• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===- BBVectorize.cpp - A Basic-Block Vectorizer -------------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
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
13 // pairing them.
14 //
15 //===----------------------------------------------------------------------===//
16 
17 #define BBV_NAME "bb-vectorize"
18 #include "llvm/Transforms/Vectorize.h"
19 #include "llvm/ADT/DenseMap.h"
20 #include "llvm/ADT/DenseSet.h"
21 #include "llvm/ADT/STLExtras.h"
22 #include "llvm/ADT/SmallSet.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/ADT/StringExtras.h"
26 #include "llvm/Analysis/AliasAnalysis.h"
27 #include "llvm/Analysis/AliasSetTracker.h"
28 #include "llvm/Analysis/GlobalsModRef.h"
29 #include "llvm/Analysis/ScalarEvolution.h"
30 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
31 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
32 #include "llvm/Analysis/TargetLibraryInfo.h"
33 #include "llvm/Analysis/TargetTransformInfo.h"
34 #include "llvm/Analysis/ValueTracking.h"
35 #include "llvm/IR/Constants.h"
36 #include "llvm/IR/DataLayout.h"
37 #include "llvm/IR/DerivedTypes.h"
38 #include "llvm/IR/Dominators.h"
39 #include "llvm/IR/Function.h"
40 #include "llvm/IR/Instructions.h"
41 #include "llvm/IR/IntrinsicInst.h"
42 #include "llvm/IR/Intrinsics.h"
43 #include "llvm/IR/LLVMContext.h"
44 #include "llvm/IR/Metadata.h"
45 #include "llvm/IR/Module.h"
46 #include "llvm/IR/Type.h"
47 #include "llvm/IR/ValueHandle.h"
48 #include "llvm/Pass.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/Support/raw_ostream.h"
52 #include "llvm/Transforms/Utils/Local.h"
53 #include <algorithm>
54 using namespace llvm;
55 
56 #define DEBUG_TYPE BBV_NAME
57 
58 static cl::opt<bool>
59 IgnoreTargetInfo("bb-vectorize-ignore-target-info",  cl::init(false),
60   cl::Hidden, cl::desc("Ignore target information"));
61 
62 static cl::opt<unsigned>
63 ReqChainDepth("bb-vectorize-req-chain-depth", cl::init(6), cl::Hidden,
64   cl::desc("The required chain depth for vectorization"));
65 
66 static cl::opt<bool>
67 UseChainDepthWithTI("bb-vectorize-use-chain-depth",  cl::init(false),
68   cl::Hidden, cl::desc("Use the chain depth requirement with"
69                        " target information"));
70 
71 static cl::opt<unsigned>
72 SearchLimit("bb-vectorize-search-limit", cl::init(400), cl::Hidden,
73   cl::desc("The maximum search distance for instruction pairs"));
74 
75 static cl::opt<bool>
76 SplatBreaksChain("bb-vectorize-splat-breaks-chain", cl::init(false), cl::Hidden,
77   cl::desc("Replicating one element to a pair breaks the chain"));
78 
79 static cl::opt<unsigned>
80 VectorBits("bb-vectorize-vector-bits", cl::init(128), cl::Hidden,
81   cl::desc("The size of the native vector registers"));
82 
83 static cl::opt<unsigned>
84 MaxIter("bb-vectorize-max-iter", cl::init(0), cl::Hidden,
85   cl::desc("The maximum number of pairing iterations"));
86 
87 static cl::opt<bool>
88 Pow2LenOnly("bb-vectorize-pow2-len-only", cl::init(false), cl::Hidden,
89   cl::desc("Don't try to form non-2^n-length vectors"));
90 
91 static cl::opt<unsigned>
92 MaxInsts("bb-vectorize-max-instr-per-group", cl::init(500), cl::Hidden,
93   cl::desc("The maximum number of pairable instructions per group"));
94 
95 static cl::opt<unsigned>
96 MaxPairs("bb-vectorize-max-pairs-per-group", cl::init(3000), cl::Hidden,
97   cl::desc("The maximum number of candidate instruction pairs per group"));
98 
99 static cl::opt<unsigned>
100 MaxCandPairsForCycleCheck("bb-vectorize-max-cycle-check-pairs", cl::init(200),
101   cl::Hidden, cl::desc("The maximum number of candidate pairs with which to use"
102                        " a full cycle check"));
103 
104 static cl::opt<bool>
105 NoBools("bb-vectorize-no-bools", cl::init(false), cl::Hidden,
106   cl::desc("Don't try to vectorize boolean (i1) values"));
107 
108 static cl::opt<bool>
109 NoInts("bb-vectorize-no-ints", cl::init(false), cl::Hidden,
110   cl::desc("Don't try to vectorize integer values"));
111 
112 static cl::opt<bool>
113 NoFloats("bb-vectorize-no-floats", cl::init(false), cl::Hidden,
114   cl::desc("Don't try to vectorize floating-point values"));
115 
116 // FIXME: This should default to false once pointer vector support works.
117 static cl::opt<bool>
118 NoPointers("bb-vectorize-no-pointers", cl::init(/*false*/ true), cl::Hidden,
119   cl::desc("Don't try to vectorize pointer values"));
120 
121 static cl::opt<bool>
122 NoCasts("bb-vectorize-no-casts", cl::init(false), cl::Hidden,
123   cl::desc("Don't try to vectorize casting (conversion) operations"));
124 
125 static cl::opt<bool>
126 NoMath("bb-vectorize-no-math", cl::init(false), cl::Hidden,
127   cl::desc("Don't try to vectorize floating-point math intrinsics"));
128 
129 static cl::opt<bool>
130   NoBitManipulation("bb-vectorize-no-bitmanip", cl::init(false), cl::Hidden,
131   cl::desc("Don't try to vectorize BitManipulation intrinsics"));
132 
133 static cl::opt<bool>
134 NoFMA("bb-vectorize-no-fma", cl::init(false), cl::Hidden,
135   cl::desc("Don't try to vectorize the fused-multiply-add intrinsic"));
136 
137 static cl::opt<bool>
138 NoSelect("bb-vectorize-no-select", cl::init(false), cl::Hidden,
139   cl::desc("Don't try to vectorize select instructions"));
140 
141 static cl::opt<bool>
142 NoCmp("bb-vectorize-no-cmp", cl::init(false), cl::Hidden,
143   cl::desc("Don't try to vectorize comparison instructions"));
144 
145 static cl::opt<bool>
146 NoGEP("bb-vectorize-no-gep", cl::init(false), cl::Hidden,
147   cl::desc("Don't try to vectorize getelementptr instructions"));
148 
149 static cl::opt<bool>
150 NoMemOps("bb-vectorize-no-mem-ops", cl::init(false), cl::Hidden,
151   cl::desc("Don't try to vectorize loads and stores"));
152 
153 static cl::opt<bool>
154 AlignedOnly("bb-vectorize-aligned-only", cl::init(false), cl::Hidden,
155   cl::desc("Only generate aligned loads and stores"));
156 
157 static cl::opt<bool>
158 NoMemOpBoost("bb-vectorize-no-mem-op-boost",
159   cl::init(false), cl::Hidden,
160   cl::desc("Don't boost the chain-depth contribution of loads and stores"));
161 
162 static cl::opt<bool>
163 FastDep("bb-vectorize-fast-dep", cl::init(false), cl::Hidden,
164   cl::desc("Use a fast instruction dependency analysis"));
165 
166 #ifndef NDEBUG
167 static cl::opt<bool>
168 DebugInstructionExamination("bb-vectorize-debug-instruction-examination",
169   cl::init(false), cl::Hidden,
170   cl::desc("When debugging is enabled, output information on the"
171            " instruction-examination process"));
172 static cl::opt<bool>
173 DebugCandidateSelection("bb-vectorize-debug-candidate-selection",
174   cl::init(false), cl::Hidden,
175   cl::desc("When debugging is enabled, output information on the"
176            " candidate-selection process"));
177 static cl::opt<bool>
178 DebugPairSelection("bb-vectorize-debug-pair-selection",
179   cl::init(false), cl::Hidden,
180   cl::desc("When debugging is enabled, output information on the"
181            " pair-selection process"));
182 static cl::opt<bool>
183 DebugCycleCheck("bb-vectorize-debug-cycle-check",
184   cl::init(false), cl::Hidden,
185   cl::desc("When debugging is enabled, output information on the"
186            " cycle-checking process"));
187 
188 static cl::opt<bool>
189 PrintAfterEveryPair("bb-vectorize-debug-print-after-every-pair",
190   cl::init(false), cl::Hidden,
191   cl::desc("When debugging is enabled, dump the basic block after"
192            " every pair is fused"));
193 #endif
194 
195 STATISTIC(NumFusedOps, "Number of operations fused by bb-vectorize");
196 
197 namespace {
198   struct BBVectorize : public BasicBlockPass {
199     static char ID; // Pass identification, replacement for typeid
200 
201     const VectorizeConfig Config;
202 
BBVectorize__anon1ee8cad80111::BBVectorize203     BBVectorize(const VectorizeConfig &C = VectorizeConfig())
204       : BasicBlockPass(ID), Config(C) {
205       initializeBBVectorizePass(*PassRegistry::getPassRegistry());
206     }
207 
BBVectorize__anon1ee8cad80111::BBVectorize208     BBVectorize(Pass *P, Function &F, const VectorizeConfig &C)
209       : BasicBlockPass(ID), Config(C) {
210       AA = &P->getAnalysis<AAResultsWrapperPass>().getAAResults();
211       DT = &P->getAnalysis<DominatorTreeWrapperPass>().getDomTree();
212       SE = &P->getAnalysis<ScalarEvolutionWrapperPass>().getSE();
213       TLI = &P->getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
214       TTI = IgnoreTargetInfo
215                 ? nullptr
216                 : &P->getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
217     }
218 
219     typedef std::pair<Value *, Value *> ValuePair;
220     typedef std::pair<ValuePair, int> ValuePairWithCost;
221     typedef std::pair<ValuePair, size_t> ValuePairWithDepth;
222     typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair
223     typedef std::pair<VPPair, unsigned> VPPairWithType;
224 
225     AliasAnalysis *AA;
226     DominatorTree *DT;
227     ScalarEvolution *SE;
228     const TargetLibraryInfo *TLI;
229     const TargetTransformInfo *TTI;
230 
231     // FIXME: const correct?
232 
233     bool vectorizePairs(BasicBlock &BB, bool NonPow2Len = false);
234 
235     bool getCandidatePairs(BasicBlock &BB,
236                        BasicBlock::iterator &Start,
237                        DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
238                        DenseSet<ValuePair> &FixedOrderPairs,
239                        DenseMap<ValuePair, int> &CandidatePairCostSavings,
240                        std::vector<Value *> &PairableInsts, bool NonPow2Len);
241 
242     // FIXME: The current implementation does not account for pairs that
243     // are connected in multiple ways. For example:
244     //   C1 = A1 / A2; C2 = A2 / A1 (which may be both direct and a swap)
245     enum PairConnectionType {
246       PairConnectionDirect,
247       PairConnectionSwap,
248       PairConnectionSplat
249     };
250 
251     void computeConnectedPairs(
252              DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
253              DenseSet<ValuePair> &CandidatePairsSet,
254              std::vector<Value *> &PairableInsts,
255              DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
256              DenseMap<VPPair, unsigned> &PairConnectionTypes);
257 
258     void buildDepMap(BasicBlock &BB,
259              DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
260              std::vector<Value *> &PairableInsts,
261              DenseSet<ValuePair> &PairableInstUsers);
262 
263     void choosePairs(DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
264              DenseSet<ValuePair> &CandidatePairsSet,
265              DenseMap<ValuePair, int> &CandidatePairCostSavings,
266              std::vector<Value *> &PairableInsts,
267              DenseSet<ValuePair> &FixedOrderPairs,
268              DenseMap<VPPair, unsigned> &PairConnectionTypes,
269              DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
270              DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
271              DenseSet<ValuePair> &PairableInstUsers,
272              DenseMap<Value *, Value *>& ChosenPairs);
273 
274     void fuseChosenPairs(BasicBlock &BB,
275              std::vector<Value *> &PairableInsts,
276              DenseMap<Value *, Value *>& ChosenPairs,
277              DenseSet<ValuePair> &FixedOrderPairs,
278              DenseMap<VPPair, unsigned> &PairConnectionTypes,
279              DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
280              DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps);
281 
282 
283     bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore);
284 
285     bool areInstsCompatible(Instruction *I, Instruction *J,
286                        bool IsSimpleLoadStore, bool NonPow2Len,
287                        int &CostSavings, int &FixedOrder);
288 
289     bool trackUsesOfI(DenseSet<Value *> &Users,
290                       AliasSetTracker &WriteSet, Instruction *I,
291                       Instruction *J, bool UpdateUsers = true,
292                       DenseSet<ValuePair> *LoadMoveSetPairs = nullptr);
293 
294   void computePairsConnectedTo(
295              DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
296              DenseSet<ValuePair> &CandidatePairsSet,
297              std::vector<Value *> &PairableInsts,
298              DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
299              DenseMap<VPPair, unsigned> &PairConnectionTypes,
300              ValuePair P);
301 
302     bool pairsConflict(ValuePair P, ValuePair Q,
303              DenseSet<ValuePair> &PairableInstUsers,
304              DenseMap<ValuePair, std::vector<ValuePair> >
305                *PairableInstUserMap = nullptr,
306              DenseSet<VPPair> *PairableInstUserPairSet = nullptr);
307 
308     bool pairWillFormCycle(ValuePair P,
309              DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUsers,
310              DenseSet<ValuePair> &CurrentPairs);
311 
312     void pruneDAGFor(
313              DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
314              std::vector<Value *> &PairableInsts,
315              DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
316              DenseSet<ValuePair> &PairableInstUsers,
317              DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
318              DenseSet<VPPair> &PairableInstUserPairSet,
319              DenseMap<Value *, Value *> &ChosenPairs,
320              DenseMap<ValuePair, size_t> &DAG,
321              DenseSet<ValuePair> &PrunedDAG, ValuePair J,
322              bool UseCycleCheck);
323 
324     void buildInitialDAGFor(
325              DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
326              DenseSet<ValuePair> &CandidatePairsSet,
327              std::vector<Value *> &PairableInsts,
328              DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
329              DenseSet<ValuePair> &PairableInstUsers,
330              DenseMap<Value *, Value *> &ChosenPairs,
331              DenseMap<ValuePair, size_t> &DAG, ValuePair J);
332 
333     void findBestDAGFor(
334              DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
335              DenseSet<ValuePair> &CandidatePairsSet,
336              DenseMap<ValuePair, int> &CandidatePairCostSavings,
337              std::vector<Value *> &PairableInsts,
338              DenseSet<ValuePair> &FixedOrderPairs,
339              DenseMap<VPPair, unsigned> &PairConnectionTypes,
340              DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
341              DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
342              DenseSet<ValuePair> &PairableInstUsers,
343              DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
344              DenseSet<VPPair> &PairableInstUserPairSet,
345              DenseMap<Value *, Value *> &ChosenPairs,
346              DenseSet<ValuePair> &BestDAG, size_t &BestMaxDepth,
347              int &BestEffSize, Value *II, std::vector<Value *>&JJ,
348              bool UseCycleCheck);
349 
350     Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I,
351                      Instruction *J, unsigned o);
352 
353     void fillNewShuffleMask(LLVMContext& Context, Instruction *J,
354                      unsigned MaskOffset, unsigned NumInElem,
355                      unsigned NumInElem1, unsigned IdxOffset,
356                      std::vector<Constant*> &Mask);
357 
358     Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I,
359                      Instruction *J);
360 
361     bool expandIEChain(LLVMContext& Context, Instruction *I, Instruction *J,
362                        unsigned o, Value *&LOp, unsigned numElemL,
363                        Type *ArgTypeL, Type *ArgTypeR, bool IBeforeJ,
364                        unsigned IdxOff = 0);
365 
366     Value *getReplacementInput(LLVMContext& Context, Instruction *I,
367                      Instruction *J, unsigned o, bool IBeforeJ);
368 
369     void getReplacementInputsForPair(LLVMContext& Context, Instruction *I,
370                      Instruction *J, SmallVectorImpl<Value *> &ReplacedOperands,
371                      bool IBeforeJ);
372 
373     void replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
374                      Instruction *J, Instruction *K,
375                      Instruction *&InsertionPt, Instruction *&K1,
376                      Instruction *&K2);
377 
378     void collectPairLoadMoveSet(BasicBlock &BB,
379                      DenseMap<Value *, Value *> &ChosenPairs,
380                      DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
381                      DenseSet<ValuePair> &LoadMoveSetPairs,
382                      Instruction *I);
383 
384     void collectLoadMoveSet(BasicBlock &BB,
385                      std::vector<Value *> &PairableInsts,
386                      DenseMap<Value *, Value *> &ChosenPairs,
387                      DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
388                      DenseSet<ValuePair> &LoadMoveSetPairs);
389 
390     bool canMoveUsesOfIAfterJ(BasicBlock &BB,
391                      DenseSet<ValuePair> &LoadMoveSetPairs,
392                      Instruction *I, Instruction *J);
393 
394     void moveUsesOfIAfterJ(BasicBlock &BB,
395                      DenseSet<ValuePair> &LoadMoveSetPairs,
396                      Instruction *&InsertionPt,
397                      Instruction *I, Instruction *J);
398 
vectorizeBB__anon1ee8cad80111::BBVectorize399     bool vectorizeBB(BasicBlock &BB) {
400       if (skipBasicBlock(BB))
401         return false;
402       if (!DT->isReachableFromEntry(&BB)) {
403         DEBUG(dbgs() << "BBV: skipping unreachable " << BB.getName() <<
404               " in " << BB.getParent()->getName() << "\n");
405         return false;
406       }
407 
408       DEBUG(if (TTI) dbgs() << "BBV: using target information\n");
409 
410       bool changed = false;
411       // Iterate a sufficient number of times to merge types of size 1 bit,
412       // then 2 bits, then 4, etc. up to half of the target vector width of the
413       // target vector register.
