1 //===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops
11 // and generates target-independent LLVM-IR.
12 // The vectorizer uses the TargetTransformInfo analysis to estimate the costs
13 // of instructions in order to estimate the profitability of vectorization.
15 // The loop vectorizer combines consecutive loop iterations into a single
16 // 'wide' iteration. After this transformation the index is incremented
17 // by the SIMD vector width, and not by one.
19 // This pass has three parts:
20 // 1. The main loop pass that drives the different parts.
21 // 2. LoopVectorizationLegality - A unit that checks for the legality
22 // of the vectorization.
23 // 3. InnerLoopVectorizer - A unit that performs the actual
24 // widening of instructions.
25 // 4. LoopVectorizationCostModel - A unit that checks for the profitability
26 // of vectorization. It decides on the optimal vector width, which
27 // can be one, if vectorization is not profitable.
29 //===----------------------------------------------------------------------===//
31 // The reduction-variable vectorization is based on the paper:
32 // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
34 // Variable uniformity checks are inspired by:
35 // Karrenberg, R. and Hack, S. Whole Function Vectorization.
37 // The interleaved access vectorization is based on the paper:
38 // Dorit Nuzman, Ira Rosen and Ayal Zaks. Auto-Vectorization of Interleaved
41 // Other ideas/concepts are from:
42 // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
44 // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of
45 // Vectorizing Compilers.
47 //===----------------------------------------------------------------------===//
49 #include "llvm/Transforms/Vectorize/LoopVectorize.h"
50 #include "llvm/ADT/DenseMap.h"
51 #include "llvm/ADT/Hashing.h"
52 #include "llvm/ADT/MapVector.h"
53 #include "llvm/ADT/SCCIterator.h"
54 #include "llvm/ADT/SetVector.h"
55 #include "llvm/ADT/SmallPtrSet.h"
56 #include "llvm/ADT/SmallSet.h"
57 #include "llvm/ADT/SmallVector.h"
58 #include "llvm/ADT/Statistic.h"
59 #include "llvm/ADT/StringExtras.h"
60 #include "llvm/Analysis/CodeMetrics.h"
61 #include "llvm/Analysis/GlobalsModRef.h"
62 #include "llvm/Analysis/LoopInfo.h"
63 #include "llvm/Analysis/LoopIterator.h"
64 #include "llvm/Analysis/LoopPass.h"
65 #include "llvm/Analysis/ScalarEvolutionExpander.h"
66 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
67 #include "llvm/Analysis/ValueTracking.h"
68 #include "llvm/Analysis/VectorUtils.h"
69 #include "llvm/IR/Constants.h"
70 #include "llvm/IR/DataLayout.h"
71 #include "llvm/IR/DebugInfo.h"
72 #include "llvm/IR/DerivedTypes.h"
73 #include "llvm/IR/DiagnosticInfo.h"
74 #include "llvm/IR/Dominators.h"
75 #include "llvm/IR/Function.h"
76 #include "llvm/IR/IRBuilder.h"
77 #include "llvm/IR/Instructions.h"
78 #include "llvm/IR/IntrinsicInst.h"
79 #include "llvm/IR/LLVMContext.h"
80 #include "llvm/IR/Module.h"
81 #include "llvm/IR/PatternMatch.h"
82 #include "llvm/IR/Type.h"
83 #include "llvm/IR/User.h"
84 #include "llvm/IR/Value.h"
85 #include "llvm/IR/ValueHandle.h"
86 #include "llvm/IR/Verifier.h"
87 #include "llvm/Pass.h"
88 #include "llvm/Support/BranchProbability.h"
89 #include "llvm/Support/CommandLine.h"
90 #include "llvm/Support/Debug.h"
91 #include "llvm/Support/raw_ostream.h"
92 #include "llvm/Transforms/Scalar.h"
93 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
94 #include "llvm/Transforms/Utils/Local.h"
95 #include "llvm/Transforms/Utils/LoopUtils.h"
96 #include "llvm/Transforms/Utils/LoopVersioning.h"
97 #include "llvm/Transforms/Vectorize.h"
102 using namespace llvm;
103 using namespace llvm::PatternMatch;
105 #define LV_NAME "loop-vectorize"
106 #define DEBUG_TYPE LV_NAME
108 STATISTIC(LoopsVectorized, "Number of loops vectorized");
109 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
112 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
113 cl::desc("Enable if-conversion during vectorization."));
115 /// We don't vectorize loops with a known constant trip count below this number.
116 static cl::opt<unsigned> TinyTripCountVectorThreshold(
117 "vectorizer-min-trip-count", cl::init(16), cl::Hidden,
118 cl::desc("Don't vectorize loops with a constant "
119 "trip count that is smaller than this "
122 static cl::opt<bool> MaximizeBandwidth(
123 "vectorizer-maximize-bandwidth", cl::init(false), cl::Hidden,
124 cl::desc("Maximize bandwidth when selecting vectorization factor which "
125 "will be determined by the smallest type in loop."));
127 static cl::opt<bool> EnableInterleavedMemAccesses(
128 "enable-interleaved-mem-accesses", cl::init(false), cl::Hidden,
129 cl::desc("Enable vectorization on interleaved memory accesses in a loop"));
131 /// Maximum factor for an interleaved memory access.
132 static cl::opt<unsigned> MaxInterleaveGroupFactor(
133 "max-interleave-group-factor", cl::Hidden,
134 cl::desc("Maximum factor for an interleaved access group (default = 8)"),
137 /// We don't interleave loops with a known constant trip count below this
139 static const unsigned TinyTripCountInterleaveThreshold = 128;
141 static cl::opt<unsigned> ForceTargetNumScalarRegs(
142 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
143 cl::desc("A flag that overrides the target's number of scalar registers."));
145 static cl::opt<unsigned> ForceTargetNumVectorRegs(
146 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
147 cl::desc("A flag that overrides the target's number of vector registers."));
149 /// Maximum vectorization interleave count.
150 static const unsigned MaxInterleaveFactor = 16;
152 static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
153 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
154 cl::desc("A flag that overrides the target's max interleave factor for "
157 static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
158 "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
159 cl::desc("A flag that overrides the target's max interleave factor for "
160 "vectorized loops."));
162 static cl::opt<unsigned> ForceTargetInstructionCost(
163 "force-target-instruction-cost", cl::init(0), cl::Hidden,
164 cl::desc("A flag that overrides the target's expected cost for "
165 "an instruction to a single constant value. Mostly "
166 "useful for getting consistent testing."));
168 static cl::opt<unsigned> SmallLoopCost(
169 "small-loop-cost", cl::init(20), cl::Hidden,
171 "The cost of a loop that is considered 'small' by the interleaver."));
173 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
174 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
175 cl::desc("Enable the use of the block frequency analysis to access PGO "
176 "heuristics minimizing code growth in cold regions and being more "
177 "aggressive in hot regions."));
179 // Runtime interleave loops for load/store throughput.
180 static cl::opt<bool> EnableLoadStoreRuntimeInterleave(
181 "enable-loadstore-runtime-interleave", cl::init(true), cl::Hidden,
183 "Enable runtime interleaving until load/store ports are saturated"));
185 /// The number of stores in a loop that are allowed to need predication.
