePairs.equal_range(*I);
983 for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) {
984 if ((SJ = dyn_cast<StoreInst>(*J)) &&
985 P.second == SJ->getPointerOperand())
988 if (isSecondInIteratorPair<Value*>(*J, IPairRange))
989 ConnectedPairs.insert(VPPair(P, ValuePair(*I, *J)));
994 // This function figures out which pairs are connected. Two pairs are
995 // connected if some output of the first pair forms an input to both members
996 // of the second pair.
997 void BBVectorize::computeConnectedPairs(
998 std::multimap<Value *, Value *> &CandidatePairs,
999 std::vector<Value *> &PairableInsts,
1000 std::multimap<ValuePair, ValuePair> &ConnectedPairs) {
1002 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
1003 PE = PairableInsts.end(); PI != PE; ++PI) {
1004 VPIteratorPair choiceRange = CandidatePairs.equal_range(*PI);
1006 for (std::multimap<Value *, Value *>::iterator P = choiceRange.first;
1007 P != choiceRange.second; ++P)
1008 computePairsConnectedTo(CandidatePairs, PairableInsts,
1009 ConnectedPairs, *P);
1012 DEBUG(dbgs() << "BBV: found " << ConnectedPairs.size()
1013 << " pair connections.\n");
1016 // This function builds a set of use tuples such that <A, B> is in the set
1017 // if B is in the use tree of A. If B is in the use tree of A, then B
1018 // depends on the output of A.
1019 void BBVectorize::buildDepMap(
1021 std::multimap<Value *, Value *> &CandidatePairs,
1022 std::vector<Value *> &PairableInsts,
1023 DenseSet<ValuePair> &PairableInstUsers) {
1024 DenseSet<Value *> IsInPair;
1025 for (std::multimap<Value *, Value *>::iterator C = CandidatePairs.begin(),
1026 E = CandidatePairs.end(); C != E; ++C) {
1027 IsInPair.insert(C->first);
1028 IsInPair.insert(C->second);
1031 // Iterate through the basic block, recording all Users of each
1032 // pairable instruction.
1034 BasicBlock::iterator E = BB.end();
1035 for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) {
1036 if (IsInPair.find(I) == IsInPair.end()) continue;
1038 DenseSet<Value *> Users;
1039 AliasSetTracker WriteSet(*AA);
1040 for (BasicBlock::iterator J = llvm::next(I); J != E; ++J)
1041 (void) trackUsesOfI(Users, WriteSet, I, J);
1043 for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end();
1045 PairableInstUsers.insert(ValuePair(I, *U));
1049 // Returns true if an input to pair P is an output of pair Q and also an
1050 // input of pair Q is an output of pair P. If this is the case, then these
1051 // two pairs cannot be simultaneously fused.
1052 bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q,
1053 DenseSet<ValuePair> &PairableInstUsers,
1054 std::multimap<ValuePair, ValuePair> *PairableInstUserMap) {
1055 // Two pairs are in conflict if they are mutual Users of eachother.
1056 bool QUsesP = PairableInstUsers.count(ValuePair(P.first, Q.first)) ||
1057 PairableInstUsers.count(ValuePair(P.first, Q.second)) ||
1058 PairableInstUsers.count(ValuePair(P.second, Q.first)) ||
1059 PairableInstUsers.count(ValuePair(P.second, Q.second));
1060 bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first, P.first)) ||
1061 PairableInstUsers.count(ValuePair(Q.first, P.second)) ||
1062 PairableInstUsers.count(ValuePair(Q.second, P.first)) ||
1063 PairableInstUsers.count(ValuePair(Q.second, P.second));
1064 if (PairableInstUserMap) {
1065 // FIXME: The expensive part of the cycle check is not so much the cycle
1066 // check itself but this edge insertion procedure. This needs some
1067 // profiling and probably a different data structure (same is true of
1068 // most uses of std::multimap).
1070 VPPIteratorPair QPairRange = PairableInstUserMap->equal_range(Q);
1071 if (!isSecondInIteratorPair(P, QPairRange))
1072 PairableInstUserMap->insert(VPPair(Q, P));
1075 VPPIteratorPair PPairRange = PairableInstUserMap->equal_range(P);
1076 if (!isSecondInIteratorPair(Q, PPairRange))
1077 PairableInstUserMap->insert(VPPair(P, Q));
1081 return (QUsesP && PUsesQ);
1084 // This function walks the use graph of current pairs to see if, starting
1085 // from P, the walk returns to P.
1086 bool BBVectorize::pairWillFormCycle(ValuePair P,
1087 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1088 DenseSet<ValuePair> &CurrentPairs) {
1089 DEBUG(if (DebugCycleCheck)
1090 dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> "
1091 << *P.second << "\n");
1092 // A lookup table of visisted pairs is kept because the PairableInstUserMap
1093 // contains non-direct associations.
1094 DenseSet<ValuePair> Visited;
1095 SmallVector<ValuePair, 32> Q;
1096 // General depth-first post-order traversal:
1099 ValuePair QTop = Q.pop_back_val();
1100 Visited.insert(QTop);
1102 DEBUG(if (DebugCycleCheck)
1103 dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> "
1104 << *QTop.second << "\n");
1105 VPPIteratorPair QPairRange = PairableInstUserMap.equal_range(QTop);
1106 for (std::multimap<ValuePair, ValuePair>::iterator C = QPairRange.first;
1107 C != QPairRange.second; ++C) {
1108 if (C->second == P) {
1110 << "BBV: rejected to prevent non-trivial cycle formation: "
1111 << *C->first.first << " <-> " << *C->first.second << "\n");
1115 if (CurrentPairs.count(C->second) && !Visited.count(C->second))
1116 Q.push_back(C->second);
1118 } while (!Q.empty());
1123 // This function builds the initial tree of connected pairs with the
1124 // pair J at the root.
1125 void BBVectorize::buildInitialTreeFor(
1126 std::multimap<Value *, Value *> &CandidatePairs,
1127 std::vector<Value *> &PairableInsts,
1128 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1129 DenseSet<ValuePair> &PairableInstUsers,
1130 DenseMap<Value *, Value *> &ChosenPairs,
1131 DenseMap<ValuePair, size_t> &Tree, ValuePair J) {
1132 // Each of these pairs is viewed as the root node of a Tree. The Tree
1133 // is then walked (depth-first). As this happens, we keep track of
1134 // the pairs that compose the Tree and the maximum depth of the Tree.