414       unsigned n = 1;
415       for (unsigned v = 2;
416            (TTI || v <= Config.VectorBits) &&
417            (!Config.MaxIter || n <= Config.MaxIter);
418            v *= 2, ++n) {
419         DEBUG(dbgs() << "BBV: fusing loop #" << n <<
420               " for " << BB.getName() << " in " <<
421               BB.getParent()->getName() << "...\n");
422         if (vectorizePairs(BB))
423           changed = true;
424         else
425           break;
426       }
427 
428       if (changed && !Pow2LenOnly) {
429         ++n;
430         for (; !Config.MaxIter || n <= Config.MaxIter; ++n) {
431           DEBUG(dbgs() << "BBV: fusing for non-2^n-length vectors loop #: " <<
432                 n << " for " << BB.getName() << " in " <<
433                 BB.getParent()->getName() << "...\n");
434           if (!vectorizePairs(BB, true)) break;
435         }
436       }
437 
438       DEBUG(dbgs() << "BBV: done!\n");
439       return changed;
440     }
441 
runOnBasicBlock__anon1ee8cad80111::BBVectorize442     bool runOnBasicBlock(BasicBlock &BB) override {
443       // OptimizeNone check deferred to vectorizeBB().
444 
445       AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
446       DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
447       SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
448       TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
449       TTI = IgnoreTargetInfo
450                 ? nullptr
451                 : &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
452                       *BB.getParent());
453 
454       return vectorizeBB(BB);
455     }
456 
getAnalysisUsage__anon1ee8cad80111::BBVectorize457     void getAnalysisUsage(AnalysisUsage &AU) const override {
458       BasicBlockPass::getAnalysisUsage(AU);
459       AU.addRequired<AAResultsWrapperPass>();
460       AU.addRequired<DominatorTreeWrapperPass>();
461       AU.addRequired<ScalarEvolutionWrapperPass>();
462       AU.addRequired<TargetLibraryInfoWrapperPass>();
463       AU.addRequired<TargetTransformInfoWrapperPass>();
464       AU.addPreserved<DominatorTreeWrapperPass>();
465       AU.addPreserved<GlobalsAAWrapperPass>();
466       AU.addPreserved<ScalarEvolutionWrapperPass>();
467       AU.addPreserved<SCEVAAWrapperPass>();
468       AU.setPreservesCFG();
469     }
470 
getVecTypeForPair__anon1ee8cad80111::BBVectorize471     static inline VectorType *getVecTypeForPair(Type *ElemTy, Type *Elem2Ty) {
472       assert(ElemTy->getScalarType() == Elem2Ty->getScalarType() &&
473              "Cannot form vector from incompatible scalar types");
474       Type *STy = ElemTy->getScalarType();
475 
476       unsigned numElem;
477       if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) {
478         numElem = VTy->getNumElements();
479       } else {
480         numElem = 1;
481       }
482 
483       if (VectorType *VTy = dyn_cast<VectorType>(Elem2Ty)) {
484         numElem += VTy->getNumElements();
485       } else {
486         numElem += 1;
487       }
488 
489       return VectorType::get(STy, numElem);
490     }
491 
getInstructionTypes__anon1ee8cad80111::BBVectorize492     static inline void getInstructionTypes(Instruction *I,
493                                            Type *&T1, Type *&T2) {
494       if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
495         // For stores, it is the value type, not the pointer type that matters
496         // because the value is what will come from a vector register.
497 
498         Value *IVal = SI->getValueOperand();
499         T1 = IVal->getType();
500       } else {
501         T1 = I->getType();
502       }
503 
504       if (CastInst *CI = dyn_cast<CastInst>(I))
505         T2 = CI->getSrcTy();
506       else
507         T2 = T1;
508 
509       if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
510         T2 = SI->getCondition()->getType();
511       } else if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(I)) {
512         T2 = SI->getOperand(0)->getType();
513       } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) {
514         T2 = CI->getOperand(0)->getType();
515       }
516     }
517 
518     // Returns the weight associated with the provided value. A chain of
519     // candidate pairs has a length given by the sum of the weights of its
520     // members (one weight per pair; the weight of each member of the pair
521     // is assumed to be the same). This length is then compared to the
522     // chain-length threshold to determine if a given chain is significant
523     // enough to be vectorized. The length is also used in comparing
524     // candidate chains where longer chains are considered to be better.
525     // Note: when this function returns 0, the resulting instructions are
526     // not actually fused.
getDepthFactor__anon1ee8cad80111::BBVectorize527     inline size_t getDepthFactor(Value *V) {
528       // InsertElement and ExtractElement have a depth factor of zero. This is
529       // for two reasons: First, they cannot be usefully fused. Second, because
530       // the pass generates a lot of these, they can confuse the simple metric
531       // used to compare the dags in the next iteration. Thus, giving them a
532       // weight of zero allows the pass to essentially ignore them in
533       // subsequent iterations when looking for vectorization opportunities
534       // while still tracking dependency chains that flow through those
535       // instructions.
536       if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V))
537         return 0;
538 
539       // Give a load or store half of the required depth so that load/store
540       // pairs will vectorize.
541       if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V)))
542         return Config.ReqChainDepth/2;
543 
544       return 1;
545     }
546 
547     // Returns the cost of the provided instruction using TTI.
548     // This does not handle loads and stores.
getInstrCost__anon1ee8cad80111::BBVectorize549     unsigned getInstrCost(unsigned Opcode, Type *T1, Type *T2,
550                           TargetTransformInfo::OperandValueKind Op1VK =
551                               TargetTransformInfo::OK_AnyValue,
552                           TargetTransformInfo::OperandValueKind Op2VK =
553                               TargetTransformInfo::OK_AnyValue) {
554       switch (Opcode) {
555       default: break;
556       case Instruction::GetElementPtr:
557         // We mark this instruction as zero-cost because scalar GEPs are usually
558         // lowered to the instruction addressing mode. At the moment we don't
559         // generate vector GEPs.
560         return 0;
561       case Instruction::Br:
562         return TTI->getCFInstrCost(Opcode);
563       case Instruction::PHI:
564         return 0;
565       case Instruction::Add:
566       case Instruction::FAdd:
567       case Instruction::Sub:
568       case Instruction::FSub:
569       case Instruction::Mul:
570       case Instruction::FMul:
571       case Instruction::UDiv:
572       case Instruction::SDiv:
573       case Instruction::FDiv:
574       case Instruction::URem:
575       case Instruction::SRem:
576       case Instruction::FRem:
577       case Instruction::Shl:
578       case Instruction::LShr:
579       case Instruction::AShr:
580       case Instruction::And:
581       case Instruction::Or:
582       case Instruction::Xor:
583         return TTI->getArithmeticInstrCost(Opcode, T1, Op1VK, Op2VK);
584       case Instruction::Select:
585       case Instruction::ICmp:
586       case Instruction::FCmp:
587         return TTI->getCmpSelInstrCost(Opcode, T1, T2);
588       case Instruction::ZExt:
589       case Instruction::SExt:
590       case Instruction::FPToUI:
591       case Instruction::FPToSI:
592       case Instruction::FPExt:
593       case Instruction::PtrToInt:
594       case Instruction::IntToPtr:
595       case Instruction::SIToFP:
596       case Instruction::UIToFP:
597       case Instruction::Trunc:
598       case Instruction::FPTrunc:
599       case Instruction::BitCast:
600       case Instruction::ShuffleVector:
601         return TTI->getCastInstrCost(Opcode, T1, T2);
602       }
603 
604       return 1;
605     }
606 
607     // This determines the relative offset of two loads or stores, returning
608     // true if the offset could be determined to be some constant value.
609     // For example, if OffsetInElmts == 1, then J accesses the memory directly
610     // after I; if OffsetInElmts == -1 then I accesses the memory
611     // directly after J.
getPairPtrInfo__anon1ee8cad80111::BBVectorize612     bool getPairPtrInfo(Instruction *I, Instruction *J,
613         Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment,
614         unsigned &IAddressSpace, unsigned &JAddressSpace,
615         int64_t &OffsetInElmts, bool ComputeOffset = true) {
616       OffsetInElmts = 0;
617       if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
618         LoadInst *LJ = cast<LoadInst>(J);
619         IPtr = LI->getPointerOperand();
620         JPtr = LJ->getPointerOperand();
621         IAlignment = LI->getAlignment();
622         JAlignment = LJ->getAlignment();
623         IAddressSpace = LI->getPointerAddressSpace();
624         JAddressSpace = LJ->getPointerAddressSpace();
625       } else {
626         StoreInst *SI = cast<StoreInst>(I), *SJ = cast<StoreInst>(J);
627         IPtr = SI->getPointerOperand();
628         JPtr = SJ->getPointerOperand();
629         IAlignment = SI->getAlignment();
630         JAlignment = SJ->getAlignment();
631         IAddressSpace = SI->getPointerAddressSpace();
632         JAddressSpace = SJ->getPointerAddressSpace();
633       }
634 
635       if (!ComputeOffset)
636         return true;
637 
638       const SCEV *IPtrSCEV = SE->getSCEV(IPtr);
639       const SCEV *JPtrSCEV = SE->getSCEV(JPtr);
640 
641       // If this is a trivial offset, then we'll get something like
642       // 1*sizeof(type). With target data, which we need anyway, this will get
643       // constant folded into a number.
644       const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV);
645       if (const SCEVConstant *ConstOffSCEV =
646             dyn_cast<SCEVConstant>(OffsetSCEV)) {
647         ConstantInt *IntOff = ConstOffSCEV->getValue();
648         int64_t Offset = IntOff->getSExtValue();
649         const DataLayout &DL = I->getModule()->getDataLayout();
650         Type *VTy = IPtr->getType()->getPointerElementType();
651         int64_t VTyTSS = (int64_t)DL.getTypeStoreSize(VTy);
652 
653         Type *VTy2 = JPtr->getType()->getPointerElementType();
654         if (VTy != VTy2 && Offset < 0) {
655           int64_t VTy2TSS = (int64_t)DL.getTypeStoreSize(VTy2);
656           OffsetInElmts = Offset/VTy2TSS;
657           return (std::abs(Offset) % VTy2TSS) == 0;
658         }
659 
660         OffsetInElmts = Offset/VTyTSS;
661         return (std::abs(Offset) % VTyTSS) == 0;
662       }
663 
664       return false;
665     }
666 
667     // Returns true if the provided CallInst represents an intrinsic that can
668     // be vectorized.
isVectorizableIntrinsic__anon1ee8cad80111::BBVectorize669     bool isVectorizableIntrinsic(CallInst* I) {
670       Function *F = I->getCalledFunction();
671       if (!F) return false;
672 
673       Intrinsic::ID IID = F->getIntrinsicID();
674       if (!IID) return false;
675 
676       switch(IID) {
677       default:
678         return false;
679       case Intrinsic::sqrt:
680       case Intrinsic::powi:
681       case Intrinsic::sin:
682       case Intrinsic::cos:
683       case Intrinsic::log:
684       case Intrinsic::log2:
685       case Intrinsic::log10:
686       case Intrinsic::exp:
687       case Intrinsic::exp2:
688       case Intrinsic::pow:
689       case Intrinsic::round:
690       case Intrinsic::copysign:
691       case Intrinsic::ceil:
692       case Intrinsic::nearbyint:
693       case Intrinsic::rint:
694       case Intrinsic::trunc:
695       case Intrinsic::floor:
696       case Intrinsic::fabs:
697       case Intrinsic::minnum:
698       case Intrinsic::maxnum:
699         return Config.VectorizeMath;
700       case Intrinsic::bswap:
701       case Intrinsic::ctpop:
702       case Intrinsic::ctlz:
703       case Intrinsic::cttz:
704         return Config.VectorizeBitManipulations;
705       case Intrinsic::fma:
706       case Intrinsic::fmuladd:
707         return Config.VectorizeFMA;
708       }
709     }
710 
isPureIEChain__anon1ee8cad80111::BBVectorize711     bool isPureIEChain(InsertElementInst *IE) {
712       InsertElementInst *IENext = IE;
713       do {
714         if (!isa<UndefValue>(IENext->getOperand(0)) &&
715             !isa<InsertElementInst>(IENext->getOperand(0))) {
716           return false;
717         }
718       } while ((IENext =
719                  dyn_cast<InsertElementInst>(IENext->getOperand(0))));
720 
721       return true;
722     }
723   };
724 
725   // This function implements one vectorization iteration on the provided
726   // basic block. It returns true if the block is changed.
vectorizePairs(BasicBlock & BB,bool NonPow2Len)727   bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) {
728     bool ShouldContinue;
729     BasicBlock::iterator Start = BB.getFirstInsertionPt();
730 
731     std::vector<Value *> AllPairableInsts;
732     DenseMap<Value *, Value *> AllChosenPairs;
733     DenseSet<ValuePair> AllFixedOrderPairs;
734     DenseMap<VPPair, unsigned> AllPairConnectionTypes;
735     DenseMap<ValuePair, std::vector<ValuePair> > AllConnectedPairs,
736                                                  AllConnectedPairDeps;
737 
738     do {
739       std::vector<Value *> PairableInsts;
740       DenseMap<Value *, std::vector<Value *> > CandidatePairs;
741       DenseSet<ValuePair> FixedOrderPairs;
742       DenseMap<ValuePair, int> CandidatePairCostSavings;
743       ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs,
744                                          FixedOrderPairs,
745                                          CandidatePairCostSavings,
746                                          PairableInsts, NonPow2Len);
747       if (PairableInsts.empty()) continue;
748 
749       // Build the candidate pair set for faster lookups.
750       DenseSet<ValuePair> CandidatePairsSet;
751       for (DenseMap<Value *, std::vector<Value *> >::iterator I =
752            CandidatePairs.begin(), E = CandidatePairs.end(); I != E; ++I)
753         for (std::vector<Value *>::iterator J = I->second.begin(),
754              JE = I->second.end(); J != JE; ++J)
755           CandidatePairsSet.insert(ValuePair(I->first, *J));
756 
757       // Now we have a map of all of the pairable instructions and we need to
758       // select the best possible pairing. A good pairing is one such that the
759       // users of the pair are also paired. This defines a (directed) forest
760       // over the pairs such that two pairs are connected iff the second pair
761       // uses the first.
762 
763       // Note that it only matters that both members of the second pair use some
764       // element of the first pair (to allow for splatting).
765 
766       DenseMap<ValuePair, std::vector<ValuePair> > ConnectedPairs,
767                                                    ConnectedPairDeps;
768       DenseMap<VPPair, unsigned> PairConnectionTypes;
769       computeConnectedPairs(CandidatePairs, CandidatePairsSet,
770                             PairableInsts, ConnectedPairs, PairConnectionTypes);
771       if (ConnectedPairs.empty()) continue;
772 
773       for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator
774            I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
775            I != IE; ++I)
776         for (std::vector<ValuePair>::iterator J = I->second.begin(),
777              JE = I->second.end(); J != JE; ++J)
778           ConnectedPairDeps[*J].push_back(I->first);
779 
780       // Build the pairable-instruction dependency map
781       DenseSet<ValuePair> PairableInstUsers;
782       buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers);
783 
784       // There is now a graph of the connected pairs. For each variable, pick
785       // the pairing with the largest dag meeting the depth requirement on at
786       // least one branch. Then select all pairings that are part of that dag
787       // and remove them from the list of available pairings and pairable
788       // variables.
789 
790       DenseMap<Value *, Value *> ChosenPairs;
791       choosePairs(CandidatePairs, CandidatePairsSet,
792         CandidatePairCostSavings,
793         PairableInsts, FixedOrderPairs, PairConnectionTypes,
794         ConnectedPairs, ConnectedPairDeps,
795         PairableInstUsers, ChosenPairs);
796 
797       if (ChosenPairs.empty()) continue;
798       AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(),
799                               PairableInsts.end());
800       AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end());
801 
802       // Only for the chosen pairs, propagate information on fixed-order pairs,
803       // pair connections, and their types to the data structures used by the
804       // pair fusion procedures.
805       for (DenseMap<Value *, Value *>::iterator I = ChosenPairs.begin(),
806            IE = ChosenPairs.end(); I != IE; ++I) {
807         if (FixedOrderPairs.count(*I))
808           AllFixedOrderPairs.insert(*I);
809         else if (FixedOrderPairs.count(ValuePair(I->second, I->first)))
810           AllFixedOrderPairs.insert(ValuePair(I->second, I->first));
811 
812         for (DenseMap<Value *, Value *>::iterator J = ChosenPairs.begin();
813              J != IE; ++J) {
814           DenseMap<VPPair, unsigned>::iterator K =
815             PairConnectionTypes.find(VPPair(*I, *J));
816           if (K != PairConnectionTypes.end()) {
817             AllPairConnectionTypes.insert(*K);
818           } else {
819             K = PairConnectionTypes.find(VPPair(*J, *I));
820             if (K != PairConnectionTypes.end())
821               AllPairConnectionTypes.insert(*K);
822           }
823         }
824       }
825 
826       for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator
827            I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
828            I != IE; ++I)
829         for (std::vector<ValuePair>::iterator J = I->second.begin(),
830           JE = I->second.end(); J != JE; ++J)
831           if (AllPairConnectionTypes.count(VPPair(I->first, *J))) {
832             AllConnectedPairs[I->first].push_back(*J);
833             AllConnectedPairDeps[*J].push_back(I->first);
834           }
835     } while (ShouldContinue);
836 
837     if (AllChosenPairs.empty()) return false;
838     NumFusedOps += AllChosenPairs.size();
839 
840     // A set of pairs has now been selected. It is now necessary to replace the
841     // paired instructions with vector instructions. For this procedure each
842     // operand must be replaced with a vector operand. This vector is formed
843     // by using build_vector on the old operands. The replaced values are then
844     // replaced with a vector_extract on the result.  Subsequent optimization
845     // passes should coalesce the build/extract combinations.
846 
847     fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs, AllFixedOrderPairs,
848                     AllPairConnectionTypes,
849                     AllConnectedPairs, AllConnectedPairDeps);
850 
851     // It is important to cleanup here so that future iterations of this
852     // function have less work to do.