186 static cl::opt<unsigned> NumberOfStoresToPredicate(
187 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
188 cl::desc("Max number of stores to be predicated behind an if."));
190 static cl::opt<bool> EnableIndVarRegisterHeur(
191 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
192 cl::desc("Count the induction variable only once when interleaving"));
194 static cl::opt<bool> EnableCondStoresVectorization(
195 "enable-cond-stores-vec", cl::init(true), cl::Hidden,
196 cl::desc("Enable if predication of stores during vectorization."));
198 static cl::opt<unsigned> MaxNestedScalarReductionIC(
199 "max-nested-scalar-reduction-interleave", cl::init(2), cl::Hidden,
200 cl::desc("The maximum interleave count to use when interleaving a scalar "
201 "reduction in a nested loop."));
203 static cl::opt<unsigned> PragmaVectorizeMemoryCheckThreshold(
204 "pragma-vectorize-memory-check-threshold", cl::init(128), cl::Hidden,
205 cl::desc("The maximum allowed number of runtime memory checks with a "
206 "vectorize(enable) pragma."));
208 static cl::opt<unsigned> VectorizeSCEVCheckThreshold(
209 "vectorize-scev-check-threshold", cl::init(16), cl::Hidden,
210 cl::desc("The maximum number of SCEV checks allowed."));
212 static cl::opt<unsigned> PragmaVectorizeSCEVCheckThreshold(
213 "pragma-vectorize-scev-check-threshold", cl::init(128), cl::Hidden,
214 cl::desc("The maximum number of SCEV checks allowed with a "
215 "vectorize(enable) pragma"));
217 /// Create an analysis remark that explains why vectorization failed
219 /// \p PassName is the name of the pass (e.g. can be AlwaysPrint). \p
220 /// RemarkName is the identifier for the remark. If \p I is passed it is an
221 /// instruction that prevents vectorization. Otherwise \p TheLoop is used for
222 /// the location of the remark. \return the remark object that can be
224 static OptimizationRemarkAnalysis
225 createMissedAnalysis(const char *PassName, StringRef RemarkName, Loop *TheLoop,
226 Instruction *I = nullptr) {
227 Value *CodeRegion = TheLoop->getHeader();
228 DebugLoc DL = TheLoop->getStartLoc();
231 CodeRegion = I->getParent();
232 // If there is no debug location attached to the instruction, revert back to
234 if (I->getDebugLoc())
235 DL = I->getDebugLoc();
238 OptimizationRemarkAnalysis R(PassName, RemarkName, DL, CodeRegion);
239 R << "loop not vectorized: ";
245 // Forward declarations.
246 class LoopVectorizeHints;
247 class LoopVectorizationLegality;
248 class LoopVectorizationCostModel;
249 class LoopVectorizationRequirements;
251 /// Returns true if the given loop body has a cycle, excluding the loop
253 static bool hasCyclesInLoopBody(const Loop &L) {
257 for (const auto &SCC :
258 make_range(scc_iterator<Loop, LoopBodyTraits>::begin(L),
259 scc_iterator<Loop, LoopBodyTraits>::end(L))) {
260 if (SCC.size() > 1) {
261 DEBUG(dbgs() << "LVL: Detected a cycle in the loop body:\n");
269 /// \brief This modifies LoopAccessReport to initialize message with
270 /// loop-vectorizer-specific part.
271 class VectorizationReport : public LoopAccessReport {
273 VectorizationReport(Instruction *I = nullptr)
274 : LoopAccessReport("loop not vectorized: ", I) {}
276 /// \brief This allows promotion of the loop-access analysis report into the
277 /// loop-vectorizer report. It modifies the message to add the
278 /// loop-vectorizer-specific part of the message.
279 explicit VectorizationReport(const LoopAccessReport &R)
280 : LoopAccessReport(Twine("loop not vectorized: ") + R.str(),
284 /// A helper function for converting Scalar types to vector types.
285 /// If the incoming type is void, we return void. If the VF is 1, we return
287 static Type *ToVectorTy(Type *Scalar, unsigned VF) {
288 if (Scalar->isVoidTy() || VF == 1)
290 return VectorType::get(Scalar, VF);
293 /// A helper function that returns GEP instruction and knows to skip a
294 /// 'bitcast'. The 'bitcast' may be skipped if the source and the destination
295 /// pointee types of the 'bitcast' have the same size.
297 /// bitcast double** %var to i64* - can be skipped
298 /// bitcast double** %var to i8* - can not
299 static GetElementPtrInst *getGEPInstruction(Value *Ptr) {
301 if (isa<GetElementPtrInst>(Ptr))
302 return cast<GetElementPtrInst>(Ptr);
304 if (isa<BitCastInst>(Ptr) &&
305 isa<GetElementPtrInst>(cast<BitCastInst>(Ptr)->getOperand(0))) {
306 Type *BitcastTy = Ptr->getType();
307 Type *GEPTy = cast<BitCastInst>(Ptr)->getSrcTy();
308 if (!isa<PointerType>(BitcastTy) || !isa<PointerType>(GEPTy))
310 Type *Pointee1Ty = cast<PointerType>(BitcastTy)->getPointerElementType();
311 Type *Pointee2Ty = cast<PointerType>(GEPTy)->getPointerElementType();
312 const DataLayout &DL = cast<BitCastInst>(Ptr)->getModule()->getDataLayout();
313 if (DL.getTypeSizeInBits(Pointee1Ty) == DL.getTypeSizeInBits(Pointee2Ty))
314 return cast<GetElementPtrInst>(cast<BitCastInst>(Ptr)->getOperand(0));
319 /// A helper function that returns the pointer operand of a load or store
321 static Value *getPointerOperand(Value *I) {
322 if (auto *LI = dyn_cast<LoadInst>(I))
323 return LI->getPointerOperand();
324 if (auto *SI = dyn_cast<StoreInst>(I))
325 return SI->getPointerOperand();
329 /// A helper function that returns true if the given type is irregular. The
330 /// type is irregular if its allocated size doesn't equal the store size of an
331 /// element of the corresponding vector type at the given vectorization factor.
332 static bool hasIrregularType(Type *Ty, const DataLayout &DL, unsigned VF) {
334 // Determine if an array of VF elements of type Ty is "bitcast compatible"
335 // with a <VF x Ty> vector.
337 auto *VectorTy = VectorType::get(Ty, VF);
338 return VF * DL.getTypeAllocSize(Ty) != DL.getTypeStoreSize(VectorTy);
341 // If the vectorization factor is one, we just check if an array of type Ty
342 // requires padding between elements.
343 return DL.getTypeAllocSizeInBits(Ty) != DL.getTypeSizeInBits(Ty);
346 /// A helper function that returns the reciprocal of the block probability of
347 /// predicated blocks. If we return X, we are assuming the predicated block
348 /// will execute once for for every X iterations of the loop header.
350 /// TODO: We should use actual block probability here, if available. Currently,
351 /// we always assume predicated blocks have a 50% chance of executing.
352 static unsigned getReciprocalPredBlockProb() { return 2; }
354 /// InnerLoopVectorizer vectorizes loops which contain only one basic
355 /// block to a specified vectorization factor (VF).
356 /// This class performs the widening of scalars into vectors, or multiple
357 /// scalars. This class also implements the following features:
358 /// * It inserts an epilogue loop for handling loops that don't have iteration
359 /// counts that are known to be a multiple of the vectorization factor.
360 /// * It handles the code generation for reduction variables.
361 /// * Scalarization (implementation using scalars) of un-vectorizable
363 /// InnerLoopVectorizer does not perform any vectorization-legality
364 /// checks, and relies on the caller to check for the different legality
365 /// aspects. The InnerLoopVectorizer relies on the
366 /// LoopVectorizationLegality class to provide information about the induction
367 /// and reduction variables that were found to a given vectorization factor.
368 class InnerLoopVectorizer {
370 InnerLoopVectorizer(Loop *OrigLoop, PredicatedScalarEvolution &PSE,
371 LoopInfo *LI, DominatorTree *DT,
372 const TargetLibraryInfo *TLI,
373 const TargetTransformInfo *TTI, AssumptionCache *AC,
374 OptimizationRemarkEmitter *ORE, unsigned VecWidth,
375 unsigned UnrollFactor, LoopVectorizationLegality *LVL,
376 LoopVectorizationCostModel *CM)
377 : OrigLoop(OrigLoop), PSE(PSE), LI(LI), DT(DT), TLI(TLI), TTI(TTI),
378 AC(AC), ORE(ORE), VF(VecWidth), UF(UnrollFactor),
379 Builder(PSE.getSE()->getContext()), Induction(nullptr),
380 OldInduction(nullptr), VectorLoopValueMap(UnrollFactor, VecWidth),
381 TripCount(nullptr), VectorTripCount(nullptr), Legal(LVL), Cost(CM),
382 AddedSafetyChecks(false) {}
384 // Perform the actual loop widening (vectorization).