1135 SmallVector<ValuePairWithDepth, 32> Q;
1136 // General depth-first post-order traversal:
1137 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1139 ValuePairWithDepth QTop = Q.back();
1141 // Push each child onto the queue:
1142 bool MoreChildren = false;
1143 size_t MaxChildDepth = QTop.second;
1144 VPPIteratorPair qtRange = ConnectedPairs.equal_range(QTop.first);
1145 for (std::multimap<ValuePair, ValuePair>::iterator k = qtRange.first;
1146 k != qtRange.second; ++k) {
1147 // Make sure that this child pair is still a candidate:
1148 bool IsStillCand = false;
1149 VPIteratorPair checkRange =
1150 CandidatePairs.equal_range(k->second.first);
1151 for (std::multimap<Value *, Value *>::iterator m = checkRange.first;
1152 m != checkRange.second; ++m) {
1153 if (m->second == k->second.second) {
1160 DenseMap<ValuePair, size_t>::iterator C = Tree.find(k->second);
1161 if (C == Tree.end()) {
1162 size_t d = getDepthFactor(k->second.first);
1163 Q.push_back(ValuePairWithDepth(k->second, QTop.second+d));
1164 MoreChildren = true;
1166 MaxChildDepth = std::max(MaxChildDepth, C->second);
1171 if (!MoreChildren) {
1172 // Record the current pair as part of the Tree:
1173 Tree.insert(ValuePairWithDepth(QTop.first, MaxChildDepth));
1176 } while (!Q.empty());
1179 // Given some initial tree, prune it by removing conflicting pairs (pairs
1180 // that cannot be simultaneously chosen for vectorization).
1181 void BBVectorize::pruneTreeFor(
1182 std::multimap<Value *, Value *> &CandidatePairs,
1183 std::vector<Value *> &PairableInsts,
1184 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1185 DenseSet<ValuePair> &PairableInstUsers,
1186 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1187 DenseMap<Value *, Value *> &ChosenPairs,
1188 DenseMap<ValuePair, size_t> &Tree,
1189 DenseSet<ValuePair> &PrunedTree, ValuePair J,
1190 bool UseCycleCheck) {
1191 SmallVector<ValuePairWithDepth, 32> Q;
1192 // General depth-first post-order traversal:
1193 Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1195 ValuePairWithDepth QTop = Q.pop_back_val();
1196 PrunedTree.insert(QTop.first);
1198 // Visit each child, pruning as necessary...
1199 DenseMap<ValuePair, size_t> BestChildren;
1200 VPPIteratorPair QTopRange = ConnectedPairs.equal_range(QTop.first);
1201 for (std::multimap<ValuePair, ValuePair>::iterator K = QTopRange.first;
1202 K != QTopRange.second; ++K) {
1203 DenseMap<ValuePair, size_t>::iterator C = Tree.find(K->second);
1204 if (C == Tree.end()) continue;
1206 // This child is in the Tree, now we need to make sure it is the
1207 // best of any conflicting children. There could be multiple
1208 // conflicting children, so first, determine if we're keeping
1209 // this child, then delete conflicting children as necessary.
1211 // It is also necessary to guard against pairing-induced
1212 // dependencies. Consider instructions a .. x .. y .. b
1213 // such that (a,b) are to be fused and (x,y) are to be fused
1214 // but a is an input to x and b is an output from y. This
1215 // means that y cannot be moved after b but x must be moved
1216 // after b for (a,b) to be fused. In other words, after
1217 // fusing (a,b) we have y .. a/b .. x where y is an input
1218 // to a/b and x is an output to a/b: x and y can no longer
1219 // be legally fused. To prevent this condition, we must
1220 // make sure that a child pair added to the Tree is not
1221 // both an input and output of an already-selected pair.
1223 // Pairing-induced dependencies can also form from more complicated
1224 // cycles. The pair vs. pair conflicts are easy to check, and so
1225 // that is done explicitly for "fast rejection", and because for
1226 // child vs. child conflicts, we may prefer to keep the current
1227 // pair in preference to the already-selected child.
1228 DenseSet<ValuePair> CurrentPairs;
1231 for (DenseMap<ValuePair, size_t>::iterator C2
1232 = BestChildren.begin(), E2 = BestChildren.end();
1234 if (C2->first.first == C->first.first ||
1235 C2->first.first == C->first.second ||
1236 C2->first.second == C->first.first ||
1237 C2->first.second == C->first.second ||
1238 pairsConflict(C2->first, C->first, PairableInstUsers,
1239 UseCycleCheck ? &PairableInstUserMap : 0)) {
1240 if (C2->second >= C->second) {
1245 CurrentPairs.insert(C2->first);
1248 if (!CanAdd) continue;
1250 // Even worse, this child could conflict with another node already
1251 // selected for the Tree. If that is the case, ignore this child.
1252 for (DenseSet<ValuePair>::iterator T = PrunedTree.begin(),
1253 E2 = PrunedTree.end(); T != E2; ++T) {
1254 if (T->first == C->first.first ||
1255 T->first == C->first.second ||
1256 T->second == C->first.first ||
1257 T->second == C->first.second ||
1258 pairsConflict(*T, C->first, PairableInstUsers,
1259 UseCycleCheck ? &PairableInstUserMap : 0)) {
1264 CurrentPairs.insert(*T);
1266 if (!CanAdd) continue;
1268 // And check the queue too...
1269 for (SmallVector<ValuePairWithDepth, 32>::iterator C2 = Q.begin(),
1270 E2 = Q.end(); C2 != E2; ++C2) {
1271 if (C2->first.first == C->first.first ||
1272 C2->first.first == C->first.second ||
1273 C2->first.second == C->first.first ||
1274 C2->first.second == C->first.second ||
1275 pairsConflict(C2->first, C->first, PairableInstUsers,
1276 UseCycleCheck ? &PairableInstUserMap : 0)) {
1281 CurrentPairs.insert(C2->first);
1283 if (!CanAdd) continue;
1285 // Last but not least, check for a conflict with any of the
1286 // already-chosen pairs.
1287 for (DenseMap<Value *, Value *>::iterator C2 =
1288 ChosenPairs.begin(), E2 = ChosenPairs.end();
1290 if (pairsConflict(*C2, C->first, PairableInstUsers,
1291 UseCycleCheck ? &PairableInstUserMap : 0)) {
1296 CurrentPairs.insert(*C2);
1298 if (!CanAdd) continue;
1300 // To check for non-trivial cycles formed by the addition of the
1301 // current pair we've formed a list of all relevant pairs, now use a
1302 // graph walk to check for a cycle. We start from the current pair and
1303 // walk the use tree to see if we again reach the current pair. If we
1304 // do, then the current pair is rejected.