853     (void)SimplifyInstructionsInBlock(&BB, TLI);
854     return true;
855   }
856 
857   // This function returns true if the provided instruction is capable of being
858   // fused into a vector instruction. This determination is based only on the
859   // type and other attributes of the instruction.
isInstVectorizable(Instruction * I,bool & IsSimpleLoadStore)860   bool BBVectorize::isInstVectorizable(Instruction *I,
861                                          bool &IsSimpleLoadStore) {
862     IsSimpleLoadStore = false;
863 
864     if (CallInst *C = dyn_cast<CallInst>(I)) {
865       if (!isVectorizableIntrinsic(C))
866         return false;
867     } else if (LoadInst *L = dyn_cast<LoadInst>(I)) {
868       // Vectorize simple loads if possbile:
869       IsSimpleLoadStore = L->isSimple();
870       if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
871         return false;
872     } else if (StoreInst *S = dyn_cast<StoreInst>(I)) {
873       // Vectorize simple stores if possbile:
874       IsSimpleLoadStore = S->isSimple();
875       if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
876         return false;
877     } else if (CastInst *C = dyn_cast<CastInst>(I)) {
878       // We can vectorize casts, but not casts of pointer types, etc.
879       if (!Config.VectorizeCasts)
880         return false;
881 
882       Type *SrcTy = C->getSrcTy();
883       if (!SrcTy->isSingleValueType())
884         return false;
885 
886       Type *DestTy = C->getDestTy();
887       if (!DestTy->isSingleValueType())
888         return false;
889     } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
890       if (!Config.VectorizeSelect)
891         return false;
892       // We can vectorize a select if either all operands are scalars,
893       // or all operands are vectors. Trying to "widen" a select between
894       // vectors that has a scalar condition results in a malformed select.
895       // FIXME: We could probably be smarter about this by rewriting the select
896       // with different types instead.
897       return (SI->getCondition()->getType()->isVectorTy() ==
898               SI->getTrueValue()->getType()->isVectorTy());
899     } else if (isa<CmpInst>(I)) {
900       if (!Config.VectorizeCmp)
901         return false;
902     } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) {
903       if (!Config.VectorizeGEP)
904         return false;
905 
906       // Currently, vector GEPs exist only with one index.
907       if (G->getNumIndices() != 1)
908         return false;
909     } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) ||
910         isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) {
911       return false;
912     }
913 
914     Type *T1, *T2;
915     getInstructionTypes(I, T1, T2);
916 
917     // Not every type can be vectorized...
918     if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) ||
919         !(VectorType::isValidElementType(T2) || T2->isVectorTy()))
920       return false;
921 
922     if (T1->getScalarSizeInBits() == 1) {
923       if (!Config.VectorizeBools)
924         return false;
925     } else {
926       if (!Config.VectorizeInts && T1->isIntOrIntVectorTy())
927         return false;
928     }
929 
930     if (T2->getScalarSizeInBits() == 1) {
931       if (!Config.VectorizeBools)
932         return false;
933     } else {
934       if (!Config.VectorizeInts && T2->isIntOrIntVectorTy())
935         return false;
936     }
937 
938     if (!Config.VectorizeFloats
939         && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy()))
940       return false;
941 
942     // Don't vectorize target-specific types.
943     if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy())
944       return false;
945     if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy())
946       return false;
947 
948     if (!Config.VectorizePointers && (T1->getScalarType()->isPointerTy() ||
949                                       T2->getScalarType()->isPointerTy()))
950       return false;
951 
952     if (!TTI && (T1->getPrimitiveSizeInBits() >= Config.VectorBits ||
953                  T2->getPrimitiveSizeInBits() >= Config.VectorBits))
954       return false;
955 
956     return true;
957   }
958 
959   // This function returns true if the two provided instructions are compatible
960   // (meaning that they can be fused into a vector instruction). This assumes
961   // that I has already been determined to be vectorizable and that J is not
962   // in the use dag of I.
areInstsCompatible(Instruction * I,Instruction * J,bool IsSimpleLoadStore,bool NonPow2Len,int & CostSavings,int & FixedOrder)963   bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J,
964                        bool IsSimpleLoadStore, bool NonPow2Len,
965                        int &CostSavings, int &FixedOrder) {
966     DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I <<
967                      " <-> " << *J << "\n");
968 
969     CostSavings = 0;
970     FixedOrder = 0;
971 
972     // Loads and stores can be merged if they have different alignments,
973     // but are otherwise the same.
974     if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment |
975                       (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0)))
976       return false;
977 
978     Type *IT1, *IT2, *JT1, *JT2;
979     getInstructionTypes(I, IT1, IT2);
980     getInstructionTypes(J, JT1, JT2);
981     unsigned MaxTypeBits = std::max(
982       IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(),
983       IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits());
984     if (!TTI && MaxTypeBits > Config.VectorBits)
985       return false;
986 
987     // FIXME: handle addsub-type operations!
988 
989     if (IsSimpleLoadStore) {
990       Value *IPtr, *JPtr;
991       unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
992       int64_t OffsetInElmts = 0;
993       if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
994                          IAddressSpace, JAddressSpace, OffsetInElmts) &&
995           std::abs(OffsetInElmts) == 1) {
996         FixedOrder = (int) OffsetInElmts;
997         unsigned BottomAlignment = IAlignment;
998         if (OffsetInElmts < 0) BottomAlignment = JAlignment;
999 
1000         Type *aTypeI = isa<StoreInst>(I) ?
1001           cast<StoreInst>(I)->getValueOperand()->getType() : I->getType();
1002         Type *aTypeJ = isa<StoreInst>(J) ?
1003           cast<StoreInst>(J)->getValueOperand()->getType() : J->getType();
1004         Type *VType = getVecTypeForPair(aTypeI, aTypeJ);
1005 
1006         if (Config.AlignedOnly) {
1007           // An aligned load or store is possible only if the instruction
1008           // with the lower offset has an alignment suitable for the
1009           // vector type.
1010           const DataLayout &DL = I->getModule()->getDataLayout();
1011           unsigned VecAlignment = DL.getPrefTypeAlignment(VType);
1012           if (BottomAlignment < VecAlignment)
1013             return false;
1014         }
1015 
1016         if (TTI) {
1017           unsigned ICost = TTI->getMemoryOpCost(I->getOpcode(), aTypeI,
1018                                                 IAlignment, IAddressSpace);
1019           unsigned JCost = TTI->getMemoryOpCost(J->getOpcode(), aTypeJ,
1020                                                 JAlignment, JAddressSpace);
1021           unsigned VCost = TTI->getMemoryOpCost(I->getOpcode(), VType,
1022                                                 BottomAlignment,
1023                                                 IAddressSpace);
1024 
1025           ICost += TTI->getAddressComputationCost(aTypeI);
1026           JCost += TTI->getAddressComputationCost(aTypeJ);
1027           VCost += TTI->getAddressComputationCost(VType);
1028 
1029           if (VCost > ICost + JCost)
1030             return false;
1031 
1032           // We don't want to fuse to a type that will be split, even
1033           // if the two input types will also be split and there is no other
1034           // associated cost.
1035           unsigned VParts = TTI->getNumberOfParts(VType);
1036           if (VParts > 1)
1037             return false;
1038           else if (!VParts && VCost == ICost + JCost)
1039             return false;
1040 
1041           CostSavings = ICost + JCost - VCost;
1042         }
1043       } else {
1044         return false;
1045       }
1046     } else if (TTI) {
1047       unsigned ICost = getInstrCost(I->getOpcode(), IT1, IT2);
1048       unsigned JCost = getInstrCost(J->getOpcode(), JT1, JT2);
1049       Type *VT1 = getVecTypeForPair(IT1, JT1),
1050            *VT2 = getVecTypeForPair(IT2, JT2);
1051       TargetTransformInfo::OperandValueKind Op1VK =
1052           TargetTransformInfo::OK_AnyValue;
1053       TargetTransformInfo::OperandValueKind Op2VK =
1054           TargetTransformInfo::OK_AnyValue;
1055 
1056       // On some targets (example X86) the cost of a vector shift may vary
1057       // depending on whether the second operand is a Uniform or
1058       // NonUniform Constant.
1059       switch (I->getOpcode()) {
1060       default : break;
1061       case Instruction::Shl:
1062       case Instruction::LShr:
1063       case Instruction::AShr:
1064 
1065         // If both I and J are scalar shifts by constant, then the
1066         // merged vector shift count would be either a constant splat value
1067         // or a non-uniform vector of constants.
1068         if (ConstantInt *CII = dyn_cast<ConstantInt>(I->getOperand(1))) {
1069           if (ConstantInt *CIJ = dyn_cast<ConstantInt>(J->getOperand(1)))
1070             Op2VK = CII == CIJ ? TargetTransformInfo::OK_UniformConstantValue :
1071                                TargetTransformInfo::OK_NonUniformConstantValue;
1072         } else {
1073           // Check for a splat of a constant or for a non uniform vector
1074           // of constants.
1075           Value *IOp = I->getOperand(1);
1076           Value *JOp = J->getOperand(1);
1077           if ((isa<ConstantVector>(IOp) || isa<ConstantDataVector>(IOp)) &&
1078               (isa<ConstantVector>(JOp) || isa<ConstantDataVector>(JOp))) {
1079             Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
1080             Constant *SplatValue = cast<Constant>(IOp)->getSplatValue();
1081             if (SplatValue != nullptr &&
1082                 SplatValue == cast<Constant>(JOp)->getSplatValue())
1083               Op2VK = TargetTransformInfo::OK_UniformConstantValue;
1084           }
1085         }
1086       }
1087 
1088       // Note that this procedure is incorrect for insert and extract element
1089       // instructions (because combining these often results in a shuffle),
1090       // but this cost is ignored (because insert and extract element
1091       // instructions are assigned a zero depth factor and are not really
1092       // fused in general).
1093       unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2, Op1VK, Op2VK);
1094 
1095       if (VCost > ICost + JCost)
1096         return false;
1097 
1098       // We don't want to fuse to a type that will be split, even
1099       // if the two input types will also be split and there is no other
1100       // associated cost.
1101       unsigned VParts1 = TTI->getNumberOfParts(VT1),
1102                VParts2 = TTI->getNumberOfParts(VT2);
1103       if (VParts1 > 1 || VParts2 > 1)
1104         return false;
1105       else if ((!VParts1 || !VParts2) && VCost == ICost + JCost)
1106         return false;
1107 
1108       CostSavings = ICost + JCost - VCost;
1109     }
1110 
1111     // The powi,ctlz,cttz intrinsics are special because only the first
1112     // argument is vectorized, the second arguments must be equal.
1113     CallInst *CI = dyn_cast<CallInst>(I);
1114     Function *FI;
1115     if (CI && (FI = CI->getCalledFunction())) {
1116       Intrinsic::ID IID = FI->getIntrinsicID();
1117       if (IID == Intrinsic::powi || IID == Intrinsic::ctlz ||
1118           IID == Intrinsic::cttz) {
1119         Value *A1I = CI->getArgOperand(1),
1120               *A1J = cast<CallInst>(J)->getArgOperand(1);
1121         const SCEV *A1ISCEV = SE->getSCEV(A1I),
1122                    *A1JSCEV = SE->getSCEV(A1J);
1123         return (A1ISCEV == A1JSCEV);
1124       }
1125 
1126       if (IID && TTI) {
1127         FastMathFlags FMFCI;
1128         if (auto *FPMOCI = dyn_cast<FPMathOperator>(CI))
1129           FMFCI = FPMOCI->getFastMathFlags();
1130 
1131         SmallVector<Type*, 4> Tys;
1132         for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
1133           Tys.push_back(CI->getArgOperand(i)->getType());
1134         unsigned ICost = TTI->getIntrinsicInstrCost(IID, IT1, Tys, FMFCI);
1135 
1136         Tys.clear();
1137         CallInst *CJ = cast<CallInst>(J);
1138 
1139         FastMathFlags FMFCJ;
1140         if (auto *FPMOCJ = dyn_cast<FPMathOperator>(CJ))
1141           FMFCJ = FPMOCJ->getFastMathFlags();
1142 
1143         for (unsigned i = 0, ie = CJ->getNumArgOperands(); i != ie; ++i)
1144           Tys.push_back(CJ->getArgOperand(i)->getType());
1145         unsigned JCost = TTI->getIntrinsicInstrCost(IID, JT1, Tys, FMFCJ);
1146 
1147         Tys.clear();
1148         assert(CI->getNumArgOperands() == CJ->getNumArgOperands() &&
1149                "Intrinsic argument counts differ");
1150         for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
1151           if ((IID == Intrinsic::powi || IID == Intrinsic::ctlz ||
1152                IID == Intrinsic::cttz) && i == 1)
1153             Tys.push_back(CI->getArgOperand(i)->getType());
1154           else
1155             Tys.push_back(getVecTypeForPair(CI->getArgOperand(i)->getType(),
1156                                             CJ->getArgOperand(i)->getType()));
1157         }
1158 
1159         FastMathFlags FMFV = FMFCI;
1160         FMFV &= FMFCJ;
1161         Type *RetTy = getVecTypeForPair(IT1, JT1);
1162         unsigned VCost = TTI->getIntrinsicInstrCost(IID, RetTy, Tys, FMFV);
1163 
1164         if (VCost > ICost + JCost)
1165           return false;
1166 
1167         // We don't want to fuse to a type that will be split, even
1168         // if the two input types will also be split and there is no other
1169         // associated cost.
1170         unsigned RetParts = TTI->getNumberOfParts(RetTy);
1171         if (RetParts > 1)
1172           return false;
1173         else if (!RetParts && VCost == ICost + JCost)
1174           return false;
1175 
1176         for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
1177           if (!Tys[i]->isVectorTy())
1178             continue;
1179 
1180           unsigned NumParts = TTI->getNumberOfParts(Tys[i]);
1181           if (NumParts > 1)
1182             return false;
1183           else if (!NumParts && VCost == ICost + JCost)
1184             return false;
1185         }
1186 
1187         CostSavings = ICost + JCost - VCost;
1188       }
1189     }
1190 
1191     return true;
1192   }
1193 
1194   // Figure out whether or not J uses I and update the users and write-set
1195   // structures associated with I. Specifically, Users represents the set of
1196   // instructions that depend on I. WriteSet represents the set
1197   // of memory locations that are dependent on I. If UpdateUsers is true,
1198   // and J uses I, then Users is updated to contain J and WriteSet is updated
1199   // to contain any memory locations to which J writes. The function returns
1200   // true if J uses I. By default, alias analysis is used to determine
1201   // whether J reads from memory that overlaps with a location in WriteSet.
1202   // If LoadMoveSet is not null, then it is a previously-computed map
1203   // where the key is the memory-based user instruction and the value is
1204   // the instruction to be compared with I. So, if LoadMoveSet is provided,
1205   // then the alias analysis is not used. This is necessary because this
1206   // function is called during the process of moving instructions during
1207   // vectorization and the results of the alias analysis are not stable during
1208   // that process.
trackUsesOfI(DenseSet<Value * > & Users,AliasSetTracker & WriteSet,Instruction * I,Instruction * J,bool UpdateUsers,DenseSet<ValuePair> * LoadMoveSetPairs)1209   bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users,
1210                        AliasSetTracker &WriteSet, Instruction *I,
1211                        Instruction *J, bool UpdateUsers,
1212                        DenseSet<ValuePair> *LoadMoveSetPairs) {
1213     bool UsesI = false;
1214 
1215     // This instruction may already be marked as a user due, for example, to
1216     // being a member of a selected pair.
1217     if (Users.count(J))
1218       UsesI = true;
1219 
1220     if (!UsesI)
1221       for (User::op_iterator JU = J->op_begin(), JE = J->op_end();
1222            JU != JE; ++JU) {
1223         Value *V = *JU;
1224         if (I == V || Users.count(V)) {
1225           UsesI = true;
1226           break;
1227         }
1228       }
1229     if (!UsesI && J->mayReadFromMemory()) {
1230       if (LoadMoveSetPairs) {
1231         UsesI = LoadMoveSetPairs->count(ValuePair(J, I));
1232       } else {
1233         for (AliasSetTracker::iterator W = WriteSet.begin(),
1234              WE = WriteSet.end(); W != WE; ++W) {
1235           if (W->aliasesUnknownInst(J, *AA)) {
1236             UsesI = true;
1237             break;
1238           }
1239         }
1240       }
1241     }
1242 
1243     if (UsesI && UpdateUsers) {
1244       if (J->mayWriteToMemory()) WriteSet.add(J);
1245       Users.insert(J);
1246     }
1247 
1248     return UsesI;
1249   }
1250 
1251   // This function iterates over all instruction pairs in the provided
1252   // basic block and collects all candidate pairs for vectorization.
getCandidatePairs(BasicBlock & BB,BasicBlock::iterator & Start,DenseMap<Value *,std::vector<Value * >> & CandidatePairs,DenseSet<ValuePair> & FixedOrderPairs,DenseMap<ValuePair,int> & CandidatePairCostSavings,std::vector<Value * > & PairableInsts,bool NonPow2Len)1253   bool BBVectorize::getCandidatePairs(BasicBlock &BB,
1254                        BasicBlock::iterator &Start,
1255                        DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1256                        DenseSet<ValuePair> &FixedOrderPairs,
1257                        DenseMap<ValuePair, int> &CandidatePairCostSavings,
1258                        std::vector<Value *> &PairableInsts, bool NonPow2Len) {
1259     size_t TotalPairs = 0;
1260     BasicBlock::iterator E = BB.end();
1261     if (Start == E) return false;
1262 
1263     bool ShouldContinue = false, IAfterStart = false;
1264     for (BasicBlock::iterator I = Start++; I != E; ++I) {
1265       if (I == Start) IAfterStart = true;
1266 
1267       bool IsSimpleLoadStore;
1268       if (!isInstVectorizable(&*I, IsSimpleLoadStore))
1269         continue;
1270 
1271       // Look for an instruction with which to pair instruction *I...
1272       DenseSet<Value *> Users;
1273       AliasSetTracker WriteSet(*AA);
1274       if (I->mayWriteToMemory())
1275         WriteSet.add(&*I);
1276 
1277       bool JAfterStart = IAfterStart;
1278       BasicBlock::iterator J = std::next(I);
1279       for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) {
1280         if (J == Start)
1281           JAfterStart = true;
1282 
1283         // Determine if J uses I, if so, exit the loop.