386 // Create a new empty loop. Unlink the old loop and connect the new one.
388 // Widen each instruction in the old loop to a new one in the new loop.
392 // Return true if any runtime check is added.
393 bool areSafetyChecksAdded() { return AddedSafetyChecks; }
395 virtual ~InnerLoopVectorizer() {}
398 /// A small list of PHINodes.
399 typedef SmallVector<PHINode *, 4> PhiVector;
401 /// A type for vectorized values in the new loop. Each value from the
402 /// original loop, when vectorized, is represented by UF vector values in the
403 /// new unrolled loop, where UF is the unroll factor.
404 typedef SmallVector<Value *, 2> VectorParts;
406 /// A type for scalarized values in the new loop. Each value from the
407 /// original loop, when scalarized, is represented by UF x VF scalar values
408 /// in the new unrolled loop, where UF is the unroll factor and VF is the
409 /// vectorization factor.
410 typedef SmallVector<SmallVector<Value *, 4>, 2> ScalarParts;
412 // When we if-convert we need to create edge masks. We have to cache values
413 // so that we don't end up with exponential recursion/IR.
414 typedef DenseMap<std::pair<BasicBlock *, BasicBlock *>, VectorParts>
417 /// Create an empty loop, based on the loop ranges of the old loop.
418 void createEmptyLoop();
420 /// Set up the values of the IVs correctly when exiting the vector loop.
421 void fixupIVUsers(PHINode *OrigPhi, const InductionDescriptor &II,
422 Value *CountRoundDown, Value *EndValue,
423 BasicBlock *MiddleBlock);
425 /// Create a new induction variable inside L.
426 PHINode *createInductionVariable(Loop *L, Value *Start, Value *End,
427 Value *Step, Instruction *DL);
428 /// Copy and widen the instructions from the old loop.
429 virtual void vectorizeLoop();
431 /// Fix a first-order recurrence. This is the second phase of vectorizing
433 void fixFirstOrderRecurrence(PHINode *Phi);
435 /// \brief The Loop exit block may have single value PHI nodes where the
436 /// incoming value is 'Undef'. While vectorizing we only handled real values
437 /// that were defined inside the loop. Here we fix the 'undef case'.
441 /// Iteratively sink the scalarized operands of a predicated instruction into
442 /// the block that was created for it.
443 void sinkScalarOperands(Instruction *PredInst);
445 /// Predicate conditional instructions that require predication on their
446 /// respective conditions.
447 void predicateInstructions();
449 /// Collect the instructions from the original loop that would be trivially
450 /// dead in the vectorized loop if generated.
451 void collectTriviallyDeadInstructions();
453 /// Shrinks vector element sizes to the smallest bitwidth they can be legally
455 void truncateToMinimalBitwidths();
457 /// A helper function that computes the predicate of the block BB, assuming
458 /// that the header block of the loop is set to True. It returns the *entry*
459 /// mask for the block BB.
460 VectorParts createBlockInMask(BasicBlock *BB);
461 /// A helper function that computes the predicate of the edge between SRC
463 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
465 /// A helper function to vectorize a single BB within the innermost loop.
466 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
468 /// Vectorize a single PHINode in a block. This method handles the induction
469 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
470 /// arbitrary length vectors.
471 void widenPHIInstruction(Instruction *PN, unsigned UF, unsigned VF,
474 /// Insert the new loop to the loop hierarchy and pass manager
475 /// and update the analysis passes.
476 void updateAnalysis();
478 /// This instruction is un-vectorizable. Implement it as a sequence
479 /// of scalars. If \p IfPredicateInstr is true we need to 'hide' each
480 /// scalarized instruction behind an if block predicated on the control
481 /// dependence of the instruction.
482 virtual void scalarizeInstruction(Instruction *Instr,
483 bool IfPredicateInstr = false);
485 /// Vectorize Load and Store instructions,
486 virtual void vectorizeMemoryInstruction(Instruction *Instr);
488 /// Create a broadcast instruction. This method generates a broadcast
489 /// instruction (shuffle) for loop invariant values and for the induction
490 /// value. If this is the induction variable then we extend it to N, N+1, ...
491 /// this is needed because each iteration in the loop corresponds to a SIMD
493 virtual Value *getBroadcastInstrs(Value *V);
495 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
496 /// to each vector element of Val. The sequence starts at StartIndex.
497 /// \p Opcode is relevant for FP induction variable.
498 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step,
499 Instruction::BinaryOps Opcode =
500 Instruction::BinaryOpsEnd);
502 /// Compute scalar induction steps. \p ScalarIV is the scalar induction
503 /// variable on which to base the steps, \p Step is the size of the step, and
504 /// \p EntryVal is the value from the original loop that maps to the steps.
505 /// Note that \p EntryVal doesn't have to be an induction variable (e.g., it
506 /// can be a truncate instruction).
507 void buildScalarSteps(Value *ScalarIV, Value *Step, Value *EntryVal);
509 /// Create a vector induction phi node based on an existing scalar one. This
510 /// currently only works for integer induction variables with a constant
511 /// step. \p EntryVal is the value from the original loop that maps to the
512 /// vector phi node. If \p EntryVal is a truncate instruction, instead of
513 /// widening the original IV, we widen a version of the IV truncated to \p
515 void createVectorIntInductionPHI(const InductionDescriptor &II,
516 Instruction *EntryVal);
518 /// Widen an integer induction variable \p IV. If \p Trunc is provided, the
519 /// induction variable will first be truncated to the corresponding type.
520 void widenIntInduction(PHINode *IV, TruncInst *Trunc = nullptr);
522 /// Returns true if an instruction \p I should be scalarized instead of
523 /// vectorized for the chosen vectorization factor.
524 bool shouldScalarizeInstruction(Instruction *I) const;
526 /// Returns true if we should generate a scalar version of \p IV.
527 bool needsScalarInduction(Instruction *IV) const;
529 /// Return a constant reference to the VectorParts corresponding to \p V from
530 /// the original loop. If the value has already been vectorized, the
531 /// corresponding vector entry in VectorLoopValueMap is returned. If,
532 /// however, the value has a scalar entry in VectorLoopValueMap, we construct
533 /// new vector values on-demand by inserting the scalar values into vectors
534 /// with an insertelement sequence. If the value has been neither vectorized
535 /// nor scalarized, it must be loop invariant, so we simply broadcast the
536 /// value into vectors.
537 const VectorParts &getVectorValue(Value *V);
539 /// Return a value in the new loop corresponding to \p V from the original
540 /// loop at unroll index \p Part and vector index \p Lane. If the value has
541 /// been vectorized but not scalarized, the necessary extractelement
542 /// instruction will be generated.
543 Value *getScalarValue(Value *V, unsigned Part, unsigned Lane);
545 /// Try to vectorize the interleaved access group that \p Instr belongs to.
546 void vectorizeInterleaveGroup(Instruction *Instr);
548 /// Generate a shuffle sequence that will reverse the vector Vec.
549 virtual Value *reverseVector(Value *Vec);
551 /// Returns (and creates if needed) the original loop trip count.
552 Value *getOrCreateTripCount(Loop *NewLoop);
554 /// Returns (and creates if needed) the trip count of the widened loop.
555 Value *getOrCreateVectorTripCount(Loop *NewLoop);
557 /// Emit a bypass check to see if the trip count would overflow, or we
558 /// wouldn't have enough iterations to execute one vector loop.
559 void emitMinimumIterationCountCheck(Loop *L, BasicBlock *Bypass);
560 /// Emit a bypass check to see if the vector trip count is nonzero.
561 void emitVectorLoopEnteredCheck(Loop *L, BasicBlock *Bypass);
562 /// Emit a bypass check to see if all of the SCEV assumptions we've
563 /// had to make are correct.