1306 // FIXME: It may be more efficient to use a topological-ordering
1307 // algorithm to improve the cycle check. This should be investigated.
1308 if (UseCycleCheck &&
1309 pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs))
1312 // This child can be added, but we may have chosen it in preference
1313 // to an already-selected child. Check for this here, and if a
1314 // conflict is found, then remove the previously-selected child
1315 // before adding this one in its place.
1316 for (DenseMap<ValuePair, size_t>::iterator C2
1317 = BestChildren.begin(); C2 != BestChildren.end();) {
1318 if (C2->first.first == C->first.first ||
1319 C2->first.first == C->first.second ||
1320 C2->first.second == C->first.first ||
1321 C2->first.second == C->first.second ||
1322 pairsConflict(C2->first, C->first, PairableInstUsers))
1323 BestChildren.erase(C2++);
1328 BestChildren.insert(ValuePairWithDepth(C->first, C->second));
1331 for (DenseMap<ValuePair, size_t>::iterator C
1332 = BestChildren.begin(), E2 = BestChildren.end();
1334 size_t DepthF = getDepthFactor(C->first.first);
1335 Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF));
1337 } while (!Q.empty());
1340 // This function finds the best tree of mututally-compatible connected
1341 // pairs, given the choice of root pairs as an iterator range.
1342 void BBVectorize::findBestTreeFor(
1343 std::multimap<Value *, Value *> &CandidatePairs,
1344 std::vector<Value *> &PairableInsts,
1345 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1346 DenseSet<ValuePair> &PairableInstUsers,
1347 std::multimap<ValuePair, ValuePair> &PairableInstUserMap,
1348 DenseMap<Value *, Value *> &ChosenPairs,
1349 DenseSet<ValuePair> &BestTree, size_t &BestMaxDepth,
1350 size_t &BestEffSize, VPIteratorPair ChoiceRange,
1351 bool UseCycleCheck) {
1352 for (std::multimap<Value *, Value *>::iterator J = ChoiceRange.first;
1353 J != ChoiceRange.second; ++J) {
1355 // Before going any further, make sure that this pair does not
1356 // conflict with any already-selected pairs (see comment below
1357 // near the Tree pruning for more details).
1358 DenseSet<ValuePair> ChosenPairSet;
1359 bool DoesConflict = false;
1360 for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(),
1361 E = ChosenPairs.end(); C != E; ++C) {
1362 if (pairsConflict(*C, *J, PairableInstUsers,
1363 UseCycleCheck ? &PairableInstUserMap : 0)) {
1364 DoesConflict = true;
1368 ChosenPairSet.insert(*C);
1370 if (DoesConflict) continue;
1372 if (UseCycleCheck &&
1373 pairWillFormCycle(*J, PairableInstUserMap, ChosenPairSet))
1376 DenseMap<ValuePair, size_t> Tree;
1377 buildInitialTreeFor(CandidatePairs, PairableInsts, ConnectedPairs,
1378 PairableInstUsers, ChosenPairs, Tree, *J);
1380 // Because we'll keep the child with the largest depth, the largest
1381 // depth is still the same in the unpruned Tree.
1382 size_t MaxDepth = Tree.lookup(*J);
1384 DEBUG(if (DebugPairSelection) dbgs() << "BBV: found Tree for pair {"
1385 << *J->first << " <-> " << *J->second << "} of depth " <<
1386 MaxDepth << " and size " << Tree.size() << "\n");
1388 // At this point the Tree has been constructed, but, may contain
1389 // contradictory children (meaning that different children of
1390 // some tree node may be attempting to fuse the same instruction).
1391 // So now we walk the tree again, in the case of a conflict,
1392 // keep only the child with the largest depth. To break a tie,
1393 // favor the first child.
1395 DenseSet<ValuePair> PrunedTree;
1396 pruneTreeFor(CandidatePairs, PairableInsts, ConnectedPairs,
1397 PairableInstUsers, PairableInstUserMap, ChosenPairs, Tree,
1398 PrunedTree, *J, UseCycleCheck);
1401 for (DenseSet<ValuePair>::iterator S = PrunedTree.begin(),
1402 E = PrunedTree.end(); S != E; ++S)
1403 EffSize += getDepthFactor(S->first);
1405 DEBUG(if (DebugPairSelection)
1406 dbgs() << "BBV: found pruned Tree for pair {"
1407 << *J->first << " <-> " << *J->second << "} of depth " <<
1408 MaxDepth << " and size " << PrunedTree.size() <<
1409 " (effective size: " << EffSize << ")\n");
1410 if (MaxDepth >= Config.ReqChainDepth && EffSize > BestEffSize) {
1411 BestMaxDepth = MaxDepth;
1412 BestEffSize = EffSize;
1413 BestTree = PrunedTree;
1418 // Given the list of candidate pairs, this function selects those
1419 // that will be fused into vector instructions.
1420 void BBVectorize::choosePairs(
1421 std::multimap<Value *, Value *> &CandidatePairs,
1422 std::vector<Value *> &PairableInsts,
1423 std::multimap<ValuePair, ValuePair> &ConnectedPairs,
1424 DenseSet<ValuePair> &PairableInstUsers,
1425 DenseMap<Value *, Value *>& ChosenPairs) {
1426 bool UseCycleCheck =
1427 CandidatePairs.size() <= Config.MaxCandPairsForCycleCheck;
1428 std::multimap<ValuePair, ValuePair> PairableInstUserMap;
1429 for (std::vector<Value *>::iterator I = PairableInsts.begin(),
1430 E = PairableInsts.end(); I != E; ++I) {
1431 // The number of possible pairings for this variable:
1432 size_t NumChoices = CandidatePairs.count(*I);
1433 if (!NumChoices) continue;
1435 VPIteratorPair ChoiceRange = CandidatePairs.equal_range(*I);
1437 // The best pair to choose and its tree:
1438 size_t BestMaxDepth = 0, BestEffSize = 0;
1439 DenseSet<ValuePair> BestTree;
1440 findBestTreeFor(CandidatePairs, PairableInsts, ConnectedPairs,
1441 PairableInstUsers, PairableInstUserMap, ChosenPairs,
1442 BestTree, BestMaxDepth, BestEffSize, ChoiceRange,
1445 // A tree has been chosen (or not) at this point. If no tree was
1446 // chosen, then this instruction, I, cannot be paired (and is no longer
1449 DEBUG(if (BestTree.size() > 0)
1450 dbgs() << "BBV: selected pairs in the best tree for: "
1451 << *cast<Instruction>(*I) << "\n");
1453 for (DenseSet<ValuePair>::iterator S = BestTree.begin(),
1454 SE2 = BestTree.end(); S != SE2; ++S) {
1455 // Insert the members of this tree into the list of chosen pairs.