1284         bool UsesI = trackUsesOfI(Users, WriteSet, &*I, &*J, !Config.FastDep);
1285         if (Config.FastDep) {
1286           // Note: For this heuristic to be effective, independent operations
1287           // must tend to be intermixed. This is likely to be true from some
1288           // kinds of grouped loop unrolling (but not the generic LLVM pass),
1289           // but otherwise may require some kind of reordering pass.
1290 
1291           // When using fast dependency analysis,
1292           // stop searching after first use:
1293           if (UsesI) break;
1294         } else {
1295           if (UsesI) continue;
1296         }
1297 
1298         // J does not use I, and comes before the first use of I, so it can be
1299         // merged with I if the instructions are compatible.
1300         int CostSavings, FixedOrder;
1301         if (!areInstsCompatible(&*I, &*J, IsSimpleLoadStore, NonPow2Len,
1302                                 CostSavings, FixedOrder))
1303           continue;
1304 
1305         // J is a candidate for merging with I.
1306         if (PairableInsts.empty() ||
1307             PairableInsts[PairableInsts.size() - 1] != &*I) {
1308           PairableInsts.push_back(&*I);
1309         }
1310 
1311         CandidatePairs[&*I].push_back(&*J);
1312         ++TotalPairs;
1313         if (TTI)
1314           CandidatePairCostSavings.insert(
1315               ValuePairWithCost(ValuePair(&*I, &*J), CostSavings));
1316 
1317         if (FixedOrder == 1)
1318           FixedOrderPairs.insert(ValuePair(&*I, &*J));
1319         else if (FixedOrder == -1)
1320           FixedOrderPairs.insert(ValuePair(&*J, &*I));
1321 
1322         // The next call to this function must start after the last instruction
1323         // selected during this invocation.
1324         if (JAfterStart) {
1325           Start = std::next(J);
1326           IAfterStart = JAfterStart = false;
1327         }
1328 
1329         DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair "
1330                      << *I << " <-> " << *J << " (cost savings: " <<
1331                      CostSavings << ")\n");
1332 
1333         // If we have already found too many pairs, break here and this function
1334         // will be called again starting after the last instruction selected
1335         // during this invocation.
1336         if (PairableInsts.size() >= Config.MaxInsts ||
1337             TotalPairs >= Config.MaxPairs) {
1338           ShouldContinue = true;
1339           break;
1340         }
1341       }
1342 
1343       if (ShouldContinue)
1344         break;
1345     }
1346 
1347     DEBUG(dbgs() << "BBV: found " << PairableInsts.size()
1348            << " instructions with candidate pairs\n");
1349 
1350     return ShouldContinue;
1351   }
1352 
1353   // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that
1354   // it looks for pairs such that both members have an input which is an
1355   // output of PI or PJ.
computePairsConnectedTo(DenseMap<Value *,std::vector<Value * >> & CandidatePairs,DenseSet<ValuePair> & CandidatePairsSet,std::vector<Value * > & PairableInsts,DenseMap<ValuePair,std::vector<ValuePair>> & ConnectedPairs,DenseMap<VPPair,unsigned> & PairConnectionTypes,ValuePair P)1356   void BBVectorize::computePairsConnectedTo(
1357                   DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1358                   DenseSet<ValuePair> &CandidatePairsSet,
1359                   std::vector<Value *> &PairableInsts,
1360                   DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1361                   DenseMap<VPPair, unsigned> &PairConnectionTypes,
1362                   ValuePair P) {
1363     StoreInst *SI, *SJ;
1364 
1365     // For each possible pairing for this variable, look at the uses of
1366     // the first value...
1367     for (Value::user_iterator I = P.first->user_begin(),
1368                               E = P.first->user_end();
1369          I != E; ++I) {
1370       User *UI = *I;
1371       if (isa<LoadInst>(UI)) {
1372         // A pair cannot be connected to a load because the load only takes one
1373         // operand (the address) and it is a scalar even after vectorization.
1374         continue;
1375       } else if ((SI = dyn_cast<StoreInst>(UI)) &&
1376                  P.first == SI->getPointerOperand()) {
1377         // Similarly, a pair cannot be connected to a store through its
1378         // pointer operand.
1379         continue;
1380       }
1381 
1382       // For each use of the first variable, look for uses of the second
1383       // variable...
1384       for (User *UJ : P.second->users()) {
1385         if ((SJ = dyn_cast<StoreInst>(UJ)) &&
1386             P.second == SJ->getPointerOperand())
1387           continue;
1388 
1389         // Look for <I, J>:
1390         if (CandidatePairsSet.count(ValuePair(UI, UJ))) {
1391           VPPair VP(P, ValuePair(UI, UJ));
1392           ConnectedPairs[VP.first].push_back(VP.second);
1393           PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionDirect));
1394         }
1395 
1396         // Look for <J, I>:
1397         if (CandidatePairsSet.count(ValuePair(UJ, UI))) {
1398           VPPair VP(P, ValuePair(UJ, UI));
1399           ConnectedPairs[VP.first].push_back(VP.second);
1400           PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSwap));
1401         }
1402       }
1403 
1404       if (Config.SplatBreaksChain) continue;
1405       // Look for cases where just the first value in the pair is used by
1406       // both members of another pair (splatting).
1407       for (Value::user_iterator J = P.first->user_begin(); J != E; ++J) {
1408         User *UJ = *J;
1409         if ((SJ = dyn_cast<StoreInst>(UJ)) &&
1410             P.first == SJ->getPointerOperand())
1411           continue;
1412 
1413         if (CandidatePairsSet.count(ValuePair(UI, UJ))) {
1414           VPPair VP(P, ValuePair(UI, UJ));
1415           ConnectedPairs[VP.first].push_back(VP.second);
1416           PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1417         }
1418       }
1419     }
1420 
1421     if (Config.SplatBreaksChain) return;
1422     // Look for cases where just the second value in the pair is used by
1423     // both members of another pair (splatting).
1424     for (Value::user_iterator I = P.second->user_begin(),
1425                               E = P.second->user_end();
1426          I != E; ++I) {
1427       User *UI = *I;
1428       if (isa<LoadInst>(UI))
1429         continue;
1430       else if ((SI = dyn_cast<StoreInst>(UI)) &&
1431                P.second == SI->getPointerOperand())
1432         continue;
1433 
1434       for (Value::user_iterator J = P.second->user_begin(); J != E; ++J) {
1435         User *UJ = *J;
1436         if ((SJ = dyn_cast<StoreInst>(UJ)) &&
1437             P.second == SJ->getPointerOperand())
1438           continue;
1439 
1440         if (CandidatePairsSet.count(ValuePair(UI, UJ))) {
1441           VPPair VP(P, ValuePair(UI, UJ));
1442           ConnectedPairs[VP.first].push_back(VP.second);
1443           PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
1444         }
1445       }
1446     }
1447   }
1448 
1449   // This function figures out which pairs are connected.  Two pairs are
1450   // connected if some output of the first pair forms an input to both members
1451   // of the second pair.
computeConnectedPairs(DenseMap<Value *,std::vector<Value * >> & CandidatePairs,DenseSet<ValuePair> & CandidatePairsSet,std::vector<Value * > & PairableInsts,DenseMap<ValuePair,std::vector<ValuePair>> & ConnectedPairs,DenseMap<VPPair,unsigned> & PairConnectionTypes)1452   void BBVectorize::computeConnectedPairs(
1453                   DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1454                   DenseSet<ValuePair> &CandidatePairsSet,
1455                   std::vector<Value *> &PairableInsts,
1456                   DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1457                   DenseMap<VPPair, unsigned> &PairConnectionTypes) {
1458     for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
1459          PE = PairableInsts.end(); PI != PE; ++PI) {
1460       DenseMap<Value *, std::vector<Value *> >::iterator PP =
1461         CandidatePairs.find(*PI);
1462       if (PP == CandidatePairs.end())
1463         continue;
1464 
1465       for (std::vector<Value *>::iterator P = PP->second.begin(),
1466            E = PP->second.end(); P != E; ++P)
1467         computePairsConnectedTo(CandidatePairs, CandidatePairsSet,
1468                                 PairableInsts, ConnectedPairs,
1469                                 PairConnectionTypes, ValuePair(*PI, *P));
1470     }
1471 
1472     DEBUG(size_t TotalPairs = 0;
1473           for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator I =
1474                ConnectedPairs.begin(), IE = ConnectedPairs.end(); I != IE; ++I)
1475             TotalPairs += I->second.size();
1476           dbgs() << "BBV: found " << TotalPairs
1477                  << " pair connections.\n");
1478   }
1479 
1480   // This function builds a set of use tuples such that <A, B> is in the set
1481   // if B is in the use dag of A. If B is in the use dag of A, then B
1482   // depends on the output of A.
buildDepMap(BasicBlock & BB,DenseMap<Value *,std::vector<Value * >> & CandidatePairs,std::vector<Value * > & PairableInsts,DenseSet<ValuePair> & PairableInstUsers)1483   void BBVectorize::buildDepMap(
1484                       BasicBlock &BB,
1485                       DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1486                       std::vector<Value *> &PairableInsts,
1487                       DenseSet<ValuePair> &PairableInstUsers) {
1488     DenseSet<Value *> IsInPair;
1489     for (DenseMap<Value *, std::vector<Value *> >::iterator C =
1490          CandidatePairs.begin(), E = CandidatePairs.end(); C != E; ++C) {
1491       IsInPair.insert(C->first);
1492       IsInPair.insert(C->second.begin(), C->second.end());
1493     }
1494 
1495     // Iterate through the basic block, recording all users of each
1496     // pairable instruction.
1497 
1498     BasicBlock::iterator E = BB.end(), EL =
1499       BasicBlock::iterator(cast<Instruction>(PairableInsts.back()));
1500     for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) {
1501       if (IsInPair.find(&*I) == IsInPair.end())
1502         continue;
1503 
1504       DenseSet<Value *> Users;
1505       AliasSetTracker WriteSet(*AA);
1506       if (I->mayWriteToMemory())
1507         WriteSet.add(&*I);
1508 
1509       for (BasicBlock::iterator J = std::next(I); J != E; ++J) {
1510         (void)trackUsesOfI(Users, WriteSet, &*I, &*J);
1511 
1512         if (J == EL)
1513           break;
1514       }
1515 
1516       for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end();
1517            U != E; ++U) {
1518         if (IsInPair.find(*U) == IsInPair.end()) continue;
1519         PairableInstUsers.insert(ValuePair(&*I, *U));
1520       }
1521 
1522       if (I == EL)
1523         break;
1524     }
1525   }
1526 
1527   // Returns true if an input to pair P is an output of pair Q and also an
1528   // input of pair Q is an output of pair P. If this is the case, then these
1529   // two pairs cannot be simultaneously fused.
pairsConflict(ValuePair P,ValuePair Q,DenseSet<ValuePair> & PairableInstUsers,DenseMap<ValuePair,std::vector<ValuePair>> * PairableInstUserMap,DenseSet<VPPair> * PairableInstUserPairSet)1530   bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q,
1531              DenseSet<ValuePair> &PairableInstUsers,
1532              DenseMap<ValuePair, std::vector<ValuePair> > *PairableInstUserMap,
1533              DenseSet<VPPair> *PairableInstUserPairSet) {
1534     // Two pairs are in conflict if they are mutual Users of eachother.
1535     bool QUsesP = PairableInstUsers.count(ValuePair(P.first,  Q.first))  ||
1536                   PairableInstUsers.count(ValuePair(P.first,  Q.second)) ||
1537                   PairableInstUsers.count(ValuePair(P.second, Q.first))  ||
1538                   PairableInstUsers.count(ValuePair(P.second, Q.second));
1539     bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first,  P.first))  ||
1540                   PairableInstUsers.count(ValuePair(Q.first,  P.second)) ||
1541                   PairableInstUsers.count(ValuePair(Q.second, P.first))  ||
1542                   PairableInstUsers.count(ValuePair(Q.second, P.second));
1543     if (PairableInstUserMap) {
1544       // FIXME: The expensive part of the cycle check is not so much the cycle
1545       // check itself but this edge insertion procedure. This needs some
1546       // profiling and probably a different data structure.
1547       if (PUsesQ) {
1548         if (PairableInstUserPairSet->insert(VPPair(Q, P)).second)
1549           (*PairableInstUserMap)[Q].push_back(P);
1550       }
1551       if (QUsesP) {
1552         if (PairableInstUserPairSet->insert(VPPair(P, Q)).second)
1553           (*PairableInstUserMap)[P].push_back(Q);
1554       }
1555     }
1556 
1557     return (QUsesP && PUsesQ);
1558   }
1559 
1560   // This function walks the use graph of current pairs to see if, starting
1561   // from P, the walk returns to P.
pairWillFormCycle(ValuePair P,DenseMap<ValuePair,std::vector<ValuePair>> & PairableInstUserMap,DenseSet<ValuePair> & CurrentPairs)1562   bool BBVectorize::pairWillFormCycle(ValuePair P,
1563              DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
1564              DenseSet<ValuePair> &CurrentPairs) {
1565     DEBUG(if (DebugCycleCheck)
1566             dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> "
1567                    << *P.second << "\n");
1568     // A lookup table of visisted pairs is kept because the PairableInstUserMap
1569     // contains non-direct associations.
1570     DenseSet<ValuePair> Visited;
1571     SmallVector<ValuePair, 32> Q;
1572     // General depth-first post-order traversal:
1573     Q.push_back(P);
1574     do {
1575       ValuePair QTop = Q.pop_back_val();
1576       Visited.insert(QTop);
1577 
1578       DEBUG(if (DebugCycleCheck)
1579               dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> "
1580                      << *QTop.second << "\n");
1581       DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
1582         PairableInstUserMap.find(QTop);
1583       if (QQ == PairableInstUserMap.end())
1584         continue;
1585 
1586       for (std::vector<ValuePair>::iterator C = QQ->second.begin(),
1587            CE = QQ->second.end(); C != CE; ++C) {
1588         if (*C == P) {
1589           DEBUG(dbgs()
1590                  << "BBV: rejected to prevent non-trivial cycle formation: "
1591                  << QTop.first << " <-> " << C->second << "\n");
1592           return true;
1593         }
1594 
1595         if (CurrentPairs.count(*C) && !Visited.count(*C))
1596           Q.push_back(*C);
1597       }
1598     } while (!Q.empty());
1599 
1600     return false;
1601   }
1602 
1603   // This function builds the initial dag of connected pairs with the
1604   // pair J at the root.
buildInitialDAGFor(DenseMap<Value *,std::vector<Value * >> & CandidatePairs,DenseSet<ValuePair> & CandidatePairsSet,std::vector<Value * > & PairableInsts,DenseMap<ValuePair,std::vector<ValuePair>> & ConnectedPairs,DenseSet<ValuePair> & PairableInstUsers,DenseMap<Value *,Value * > & ChosenPairs,DenseMap<ValuePair,size_t> & DAG,ValuePair J)1605   void BBVectorize::buildInitialDAGFor(
1606                   DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1607                   DenseSet<ValuePair> &CandidatePairsSet,
1608                   std::vector<Value *> &PairableInsts,
1609                   DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1610                   DenseSet<ValuePair> &PairableInstUsers,
1611                   DenseMap<Value *, Value *> &ChosenPairs,
1612                   DenseMap<ValuePair, size_t> &DAG, ValuePair J) {
1613     // Each of these pairs is viewed as the root node of a DAG. The DAG
1614     // is then walked (depth-first). As this happens, we keep track of
1615     // the pairs that compose the DAG and the maximum depth of the DAG.
1616     SmallVector<ValuePairWithDepth, 32> Q;
1617     // General depth-first post-order traversal:
1618     Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1619     do {
1620       ValuePairWithDepth QTop = Q.back();
1621 
1622       // Push each child onto the queue:
1623       bool MoreChildren = false;
1624       size_t MaxChildDepth = QTop.second;
1625       DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
1626         ConnectedPairs.find(QTop.first);
1627       if (QQ != ConnectedPairs.end())
1628         for (std::vector<ValuePair>::iterator k = QQ->second.begin(),
1629              ke = QQ->second.end(); k != ke; ++k) {
1630           // Make sure that this child pair is still a candidate:
1631           if (CandidatePairsSet.count(*k)) {
1632             DenseMap<ValuePair, size_t>::iterator C = DAG.find(*k);
1633             if (C == DAG.end()) {
1634               size_t d = getDepthFactor(k->first);
1635               Q.push_back(ValuePairWithDepth(*k, QTop.second+d));
1636               MoreChildren = true;
1637             } else {
1638               MaxChildDepth = std::max(MaxChildDepth, C->second);
1639             }
1640           }
1641         }
1642 
1643       if (!MoreChildren) {
1644         // Record the current pair as part of the DAG:
1645         DAG.insert(ValuePairWithDepth(QTop.first, MaxChildDepth));
1646         Q.pop_back();
1647       }
1648     } while (!Q.empty());
1649   }
1650 
1651   // Given some initial dag, prune it by removing conflicting pairs (pairs
1652   // that cannot be simultaneously chosen for vectorization).
pruneDAGFor(DenseMap<Value *,std::vector<Value * >> & CandidatePairs,std::vector<Value * > & PairableInsts,DenseMap<ValuePair,std::vector<ValuePair>> & ConnectedPairs,DenseSet<ValuePair> & PairableInstUsers,DenseMap<ValuePair,std::vector<ValuePair>> & PairableInstUserMap,DenseSet<VPPair> & PairableInstUserPairSet,DenseMap<Value *,Value * > & ChosenPairs,DenseMap<ValuePair,size_t> & DAG,DenseSet<ValuePair> & PrunedDAG,ValuePair J,bool UseCycleCheck)1653   void BBVectorize::pruneDAGFor(
1654               DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1655               std::vector<Value *> &PairableInsts,
1656               DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1657               DenseSet<ValuePair> &PairableInstUsers,
1658               DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
1659               DenseSet<VPPair> &PairableInstUserPairSet,
1660               DenseMap<Value *, Value *> &ChosenPairs,
1661               DenseMap<ValuePair, size_t> &DAG,
1662               DenseSet<ValuePair> &PrunedDAG, ValuePair J,
1663               bool UseCycleCheck) {
1664     SmallVector<ValuePairWithDepth, 32> Q;
1665     // General depth-first post-order traversal:
1666     Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
1667     do {
1668       ValuePairWithDepth QTop = Q.pop_back_val();
1669       PrunedDAG.insert(QTop.first);
1670 
1671       // Visit each child, pruning as necessary...