564 void emitSCEVChecks(Loop *L, BasicBlock *Bypass);
565 /// Emit bypass checks to check any memory assumptions we may have made.
566 void emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass);
568 /// Add additional metadata to \p To that was not present on \p Orig.
570 /// Currently this is used to add the noalias annotations based on the
571 /// inserted memchecks. Use this for instructions that are *cloned* into the
573 void addNewMetadata(Instruction *To, const Instruction *Orig);
575 /// Add metadata from one instruction to another.
577 /// This includes both the original MDs from \p From and additional ones (\see
578 /// addNewMetadata). Use this for *newly created* instructions in the vector
580 void addMetadata(Instruction *To, Instruction *From);
582 /// \brief Similar to the previous function but it adds the metadata to a
583 /// vector of instructions.
584 void addMetadata(ArrayRef<Value *> To, Instruction *From);
586 /// This is a helper class for maintaining vectorization state. It's used for
587 /// mapping values from the original loop to their corresponding values in
588 /// the new loop. Two mappings are maintained: one for vectorized values and
589 /// one for scalarized values. Vectorized values are represented with UF
590 /// vector values in the new loop, and scalarized values are represented with
591 /// UF x VF scalar values in the new loop. UF and VF are the unroll and
592 /// vectorization factors, respectively.
594 /// Entries can be added to either map with initVector and initScalar, which
595 /// initialize and return a constant reference to the new entry. If a
596 /// non-constant reference to a vector entry is required, getVector can be
597 /// used to retrieve a mutable entry. We currently directly modify the mapped
598 /// values during "fix-up" operations that occur once the first phase of
599 /// widening is complete. These operations include type truncation and the
600 /// second phase of recurrence widening.
602 /// Otherwise, entries from either map should be accessed using the
603 /// getVectorValue or getScalarValue functions from InnerLoopVectorizer.
604 /// getVectorValue and getScalarValue coordinate to generate a vector or
605 /// scalar value on-demand if one is not yet available. When vectorizing a
606 /// loop, we visit the definition of an instruction before its uses. When
607 /// visiting the definition, we either vectorize or scalarize the
608 /// instruction, creating an entry for it in the corresponding map. (In some
609 /// cases, such as induction variables, we will create both vector and scalar
610 /// entries.) Then, as we encounter uses of the definition, we derive values
611 /// for each scalar or vector use unless such a value is already available.
612 /// For example, if we scalarize a definition and one of its uses is vector,
613 /// we build the required vector on-demand with an insertelement sequence
614 /// when visiting the use. Otherwise, if the use is scalar, we can use the
615 /// existing scalar definition.
618 /// Construct an empty map with the given unroll and vectorization factors.
619 ValueMap(unsigned UnrollFactor, unsigned VecWidth)
620 : UF(UnrollFactor), VF(VecWidth) {
621 // The unroll and vectorization factors are only used in asserts builds
622 // to verify map entries are sized appropriately.
627 /// \return True if the map has a vector entry for \p Key.
628 bool hasVector(Value *Key) const { return VectorMapStorage.count(Key); }
630 /// \return True if the map has a scalar entry for \p Key.
631 bool hasScalar(Value *Key) const { return ScalarMapStorage.count(Key); }
633 /// \brief Map \p Key to the given VectorParts \p Entry, and return a
634 /// constant reference to the new vector map entry. The given key should
635 /// not already be in the map, and the given VectorParts should be
636 /// correctly sized for the current unroll factor.
637 const VectorParts &initVector(Value *Key, const VectorParts &Entry) {
638 assert(!hasVector(Key) && "Vector entry already initialized");
639 assert(Entry.size() == UF && "VectorParts has wrong dimensions");
640 VectorMapStorage[Key] = Entry;
641 return VectorMapStorage[Key];
644 /// \brief Map \p Key to the given ScalarParts \p Entry, and return a
645 /// constant reference to the new scalar map entry. The given key should
646 /// not already be in the map, and the given ScalarParts should be
647 /// correctly sized for the current unroll and vectorization factors.
648 const ScalarParts &initScalar(Value *Key, const ScalarParts &Entry) {
649 assert(!hasScalar(Key) && "Scalar entry already initialized");
650 assert(Entry.size() == UF &&
651 all_of(make_range(Entry.begin(), Entry.end()),
652 [&](const SmallVectorImpl<Value *> &Values) -> bool {
653 return Values.size() == VF;
655 "ScalarParts has wrong dimensions");
656 ScalarMapStorage[Key] = Entry;
657 return ScalarMapStorage[Key];
660 /// \return A reference to the vector map entry corresponding to \p Key.
661 /// The key should already be in the map. This function should only be used
662 /// when it's necessary to update values that have already been vectorized.
663 /// This is the case for "fix-up" operations including type truncation and
664 /// the second phase of recurrence vectorization. If a non-const reference
665 /// isn't required, getVectorValue should be used instead.
666 VectorParts &getVector(Value *Key) {
667 assert(hasVector(Key) && "Vector entry not initialized");
668 return VectorMapStorage.find(Key)->second;
671 /// Retrieve an entry from the vector or scalar maps. The preferred way to
672 /// access an existing mapped entry is with getVectorValue or
673 /// getScalarValue from InnerLoopVectorizer. Until those functions can be
674 /// moved inside ValueMap, we have to declare them as friends.
675 friend const VectorParts &InnerLoopVectorizer::getVectorValue(Value *V);
676 friend Value *InnerLoopVectorizer::getScalarValue(Value *V, unsigned Part,
680 /// The unroll factor. Each entry in the vector map contains UF vector
684 /// The vectorization factor. Each entry in the scalar map contains UF x VF
688 /// The vector and scalar map storage. We use std::map and not DenseMap
689 /// because insertions to DenseMap invalidate its iterators.
690 std::map<Value *, VectorParts> VectorMapStorage;
691 std::map<Value *, ScalarParts> ScalarMapStorage;
694 /// The original loop.
696 /// A wrapper around ScalarEvolution used to add runtime SCEV checks. Applies
697 /// dynamic knowledge to simplify SCEV expressions and converts them to a
698 /// more usable form.
699 PredicatedScalarEvolution &PSE;
706 /// Target Library Info.
707 const TargetLibraryInfo *TLI;
708 /// Target Transform Info.
709 const TargetTransformInfo *TTI;
710 /// Assumption Cache.
712 /// Interface to emit optimization remarks.
713 OptimizationRemarkEmitter *ORE;
715 /// \brief LoopVersioning. It's only set up (non-null) if memchecks were
718 /// This is currently only used to add no-alias metadata based on the
719 /// memchecks. The actually versioning is performed manually.
720 std::unique_ptr<LoopVersioning> LVer;
722 /// The vectorization SIMD factor to use. Each vector will have this many
727 /// The vectorization unroll factor to use. Each scalar is vectorized to this
728 /// many different vector instructions.
731 /// The builder that we use
734 // --- Vectorization state ---
736 /// The vector-loop preheader.
737 BasicBlock *LoopVectorPreHeader;
738 /// The scalar-loop preheader.
739 BasicBlock *LoopScalarPreHeader;
740 /// Middle Block between the vector and the scalar.
741 BasicBlock *LoopMiddleBlock;
742 /// The ExitBlock of the scalar loop.
743 BasicBlock *LoopExitBlock;
744 /// The vector loop body.
745 BasicBlock *LoopVectorBody;
746 /// The scalar loop body.
747 BasicBlock *LoopScalarBody;
748 /// A list of all bypass blocks. The first block is the entry of the loop.
749 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
751 /// The new Induction variable which was added to the new block.
753 /// The induction variable of the old basic block.
754 PHINode *OldInduction;
756 /// Maps values from the original loop to their corresponding values in the
757 /// vectorized loop. A key value can map to either vector values, scalar
758 /// values or both kinds of values, depending on whether the key was
759 /// vectorized and scalarized.