1456 ChosenPairs.insert(ValuePair(S->first, S->second));
1457 DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " <<
1458 *S->second << "\n");
1460 // Remove all candidate pairs that have values in the chosen tree.
1461 for (std::multimap<Value *, Value *>::iterator K =
1462 CandidatePairs.begin(); K != CandidatePairs.end();) {
1463 if (K->first == S->first || K->second == S->first ||
1464 K->second == S->second || K->first == S->second) {
1465 // Don't remove the actual pair chosen so that it can be used
1466 // in subsequent tree selections.
1467 if (!(K->first == S->first && K->second == S->second))
1468 CandidatePairs.erase(K++);
1478 DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n");
1481 std::string getReplacementName(Instruction *I, bool IsInput, unsigned o,
1486 return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) +
1487 (n > 0 ? "." + utostr(n) : "")).str();
1490 // Returns the value that is to be used as the pointer input to the vector
1491 // instruction that fuses I with J.
1492 Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context,
1493 Instruction *I, Instruction *J, unsigned o,
1494 bool FlipMemInputs) {
1496 unsigned IAlignment, JAlignment;
1497 int64_t OffsetInElmts;
1499 // Note: the analysis might fail here, that is why FlipMemInputs has
1500 // been precomputed (OffsetInElmts must be unused here).
1501 (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
1504 // The pointer value is taken to be the one with the lowest offset.
1506 if (!FlipMemInputs) {
1512 Type *ArgTypeI = cast<PointerType>(IPtr->getType())->getElementType();
1513 Type *ArgTypeJ = cast<PointerType>(JPtr->getType())->getElementType();
1514 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
1515 Type *VArgPtrType = PointerType::get(VArgType,
1516 cast<PointerType>(IPtr->getType())->getAddressSpace());
1517 return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o),
1518 /* insert before */ FlipMemInputs ? J : I);
1521 void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J,
1522 unsigned MaskOffset, unsigned NumInElem,
1523 unsigned NumInElem1, unsigned IdxOffset,
1524 std::vector<Constant*> &Mask) {
1525 unsigned NumElem1 = cast<VectorType>(J->getType())->getNumElements();
1526 for (unsigned v = 0; v < NumElem1; ++v) {
1527 int m = cast<ShuffleVectorInst>(J)->getMaskValue(v);
1529 Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context));
1531 unsigned mm = m + (int) IdxOffset;
1532 if (m >= (int) NumInElem1)
1533 mm += (int) NumInElem;
1535 Mask[v+MaskOffset] =
1536 ConstantInt::get(Type::getInt32Ty(Context), mm);
1541 // Returns the value that is to be used as the vector-shuffle mask to the
1542 // vector instruction that fuses I with J.
1543 Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context,
1544 Instruction *I, Instruction *J) {
1545 // This is the shuffle mask. We need to append the second
1546 // mask to the first, and the numbers need to be adjusted.
1548 Type *ArgTypeI = I->getType();
1549 Type *ArgTypeJ = J->getType();
1550 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
1552 unsigned NumElemI = cast<VectorType>(ArgTypeI)->getNumElements();
1554 // Get the total number of elements in the fused vector type.
1555 // By definition, this must equal the number of elements in
1557 unsigned NumElem = cast<VectorType>(VArgType)->getNumElements();
1558 std::vector<Constant*> Mask(NumElem);
1560 Type *OpTypeI = I->getOperand(0)->getType();
1561 unsigned NumInElemI = cast<VectorType>(OpTypeI)->getNumElements();
1562 Type *OpTypeJ = J->getOperand(0)->getType();
1563 unsigned NumInElemJ = cast<VectorType>(OpTypeJ)->getNumElements();
1565 // The fused vector will be:
1566 // -----------------------------------------------------
1567 // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ |
1568 // -----------------------------------------------------
1569 // from which we'll extract NumElem total elements (where the first NumElemI
1570 // of them come from the mask in I and the remainder come from the mask
1573 // For the mask from the first pair...
1574 fillNewShuffleMask(Context, I, 0, NumInElemJ, NumInElemI,
1577 // For the mask from the second pair...
1578 fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ,
1581 return ConstantVector::get(Mask);
1584 bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I,
1585 Instruction *J, unsigned o, Value *&LOp,
1587 Type *ArgTypeL, Type *ArgTypeH,
1589 bool ExpandedIEChain = false;
1590 if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) {
1591 // If we have a pure insertelement chain, then this can be rewritten
1592 // into a chain that directly builds the larger type.
1593 bool PureChain = true;
1594 InsertElementInst *LIENext = LIE;
1596 if (!isa<UndefValue>(LIENext->getOperand(0)) &&
1597 !isa<InsertElementInst>(LIENext->getOperand(0))) {
1602 dyn_cast<InsertElementInst>(LIENext->getOperand(0))));
1605 SmallVector<Value *, 8> VectElemts(numElemL,
1606 UndefValue::get(ArgTypeL->getScalarType()));
1607 InsertElementInst *LIENext = LIE;
1610 cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue();
1611 VectElemts[Idx] = LIENext->getOperand(1);
1613 dyn_cast<InsertElementInst>(LIENext->getOperand(0))));
1616 Value *LIEPrev = UndefValue::get(ArgTypeH);
1617 for (unsigned i = 0; i < numElemL; ++i) {
1618 if (isa<UndefValue>(VectElemts[i])) continue;
1619 LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i],
1620 ConstantInt::get(Type::getInt32Ty(Context),
1622 getReplacementName(I, true, o, i+1));
1623 LIENext->insertBefore(J);
1627 LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH);
1628 ExpandedIEChain = true;
1632 return ExpandedIEChain;
1635 // Returns the value to be used as the specified operand of the vector
1636 // instruction that fuses I with J.