1672       SmallVector<ValuePairWithDepth, 8> BestChildren;
1673       DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
1674         ConnectedPairs.find(QTop.first);
1675       if (QQ == ConnectedPairs.end())
1676         continue;
1677 
1678       for (std::vector<ValuePair>::iterator K = QQ->second.begin(),
1679            KE = QQ->second.end(); K != KE; ++K) {
1680         DenseMap<ValuePair, size_t>::iterator C = DAG.find(*K);
1681         if (C == DAG.end()) continue;
1682 
1683         // This child is in the DAG, now we need to make sure it is the
1684         // best of any conflicting children. There could be multiple
1685         // conflicting children, so first, determine if we're keeping
1686         // this child, then delete conflicting children as necessary.
1687 
1688         // It is also necessary to guard against pairing-induced
1689         // dependencies. Consider instructions a .. x .. y .. b
1690         // such that (a,b) are to be fused and (x,y) are to be fused
1691         // but a is an input to x and b is an output from y. This
1692         // means that y cannot be moved after b but x must be moved
1693         // after b for (a,b) to be fused. In other words, after
1694         // fusing (a,b) we have y .. a/b .. x where y is an input
1695         // to a/b and x is an output to a/b: x and y can no longer
1696         // be legally fused. To prevent this condition, we must
1697         // make sure that a child pair added to the DAG is not
1698         // both an input and output of an already-selected pair.
1699 
1700         // Pairing-induced dependencies can also form from more complicated
1701         // cycles. The pair vs. pair conflicts are easy to check, and so
1702         // that is done explicitly for "fast rejection", and because for
1703         // child vs. child conflicts, we may prefer to keep the current
1704         // pair in preference to the already-selected child.
1705         DenseSet<ValuePair> CurrentPairs;
1706 
1707         bool CanAdd = true;
1708         for (SmallVectorImpl<ValuePairWithDepth>::iterator C2
1709               = BestChildren.begin(), E2 = BestChildren.end();
1710              C2 != E2; ++C2) {
1711           if (C2->first.first == C->first.first ||
1712               C2->first.first == C->first.second ||
1713               C2->first.second == C->first.first ||
1714               C2->first.second == C->first.second ||
1715               pairsConflict(C2->first, C->first, PairableInstUsers,
1716                             UseCycleCheck ? &PairableInstUserMap : nullptr,
1717                             UseCycleCheck ? &PairableInstUserPairSet
1718                                           : nullptr)) {
1719             if (C2->second >= C->second) {
1720               CanAdd = false;
1721               break;
1722             }
1723 
1724             CurrentPairs.insert(C2->first);
1725           }
1726         }
1727         if (!CanAdd) continue;
1728 
1729         // Even worse, this child could conflict with another node already
1730         // selected for the DAG. If that is the case, ignore this child.
1731         for (DenseSet<ValuePair>::iterator T = PrunedDAG.begin(),
1732              E2 = PrunedDAG.end(); T != E2; ++T) {
1733           if (T->first == C->first.first ||
1734               T->first == C->first.second ||
1735               T->second == C->first.first ||
1736               T->second == C->first.second ||
1737               pairsConflict(*T, C->first, PairableInstUsers,
1738                             UseCycleCheck ? &PairableInstUserMap : nullptr,
1739                             UseCycleCheck ? &PairableInstUserPairSet
1740                                           : nullptr)) {
1741             CanAdd = false;
1742             break;
1743           }
1744 
1745           CurrentPairs.insert(*T);
1746         }
1747         if (!CanAdd) continue;
1748 
1749         // And check the queue too...
1750         for (SmallVectorImpl<ValuePairWithDepth>::iterator C2 = Q.begin(),
1751              E2 = Q.end(); C2 != E2; ++C2) {
1752           if (C2->first.first == C->first.first ||
1753               C2->first.first == C->first.second ||
1754               C2->first.second == C->first.first ||
1755               C2->first.second == C->first.second ||
1756               pairsConflict(C2->first, C->first, PairableInstUsers,
1757                             UseCycleCheck ? &PairableInstUserMap : nullptr,
1758                             UseCycleCheck ? &PairableInstUserPairSet
1759                                           : nullptr)) {
1760             CanAdd = false;
1761             break;
1762           }
1763 
1764           CurrentPairs.insert(C2->first);
1765         }
1766         if (!CanAdd) continue;
1767 
1768         // Last but not least, check for a conflict with any of the
1769         // already-chosen pairs.
1770         for (DenseMap<Value *, Value *>::iterator C2 =
1771               ChosenPairs.begin(), E2 = ChosenPairs.end();
1772              C2 != E2; ++C2) {
1773           if (pairsConflict(*C2, C->first, PairableInstUsers,
1774                             UseCycleCheck ? &PairableInstUserMap : nullptr,
1775                             UseCycleCheck ? &PairableInstUserPairSet
1776                                           : nullptr)) {
1777             CanAdd = false;
1778             break;
1779           }
1780 
1781           CurrentPairs.insert(*C2);
1782         }
1783         if (!CanAdd) continue;
1784 
1785         // To check for non-trivial cycles formed by the addition of the
1786         // current pair we've formed a list of all relevant pairs, now use a
1787         // graph walk to check for a cycle. We start from the current pair and
1788         // walk the use dag to see if we again reach the current pair. If we
1789         // do, then the current pair is rejected.
1790 
1791         // FIXME: It may be more efficient to use a topological-ordering
1792         // algorithm to improve the cycle check. This should be investigated.
1793         if (UseCycleCheck &&
1794             pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs))
1795           continue;
1796 
1797         // This child can be added, but we may have chosen it in preference
1798         // to an already-selected child. Check for this here, and if a
1799         // conflict is found, then remove the previously-selected child
1800         // before adding this one in its place.
1801         for (SmallVectorImpl<ValuePairWithDepth>::iterator C2
1802               = BestChildren.begin(); C2 != BestChildren.end();) {
1803           if (C2->first.first == C->first.first ||
1804               C2->first.first == C->first.second ||
1805               C2->first.second == C->first.first ||
1806               C2->first.second == C->first.second ||
1807               pairsConflict(C2->first, C->first, PairableInstUsers))
1808             C2 = BestChildren.erase(C2);
1809           else
1810             ++C2;
1811         }
1812 
1813         BestChildren.push_back(ValuePairWithDepth(C->first, C->second));
1814       }
1815 
1816       for (SmallVectorImpl<ValuePairWithDepth>::iterator C
1817             = BestChildren.begin(), E2 = BestChildren.end();
1818            C != E2; ++C) {
1819         size_t DepthF = getDepthFactor(C->first.first);
1820         Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF));
1821       }
1822     } while (!Q.empty());
1823   }
1824 
1825   // This function finds the best dag of mututally-compatible connected
1826   // pairs, given the choice of root pairs as an iterator range.
findBestDAGFor(DenseMap<Value *,std::vector<Value * >> & CandidatePairs,DenseSet<ValuePair> & CandidatePairsSet,DenseMap<ValuePair,int> & CandidatePairCostSavings,std::vector<Value * > & PairableInsts,DenseSet<ValuePair> & FixedOrderPairs,DenseMap<VPPair,unsigned> & PairConnectionTypes,DenseMap<ValuePair,std::vector<ValuePair>> & ConnectedPairs,DenseMap<ValuePair,std::vector<ValuePair>> & ConnectedPairDeps,DenseSet<ValuePair> & PairableInstUsers,DenseMap<ValuePair,std::vector<ValuePair>> & PairableInstUserMap,DenseSet<VPPair> & PairableInstUserPairSet,DenseMap<Value *,Value * > & ChosenPairs,DenseSet<ValuePair> & BestDAG,size_t & BestMaxDepth,int & BestEffSize,Value * II,std::vector<Value * > & JJ,bool UseCycleCheck)1827   void BBVectorize::findBestDAGFor(
1828               DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
1829               DenseSet<ValuePair> &CandidatePairsSet,
1830               DenseMap<ValuePair, int> &CandidatePairCostSavings,
1831               std::vector<Value *> &PairableInsts,
1832               DenseSet<ValuePair> &FixedOrderPairs,
1833               DenseMap<VPPair, unsigned> &PairConnectionTypes,
1834               DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
1835               DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
1836               DenseSet<ValuePair> &PairableInstUsers,
1837               DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
1838               DenseSet<VPPair> &PairableInstUserPairSet,
1839               DenseMap<Value *, Value *> &ChosenPairs,
1840               DenseSet<ValuePair> &BestDAG, size_t &BestMaxDepth,
1841               int &BestEffSize, Value *II, std::vector<Value *>&JJ,
1842               bool UseCycleCheck) {
1843     for (std::vector<Value *>::iterator J = JJ.begin(), JE = JJ.end();
1844          J != JE; ++J) {
1845       ValuePair IJ(II, *J);
1846       if (!CandidatePairsSet.count(IJ))
1847         continue;
1848 
1849       // Before going any further, make sure that this pair does not
1850       // conflict with any already-selected pairs (see comment below
1851       // near the DAG pruning for more details).
1852       DenseSet<ValuePair> ChosenPairSet;
1853       bool DoesConflict = false;
1854       for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(),
1855            E = ChosenPairs.end(); C != E; ++C) {
1856         if (pairsConflict(*C, IJ, PairableInstUsers,
1857                           UseCycleCheck ? &PairableInstUserMap : nullptr,
1858                           UseCycleCheck ? &PairableInstUserPairSet : nullptr)) {
1859           DoesConflict = true;
1860           break;
1861         }
1862 
1863         ChosenPairSet.insert(*C);
1864       }
1865       if (DoesConflict) continue;
1866 
1867       if (UseCycleCheck &&
1868           pairWillFormCycle(IJ, PairableInstUserMap, ChosenPairSet))
1869         continue;
1870 
1871       DenseMap<ValuePair, size_t> DAG;
1872       buildInitialDAGFor(CandidatePairs, CandidatePairsSet,
1873                           PairableInsts, ConnectedPairs,
1874                           PairableInstUsers, ChosenPairs, DAG, IJ);
1875 
1876       // Because we'll keep the child with the largest depth, the largest
1877       // depth is still the same in the unpruned DAG.
1878       size_t MaxDepth = DAG.lookup(IJ);
1879 
1880       DEBUG(if (DebugPairSelection) dbgs() << "BBV: found DAG for pair {"
1881                    << *IJ.first << " <-> " << *IJ.second << "} of depth " <<
1882                    MaxDepth << " and size " << DAG.size() << "\n");
1883 
1884       // At this point the DAG has been constructed, but, may contain
1885       // contradictory children (meaning that different children of
1886       // some dag node may be attempting to fuse the same instruction).
1887       // So now we walk the dag again, in the case of a conflict,
1888       // keep only the child with the largest depth. To break a tie,
1889       // favor the first child.
1890 
1891       DenseSet<ValuePair> PrunedDAG;
1892       pruneDAGFor(CandidatePairs, PairableInsts, ConnectedPairs,
1893                    PairableInstUsers, PairableInstUserMap,
1894                    PairableInstUserPairSet,
1895                    ChosenPairs, DAG, PrunedDAG, IJ, UseCycleCheck);
1896 
1897       int EffSize = 0;
1898       if (TTI) {
1899         DenseSet<Value *> PrunedDAGInstrs;
1900         for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(),
1901              E = PrunedDAG.end(); S != E; ++S) {
1902           PrunedDAGInstrs.insert(S->first);
1903           PrunedDAGInstrs.insert(S->second);
1904         }
1905 
1906         // The set of pairs that have already contributed to the total cost.
1907         DenseSet<ValuePair> IncomingPairs;
1908 
1909         // If the cost model were perfect, this might not be necessary; but we
1910         // need to make sure that we don't get stuck vectorizing our own
1911         // shuffle chains.
1912         bool HasNontrivialInsts = false;
1913 
1914         // The node weights represent the cost savings associated with
1915         // fusing the pair of instructions.
1916         for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(),
1917              E = PrunedDAG.end(); S != E; ++S) {
1918           if (!isa<ShuffleVectorInst>(S->first) &&
1919               !isa<InsertElementInst>(S->first) &&
1920               !isa<ExtractElementInst>(S->first))
1921             HasNontrivialInsts = true;
1922 
1923           bool FlipOrder = false;
1924 
1925           if (getDepthFactor(S->first)) {
1926             int ESContrib = CandidatePairCostSavings.find(*S)->second;
1927             DEBUG(if (DebugPairSelection) dbgs() << "\tweight {"
1928                    << *S->first << " <-> " << *S->second << "} = " <<
1929                    ESContrib << "\n");
1930             EffSize += ESContrib;
1931           }
1932 
1933           // The edge weights contribute in a negative sense: they represent
1934           // the cost of shuffles.
1935           DenseMap<ValuePair, std::vector<ValuePair> >::iterator SS =
1936             ConnectedPairDeps.find(*S);
1937           if (SS != ConnectedPairDeps.end()) {
1938             unsigned NumDepsDirect = 0, NumDepsSwap = 0;
1939             for (std::vector<ValuePair>::iterator T = SS->second.begin(),
1940                  TE = SS->second.end(); T != TE; ++T) {
1941               VPPair Q(*S, *T);
1942               if (!PrunedDAG.count(Q.second))
1943                 continue;
1944               DenseMap<VPPair, unsigned>::iterator R =
1945                 PairConnectionTypes.find(VPPair(Q.second, Q.first));
1946               assert(R != PairConnectionTypes.end() &&
1947                      "Cannot find pair connection type");
1948               if (R->second == PairConnectionDirect)
1949                 ++NumDepsDirect;
1950               else if (R->second == PairConnectionSwap)
1951                 ++NumDepsSwap;
1952             }
1953 
1954             // If there are more swaps than direct connections, then
1955             // the pair order will be flipped during fusion. So the real
1956             // number of swaps is the minimum number.
1957             FlipOrder = !FixedOrderPairs.count(*S) &&
1958               ((NumDepsSwap > NumDepsDirect) ||
1959                 FixedOrderPairs.count(ValuePair(S->second, S->first)));
1960 
1961             for (std::vector<ValuePair>::iterator T = SS->second.begin(),
1962                  TE = SS->second.end(); T != TE; ++T) {
1963               VPPair Q(*S, *T);
1964               if (!PrunedDAG.count(Q.second))
1965                 continue;
1966               DenseMap<VPPair, unsigned>::iterator R =
1967                 PairConnectionTypes.find(VPPair(Q.second, Q.first));
1968               assert(R != PairConnectionTypes.end() &&
1969                      "Cannot find pair connection type");
1970               Type *Ty1 = Q.second.first->getType(),
1971                    *Ty2 = Q.second.second->getType();
1972               Type *VTy = getVecTypeForPair(Ty1, Ty2);
1973               if ((R->second == PairConnectionDirect && FlipOrder) ||
1974                   (R->second == PairConnectionSwap && !FlipOrder)  ||
1975                   R->second == PairConnectionSplat) {
1976                 int ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
1977                                                    VTy, VTy);
1978 
1979                 if (VTy->getVectorNumElements() == 2) {
1980                   if (R->second == PairConnectionSplat)
1981                     ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1982                       TargetTransformInfo::SK_Broadcast, VTy));
1983                   else
1984                     ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
1985                       TargetTransformInfo::SK_Reverse, VTy));
1986                 }
1987 
1988                 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
1989                   *Q.second.first << " <-> " << *Q.second.second <<
1990                     "} -> {" <<
1991                   *S->first << " <-> " << *S->second << "} = " <<
1992                    ESContrib << "\n");
1993                 EffSize -= ESContrib;
1994               }
1995             }
1996           }
1997 
1998           // Compute the cost of outgoing edges. We assume that edges outgoing
1999           // to shuffles, inserts or extracts can be merged, and so contribute
2000           // no additional cost.
2001           if (!S->first->getType()->isVoidTy()) {
2002             Type *Ty1 = S->first->getType(),
2003                  *Ty2 = S->second->getType();
2004             Type *VTy = getVecTypeForPair(Ty1, Ty2);
2005 
2006             bool NeedsExtraction = false;
2007             for (User *U : S->first->users()) {
2008               if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(U)) {
2009                 // Shuffle can be folded if it has no other input
2010                 if (isa<UndefValue>(SI->getOperand(1)))
2011                   continue;
2012               }
2013               if (isa<ExtractElementInst>(U))
2014                 continue;
2015               if (PrunedDAGInstrs.count(U))
2016                 continue;
2017               NeedsExtraction = true;
2018               break;
2019             }
2020 
2021             if (NeedsExtraction) {
2022               int ESContrib;
2023               if (Ty1->isVectorTy()) {
2024                 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2025                                                Ty1, VTy);
2026                 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
2027                   TargetTransformInfo::SK_ExtractSubvector, VTy, 0, Ty1));
2028               } else
2029                 ESContrib = (int) TTI->getVectorInstrCost(
2030                                     Instruction::ExtractElement, VTy, 0);
2031 
2032               DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
2033                 *S->first << "} = " << ESContrib << "\n");
2034               EffSize -= ESContrib;
2035             }
2036 
2037             NeedsExtraction = false;
2038             for (User *U : S->second->users()) {
2039               if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(U)) {
2040                 // Shuffle can be folded if it has no other input
2041                 if (isa<UndefValue>(SI->getOperand(1)))
2042                   continue;
2043               }
2044               if (isa<ExtractElementInst>(U))
2045                 continue;
2046               if (PrunedDAGInstrs.count(U))
2047                 continue;
2048               NeedsExtraction = true;
2049               break;
2050             }
2051 
2052             if (NeedsExtraction) {
2053               int ESContrib;
2054               if (Ty2->isVectorTy()) {
2055                 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2056                                                Ty2, VTy);
2057                 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
2058                   TargetTransformInfo::SK_ExtractSubvector, VTy,
2059                   Ty1->isVectorTy() ? Ty1->getVectorNumElements() : 1, Ty2));
2060               } else
2061                 ESContrib = (int) TTI->getVectorInstrCost(
2062                                     Instruction::ExtractElement, VTy, 1);
2063               DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
2064                 *S->second << "} = " << ESContrib << "\n");
2065               EffSize -= ESContrib;
2066             }
2067           }
2068 
2069           // Compute the cost of incoming edges.