760 ValueMap VectorLoopValueMap;
762 /// Store instructions that should be predicated, as a pair
763 /// <StoreInst, Predicate>
764 SmallVector<std::pair<Instruction *, Value *>, 4> PredicatedInstructions;
765 EdgeMaskCache MaskCache;
766 /// Trip count of the original loop.
768 /// Trip count of the widened loop (TripCount - TripCount % (VF*UF))
769 Value *VectorTripCount;
771 /// The legality analysis.
772 LoopVectorizationLegality *Legal;
774 /// The profitablity analysis.
775 LoopVectorizationCostModel *Cost;
777 // Record whether runtime checks are added.
778 bool AddedSafetyChecks;
780 // Holds instructions from the original loop whose counterparts in the
781 // vectorized loop would be trivially dead if generated. For example,
782 // original induction update instructions can become dead because we
783 // separately emit induction "steps" when generating code for the new loop.
784 // Similarly, we create a new latch condition when setting up the structure
785 // of the new loop, so the old one can become dead.
786 SmallPtrSet<Instruction *, 4> DeadInstructions;
788 // Holds the end values for each induction variable. We save the end values
789 // so we can later fix-up the external users of the induction variables.
790 DenseMap<PHINode *, Value *> IVEndValues;
793 class InnerLoopUnroller : public InnerLoopVectorizer {
795 InnerLoopUnroller(Loop *OrigLoop, PredicatedScalarEvolution &PSE,
796 LoopInfo *LI, DominatorTree *DT,
797 const TargetLibraryInfo *TLI,
798 const TargetTransformInfo *TTI, AssumptionCache *AC,
799 OptimizationRemarkEmitter *ORE, unsigned UnrollFactor,
800 LoopVectorizationLegality *LVL,
801 LoopVectorizationCostModel *CM)
802 : InnerLoopVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, 1,
803 UnrollFactor, LVL, CM) {}
806 void scalarizeInstruction(Instruction *Instr,
807 bool IfPredicateInstr = false) override;
808 void vectorizeMemoryInstruction(Instruction *Instr) override;
809 Value *getBroadcastInstrs(Value *V) override;
810 Value *getStepVector(Value *Val, int StartIdx, Value *Step,
811 Instruction::BinaryOps Opcode =
812 Instruction::BinaryOpsEnd) override;
813 Value *reverseVector(Value *Vec) override;
816 /// \brief Look for a meaningful debug location on the instruction or it's
818 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
823 if (I->getDebugLoc() != Empty)
826 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
827 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
828 if (OpInst->getDebugLoc() != Empty)
835 /// \brief Set the debug location in the builder using the debug location in the
837 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
838 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
839 B.SetCurrentDebugLocation(Inst->getDebugLoc());
841 B.SetCurrentDebugLocation(DebugLoc());
845 /// \return string containing a file name and a line # for the given loop.
846 static std::string getDebugLocString(const Loop *L) {
849 raw_string_ostream OS(Result);
850 if (const DebugLoc LoopDbgLoc = L->getStartLoc())
851 LoopDbgLoc.print(OS);
853 // Just print the module name.
854 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
861 void InnerLoopVectorizer::addNewMetadata(Instruction *To,
862 const Instruction *Orig) {
863 // If the loop was versioned with memchecks, add the corresponding no-alias
865 if (LVer && (isa<LoadInst>(Orig) || isa<StoreInst>(Orig)))
866 LVer->annotateInstWithNoAlias(To, Orig);
869 void InnerLoopVectorizer::addMetadata(Instruction *To,
871 propagateMetadata(To, From);
872 addNewMetadata(To, From);
875 void InnerLoopVectorizer::addMetadata(ArrayRef<Value *> To,
877 for (Value *V : To) {
878 if (Instruction *I = dyn_cast<Instruction>(V))
879 addMetadata(I, From);
883 /// \brief The group of interleaved loads/stores sharing the same stride and
884 /// close to each other.
886 /// Each member in this group has an index starting from 0, and the largest
887 /// index should be less than interleaved factor, which is equal to the absolute
888 /// value of the access's stride.
890 /// E.g. An interleaved load group of factor 4:
891 /// for (unsigned i = 0; i < 1024; i+=4) {
892 /// a = A[i]; // Member of index 0
893 /// b = A[i+1]; // Member of index 1
894 /// d = A[i+3]; // Member of index 3
898 /// An interleaved store group of factor 4:
899 /// for (unsigned i = 0; i < 1024; i+=4) {
901 /// A[i] = a; // Member of index 0
902 /// A[i+1] = b; // Member of index 1
903 /// A[i+2] = c; // Member of index 2
904 /// A[i+3] = d; // Member of index 3
907 /// Note: the interleaved load group could have gaps (missing members), but
908 /// the interleaved store group doesn't allow gaps.
909 class InterleaveGroup {
911 InterleaveGroup(Instruction *Instr, int Stride, unsigned Align)
912 : Align(Align), SmallestKey(0), LargestKey(0), InsertPos(Instr) {
913 assert(Align && "The alignment should be non-zero");
915 Factor = std::abs(Stride);
916 assert(Factor > 1 && "Invalid interleave factor");
918 Reverse = Stride < 0;
922 bool isReverse() const { return Reverse; }
923 unsigned getFactor() const { return Factor; }
924 unsigned getAlignment() const { return Align; }
925 unsigned getNumMembers() const { return Members.size(); }
927 /// \brief Try to insert a new member \p Instr with index \p Index and
928 /// alignment \p NewAlign. The index is related to the leader and it could be
929 /// negative if it is the new leader.
931 /// \returns false if the instruction doesn't belong to the group.
932 bool insertMember(Instruction *Instr, int Index, unsigned NewAlign) {
933 assert(NewAlign && "The new member's alignment should be non-zero");
935 int Key = Index + SmallestKey;
937 // Skip if there is already a member with the same index.
938 if (Members.count(Key))
941 if (Key > LargestKey) {
942 // The largest index is always less than the interleave factor.
943 if (Index >= static_cast<int>(Factor))
947 } else if (Key < SmallestKey) {
948 // The largest index is always less than the interleave factor.
949 if (LargestKey - Key >= static_cast<int>(Factor))
955 // It's always safe to select the minimum alignment.
956 Align = std::min(Align, NewAlign);
957 Members[Key] = Instr;
961 /// \brief Get the member with the given index \p Index
963 /// \returns nullptr if contains no such member.
964 Instruction *getMember(unsigned Index) const {
965 int Key = SmallestKey + Index;
966 if (!Members.count(Key))
969 return Members.find(Key)->second;
972 /// \brief Get the index for the given member. Unlike the key in the member
973 /// map, the index starts from 0.
974 unsigned getIndex(Instruction *Instr) const {
975 for (auto I : Members)
976 if (I.second == Instr)
977 return I.first - SmallestKey;
979 llvm_unreachable("InterleaveGroup contains no such member");
982 Instruction *getInsertPos() const { return InsertPos; }
983 void setInsertPos(Instruction *Inst) { InsertPos = Inst; }
986 unsigned Factor; // Interleave Factor.
989 DenseMap<int, Instruction *> Members;
993 // To avoid breaking dependences, vectorized instructions of an interleave
994 // group should be inserted at either the first load or the last store in
997 // E.g. %even = load i32 // Insert Position
998 // %add = add i32 %even // Use of %even
1002 // %odd = add i32 // Def of %odd
1003 // store i32 %odd // Insert Position
1004 Instruction *InsertPos;
1007 /// \brief Drive the analysis of interleaved memory accesses in the loop.
1009 /// Use this class to analyze interleaved accesses only when we can vectorize
1010 /// a loop. Otherwise it's meaningless to do analysis as the vectorization
1011 /// on interleaved accesses is unsafe.
1013 /// The analysis collects interleave groups and records the relationships
1014 /// between the member and the group in a map.