1637 Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I,
1638 Instruction *J, unsigned o, bool FlipMemInputs) {
1639 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
1640 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1);
1642 // Compute the fused vector type for this operand
1643 Type *ArgTypeI = I->getOperand(o)->getType();
1644 Type *ArgTypeJ = J->getOperand(o)->getType();
1645 VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
1647 Instruction *L = I, *H = J;
1648 Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ;
1649 if (FlipMemInputs) {
1652 ArgTypeL = ArgTypeJ;
1653 ArgTypeH = ArgTypeI;
1657 if (ArgTypeL->isVectorTy())
1658 numElemL = cast<VectorType>(ArgTypeL)->getNumElements();
1663 if (ArgTypeH->isVectorTy())
1664 numElemH = cast<VectorType>(ArgTypeH)->getNumElements();
1668 Value *LOp = L->getOperand(o);
1669 Value *HOp = H->getOperand(o);
1670 unsigned numElem = VArgType->getNumElements();
1672 // First, we check if we can reuse the "original" vector outputs (if these
1673 // exist). We might need a shuffle.
1674 ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp);
1675 ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp);
1676 ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp);
1677 ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp);
1679 // FIXME: If we're fusing shuffle instructions, then we can't apply this
1680 // optimization. The input vectors to the shuffle might be a different
1681 // length from the shuffle outputs. Unfortunately, the replacement
1682 // shuffle mask has already been formed, and the mask entries are sensitive
1683 // to the sizes of the inputs.
1684 bool IsSizeChangeShuffle =
1685 isa<ShuffleVectorInst>(L) &&
1686 (LOp->getType() != L->getType() || HOp->getType() != H->getType());
1688 if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) {
1689 // We can have at most two unique vector inputs.
1690 bool CanUseInputs = true;
1693 I1 = LEE->getOperand(0);
1695 I1 = LSV->getOperand(0);
1696 I2 = LSV->getOperand(1);
1697 if (I2 == I1 || isa<UndefValue>(I2))
1702 Value *I3 = HEE->getOperand(0);
1703 if (!I2 && I3 != I1)
1705 else if (I3 != I1 && I3 != I2)
1706 CanUseInputs = false;
1708 Value *I3 = HSV->getOperand(0);
1709 if (!I2 && I3 != I1)
1711 else if (I3 != I1 && I3 != I2)
1712 CanUseInputs = false;
1715 Value *I4 = HSV->getOperand(1);
1716 if (!isa<UndefValue>(I4)) {
1717 if (!I2 && I4 != I1)
1719 else if (I4 != I1 && I4 != I2)
1720 CanUseInputs = false;
1727 cast<VectorType>(cast<Instruction>(LOp)->getOperand(0)->getType())
1730 cast<VectorType>(cast<Instruction>(HOp)->getOperand(0)->getType())
1733 // We have one or two input vectors. We need to map each index of the
1734 // operands to the index of the original vector.
1735 SmallVector<std::pair<int, int>, 8> II(numElem);
1736 for (unsigned i = 0; i < numElemL; ++i) {
1740 cast<ConstantInt>(LEE->getOperand(1))->getSExtValue();
1741 INum = LEE->getOperand(0) == I1 ? 0 : 1;
1743 Idx = LSV->getMaskValue(i);
1744 if (Idx < (int) LOpElem) {
1745 INum = LSV->getOperand(0) == I1 ? 0 : 1;
1748 INum = LSV->getOperand(1) == I1 ? 0 : 1;
1752 II[i] = std::pair<int, int>(Idx, INum);
1754 for (unsigned i = 0; i < numElemH; ++i) {
1758 cast<ConstantInt>(HEE->getOperand(1))->getSExtValue();
1759 INum = HEE->getOperand(0) == I1 ? 0 : 1;
1761 Idx = HSV->getMaskValue(i);
1762 if (Idx < (int) HOpElem) {
1763 INum = HSV->getOperand(0) == I1 ? 0 : 1;
1766 INum = HSV->getOperand(1) == I1 ? 0 : 1;
1770 II[i + numElemL] = std::pair<int, int>(Idx, INum);
1773 // We now have an array which tells us from which index of which
1774 // input vector each element of the operand comes.
1775 VectorType *I1T = cast<VectorType>(I1->getType());
1776 unsigned I1Elem = I1T->getNumElements();
1779 // In this case there is only one underlying vector input. Check for
1780 // the trivial case where we can use the input directly.
1781 if (I1Elem == numElem) {
1782 bool ElemInOrder = true;
1783 for (unsigned i = 0; i < numElem; ++i) {
1784 if (II[i].first != (int) i && II[i].first != -1) {
1785 ElemInOrder = false;
1794 // A shuffle is needed.
1795 std::vector<Constant *> Mask(numElem);
1796 for (unsigned i = 0; i < numElem; ++i) {
1797 int Idx = II[i].first;
1799 Mask[i] = UndefValue::get(Type::getInt32Ty(Context));
1801 Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
1805 new ShuffleVectorInst(I1, UndefValue::get(I1T),
1806 ConstantVector::get(Mask),
1807 getReplacementName(I, true, o));
1812 VectorType *I2T = cast<VectorType>(I2->getType());
1813 unsigned I2Elem = I2T->getNumElements();
1815 // This input comes from two distinct vectors. The first step is to
1816 // make sure that both vectors are the same length. If not, the
1817 // smaller one will need to grow before they can be shuffled together.
1818 if (I1Elem < I2Elem) {
1819 std::vector<Constant *> Mask(I2Elem);
1821 for (; v < I1Elem; ++v)
1822 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
1823 for (; v < I2Elem; ++v)
1824 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
1826 Instruction *NewI1 =
1827 new ShuffleVectorInst(I1, UndefValue::get(I1T),
1828 ConstantVector::get(Mask),
1829 getReplacementName(I, true, o, 1));
1830 NewI1->insertBefore(J);
1834 } else if (I1Elem > I2Elem) {
1835 std::vector<Constant *> Mask(I1Elem);
1837 for (; v < I2Elem; ++v)
1838 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
1839 for (; v < I1Elem; ++v)
1840 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
1842 Instruction *NewI2 =
1843 new ShuffleVectorInst(I2, UndefValue::get(I2T),
1844 ConstantVector::get(Mask),
1845 getReplacementName(I, true, o, 1));
1846 NewI2->insertBefore(J);
1852 // Now that both I1 and I2 are the same length we can shuffle them
1853 // together (and use the result).
1854 std::vector<Constant *> Mask(numElem);
1855 for (unsigned v = 0; v < numElem; ++v) {
1856 if (II[v].first == -1) {
1857 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
1859 int Idx = II[v].first + II[v].second * I1Elem;
1860 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
1864 Instruction *NewOp =
1865 new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask),
1866 getReplacementName(I, true, o));
1867 NewOp->insertBefore(J);
1872 Type *ArgType = ArgTypeL;
1873 if (numElemL < numElemH) {
1874 if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH,
1875 ArgTypeL, VArgType, 1)) {
1876 // This is another short-circuit case: we're combining a scalar into
1877 // a vector that is formed by an IE chain. We've just expanded the IE
1878 // chain, now insert the scalar and we're done.