2070           if (!isa<LoadInst>(S->first) && !isa<StoreInst>(S->first)) {
2071             Instruction *S1 = cast<Instruction>(S->first),
2072                         *S2 = cast<Instruction>(S->second);
2073             for (unsigned o = 0; o < S1->getNumOperands(); ++o) {
2074               Value *O1 = S1->getOperand(o), *O2 = S2->getOperand(o);
2075 
2076               // Combining constants into vector constants (or small vector
2077               // constants into larger ones are assumed free).
2078               if (isa<Constant>(O1) && isa<Constant>(O2))
2079                 continue;
2080 
2081               if (FlipOrder)
2082                 std::swap(O1, O2);
2083 
2084               ValuePair VP  = ValuePair(O1, O2);
2085               ValuePair VPR = ValuePair(O2, O1);
2086 
2087               // Internal edges are not handled here.
2088               if (PrunedDAG.count(VP) || PrunedDAG.count(VPR))
2089                 continue;
2090 
2091               Type *Ty1 = O1->getType(),
2092                    *Ty2 = O2->getType();
2093               Type *VTy = getVecTypeForPair(Ty1, Ty2);
2094 
2095               // Combining vector operations of the same type is also assumed
2096               // folded with other operations.
2097               if (Ty1 == Ty2) {
2098                 // If both are insert elements, then both can be widened.
2099                 InsertElementInst *IEO1 = dyn_cast<InsertElementInst>(O1),
2100                                   *IEO2 = dyn_cast<InsertElementInst>(O2);
2101                 if (IEO1 && IEO2 && isPureIEChain(IEO1) && isPureIEChain(IEO2))
2102                   continue;
2103                 // If both are extract elements, and both have the same input
2104                 // type, then they can be replaced with a shuffle
2105                 ExtractElementInst *EIO1 = dyn_cast<ExtractElementInst>(O1),
2106                                    *EIO2 = dyn_cast<ExtractElementInst>(O2);
2107                 if (EIO1 && EIO2 &&
2108                     EIO1->getOperand(0)->getType() ==
2109                       EIO2->getOperand(0)->getType())
2110                   continue;
2111                 // If both are a shuffle with equal operand types and only two
2112                 // unqiue operands, then they can be replaced with a single
2113                 // shuffle
2114                 ShuffleVectorInst *SIO1 = dyn_cast<ShuffleVectorInst>(O1),
2115                                   *SIO2 = dyn_cast<ShuffleVectorInst>(O2);
2116                 if (SIO1 && SIO2 &&
2117                     SIO1->getOperand(0)->getType() ==
2118                       SIO2->getOperand(0)->getType()) {
2119                   SmallSet<Value *, 4> SIOps;
2120                   SIOps.insert(SIO1->getOperand(0));
2121                   SIOps.insert(SIO1->getOperand(1));
2122                   SIOps.insert(SIO2->getOperand(0));
2123                   SIOps.insert(SIO2->getOperand(1));
2124                   if (SIOps.size() <= 2)
2125                     continue;
2126                 }
2127               }
2128 
2129               int ESContrib;
2130               // This pair has already been formed.
2131               if (IncomingPairs.count(VP)) {
2132                 continue;
2133               } else if (IncomingPairs.count(VPR)) {
2134                 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2135                                                VTy, VTy);
2136 
2137                 if (VTy->getVectorNumElements() == 2)
2138                   ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
2139                     TargetTransformInfo::SK_Reverse, VTy));
2140               } else if (!Ty1->isVectorTy() && !Ty2->isVectorTy()) {
2141                 ESContrib = (int) TTI->getVectorInstrCost(
2142                                     Instruction::InsertElement, VTy, 0);
2143                 ESContrib += (int) TTI->getVectorInstrCost(
2144                                      Instruction::InsertElement, VTy, 1);
2145               } else if (!Ty1->isVectorTy()) {
2146                 // O1 needs to be inserted into a vector of size O2, and then
2147                 // both need to be shuffled together.
2148                 ESContrib = (int) TTI->getVectorInstrCost(
2149                                     Instruction::InsertElement, Ty2, 0);
2150                 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2151                                                 VTy, Ty2);
2152               } else if (!Ty2->isVectorTy()) {
2153                 // O2 needs to be inserted into a vector of size O1, and then
2154                 // both need to be shuffled together.
2155                 ESContrib = (int) TTI->getVectorInstrCost(
2156                                     Instruction::InsertElement, Ty1, 0);
2157                 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2158                                                 VTy, Ty1);
2159               } else {
2160                 Type *TyBig = Ty1, *TySmall = Ty2;
2161                 if (Ty2->getVectorNumElements() > Ty1->getVectorNumElements())
2162                   std::swap(TyBig, TySmall);
2163 
2164                 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
2165                                                VTy, TyBig);
2166                 if (TyBig != TySmall)
2167                   ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
2168                                                   TyBig, TySmall);
2169               }
2170 
2171               DEBUG(if (DebugPairSelection) dbgs() << "\tcost {"
2172                      << *O1 << " <-> " << *O2 << "} = " <<
2173                      ESContrib << "\n");
2174               EffSize -= ESContrib;
2175               IncomingPairs.insert(VP);
2176             }
2177           }
2178         }
2179 
2180         if (!HasNontrivialInsts) {
2181           DEBUG(if (DebugPairSelection) dbgs() <<
2182                 "\tNo non-trivial instructions in DAG;"
2183                 " override to zero effective size\n");
2184           EffSize = 0;
2185         }
2186       } else {
2187         for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(),
2188              E = PrunedDAG.end(); S != E; ++S)
2189           EffSize += (int) getDepthFactor(S->first);
2190       }
2191 
2192       DEBUG(if (DebugPairSelection)
2193              dbgs() << "BBV: found pruned DAG for pair {"
2194              << *IJ.first << " <-> " << *IJ.second << "} of depth " <<
2195              MaxDepth << " and size " << PrunedDAG.size() <<
2196             " (effective size: " << EffSize << ")\n");
2197       if (((TTI && !UseChainDepthWithTI) ||
2198             MaxDepth >= Config.ReqChainDepth) &&
2199           EffSize > 0 && EffSize > BestEffSize) {
2200         BestMaxDepth = MaxDepth;
2201         BestEffSize = EffSize;
2202         BestDAG = PrunedDAG;
2203       }
2204     }
2205   }
2206 
2207   // Given the list of candidate pairs, this function selects those
2208   // that will be fused into vector instructions.
choosePairs(DenseMap<Value *,std::vector<Value * >> & CandidatePairs,DenseSet<ValuePair> & CandidatePairsSet,DenseMap<ValuePair,int> & CandidatePairCostSavings,std::vector<Value * > & PairableInsts,DenseSet<ValuePair> & FixedOrderPairs,DenseMap<VPPair,unsigned> & PairConnectionTypes,DenseMap<ValuePair,std::vector<ValuePair>> & ConnectedPairs,DenseMap<ValuePair,std::vector<ValuePair>> & ConnectedPairDeps,DenseSet<ValuePair> & PairableInstUsers,DenseMap<Value *,Value * > & ChosenPairs)2209   void BBVectorize::choosePairs(
2210                 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
2211                 DenseSet<ValuePair> &CandidatePairsSet,
2212                 DenseMap<ValuePair, int> &CandidatePairCostSavings,
2213                 std::vector<Value *> &PairableInsts,
2214                 DenseSet<ValuePair> &FixedOrderPairs,
2215                 DenseMap<VPPair, unsigned> &PairConnectionTypes,
2216                 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
2217                 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
2218                 DenseSet<ValuePair> &PairableInstUsers,
2219                 DenseMap<Value *, Value *>& ChosenPairs) {
2220     bool UseCycleCheck =
2221      CandidatePairsSet.size() <= Config.MaxCandPairsForCycleCheck;
2222 
2223     DenseMap<Value *, std::vector<Value *> > CandidatePairs2;
2224     for (DenseSet<ValuePair>::iterator I = CandidatePairsSet.begin(),
2225          E = CandidatePairsSet.end(); I != E; ++I) {
2226       std::vector<Value *> &JJ = CandidatePairs2[I->second];
2227       if (JJ.empty()) JJ.reserve(32);
2228       JJ.push_back(I->first);
2229     }
2230 
2231     DenseMap<ValuePair, std::vector<ValuePair> > PairableInstUserMap;
2232     DenseSet<VPPair> PairableInstUserPairSet;
2233     for (std::vector<Value *>::iterator I = PairableInsts.begin(),
2234          E = PairableInsts.end(); I != E; ++I) {
2235       // The number of possible pairings for this variable:
2236       size_t NumChoices = CandidatePairs.lookup(*I).size();
2237       if (!NumChoices) continue;
2238 
2239       std::vector<Value *> &JJ = CandidatePairs[*I];
2240 
2241       // The best pair to choose and its dag:
2242       size_t BestMaxDepth = 0;
2243       int BestEffSize = 0;
2244       DenseSet<ValuePair> BestDAG;
2245       findBestDAGFor(CandidatePairs, CandidatePairsSet,
2246                       CandidatePairCostSavings,
2247                       PairableInsts, FixedOrderPairs, PairConnectionTypes,
2248                       ConnectedPairs, ConnectedPairDeps,
2249                       PairableInstUsers, PairableInstUserMap,
2250                       PairableInstUserPairSet, ChosenPairs,
2251                       BestDAG, BestMaxDepth, BestEffSize, *I, JJ,
2252                       UseCycleCheck);
2253 
2254       if (BestDAG.empty())
2255         continue;
2256 
2257       // A dag has been chosen (or not) at this point. If no dag was
2258       // chosen, then this instruction, I, cannot be paired (and is no longer
2259       // considered).
2260 
2261       DEBUG(dbgs() << "BBV: selected pairs in the best DAG for: "
2262                    << *cast<Instruction>(*I) << "\n");
2263 
2264       for (DenseSet<ValuePair>::iterator S = BestDAG.begin(),
2265            SE2 = BestDAG.end(); S != SE2; ++S) {
2266         // Insert the members of this dag into the list of chosen pairs.
2267         ChosenPairs.insert(ValuePair(S->first, S->second));
2268         DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " <<
2269                *S->second << "\n");
2270 
2271         // Remove all candidate pairs that have values in the chosen dag.
2272         std::vector<Value *> &KK = CandidatePairs[S->first];
2273         for (std::vector<Value *>::iterator K = KK.begin(), KE = KK.end();
2274              K != KE; ++K) {
2275           if (*K == S->second)
2276             continue;
2277 
2278           CandidatePairsSet.erase(ValuePair(S->first, *K));
2279         }
2280 
2281         std::vector<Value *> &LL = CandidatePairs2[S->second];
2282         for (std::vector<Value *>::iterator L = LL.begin(), LE = LL.end();
2283              L != LE; ++L) {
2284           if (*L == S->first)
2285             continue;
2286 
2287           CandidatePairsSet.erase(ValuePair(*L, S->second));
2288         }
2289 
2290         std::vector<Value *> &MM = CandidatePairs[S->second];
2291         for (std::vector<Value *>::iterator M = MM.begin(), ME = MM.end();
2292              M != ME; ++M) {
2293           assert(*M != S->first && "Flipped pair in candidate list?");
2294           CandidatePairsSet.erase(ValuePair(S->second, *M));
2295         }
2296 
2297         std::vector<Value *> &NN = CandidatePairs2[S->first];
2298         for (std::vector<Value *>::iterator N = NN.begin(), NE = NN.end();
2299              N != NE; ++N) {
2300           assert(*N != S->second && "Flipped pair in candidate list?");
2301           CandidatePairsSet.erase(ValuePair(*N, S->first));
2302         }
2303       }
2304     }
2305 
2306     DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n");
2307   }
2308 
getReplacementName(Instruction * I,bool IsInput,unsigned o,unsigned n=0)2309   std::string getReplacementName(Instruction *I, bool IsInput, unsigned o,
2310                      unsigned n = 0) {
2311     if (!I->hasName())
2312       return "";
2313 
2314     return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) +
2315              (n > 0 ? "." + utostr(n) : "")).str();
2316   }
2317 
2318   // Returns the value that is to be used as the pointer input to the vector
2319   // instruction that fuses I with J.
getReplacementPointerInput(LLVMContext & Context,Instruction * I,Instruction * J,unsigned o)2320   Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context,
2321                      Instruction *I, Instruction *J, unsigned o) {
2322     Value *IPtr, *JPtr;
2323     unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
2324     int64_t OffsetInElmts;
2325 
2326     // Note: the analysis might fail here, that is why the pair order has
2327     // been precomputed (OffsetInElmts must be unused here).
2328     (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
2329                           IAddressSpace, JAddressSpace,
2330                           OffsetInElmts, false);
2331 
2332     // The pointer value is taken to be the one with the lowest offset.
2333     Value *VPtr = IPtr;
2334 
2335     Type *ArgTypeI = IPtr->getType()->getPointerElementType();
2336     Type *ArgTypeJ = JPtr->getType()->getPointerElementType();
2337     Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2338     Type *VArgPtrType
2339       = PointerType::get(VArgType,
2340                          IPtr->getType()->getPointerAddressSpace());
2341     return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o),
2342                         /* insert before */ I);
2343   }
2344 
fillNewShuffleMask(LLVMContext & Context,Instruction * J,unsigned MaskOffset,unsigned NumInElem,unsigned NumInElem1,unsigned IdxOffset,std::vector<Constant * > & Mask)2345   void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J,
2346                      unsigned MaskOffset, unsigned NumInElem,
2347                      unsigned NumInElem1, unsigned IdxOffset,
2348                      std::vector<Constant*> &Mask) {
2349     unsigned NumElem1 = J->getType()->getVectorNumElements();
2350     for (unsigned v = 0; v < NumElem1; ++v) {
2351       int m = cast<ShuffleVectorInst>(J)->getMaskValue(v);
2352       if (m < 0) {
2353         Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context));
2354       } else {
2355         unsigned mm = m + (int) IdxOffset;
2356         if (m >= (int) NumInElem1)
2357           mm += (int) NumInElem;
2358 
2359         Mask[v+MaskOffset] =
2360           ConstantInt::get(Type::getInt32Ty(Context), mm);
2361       }
2362     }
2363   }
2364 
2365   // Returns the value that is to be used as the vector-shuffle mask to the
2366   // vector instruction that fuses I with J.
getReplacementShuffleMask(LLVMContext & Context,Instruction * I,Instruction * J)2367   Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context,
2368                      Instruction *I, Instruction *J) {
2369     // This is the shuffle mask. We need to append the second
2370     // mask to the first, and the numbers need to be adjusted.
2371 
2372     Type *ArgTypeI = I->getType();
2373     Type *ArgTypeJ = J->getType();
2374     Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2375 
2376     unsigned NumElemI = ArgTypeI->getVectorNumElements();
2377 
2378     // Get the total number of elements in the fused vector type.
2379     // By definition, this must equal the number of elements in
2380     // the final mask.
2381     unsigned NumElem = VArgType->getVectorNumElements();
2382     std::vector<Constant*> Mask(NumElem);
2383 
2384     Type *OpTypeI = I->getOperand(0)->getType();
2385     unsigned NumInElemI = OpTypeI->getVectorNumElements();
2386     Type *OpTypeJ = J->getOperand(0)->getType();
2387     unsigned NumInElemJ = OpTypeJ->getVectorNumElements();
2388 
2389     // The fused vector will be:
2390     // -----------------------------------------------------
2391     // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ |
2392     // -----------------------------------------------------
2393     // from which we'll extract NumElem total elements (where the first NumElemI
2394     // of them come from the mask in I and the remainder come from the mask
2395     // in J.
2396 
2397     // For the mask from the first pair...
2398     fillNewShuffleMask(Context, I, 0,        NumInElemJ, NumInElemI,
2399                        0,          Mask);
2400 
2401     // For the mask from the second pair...
2402     fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ,
2403                        NumInElemI, Mask);
2404 
2405     return ConstantVector::get(Mask);
2406   }
2407 
expandIEChain(LLVMContext & Context,Instruction * I,Instruction * J,unsigned o,Value * & LOp,unsigned numElemL,Type * ArgTypeL,Type * ArgTypeH,bool IBeforeJ,unsigned IdxOff)2408   bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I,
2409                                   Instruction *J, unsigned o, Value *&LOp,
2410                                   unsigned numElemL,
2411                                   Type *ArgTypeL, Type *ArgTypeH,
2412                                   bool IBeforeJ, unsigned IdxOff) {
2413     bool ExpandedIEChain = false;
2414     if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) {
2415       // If we have a pure insertelement chain, then this can be rewritten
2416       // into a chain that directly builds the larger type.