1015 class InterleavedAccessInfo {
1017 InterleavedAccessInfo(PredicatedScalarEvolution &PSE, Loop *L,
1018 DominatorTree *DT, LoopInfo *LI)
1019 : PSE(PSE), TheLoop(L), DT(DT), LI(LI), LAI(nullptr),
1020 RequiresScalarEpilogue(false) {}
1022 ~InterleavedAccessInfo() {
1023 SmallSet<InterleaveGroup *, 4> DelSet;
1024 // Avoid releasing a pointer twice.
1025 for (auto &I : InterleaveGroupMap)
1026 DelSet.insert(I.second);
1027 for (auto *Ptr : DelSet)
1031 /// \brief Analyze the interleaved accesses and collect them in interleave
1032 /// groups. Substitute symbolic strides using \p Strides.
1033 void analyzeInterleaving(const ValueToValueMap &Strides);
1035 /// \brief Check if \p Instr belongs to any interleave group.
1036 bool isInterleaved(Instruction *Instr) const {
1037 return InterleaveGroupMap.count(Instr);
1040 /// \brief Return the maximum interleave factor of all interleaved groups.
1041 unsigned getMaxInterleaveFactor() const {
1042 unsigned MaxFactor = 1;
1043 for (auto &Entry : InterleaveGroupMap)
1044 MaxFactor = std::max(MaxFactor, Entry.second->getFactor());
1048 /// \brief Get the interleave group that \p Instr belongs to.
1050 /// \returns nullptr if doesn't have such group.
1051 InterleaveGroup *getInterleaveGroup(Instruction *Instr) const {
1052 if (InterleaveGroupMap.count(Instr))
1053 return InterleaveGroupMap.find(Instr)->second;
1057 /// \brief Returns true if an interleaved group that may access memory
1058 /// out-of-bounds requires a scalar epilogue iteration for correctness.
1059 bool requiresScalarEpilogue() const { return RequiresScalarEpilogue; }
1061 /// \brief Initialize the LoopAccessInfo used for dependence checking.
1062 void setLAI(const LoopAccessInfo *Info) { LAI = Info; }
1065 /// A wrapper around ScalarEvolution, used to add runtime SCEV checks.
1066 /// Simplifies SCEV expressions in the context of existing SCEV assumptions.
1067 /// The interleaved access analysis can also add new predicates (for example
1068 /// by versioning strides of pointers).
1069 PredicatedScalarEvolution &PSE;
1073 const LoopAccessInfo *LAI;
1075 /// True if the loop may contain non-reversed interleaved groups with
1076 /// out-of-bounds accesses. We ensure we don't speculatively access memory
1077 /// out-of-bounds by executing at least one scalar epilogue iteration.
1078 bool RequiresScalarEpilogue;
1080 /// Holds the relationships between the members and the interleave group.
1081 DenseMap<Instruction *, InterleaveGroup *> InterleaveGroupMap;
1083 /// Holds dependences among the memory accesses in the loop. It maps a source
1084 /// access to a set of dependent sink accesses.
1085 DenseMap<Instruction *, SmallPtrSet<Instruction *, 2>> Dependences;
1087 /// \brief The descriptor for a strided memory access.
1088 struct StrideDescriptor {
1089 StrideDescriptor(int64_t Stride, const SCEV *Scev, uint64_t Size,
1091 : Stride(Stride), Scev(Scev), Size(Size), Align(Align) {}
1093 StrideDescriptor() = default;
1095 // The access's stride. It is negative for a reverse access.
1097 const SCEV *Scev = nullptr; // The scalar expression of this access
1098 uint64_t Size = 0; // The size of the memory object.
1099 unsigned Align = 0; // The alignment of this access.
1102 /// \brief A type for holding instructions and their stride descriptors.
1103 typedef std::pair<Instruction *, StrideDescriptor> StrideEntry;
1105 /// \brief Create a new interleave group with the given instruction \p Instr,
1106 /// stride \p Stride and alignment \p Align.
1108 /// \returns the newly created interleave group.
1109 InterleaveGroup *createInterleaveGroup(Instruction *Instr, int Stride,
1111 assert(!InterleaveGroupMap.count(Instr) &&
1112 "Already in an interleaved access group");
1113 InterleaveGroupMap[Instr] = new InterleaveGroup(Instr, Stride, Align);
1114 return InterleaveGroupMap[Instr];
1117 /// \brief Release the group and remove all the relationships.
1118 void releaseGroup(InterleaveGroup *Group) {
1119 for (unsigned i = 0; i < Group->getFactor(); i++)
1120 if (Instruction *Member = Group->getMember(i))
1121 InterleaveGroupMap.erase(Member);
1126 /// \brief Collect all the accesses with a constant stride in program order.
1127 void collectConstStrideAccesses(
1128 MapVector<Instruction *, StrideDescriptor> &AccessStrideInfo,
1129 const ValueToValueMap &Strides);
1131 /// \brief Returns true if \p Stride is allowed in an interleaved group.
1132 static bool isStrided(int Stride) {
1133 unsigned Factor = std::abs(Stride);
1134 return Factor >= 2 && Factor <= MaxInterleaveGroupFactor;
1137 /// \brief Returns true if \p BB is a predicated block.
1138 bool isPredicated(BasicBlock *BB) const {
1139 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
1142 /// \brief Returns true if LoopAccessInfo can be used for dependence queries.
1143 bool areDependencesValid() const {
1144 return LAI && LAI->getDepChecker().getDependences();
1147 /// \brief Returns true if memory accesses \p A and \p B can be reordered, if
1148 /// necessary, when constructing interleaved groups.
1150 /// \p A must precede \p B in program order. We return false if reordering is
1151 /// not necessary or is prevented because \p A and \p B may be dependent.
1152 bool canReorderMemAccessesForInterleavedGroups(StrideEntry *A,
1153 StrideEntry *B) const {
1155 // Code motion for interleaved accesses can potentially hoist strided loads
1156 // and sink strided stores. The code below checks the legality of the
1157 // following two conditions:
1159 // 1. Potentially moving a strided load (B) before any store (A) that
1162 // 2. Potentially moving a strided store (A) after any load or store (B)
1165 // It's legal to reorder A and B if we know there isn't a dependence from A
1166 // to B. Note that this determination is conservative since some
1167 // dependences could potentially be reordered safely.
1169 // A is potentially the source of a dependence.
1170 auto *Src = A->first;
1171 auto SrcDes = A->second;
1173 // B is potentially the sink of a dependence.
1174 auto *Sink = B->first;
1175 auto SinkDes = B->second;
1177 // Code motion for interleaved accesses can't violate WAR dependences.
1178 // Thus, reordering is legal if the source isn't a write.
1179 if (!Src->mayWriteToMemory())
1182 // At least one of the accesses must be strided.
1183 if (!isStrided(SrcDes.Stride) && !isStrided(SinkDes.Stride))
1186 // If dependence information is not available from LoopAccessInfo,
1187 // conservatively assume the instructions can't be reordered.
1188 if (!areDependencesValid())
1191 // If we know there is a dependence from source to sink, assume the
1192 // instructions can't be reordered. Otherwise, reordering is legal.
1193 return !Dependences.count(Src) || !Dependences.lookup(Src).count(Sink);
1196 /// \brief Collect the dependences from LoopAccessInfo.
1198 /// We process the dependences once during the interleaved access analysis to
1199 /// enable constant-time dependence queries.
1200 void collectDependences() {
1201 if (!areDependencesValid())
1203 auto *Deps = LAI->getDepChecker().getDependences();
1204 for (auto Dep : *Deps)
1205 Dependences[Dep.getSource(*LAI)].insert(Dep.getDestination(*LAI));
1209 /// Utility class for getting and setting loop vectorizer hints in the form
1210 /// of loop metadata.
1211 /// This class keeps a number of loop annotations locally (as member variables)
1212 /// and can, upon request, write them back as metadata on the loop. It will
1213 /// initially scan the loop for existing metadata, and will update the local
1214 /// values based on information in the loop.