1880 Instruction *S = InsertElementInst::Create(HOp, LOp, CV0,
1881 getReplacementName(I, true, o));
1884 } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL,
1886 // The two vector inputs to the shuffle must be the same length,
1887 // so extend the smaller vector to be the same length as the larger one.
1891 std::vector<Constant *> Mask(numElemH);
1893 for (; v < numElemL; ++v)
1894 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
1895 for (; v < numElemH; ++v)
1896 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
1898 NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL),
1899 ConstantVector::get(Mask),
1900 getReplacementName(I, true, o, 1));
1902 NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0,
1903 getReplacementName(I, true, o, 1));
1906 NLOp->insertBefore(J);
1911 } else if (numElemL > numElemH) {
1912 if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL,
1913 ArgTypeH, VArgType)) {
1915 InsertElementInst::Create(LOp, HOp,
1916 ConstantInt::get(Type::getInt32Ty(Context),
1918 getReplacementName(I, true, o));
1921 } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH,
1925 std::vector<Constant *> Mask(numElemL);
1927 for (; v < numElemH; ++v)
1928 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
1929 for (; v < numElemL; ++v)
1930 Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
1932 NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH),
1933 ConstantVector::get(Mask),
1934 getReplacementName(I, true, o, 1));
1936 NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0,
1937 getReplacementName(I, true, o, 1));
1940 NHOp->insertBefore(J);
1945 if (ArgType->isVectorTy()) {
1946 unsigned numElem = cast<VectorType>(VArgType)->getNumElements();
1947 std::vector<Constant*> Mask(numElem);
1948 for (unsigned v = 0; v < numElem; ++v) {
1950 // If the low vector was expanded, we need to skip the extra
1951 // undefined entries.
1952 if (v >= numElemL && numElemH > numElemL)
1953 Idx += (numElemH - numElemL);
1954 Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
1957 Instruction *BV = new ShuffleVectorInst(LOp, HOp,
1958 ConstantVector::get(Mask),
1959 getReplacementName(I, true, o));
1960 BV->insertBefore(J);
1964 Instruction *BV1 = InsertElementInst::Create(
1965 UndefValue::get(VArgType), LOp, CV0,
1966 getReplacementName(I, true, o, 1));
1967 BV1->insertBefore(I);
1968 Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1,
1969 getReplacementName(I, true, o, 2));
1970 BV2->insertBefore(J);
1974 // This function creates an array of values that will be used as the inputs
1975 // to the vector instruction that fuses I with J.
1976 void BBVectorize::getReplacementInputsForPair(LLVMContext& Context,
1977 Instruction *I, Instruction *J,
1978 SmallVector<Value *, 3> &ReplacedOperands,
1979 bool FlipMemInputs) {
1980 unsigned NumOperands = I->getNumOperands();
1982 for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) {
1983 // Iterate backward so that we look at the store pointer
1984 // first and know whether or not we need to flip the inputs.
1986 if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) {
1987 // This is the pointer for a load/store instruction.
1988 ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o,
1991 } else if (isa<CallInst>(I)) {
1992 Function *F = cast<CallInst>(I)->getCalledFunction();
1993 unsigned IID = F->getIntrinsicID();
1994 if (o == NumOperands-1) {
1995 BasicBlock &BB = *I->getParent();
1997 Module *M = BB.getParent()->getParent();
1998 Type *ArgTypeI = I->getType();
1999 Type *ArgTypeJ = J->getType();
2000 Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2002 ReplacedOperands[o] = Intrinsic::getDeclaration(M,
2003 (Intrinsic::ID) IID, VArgType);
2005 } else if (IID == Intrinsic::powi && o == 1) {
2006 // The second argument of powi is a single integer and we've already
2007 // checked that both arguments are equal. As a result, we just keep
2008 // I's second argument.
2009 ReplacedOperands[o] = I->getOperand(o);
2012 } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) {
2013 ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J);
2017 ReplacedOperands[o] =
2018 getReplacementInput(Context, I, J, o, FlipMemInputs);
2022 // This function creates two values that represent the outputs of the
2023 // original I and J instructions. These are generally vector shuffles
2024 // or extracts. In many cases, these will end up being unused and, thus,
2025 // eliminated by later passes.
2026 void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
2027 Instruction *J, Instruction *K,
2028 Instruction *&InsertionPt,
2029 Instruction *&K1, Instruction *&K2,
2030 bool FlipMemInputs) {
2031 if (isa<StoreInst>(I)) {
2032 AA->replaceWithNewValue(I, K);
2033 AA->replaceWithNewValue(J, K);
2035 Type *IType = I->getType();
2036 Type *JType = J->getType();
2038 VectorType *VType = getVecTypeForPair(IType, JType);
2039 unsigned numElem = VType->getNumElements();
2041 unsigned numElemI, numElemJ;
2042 if (IType->isVectorTy())
2043 numElemI = cast<VectorType>(IType)->getNumElements();
2047 if (JType->isVectorTy())
2048 numElemJ = cast<VectorType>(JType)->getNumElements();
2052 if (IType->isVectorTy()) {
2053 std::vector<Constant*> Mask1(numElemI), Mask2(numElemI);
2054 for (unsigned v = 0; v < numElemI; ++v) {
2055 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2056 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v);
2059 K1 = new ShuffleVectorInst(K, UndefValue::get(VType),
2060 ConstantVector::get(
2061 FlipMemInputs ? Mask2 : Mask1),
2062 getReplacementName(K, false, 1));
2064 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2065 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1);
2066 K1 = ExtractElementInst::Create(K, FlipMemInputs ? CV1 : CV0,
2067 getReplacementName(K, false, 1));
2070 if (JType->isVectorTy()) {
2071 std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ);
2072 for (unsigned v = 0; v < numElemJ; ++v) {
2073 Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2074 Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v);
2077 K2 = new ShuffleVectorInst(K, UndefValue::get(VType),
2078 ConstantVector::get(
2079 FlipMemInputs ? Mask1 : Mask2),
2080 getReplacementName(K, false, 2));
2082 Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2083 Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1);
2084 K2 = ExtractElementInst::Create(K, FlipMemInputs ? CV0 : CV1,
2085 getReplacementName(K, false, 2));
2089 K2->insertAfter(K1);
2094 // Move all uses of the function I (including pairing-induced uses) after J.