2417       if (isPureIEChain(LIE)) {
2418         SmallVector<Value *, 8> VectElemts(numElemL,
2419           UndefValue::get(ArgTypeL->getScalarType()));
2420         InsertElementInst *LIENext = LIE;
2421         do {
2422           unsigned Idx =
2423             cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue();
2424           VectElemts[Idx] = LIENext->getOperand(1);
2425         } while ((LIENext =
2426                    dyn_cast<InsertElementInst>(LIENext->getOperand(0))));
2427 
2428         LIENext = nullptr;
2429         Value *LIEPrev = UndefValue::get(ArgTypeH);
2430         for (unsigned i = 0; i < numElemL; ++i) {
2431           if (isa<UndefValue>(VectElemts[i])) continue;
2432           LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i],
2433                              ConstantInt::get(Type::getInt32Ty(Context),
2434                                               i + IdxOff),
2435                              getReplacementName(IBeforeJ ? I : J,
2436                                                 true, o, i+1));
2437           LIENext->insertBefore(IBeforeJ ? J : I);
2438           LIEPrev = LIENext;
2439         }
2440 
2441         LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH);
2442         ExpandedIEChain = true;
2443       }
2444     }
2445 
2446     return ExpandedIEChain;
2447   }
2448 
getNumScalarElements(Type * Ty)2449   static unsigned getNumScalarElements(Type *Ty) {
2450     if (VectorType *VecTy = dyn_cast<VectorType>(Ty))
2451       return VecTy->getNumElements();
2452     return 1;
2453   }
2454 
2455   // Returns the value to be used as the specified operand of the vector
2456   // instruction that fuses I with J.
getReplacementInput(LLVMContext & Context,Instruction * I,Instruction * J,unsigned o,bool IBeforeJ)2457   Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I,
2458                      Instruction *J, unsigned o, bool IBeforeJ) {
2459     Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2460     Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1);
2461 
2462     // Compute the fused vector type for this operand
2463     Type *ArgTypeI = I->getOperand(o)->getType();
2464     Type *ArgTypeJ = J->getOperand(o)->getType();
2465     VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2466 
2467     Instruction *L = I, *H = J;
2468     Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ;
2469 
2470     unsigned numElemL = getNumScalarElements(ArgTypeL);
2471     unsigned numElemH = getNumScalarElements(ArgTypeH);
2472 
2473     Value *LOp = L->getOperand(o);
2474     Value *HOp = H->getOperand(o);
2475     unsigned numElem = VArgType->getNumElements();
2476 
2477     // First, we check if we can reuse the "original" vector outputs (if these
2478     // exist). We might need a shuffle.
2479     ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp);
2480     ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp);
2481     ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp);
2482     ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp);
2483 
2484     // FIXME: If we're fusing shuffle instructions, then we can't apply this
2485     // optimization. The input vectors to the shuffle might be a different
2486     // length from the shuffle outputs. Unfortunately, the replacement
2487     // shuffle mask has already been formed, and the mask entries are sensitive
2488     // to the sizes of the inputs.
2489     bool IsSizeChangeShuffle =
2490       isa<ShuffleVectorInst>(L) &&
2491         (LOp->getType() != L->getType() || HOp->getType() != H->getType());
2492 
2493     if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) {
2494       // We can have at most two unique vector inputs.
2495       bool CanUseInputs = true;
2496       Value *I1, *I2 = nullptr;
2497       if (LEE) {
2498         I1 = LEE->getOperand(0);
2499       } else {
2500         I1 = LSV->getOperand(0);
2501         I2 = LSV->getOperand(1);
2502         if (I2 == I1 || isa<UndefValue>(I2))
2503           I2 = nullptr;
2504       }
2505 
2506       if (HEE) {
2507         Value *I3 = HEE->getOperand(0);
2508         if (!I2 && I3 != I1)
2509           I2 = I3;
2510         else if (I3 != I1 && I3 != I2)
2511           CanUseInputs = false;
2512       } else {
2513         Value *I3 = HSV->getOperand(0);
2514         if (!I2 && I3 != I1)
2515           I2 = I3;
2516         else if (I3 != I1 && I3 != I2)
2517           CanUseInputs = false;
2518 
2519         if (CanUseInputs) {
2520           Value *I4 = HSV->getOperand(1);
2521           if (!isa<UndefValue>(I4)) {
2522             if (!I2 && I4 != I1)
2523               I2 = I4;
2524             else if (I4 != I1 && I4 != I2)
2525               CanUseInputs = false;
2526           }
2527         }
2528       }
2529 
2530       if (CanUseInputs) {
2531         unsigned LOpElem =
2532           cast<Instruction>(LOp)->getOperand(0)->getType()
2533             ->getVectorNumElements();
2534 
2535         unsigned HOpElem =
2536           cast<Instruction>(HOp)->getOperand(0)->getType()
2537             ->getVectorNumElements();
2538 
2539         // We have one or two input vectors. We need to map each index of the
2540         // operands to the index of the original vector.
2541         SmallVector<std::pair<int, int>, 8>  II(numElem);
2542         for (unsigned i = 0; i < numElemL; ++i) {
2543           int Idx, INum;
2544           if (LEE) {
2545             Idx =
2546               cast<ConstantInt>(LEE->getOperand(1))->getSExtValue();
2547             INum = LEE->getOperand(0) == I1 ? 0 : 1;
2548           } else {
2549             Idx = LSV->getMaskValue(i);
2550             if (Idx < (int) LOpElem) {
2551               INum = LSV->getOperand(0) == I1 ? 0 : 1;
2552             } else {
2553               Idx -= LOpElem;
2554               INum = LSV->getOperand(1) == I1 ? 0 : 1;
2555             }
2556           }
2557 
2558           II[i] = std::pair<int, int>(Idx, INum);
2559         }
2560         for (unsigned i = 0; i < numElemH; ++i) {
2561           int Idx, INum;
2562           if (HEE) {
2563             Idx =
2564               cast<ConstantInt>(HEE->getOperand(1))->getSExtValue();
2565             INum = HEE->getOperand(0) == I1 ? 0 : 1;
2566           } else {
2567             Idx = HSV->getMaskValue(i);
2568             if (Idx < (int) HOpElem) {
2569               INum = HSV->getOperand(0) == I1 ? 0 : 1;
2570             } else {
2571               Idx -= HOpElem;
2572               INum = HSV->getOperand(1) == I1 ? 0 : 1;
2573             }
2574           }
2575 
2576           II[i + numElemL] = std::pair<int, int>(Idx, INum);
2577         }
2578 
2579         // We now have an array which tells us from which index of which
2580         // input vector each element of the operand comes.
2581         VectorType *I1T = cast<VectorType>(I1->getType());
2582         unsigned I1Elem = I1T->getNumElements();
2583 
2584         if (!I2) {
2585           // In this case there is only one underlying vector input. Check for
2586           // the trivial case where we can use the input directly.
2587           if (I1Elem == numElem) {
2588             bool ElemInOrder = true;
2589             for (unsigned i = 0; i < numElem; ++i) {
2590               if (II[i].first != (int) i && II[i].first != -1) {
2591                 ElemInOrder = false;
2592                 break;
2593               }
2594             }
2595 
2596             if (ElemInOrder)
2597               return I1;
2598           }
2599 
2600           // A shuffle is needed.
2601           std::vector<Constant *> Mask(numElem);
2602           for (unsigned i = 0; i < numElem; ++i) {
2603             int Idx = II[i].first;
2604             if (Idx == -1)
2605               Mask[i] = UndefValue::get(Type::getInt32Ty(Context));
2606             else
2607               Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2608           }
2609 
2610           Instruction *S =
2611             new ShuffleVectorInst(I1, UndefValue::get(I1T),
2612                                   ConstantVector::get(Mask),
2613                                   getReplacementName(IBeforeJ ? I : J,
2614                                                      true, o));
2615           S->insertBefore(IBeforeJ ? J : I);
2616           return S;
2617         }
2618 
2619         VectorType *I2T = cast<VectorType>(I2->getType());
2620         unsigned I2Elem = I2T->getNumElements();
2621 
2622         // This input comes from two distinct vectors. The first step is to
2623         // make sure that both vectors are the same length. If not, the
2624         // smaller one will need to grow before they can be shuffled together.
2625         if (I1Elem < I2Elem) {
2626           std::vector<Constant *> Mask(I2Elem);
2627           unsigned v = 0;
2628           for (; v < I1Elem; ++v)
2629             Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2630           for (; v < I2Elem; ++v)
2631             Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2632 
2633           Instruction *NewI1 =
2634             new ShuffleVectorInst(I1, UndefValue::get(I1T),
2635                                   ConstantVector::get(Mask),
2636                                   getReplacementName(IBeforeJ ? I : J,
2637                                                      true, o, 1));
2638           NewI1->insertBefore(IBeforeJ ? J : I);
2639           I1 = NewI1;
2640           I1Elem = I2Elem;
2641         } else if (I1Elem > I2Elem) {
2642           std::vector<Constant *> Mask(I1Elem);
2643           unsigned v = 0;
2644           for (; v < I2Elem; ++v)
2645             Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2646           for (; v < I1Elem; ++v)
2647             Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2648 
2649           Instruction *NewI2 =
2650             new ShuffleVectorInst(I2, UndefValue::get(I2T),
2651                                   ConstantVector::get(Mask),
2652                                   getReplacementName(IBeforeJ ? I : J,
2653                                                      true, o, 1));
2654           NewI2->insertBefore(IBeforeJ ? J : I);
2655           I2 = NewI2;
2656         }
2657 
2658         // Now that both I1 and I2 are the same length we can shuffle them
2659         // together (and use the result).
2660         std::vector<Constant *> Mask(numElem);
2661         for (unsigned v = 0; v < numElem; ++v) {
2662           if (II[v].first == -1) {
2663             Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2664           } else {
2665             int Idx = II[v].first + II[v].second * I1Elem;
2666             Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2667           }
2668         }
2669 
2670         Instruction *NewOp =
2671           new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask),
2672                                 getReplacementName(IBeforeJ ? I : J, true, o));
2673         NewOp->insertBefore(IBeforeJ ? J : I);
2674         return NewOp;
2675       }
2676     }
2677 
2678     Type *ArgType = ArgTypeL;
2679     if (numElemL < numElemH) {
2680       if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH,
2681                                          ArgTypeL, VArgType, IBeforeJ, 1)) {
2682         // This is another short-circuit case: we're combining a scalar into
2683         // a vector that is formed by an IE chain. We've just expanded the IE
2684         // chain, now insert the scalar and we're done.
2685 
2686         Instruction *S = InsertElementInst::Create(HOp, LOp, CV0,
2687                            getReplacementName(IBeforeJ ? I : J, true, o));
2688         S->insertBefore(IBeforeJ ? J : I);
2689         return S;
2690       } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL,
2691                                 ArgTypeH, IBeforeJ)) {
2692         // The two vector inputs to the shuffle must be the same length,
2693         // so extend the smaller vector to be the same length as the larger one.
2694         Instruction *NLOp;
2695         if (numElemL > 1) {
2696 
2697           std::vector<Constant *> Mask(numElemH);
2698           unsigned v = 0;
2699           for (; v < numElemL; ++v)
2700             Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2701           for (; v < numElemH; ++v)
2702             Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2703 
2704           NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL),
2705                                        ConstantVector::get(Mask),
2706                                        getReplacementName(IBeforeJ ? I : J,
2707                                                           true, o, 1));
2708         } else {
2709           NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0,
2710                                            getReplacementName(IBeforeJ ? I : J,
2711                                                               true, o, 1));
2712         }
2713 
2714         NLOp->insertBefore(IBeforeJ ? J : I);
2715         LOp = NLOp;
2716       }
2717 
2718       ArgType = ArgTypeH;
2719     } else if (numElemL > numElemH) {
2720       if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL,
2721                                          ArgTypeH, VArgType, IBeforeJ)) {
2722         Instruction *S =
2723           InsertElementInst::Create(LOp, HOp,
2724                                     ConstantInt::get(Type::getInt32Ty(Context),
2725                                                      numElemL),
2726                                     getReplacementName(IBeforeJ ? I : J,
2727                                                        true, o));
2728         S->insertBefore(IBeforeJ ? J : I);
2729         return S;
2730       } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH,
2731                                 ArgTypeL, IBeforeJ)) {
2732         Instruction *NHOp;
2733         if (numElemH > 1) {
2734           std::vector<Constant *> Mask(numElemL);
2735           unsigned v = 0;
2736           for (; v < numElemH; ++v)
2737             Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2738           for (; v < numElemL; ++v)
2739             Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
2740 
2741           NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH),
2742                                        ConstantVector::get(Mask),
2743                                        getReplacementName(IBeforeJ ? I : J,
2744                                                           true, o, 1));
2745         } else {
2746           NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0,
2747                                            getReplacementName(IBeforeJ ? I : J,
2748                                                               true, o, 1));
2749         }
2750 
2751         NHOp->insertBefore(IBeforeJ ? J : I);
2752         HOp = NHOp;
2753       }
2754     }
2755 
2756     if (ArgType->isVectorTy()) {
2757       unsigned numElem = VArgType->getVectorNumElements();
2758       std::vector<Constant*> Mask(numElem);
2759       for (unsigned v = 0; v < numElem; ++v) {
2760         unsigned Idx = v;
2761         // If the low vector was expanded, we need to skip the extra
2762         // undefined entries.
2763         if (v >= numElemL && numElemH > numElemL)
2764           Idx += (numElemH - numElemL);
2765         Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
2766       }
2767 
2768       Instruction *BV = new ShuffleVectorInst(LOp, HOp,
2769                           ConstantVector::get(Mask),
2770                           getReplacementName(IBeforeJ ? I : J, true, o));
2771       BV->insertBefore(IBeforeJ ? J : I);
2772       return BV;
2773     }
2774 
2775     Instruction *BV1 = InsertElementInst::Create(
2776                                           UndefValue::get(VArgType), LOp, CV0,
2777                                           getReplacementName(IBeforeJ ? I : J,
2778                                                              true, o, 1));
2779     BV1->insertBefore(IBeforeJ ? J : I);
2780     Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1,
2781                                           getReplacementName(IBeforeJ ? I : J,
2782                                                              true, o, 2));
2783     BV2->insertBefore(IBeforeJ ? J : I);
2784     return BV2;
2785   }
2786 
2787   // This function creates an array of values that will be used as the inputs
2788   // to the vector instruction that fuses I with J.
getReplacementInputsForPair(LLVMContext & Context,Instruction * I,Instruction * J,SmallVectorImpl<Value * > & ReplacedOperands,bool IBeforeJ)2789   void BBVectorize::getReplacementInputsForPair(LLVMContext& Context,
2790                      Instruction *I, Instruction *J,
2791                      SmallVectorImpl<Value *> &ReplacedOperands,
2792                      bool IBeforeJ) {
2793     unsigned NumOperands = I->getNumOperands();
2794 
2795     for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) {
2796       // Iterate backward so that we look at the store pointer
2797       // first and know whether or not we need to flip the inputs.
2798 
2799       if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) {
2800         // This is the pointer for a load/store instruction.
2801         ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o);
2802         continue;
2803       } else if (isa<CallInst>(I)) {
2804         Function *F = cast<CallInst>(I)->getCalledFunction();
2805         Intrinsic::ID IID = F->getIntrinsicID();
2806         if (o == NumOperands-1) {
2807           BasicBlock &BB = *I->getParent();
2808 
2809           Module *M = BB.getParent()->getParent();
2810           Type *ArgTypeI = I->getType();
2811           Type *ArgTypeJ = J->getType();
2812           Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
2813 
2814           ReplacedOperands[o] = Intrinsic::getDeclaration(M, IID, VArgType);
2815           continue;
2816         } else if ((IID == Intrinsic::powi || IID == Intrinsic::ctlz ||
2817                     IID == Intrinsic::cttz) && o == 1) {
2818           // The second argument of powi/ctlz/cttz is a single integer/constant
2819           // and we've already checked that both arguments are equal.
2820           // As a result, we just keep I's second argument.
2821           ReplacedOperands[o] = I->getOperand(o);
2822           continue;
2823         }
2824       } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) {
2825         ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J);
2826         continue;
2827       }
2828 
2829       ReplacedOperands[o] = getReplacementInput(Context, I, J, o, IBeforeJ);
2830     }
2831   }
2832 
2833   // This function creates two values that represent the outputs of the
2834   // original I and J instructions. These are generally vector shuffles
2835   // or extracts. In many cases, these will end up being unused and, thus,
2836   // eliminated by later passes.
replaceOutputsOfPair(LLVMContext & Context,Instruction * I,Instruction * J,Instruction * K,Instruction * & InsertionPt,Instruction * & K1,Instruction * & K2)2837   void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
2838                      Instruction *J, Instruction *K,
2839                      Instruction *&InsertionPt,
2840                      Instruction *&K1, Instruction *&K2) {
2841     if (isa<StoreInst>(I))
2842       return;
2843 
2844     Type *IType = I->getType();
2845     Type *JType = J->getType();
2846 
2847     VectorType *VType = getVecTypeForPair(IType, JType);
2848     unsigned numElem = VType->getNumElements();
2849 
2850     unsigned numElemI = getNumScalarElements(IType);
2851     unsigned numElemJ = getNumScalarElements(JType);
2852 
2853     if (IType->isVectorTy()) {
2854       std::vector<Constant *> Mask1(numElemI), Mask2(numElemI);
2855       for (unsigned v = 0; v < numElemI; ++v) {
2856         Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2857         Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ + v);
2858       }
2859 
2860       K1 = new ShuffleVectorInst(K, UndefValue::get(VType),
2861                                  ConstantVector::get(Mask1),
2862                                  getReplacementName(K, false, 1));
2863     } else {
2864       Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
2865       K1 = ExtractElementInst::Create(K, CV0, getReplacementName(K, false, 1));
2866     }
2867 
2868     if (JType->isVectorTy()) {
2869       std::vector<Constant *> Mask1(numElemJ), Mask2(numElemJ);
2870       for (unsigned v = 0; v < numElemJ; ++v) {
2871         Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
2872         Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI + v);
2873       }
2874 
2875       K2 = new ShuffleVectorInst(K, UndefValue::get(VType),
2876                                  ConstantVector::get(Mask2),
2877                                  getReplacementName(K, false, 2));
2878     } else {
2879       Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem - 1);
2880       K2 = ExtractElementInst::Create(K, CV1, getReplacementName(K, false, 2));
2881     }
2882 
2883     K1->insertAfter(K);
2884     K2->insertAfter(K1);
2885     InsertionPt = K2;
2886   }
2887 
2888   // Move all uses of the function I (including pairing-induced uses) after J.
canMoveUsesOfIAfterJ(BasicBlock & BB,DenseSet<ValuePair> & LoadMoveSetPairs,Instruction * I,Instruction * J)2889   bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB,
2890                      DenseSet<ValuePair> &LoadMoveSetPairs,
2891                      Instruction *I, Instruction *J) {
2892     // Skip to the first instruction past I.