1215 /// We cannot write all values to metadata, as the mere presence of some info,
1216 /// for example 'force', means a decision has been made. So, we need to be
1217 /// careful NOT to add them if the user hasn't specifically asked so.
1218 class LoopVectorizeHints {
1219 enum HintKind { HK_WIDTH, HK_UNROLL, HK_FORCE };
1221 /// Hint - associates name and validation with the hint value.
1224 unsigned Value; // This may have to change for non-numeric values.
1227 Hint(const char *Name, unsigned Value, HintKind Kind)
1228 : Name(Name), Value(Value), Kind(Kind) {}
1230 bool validate(unsigned Val) {
1233 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
1235 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1243 /// Vectorization width.
1245 /// Vectorization interleave factor.
1247 /// Vectorization forced
1250 /// Return the loop metadata prefix.
1251 static StringRef Prefix() { return "llvm.loop."; }
1253 /// True if there is any unsafe math in the loop.
1254 bool PotentiallyUnsafe;
1258 FK_Undefined = -1, ///< Not selected.
1259 FK_Disabled = 0, ///< Forcing disabled.
1260 FK_Enabled = 1, ///< Forcing enabled.
1263 LoopVectorizeHints(const Loop *L, bool DisableInterleaving,
1264 OptimizationRemarkEmitter &ORE)
1265 : Width("vectorize.width", VectorizerParams::VectorizationFactor,
1267 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1268 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1269 PotentiallyUnsafe(false), TheLoop(L), ORE(ORE) {
1270 // Populate values with existing loop metadata.
1271 getHintsFromMetadata();
1273 // force-vector-interleave overrides DisableInterleaving.
1274 if (VectorizerParams::isInterleaveForced())
1275 Interleave.Value = VectorizerParams::VectorizationInterleave;
1277 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
1278 << "LV: Interleaving disabled by the pass manager\n");
1281 /// Mark the loop L as already vectorized by setting the width to 1.
1282 void setAlreadyVectorized() {
1283 Width.Value = Interleave.Value = 1;
1284 Hint Hints[] = {Width, Interleave};
1285 writeHintsToMetadata(Hints);
1288 bool allowVectorization(Function *F, Loop *L, bool AlwaysVectorize) const {
1289 if (getForce() == LoopVectorizeHints::FK_Disabled) {
1290 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1291 emitRemarkWithHints();
1295 if (!AlwaysVectorize && getForce() != LoopVectorizeHints::FK_Enabled) {
1296 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1297 emitRemarkWithHints();
1301 if (getWidth() == 1 && getInterleave() == 1) {
1302 // FIXME: Add a separate metadata to indicate when the loop has already
1303 // been vectorized instead of setting width and count to 1.
1304 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1305 // FIXME: Add interleave.disable metadata. This will allow
1306 // vectorize.disable to be used without disabling the pass and errors
1307 // to differentiate between disabled vectorization and a width of 1.
1308 ORE.emit(OptimizationRemarkAnalysis(vectorizeAnalysisPassName(),
1309 "AllDisabled", L->getStartLoc(),
1311 << "loop not vectorized: vectorization and interleaving are "
1312 "explicitly disabled, or vectorize width and interleave "
1313 "count are both set to 1");
1320 /// Dumps all the hint information.
1321 void emitRemarkWithHints() const {
1322 using namespace ore;
1323 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1324 ORE.emit(OptimizationRemarkMissed(LV_NAME, "MissedExplicitlyDisabled",
1325 TheLoop->getStartLoc(),
1326 TheLoop->getHeader())
1327 << "loop not vectorized: vectorization is explicitly disabled");
1329 OptimizationRemarkMissed R(LV_NAME, "MissedDetails",
1330 TheLoop->getStartLoc(), TheLoop->getHeader());
1331 R << "loop not vectorized";
1332 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1333 R << " (Force=" << NV("Force", true);
1334 if (Width.Value != 0)
1335 R << ", Vector Width=" << NV("VectorWidth", Width.Value);
1336 if (Interleave.Value != 0)
1337 R << ", Interleave Count=" << NV("InterleaveCount", Interleave.Value);
1344 unsigned getWidth() const { return Width.Value; }
1345 unsigned getInterleave() const { return Interleave.Value; }
1346 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1348 /// \brief If hints are provided that force vectorization, use the AlwaysPrint
1349 /// pass name to force the frontend to print the diagnostic.
1350 const char *vectorizeAnalysisPassName() const {
1351 if (getWidth() == 1)
1353 if (getForce() == LoopVectorizeHints::FK_Disabled)
1355 if (getForce() == LoopVectorizeHints::FK_Undefined && getWidth() == 0)
1357 return OptimizationRemarkAnalysis::AlwaysPrint;
1360 bool allowReordering() const {
1361 // When enabling loop hints are provided we allow the vectorizer to change
1362 // the order of operations that is given by the scalar loop. This is not
1363 // enabled by default because can be unsafe or inefficient. For example,
1364 // reordering floating-point operations will change the way round-off
1365 // error accumulates in the loop.
1366 return getForce() == LoopVectorizeHints::FK_Enabled || getWidth() > 1;
1369 bool isPotentiallyUnsafe() const {
1370 // Avoid FP vectorization if the target is unsure about proper support.
1371 // This may be related to the SIMD unit in the target not handling
1372 // IEEE 754 FP ops properly, or bad single-to-double promotions.
1373 // Otherwise, a sequence of vectorized loops, even without reduction,
1374 // could lead to different end results on the destination vectors.
1375 return getForce() != LoopVectorizeHints::FK_Enabled && PotentiallyUnsafe;
1378 void setPotentiallyUnsafe() { PotentiallyUnsafe = true; }
1381 /// Find hints specified in the loop metadata and update local values.
1382 void getHintsFromMetadata() {
1383 MDNode *LoopID = TheLoop->getLoopID();
1387 // First operand should refer to the loop id itself.
1388 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1389 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1391 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1392 const MDString *S = nullptr;
1393 SmallVector<Metadata *, 4> Args;
1395 // The expected hint is either a MDString or a MDNode with the first
1396 // operand a MDString.
1397 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1398 if (!MD || MD->getNumOperands() == 0)
1400 S = dyn_cast<MDString>(MD->getOperand(0));
1401 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1402 Args.push_back(MD->getOperand(i));
1404 S = dyn_cast<MDString>(LoopID->getOperand(i));
1405 assert(Args.size() == 0 && "too many arguments for MDString");
1411 // Check if the hint starts with the loop metadata prefix.
1412 StringRef Name = S->getString();
1413 if (Args.size() == 1)
1414 setHint(Name, Args[0]);
1418 /// Checks string hint with one operand and set value if valid.
1419 void setHint(StringRef Name, Metadata *Arg) {
1420 if (!Name.startswith(Prefix()))
1422 Name = Name.substr(Prefix().size(), StringRef::npos);
1424 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1427 unsigned Val = C->getZExtValue();
1429 Hint *Hints[] = {&Width, &Interleave, &Force};
1430 for (auto H : Hints) {
1431 if (Name == H->Name) {
1432 if (H->validate(Val))
1435 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1441 /// Create a new hint from name / value pair.
1442 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1443 LLVMContext &Context = TheLoop->getHeader()->getContext();
1444 Metadata *MDs[] = {MDString::get(Context, Name),
1445 ConstantAsMetadata::get(
1446 ConstantInt::get(Type::getInt32Ty(Context), V))};
1447 return MDNode::get(Context, MDs);
1450 /// Matches metadata with hint name.
1451 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1452 MDString *Name = dyn_cast<MDString>(Node->getOperand(0));
1456 for (auto H : HintTypes)
1457 if (Name->getString().endswith(H.Name))
1462 /// Sets current hints into loop metadata, keeping other values intact.
1463 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1464 if (HintTypes.size() == 0)
1467 // Reserve the first element to LoopID (see below).
1468 SmallVector<Metadata *, 4> MDs(1);
1469 // If the loop already has metadata, then ignore the existing operands.