2095 bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB,
2096 std::multimap<Value *, Value *> &LoadMoveSet,
2097 Instruction *I, Instruction *J) {
2098 // Skip to the first instruction past I.
2099 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2101 DenseSet<Value *> Users;
2102 AliasSetTracker WriteSet(*AA);
2103 for (; cast<Instruction>(L) != J; ++L)
2104 (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet);
2106 assert(cast<Instruction>(L) == J &&
2107 "Tracking has not proceeded far enough to check for dependencies");
2108 // If J is now in the use set of I, then trackUsesOfI will return true
2109 // and we have a dependency cycle (and the fusing operation must abort).
2110 return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSet);
2113 // Move all uses of the function I (including pairing-induced uses) after J.
2114 void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB,
2115 std::multimap<Value *, Value *> &LoadMoveSet,
2116 Instruction *&InsertionPt,
2117 Instruction *I, Instruction *J) {
2118 // Skip to the first instruction past I.
2119 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2121 DenseSet<Value *> Users;
2122 AliasSetTracker WriteSet(*AA);
2123 for (; cast<Instruction>(L) != J;) {
2124 if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSet)) {
2125 // Move this instruction
2126 Instruction *InstToMove = L; ++L;
2128 DEBUG(dbgs() << "BBV: moving: " << *InstToMove <<
2129 " to after " << *InsertionPt << "\n");
2130 InstToMove->removeFromParent();
2131 InstToMove->insertAfter(InsertionPt);
2132 InsertionPt = InstToMove;
2139 // Collect all load instruction that are in the move set of a given first
2140 // pair member. These loads depend on the first instruction, I, and so need
2141 // to be moved after J (the second instruction) when the pair is fused.
2142 void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB,
2143 DenseMap<Value *, Value *> &ChosenPairs,
2144 std::multimap<Value *, Value *> &LoadMoveSet,
2146 // Skip to the first instruction past I.
2147 BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
2149 DenseSet<Value *> Users;
2150 AliasSetTracker WriteSet(*AA);
2152 // Note: We cannot end the loop when we reach J because J could be moved
2153 // farther down the use chain by another instruction pairing. Also, J
2154 // could be before I if this is an inverted input.
2155 for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) {
2156 if (trackUsesOfI(Users, WriteSet, I, L)) {
2157 if (L->mayReadFromMemory())
2158 LoadMoveSet.insert(ValuePair(L, I));
2163 // In cases where both load/stores and the computation of their pointers
2164 // are chosen for vectorization, we can end up in a situation where the
2165 // aliasing analysis starts returning different query results as the
2166 // process of fusing instruction pairs continues. Because the algorithm
2167 // relies on finding the same use trees here as were found earlier, we'll
2168 // need to precompute the necessary aliasing information here and then
2169 // manually update it during the fusion process.
2170 void BBVectorize::collectLoadMoveSet(BasicBlock &BB,
2171 std::vector<Value *> &PairableInsts,
2172 DenseMap<Value *, Value *> &ChosenPairs,
2173 std::multimap<Value *, Value *> &LoadMoveSet) {
2174 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
2175 PIE = PairableInsts.end(); PI != PIE; ++PI) {
2176 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI);
2177 if (P == ChosenPairs.end()) continue;
2179 Instruction *I = cast<Instruction>(P->first);
2180 collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet, I);
2184 // As with the aliasing information, SCEV can also change because of
2185 // vectorization. This information is used to compute relative pointer
2186 // offsets; the necessary information will be cached here prior to
2188 void BBVectorize::collectPtrInfo(std::vector<Value *> &PairableInsts,
2189 DenseMap<Value *, Value *> &ChosenPairs,
2190 DenseSet<Value *> &LowPtrInsts) {
2191 for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
2192 PIE = PairableInsts.end(); PI != PIE; ++PI) {
2193 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI);
2194 if (P == ChosenPairs.end()) continue;
2196 Instruction *I = cast<Instruction>(P->first);
2197 Instruction *J = cast<Instruction>(P->second);
2199 if (!isa<LoadInst>(I) && !isa<StoreInst>(I))
2203 unsigned IAlignment, JAlignment;
2204 int64_t OffsetInElmts;
2205 if (!getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
2206 OffsetInElmts) || abs64(OffsetInElmts) != 1)
2207 llvm_unreachable("Pre-fusion pointer analysis failed");
2209 Value *LowPI = (OffsetInElmts > 0) ? I : J;
2210 LowPtrInsts.insert(LowPI);
2214 // When the first instruction in each pair is cloned, it will inherit its
2215 // parent's metadata. This metadata must be combined with that of the other
2216 // instruction in a safe way.
2217 void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) {
2218 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata;
2219 K->getAllMetadataOtherThanDebugLoc(Metadata);
2220 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) {
2221 unsigned Kind = Metadata[i].first;
2222 MDNode *JMD = J->getMetadata(Kind);
2223 MDNode *KMD = Metadata[i].second;
2227 K->setMetadata(Kind, 0); // Remove unknown metadata
2229 case LLVMContext::MD_tbaa:
2230 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD));
2232 case LLVMContext::MD_fpmath:
2233 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD));
2239 // This function fuses the chosen instruction pairs into vector instructions,
2240 // taking care preserve any needed scalar outputs and, then, it reorders the
2241 // remaining instructions as needed (users of the first member of the pair
2242 // need to be moved to after the location of the second member of the pair
2243 // because the vector instruction is inserted in the location of the pair's
2245 void BBVectorize::fuseChosenPairs(BasicBlock &BB,
2246 std::vector<Value *> &PairableInsts,
2247 DenseMap<Value *, Value *> &ChosenPairs) {
2248 LLVMContext& Context = BB.getContext();
2250 // During the vectorization process, the order of the pairs to be fused
2251 // could be flipped. So we'll add each pair, flipped, into the ChosenPairs
2252 // list. After a pair is fused, the flipped pair is removed from the list.