2893     BasicBlock::iterator L = std::next(BasicBlock::iterator(I));
2894 
2895     DenseSet<Value *> Users;
2896     AliasSetTracker WriteSet(*AA);
2897     if (I->mayWriteToMemory()) WriteSet.add(I);
2898 
2899     for (; cast<Instruction>(L) != J; ++L)
2900       (void)trackUsesOfI(Users, WriteSet, I, &*L, true, &LoadMoveSetPairs);
2901 
2902     assert(cast<Instruction>(L) == J &&
2903       "Tracking has not proceeded far enough to check for dependencies");
2904     // If J is now in the use set of I, then trackUsesOfI will return true
2905     // and we have a dependency cycle (and the fusing operation must abort).
2906     return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSetPairs);
2907   }
2908 
2909   // Move all uses of the function I (including pairing-induced uses) after J.
moveUsesOfIAfterJ(BasicBlock & BB,DenseSet<ValuePair> & LoadMoveSetPairs,Instruction * & InsertionPt,Instruction * I,Instruction * J)2910   void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB,
2911                      DenseSet<ValuePair> &LoadMoveSetPairs,
2912                      Instruction *&InsertionPt,
2913                      Instruction *I, Instruction *J) {
2914     // Skip to the first instruction past I.
2915     BasicBlock::iterator L = std::next(BasicBlock::iterator(I));
2916 
2917     DenseSet<Value *> Users;
2918     AliasSetTracker WriteSet(*AA);
2919     if (I->mayWriteToMemory()) WriteSet.add(I);
2920 
2921     for (; cast<Instruction>(L) != J;) {
2922       if (trackUsesOfI(Users, WriteSet, I, &*L, true, &LoadMoveSetPairs)) {
2923         // Move this instruction
2924         Instruction *InstToMove = &*L++;
2925 
2926         DEBUG(dbgs() << "BBV: moving: " << *InstToMove <<
2927                         " to after " << *InsertionPt << "\n");
2928         InstToMove->removeFromParent();
2929         InstToMove->insertAfter(InsertionPt);
2930         InsertionPt = InstToMove;
2931       } else {
2932         ++L;
2933       }
2934     }
2935   }
2936 
2937   // Collect all load instruction that are in the move set of a given first
2938   // pair member.  These loads depend on the first instruction, I, and so need
2939   // to be moved after J (the second instruction) when the pair is fused.
collectPairLoadMoveSet(BasicBlock & BB,DenseMap<Value *,Value * > & ChosenPairs,DenseMap<Value *,std::vector<Value * >> & LoadMoveSet,DenseSet<ValuePair> & LoadMoveSetPairs,Instruction * I)2940   void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB,
2941                      DenseMap<Value *, Value *> &ChosenPairs,
2942                      DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
2943                      DenseSet<ValuePair> &LoadMoveSetPairs,
2944                      Instruction *I) {
2945     // Skip to the first instruction past I.
2946     BasicBlock::iterator L = std::next(BasicBlock::iterator(I));
2947 
2948     DenseSet<Value *> Users;
2949     AliasSetTracker WriteSet(*AA);
2950     if (I->mayWriteToMemory()) WriteSet.add(I);
2951 
2952     // Note: We cannot end the loop when we reach J because J could be moved
2953     // farther down the use chain by another instruction pairing. Also, J
2954     // could be before I if this is an inverted input.
2955     for (BasicBlock::iterator E = BB.end(); L != E; ++L) {
2956       if (trackUsesOfI(Users, WriteSet, I, &*L)) {
2957         if (L->mayReadFromMemory()) {
2958           LoadMoveSet[&*L].push_back(I);
2959           LoadMoveSetPairs.insert(ValuePair(&*L, I));
2960         }
2961       }
2962     }
2963   }
2964 
2965   // In cases where both load/stores and the computation of their pointers
2966   // are chosen for vectorization, we can end up in a situation where the
2967   // aliasing analysis starts returning different query results as the
2968   // process of fusing instruction pairs continues. Because the algorithm
2969   // relies on finding the same use dags here as were found earlier, we'll
2970   // need to precompute the necessary aliasing information here and then
2971   // manually update it during the fusion process.
collectLoadMoveSet(BasicBlock & BB,std::vector<Value * > & PairableInsts,DenseMap<Value *,Value * > & ChosenPairs,DenseMap<Value *,std::vector<Value * >> & LoadMoveSet,DenseSet<ValuePair> & LoadMoveSetPairs)2972   void BBVectorize::collectLoadMoveSet(BasicBlock &BB,
2973                      std::vector<Value *> &PairableInsts,
2974                      DenseMap<Value *, Value *> &ChosenPairs,
2975                      DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
2976                      DenseSet<ValuePair> &LoadMoveSetPairs) {
2977     for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
2978          PIE = PairableInsts.end(); PI != PIE; ++PI) {
2979       DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI);
2980       if (P == ChosenPairs.end()) continue;
2981 
2982       Instruction *I = cast<Instruction>(P->first);
2983       collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet,
2984                              LoadMoveSetPairs, I);
2985     }
2986   }
2987 
2988   // This function fuses the chosen instruction pairs into vector instructions,
2989   // taking care preserve any needed scalar outputs and, then, it reorders the
2990   // remaining instructions as needed (users of the first member of the pair
2991   // need to be moved to after the location of the second member of the pair
2992   // because the vector instruction is inserted in the location of the pair's
2993   // second member).
fuseChosenPairs(BasicBlock & BB,std::vector<Value * > & PairableInsts,DenseMap<Value *,Value * > & ChosenPairs,DenseSet<ValuePair> & FixedOrderPairs,DenseMap<VPPair,unsigned> & PairConnectionTypes,DenseMap<ValuePair,std::vector<ValuePair>> & ConnectedPairs,DenseMap<ValuePair,std::vector<ValuePair>> & ConnectedPairDeps)2994   void BBVectorize::fuseChosenPairs(BasicBlock &BB,
2995              std::vector<Value *> &PairableInsts,
2996              DenseMap<Value *, Value *> &ChosenPairs,
2997              DenseSet<ValuePair> &FixedOrderPairs,
2998              DenseMap<VPPair, unsigned> &PairConnectionTypes,
2999              DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
3000              DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps) {
3001     LLVMContext& Context = BB.getContext();
3002 
3003     // During the vectorization process, the order of the pairs to be fused
3004     // could be flipped. So we'll add each pair, flipped, into the ChosenPairs
3005     // list. After a pair is fused, the flipped pair is removed from the list.
3006     DenseSet<ValuePair> FlippedPairs;
3007     for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(),
3008          E = ChosenPairs.end(); P != E; ++P)
3009       FlippedPairs.insert(ValuePair(P->second, P->first));
3010     for (DenseSet<ValuePair>::iterator P = FlippedPairs.begin(),
3011          E = FlippedPairs.end(); P != E; ++P)
3012       ChosenPairs.insert(*P);
3013 
3014     DenseMap<Value *, std::vector<Value *> > LoadMoveSet;
3015     DenseSet<ValuePair> LoadMoveSetPairs;
3016     collectLoadMoveSet(BB, PairableInsts, ChosenPairs,
3017                        LoadMoveSet, LoadMoveSetPairs);
3018 
3019     DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n");
3020 
3021     for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) {
3022       DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(&*PI);
3023       if (P == ChosenPairs.end()) {
3024         ++PI;
3025         continue;
3026       }
3027 
3028       if (getDepthFactor(P->first) == 0) {
3029         // These instructions are not really fused, but are tracked as though
3030         // they are. Any case in which it would be interesting to fuse them
3031         // will be taken care of by InstCombine.
3032         --NumFusedOps;
3033         ++PI;
3034         continue;
3035       }
3036 
3037       Instruction *I = cast<Instruction>(P->first),
3038         *J = cast<Instruction>(P->second);
3039 
3040       DEBUG(dbgs() << "BBV: fusing: " << *I <<
3041              " <-> " << *J << "\n");
3042 
3043       // Remove the pair and flipped pair from the list.
3044       DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second);
3045       assert(FP != ChosenPairs.end() && "Flipped pair not found in list");
3046       ChosenPairs.erase(FP);
3047       ChosenPairs.erase(P);
3048 
3049       if (!canMoveUsesOfIAfterJ(BB, LoadMoveSetPairs, I, J)) {
3050         DEBUG(dbgs() << "BBV: fusion of: " << *I <<
3051                " <-> " << *J <<
3052                " aborted because of non-trivial dependency cycle\n");
3053         --NumFusedOps;
3054         ++PI;
3055         continue;
3056       }
3057 
3058       // If the pair must have the other order, then flip it.
3059       bool FlipPairOrder = FixedOrderPairs.count(ValuePair(J, I));
3060       if (!FlipPairOrder && !FixedOrderPairs.count(ValuePair(I, J))) {
3061         // This pair does not have a fixed order, and so we might want to
3062         // flip it if that will yield fewer shuffles. We count the number
3063         // of dependencies connected via swaps, and those directly connected,
3064         // and flip the order if the number of swaps is greater.
3065         bool OrigOrder = true;
3066         DenseMap<ValuePair, std::vector<ValuePair> >::iterator IJ =
3067           ConnectedPairDeps.find(ValuePair(I, J));
3068         if (IJ == ConnectedPairDeps.end()) {
3069           IJ = ConnectedPairDeps.find(ValuePair(J, I));
3070           OrigOrder = false;
3071         }
3072 
3073         if (IJ != ConnectedPairDeps.end()) {
3074           unsigned NumDepsDirect = 0, NumDepsSwap = 0;
3075           for (std::vector<ValuePair>::iterator T = IJ->second.begin(),
3076                TE = IJ->second.end(); T != TE; ++T) {
3077             VPPair Q(IJ->first, *T);
3078             DenseMap<VPPair, unsigned>::iterator R =
3079               PairConnectionTypes.find(VPPair(Q.second, Q.first));
3080             assert(R != PairConnectionTypes.end() &&
3081                    "Cannot find pair connection type");
3082             if (R->second == PairConnectionDirect)
3083               ++NumDepsDirect;
3084             else if (R->second == PairConnectionSwap)
3085               ++NumDepsSwap;
3086           }
3087 
3088           if (!OrigOrder)
3089             std::swap(NumDepsDirect, NumDepsSwap);
3090 
3091           if (NumDepsSwap > NumDepsDirect) {
3092             FlipPairOrder = true;
3093             DEBUG(dbgs() << "BBV: reordering pair: " << *I <<
3094                             " <-> " << *J << "\n");
3095           }
3096         }
3097       }
3098 
3099       Instruction *L = I, *H = J;
3100       if (FlipPairOrder)
3101         std::swap(H, L);
3102 
3103       // If the pair being fused uses the opposite order from that in the pair
3104       // connection map, then we need to flip the types.
3105       DenseMap<ValuePair, std::vector<ValuePair> >::iterator HL =
3106         ConnectedPairs.find(ValuePair(H, L));
3107       if (HL != ConnectedPairs.end())
3108         for (std::vector<ValuePair>::iterator T = HL->second.begin(),
3109              TE = HL->second.end(); T != TE; ++T) {
3110           VPPair Q(HL->first, *T);
3111           DenseMap<VPPair, unsigned>::iterator R = PairConnectionTypes.find(Q);
3112           assert(R != PairConnectionTypes.end() &&
3113                  "Cannot find pair connection type");
3114           if (R->second == PairConnectionDirect)
3115             R->second = PairConnectionSwap;
3116           else if (R->second == PairConnectionSwap)
3117             R->second = PairConnectionDirect;
3118         }
3119 
3120       bool LBeforeH = !FlipPairOrder;
3121       unsigned NumOperands = I->getNumOperands();
3122       SmallVector<Value *, 3> ReplacedOperands(NumOperands);
3123       getReplacementInputsForPair(Context, L, H, ReplacedOperands,
3124                                   LBeforeH);
3125 
3126       // Make a copy of the original operation, change its type to the vector
3127       // type and replace its operands with the vector operands.
3128       Instruction *K = L->clone();
3129       if (L->hasName())
3130         K->takeName(L);
3131       else if (H->hasName())
3132         K->takeName(H);
3133 
3134       if (auto CS = CallSite(K)) {
3135         SmallVector<Type *, 3> Tys;
3136         FunctionType *Old = CS.getFunctionType();
3137         unsigned NumOld = Old->getNumParams();
3138         assert(NumOld <= ReplacedOperands.size());
3139         for (unsigned i = 0; i != NumOld; ++i)
3140           Tys.push_back(ReplacedOperands[i]->getType());
3141         CS.mutateFunctionType(
3142             FunctionType::get(getVecTypeForPair(L->getType(), H->getType()),
3143                               Tys, Old->isVarArg()));
3144       } else if (!isa<StoreInst>(K))
3145         K->mutateType(getVecTypeForPair(L->getType(), H->getType()));
3146 
3147       unsigned KnownIDs[] = {LLVMContext::MD_tbaa, LLVMContext::MD_alias_scope,
3148                              LLVMContext::MD_noalias, LLVMContext::MD_fpmath,
3149                              LLVMContext::MD_invariant_group};
3150       combineMetadata(K, H, KnownIDs);
3151       K->intersectOptionalDataWith(H);
3152 
3153       for (unsigned o = 0; o < NumOperands; ++o)
3154         K->setOperand(o, ReplacedOperands[o]);
3155 
3156       K->insertAfter(J);
3157 
3158       // Instruction insertion point:
3159       Instruction *InsertionPt = K;
3160       Instruction *K1 = nullptr, *K2 = nullptr;
3161       replaceOutputsOfPair(Context, L, H, K, InsertionPt, K1, K2);
3162 
3163       // The use dag of the first original instruction must be moved to after
3164       // the location of the second instruction. The entire use dag of the
3165       // first instruction is disjoint from the input dag of the second
3166       // (by definition), and so commutes with it.
3167 
3168       moveUsesOfIAfterJ(BB, LoadMoveSetPairs, InsertionPt, I, J);
3169 
3170       if (!isa<StoreInst>(I)) {
3171         L->replaceAllUsesWith(K1);
3172         H->replaceAllUsesWith(K2);
3173       }
3174 
3175       // Instructions that may read from memory may be in the load move set.
3176       // Once an instruction is fused, we no longer need its move set, and so
3177       // the values of the map never need to be updated. However, when a load
3178       // is fused, we need to merge the entries from both instructions in the
3179       // pair in case those instructions were in the move set of some other
3180       // yet-to-be-fused pair. The loads in question are the keys of the map.
3181       if (I->mayReadFromMemory()) {
3182         std::vector<ValuePair> NewSetMembers;
3183         DenseMap<Value *, std::vector<Value *> >::iterator II =
3184           LoadMoveSet.find(I);
3185         if (II != LoadMoveSet.end())
3186           for (std::vector<Value *>::iterator N = II->second.begin(),
3187                NE = II->second.end(); N != NE; ++N)
3188             NewSetMembers.push_back(ValuePair(K, *N));
3189         DenseMap<Value *, std::vector<Value *> >::iterator JJ =
3190           LoadMoveSet.find(J);
3191         if (JJ != LoadMoveSet.end())
3192           for (std::vector<Value *>::iterator N = JJ->second.begin(),
3193                NE = JJ->second.end(); N != NE; ++N)
3194             NewSetMembers.push_back(ValuePair(K, *N));
3195         for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(),
3196              AE = NewSetMembers.end(); A != AE; ++A) {
3197           LoadMoveSet[A->first].push_back(A->second);
3198           LoadMoveSetPairs.insert(*A);
3199         }
3200       }
3201 
3202       // Before removing I, set the iterator to the next instruction.
3203       PI = std::next(BasicBlock::iterator(I));
3204       if (cast<Instruction>(PI) == J)
3205         ++PI;
3206 
3207       SE->forgetValue(I);
3208       SE->forgetValue(J);
3209       I->eraseFromParent();
3210       J->eraseFromParent();
3211 
3212       DEBUG(if (PrintAfterEveryPair) dbgs() << "BBV: block is now: \n" <<
3213                                                BB << "\n");
3214     }
3215 
3216     DEBUG(dbgs() << "BBV: final: \n" << BB << "\n");
3217   }
3218 }
3219 
3220 char BBVectorize::ID = 0;
3221 static const char bb_vectorize_name[] = "Basic-Block Vectorization";
INITIALIZE_PASS_BEGIN(BBVectorize,BBV_NAME,bb_vectorize_name,false,false)3222 INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
3223 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
3224 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
3225 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
3226 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
3227 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
3228 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
3229 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
3230 INITIALIZE_PASS_DEPENDENCY(SCEVAAWrapperPass)
3231 INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
3232 
3233 BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) {
3234   return new BBVectorize(C);
3235 }
3236 
3237 bool
vectorizeBasicBlock(Pass * P,BasicBlock & BB,const VectorizeConfig & C)3238 llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) {
3239   BBVectorize BBVectorizer(P, *BB.getParent(), C);
3240   return BBVectorizer.vectorizeBB(BB);
3241 }
3242 
3243 //===----------------------------------------------------------------------===//
VectorizeConfig()3244 VectorizeConfig::VectorizeConfig() {
3245   VectorBits = ::VectorBits;
3246   VectorizeBools = !::NoBools;
3247   VectorizeInts = !::NoInts;
3248   VectorizeFloats = !::NoFloats;
3249   VectorizePointers = !::NoPointers;
3250   VectorizeCasts = !::NoCasts;
3251   VectorizeMath = !::NoMath;
3252   VectorizeBitManipulations = !::NoBitManipulation;
3253   VectorizeFMA = !::NoFMA;
3254   VectorizeSelect = !::NoSelect;
3255   VectorizeCmp = !::NoCmp;
3256   VectorizeGEP = !::NoGEP;
3257   VectorizeMemOps = !::NoMemOps;
3258   AlignedOnly = ::AlignedOnly;
3259   ReqChainDepth= ::ReqChainDepth;
3260   SearchLimit = ::SearchLimit;
3261   MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck;
3262   SplatBreaksChain = ::SplatBreaksChain;
3263   MaxInsts = ::MaxInsts;
3264   MaxPairs = ::MaxPairs;
3265   MaxIter = ::MaxIter;
3266   Pow2LenOnly = ::Pow2LenOnly;
3267   NoMemOpBoost = ::NoMemOpBoost;
3268   FastDep = ::FastDep;
3269 }
3270