2253 std::vector<ValuePair> FlippedPairs;
2254 FlippedPairs.reserve(ChosenPairs.size());
2255 for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(),
2256 E = ChosenPairs.end(); P != E; ++P)
2257 FlippedPairs.push_back(ValuePair(P->second, P->first));
2258 for (std::vector<ValuePair>::iterator P = FlippedPairs.begin(),
2259 E = FlippedPairs.end(); P != E; ++P)
2260 ChosenPairs.insert(*P);
2262 std::multimap<Value *, Value *> LoadMoveSet;
2263 collectLoadMoveSet(BB, PairableInsts, ChosenPairs, LoadMoveSet);
2265 DenseSet<Value *> LowPtrInsts;
2266 collectPtrInfo(PairableInsts, ChosenPairs, LowPtrInsts);
2268 DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n");
2270 for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) {
2271 DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI);
2272 if (P == ChosenPairs.end()) {
2277 if (getDepthFactor(P->first) == 0) {
2278 // These instructions are not really fused, but are tracked as though
2279 // they are. Any case in which it would be interesting to fuse them
2280 // will be taken care of by InstCombine.
2286 Instruction *I = cast<Instruction>(P->first),
2287 *J = cast<Instruction>(P->second);
2289 DEBUG(dbgs() << "BBV: fusing: " << *I <<
2290 " <-> " << *J << "\n");
2292 // Remove the pair and flipped pair from the list.
2293 DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second);
2294 assert(FP != ChosenPairs.end() && "Flipped pair not found in list");
2295 ChosenPairs.erase(FP);
2296 ChosenPairs.erase(P);
2298 if (!canMoveUsesOfIAfterJ(BB, LoadMoveSet, I, J)) {
2299 DEBUG(dbgs() << "BBV: fusion of: " << *I <<
2301 " aborted because of non-trivial dependency cycle\n");
2307 bool FlipMemInputs = false;
2308 if (isa<LoadInst>(I) || isa<StoreInst>(I))
2309 FlipMemInputs = (LowPtrInsts.find(I) == LowPtrInsts.end());
2311 unsigned NumOperands = I->getNumOperands();
2312 SmallVector<Value *, 3> ReplacedOperands(NumOperands);
2313 getReplacementInputsForPair(Context, I, J, ReplacedOperands,
2316 // Make a copy of the original operation, change its type to the vector
2317 // type and replace its operands with the vector operands.
2318 Instruction *K = I->clone();
2319 if (I->hasName()) K->takeName(I);
2321 if (!isa<StoreInst>(K))
2322 K->mutateType(getVecTypeForPair(I->getType(), J->getType()));
2324 combineMetadata(K, J);
2326 for (unsigned o = 0; o < NumOperands; ++o)
2327 K->setOperand(o, ReplacedOperands[o]);
2329 // If we've flipped the memory inputs, make sure that we take the correct
2331 if (FlipMemInputs) {
2332 if (isa<StoreInst>(K))
2333 cast<StoreInst>(K)->setAlignment(cast<StoreInst>(J)->getAlignment());
2335 cast<LoadInst>(K)->setAlignment(cast<LoadInst>(J)->getAlignment());
2340 // Instruction insertion point:
2341 Instruction *InsertionPt = K;
2342 Instruction *K1 = 0, *K2 = 0;
2343 replaceOutputsOfPair(Context, I, J, K, InsertionPt, K1, K2,
2346 // The use tree of the first original instruction must be moved to after
2347 // the location of the second instruction. The entire use tree of the
2348 // first instruction is disjoint from the input tree of the second
2349 // (by definition), and so commutes with it.
2351 moveUsesOfIAfterJ(BB, LoadMoveSet, InsertionPt, I, J);
2353 if (!isa<StoreInst>(I)) {
2354 I->replaceAllUsesWith(K1);
2355 J->replaceAllUsesWith(K2);
2356 AA->replaceWithNewValue(I, K1);
2357 AA->replaceWithNewValue(J, K2);
2360 // Instructions that may read from memory may be in the load move set.
2361 // Once an instruction is fused, we no longer need its move set, and so
2362 // the values of the map never need to be updated. However, when a load
2363 // is fused, we need to merge the entries from both instructions in the
2364 // pair in case those instructions were in the move set of some other
2365 // yet-to-be-fused pair. The loads in question are the keys of the map.
2366 if (I->mayReadFromMemory()) {
2367 std::vector<ValuePair> NewSetMembers;
2368 VPIteratorPair IPairRange = LoadMoveSet.equal_range(I);
2369 VPIteratorPair JPairRange = LoadMoveSet.equal_range(J);
2370 for (std::multimap<Value *, Value *>::iterator N = IPairRange.first;
2371 N != IPairRange.second; ++N)
2372 NewSetMembers.push_back(ValuePair(K, N->second));
2373 for (std::multimap<Value *, Value *>::iterator N = JPairRange.first;
2374 N != JPairRange.second; ++N)
2375 NewSetMembers.push_back(ValuePair(K, N->second));
2376 for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(),
2377 AE = NewSetMembers.end(); A != AE; ++A)
2378 LoadMoveSet.insert(*A);
2381 // Before removing I, set the iterator to the next instruction.
2382 PI = llvm::next(BasicBlock::iterator(I));
2383 if (cast<Instruction>(PI) == J)
2388 I->eraseFromParent();
2389 J->eraseFromParent();
2392 DEBUG(dbgs() << "BBV: final: \n" << BB << "\n");
2396 char BBVectorize::ID = 0;
2397 static const char bb_vectorize_name[] = "Basic-Block Vectorization";
2398 INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
2399 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
2400 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
2401 INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
2403 BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) {
2404 return new BBVectorize(C);
2408 llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) {
2409 BBVectorize BBVectorizer(P, C);
2410 return BBVectorizer.vectorizeBB(BB);
2413 //===----------------------------------------------------------------------===//
2414 VectorizeConfig::VectorizeConfig() {
2415 VectorBits = ::VectorBits;
2416 VectorizeBools = !::NoBools;
2417 VectorizeInts = !::NoInts;
2418 VectorizeFloats = !::NoFloats;
2419 VectorizePointers = !::NoPointers;
2420 VectorizeCasts = !::NoCasts;
2421 VectorizeMath = !::NoMath;
2422 VectorizeFMA = !::NoFMA;
2423 VectorizeSelect = !::NoSelect;
2424 VectorizeCmp = !::NoCmp;
2425 VectorizeGEP = !::NoGEP;
2426 VectorizeMemOps = !::NoMemOps;
2427 AlignedOnly = ::AlignedOnly;
2428 ReqChainDepth= ::ReqChainDepth;
2429 SearchLimit = ::SearchLimit;
2430 MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck;
2431 SplatBreaksChain = ::SplatBreaksChain;
2432 MaxInsts = ::MaxInsts;
2433 MaxIter = ::MaxIter;
2434 Pow2LenOnly = ::Pow2LenOnly;
2435 NoMemOpBoost = ::NoMemOpBoost;
2436 FastDep = ::FastDep;