1470 MDNode *LoopID = TheLoop->getLoopID();
1472 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1473 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1474 // If node in update list, ignore old value.
1475 if (!matchesHintMetadataName(Node, HintTypes))
1476 MDs.push_back(Node);
1480 // Now, add the missing hints.
1481 for (auto H : HintTypes)
1482 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1484 // Replace current metadata node with new one.
1485 LLVMContext &Context = TheLoop->getHeader()->getContext();
1486 MDNode *NewLoopID = MDNode::get(Context, MDs);
1487 // Set operand 0 to refer to the loop id itself.
1488 NewLoopID->replaceOperandWith(0, NewLoopID);
1490 TheLoop->setLoopID(NewLoopID);
1493 /// The loop these hints belong to.
1494 const Loop *TheLoop;
1496 /// Interface to emit optimization remarks.
1497 OptimizationRemarkEmitter &ORE;
1500 static void emitAnalysisDiag(const Loop *TheLoop,
1501 const LoopVectorizeHints &Hints,
1502 OptimizationRemarkEmitter &ORE,
1503 const LoopAccessReport &Message) {
1504 const char *Name = Hints.vectorizeAnalysisPassName();
1505 LoopAccessReport::emitAnalysis(Message, TheLoop, Name, ORE);
1508 static void emitMissedWarning(Function *F, Loop *L,
1509 const LoopVectorizeHints &LH,
1510 OptimizationRemarkEmitter *ORE) {
1511 LH.emitRemarkWithHints();
1513 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1514 if (LH.getWidth() != 1)
1515 emitLoopVectorizeWarning(
1516 F->getContext(), *F, L->getStartLoc(),
1517 "failed explicitly specified loop vectorization");
1518 else if (LH.getInterleave() != 1)
1519 emitLoopInterleaveWarning(
1520 F->getContext(), *F, L->getStartLoc(),
1521 "failed explicitly specified loop interleaving");
1525 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
1526 /// to what vectorization factor.
1527 /// This class does not look at the profitability of vectorization, only the
1528 /// legality. This class has two main kinds of checks:
1529 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
1530 /// will change the order of memory accesses in a way that will change the
1531 /// correctness of the program.
1532 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
1533 /// checks for a number of different conditions, such as the availability of a
1534 /// single induction variable, that all types are supported and vectorize-able,
1535 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
1536 /// This class is also used by InnerLoopVectorizer for identifying
1537 /// induction variable and the different reduction variables.
1538 class LoopVectorizationLegality {
1540 LoopVectorizationLegality(
1541 Loop *L, PredicatedScalarEvolution &PSE, DominatorTree *DT,
1542 TargetLibraryInfo *TLI, AliasAnalysis *AA, Function *F,
1543 const TargetTransformInfo *TTI,
1544 std::function<const LoopAccessInfo &(Loop &)> *GetLAA, LoopInfo *LI,
1545 OptimizationRemarkEmitter *ORE, LoopVectorizationRequirements *R,
1546 LoopVectorizeHints *H)
1547 : NumPredStores(0), TheLoop(L), PSE(PSE), TLI(TLI), TTI(TTI), DT(DT),
1548 GetLAA(GetLAA), LAI(nullptr), ORE(ORE), InterleaveInfo(PSE, L, DT, LI),
1549 Induction(nullptr), WidestIndTy(nullptr), HasFunNoNaNAttr(false),
1550 Requirements(R), Hints(H) {}
1552 /// ReductionList contains the reduction descriptors for all
1553 /// of the reductions that were found in the loop.
1554 typedef DenseMap<PHINode *, RecurrenceDescriptor> ReductionList;
1556 /// InductionList saves induction variables and maps them to the
1557 /// induction descriptor.
1558 typedef MapVector<PHINode *, InductionDescriptor> InductionList;
1560 /// RecurrenceSet contains the phi nodes that are recurrences other than
1561 /// inductions and reductions.
1562 typedef SmallPtrSet<const PHINode *, 8> RecurrenceSet;
1564 /// Returns true if it is legal to vectorize this loop.
1565 /// This does not mean that it is profitable to vectorize this
1566 /// loop, only that it is legal to do so.
1567 bool canVectorize();
1569 /// Returns the Induction variable.
1570 PHINode *getInduction() { return Induction; }
1572 /// Returns the reduction variables found in the loop.
1573 ReductionList *getReductionVars() { return &Reductions; }
1575 /// Returns the induction variables found in the loop.
1576 InductionList *getInductionVars() { return &Inductions; }
1578 /// Return the first-order recurrences found in the loop.
1579 RecurrenceSet *getFirstOrderRecurrences() { return &FirstOrderRecurrences; }
1581 /// Returns the widest induction type.
1582 Type *getWidestInductionType() { return WidestIndTy; }
1584 /// Returns True if V is an induction variable in this loop.
1585 bool isInductionVariable(const Value *V);
1587 /// Returns True if PN is a reduction variable in this loop.
1588 bool isReductionVariable(PHINode *PN) { return Reductions.count(PN); }
1590 /// Returns True if Phi is a first-order recurrence in this loop.
1591 bool isFirstOrderRecurrence(const PHINode *Phi);
1593 /// Return true if the block BB needs to be predicated in order for the loop
1594 /// to be vectorized.
1595 bool blockNeedsPredication(BasicBlock *BB);
1597 /// Check if this pointer is consecutive when vectorizing. This happens
1598 /// when the last index of the GEP is the induction variable, or that the
1599 /// pointer itself is an induction variable.
1600 /// This check allows us to vectorize A[idx] into a wide load/store.
1602 /// 0 - Stride is unknown or non-consecutive.
1603 /// 1 - Address is consecutive.
1604 /// -1 - Address is consecutive, and decreasing.
1605 int isConsecutivePtr(Value *Ptr);
1607 /// Returns true if the value V is uniform within the loop.
1608 bool isUniform(Value *V);
1610 /// Returns true if \p I is known to be uniform after vectorization.
1611 bool isUniformAfterVectorization(Instruction *I) { return Uniforms.count(I); }
1613 /// Returns true if \p I is known to be scalar after vectorization.
1614 bool isScalarAfterVectorization(Instruction *I) { return Scalars.count(I); }
1616 /// Returns the information that we collected about runtime memory check.
1617 const RuntimePointerChecking *getRuntimePointerChecking() const {
1618 return LAI->getRuntimePointerChecking();
1621 const LoopAccessInfo *getLAI() const { return LAI; }
1623 /// \brief Check if \p Instr belongs to any interleaved access group.
1624 bool isAccessInterleaved(Instruction *Instr) {
1625 return InterleaveInfo.isInterleaved(Instr);
1628 /// \brief Return the maximum interleave factor of all interleaved groups.
1629 unsigned getMaxInterleaveFactor() const {
1630 return InterleaveInfo.getMaxInterleaveFactor();
1633 /// \brief Get the interleaved access group that \p Instr belongs to.
1634 const InterleaveGroup *getInterleavedAccessGroup(Instruction *Instr) {
1635 return InterleaveInfo.getInterleaveGroup(Instr);
1638 /// \brief Returns true if an interleaved group requires a scalar iteration
1639 /// to handle accesses with gaps.
1640 bool requiresScalarEpilogue() const {
1641 return InterleaveInfo.requiresScalarEpilogue();
1644 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
1646 bool hasStride(Value *V) { return LAI->hasStride(V); }
1648 /// Returns true if the target machine supports masked store operation
1649 /// for the given \p DataType and kind of access to \p Ptr.
1650 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
1651 return isConsecutivePtr(Ptr) && TTI->isLegalMaskedStore(DataType);
1653 /// Returns true if the target machine supports masked load operation
1654 /// for the given \p DataType and kind of access to \p Ptr.
1655 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
1656 return isConsecutivePtr(Ptr) && TTI->isLegalMaskedLoad(DataType);
1658 /// Returns true if the target machine supports masked scatter operation
1659 /// for the given \p DataType.
1660 bool isLegalMaskedScatter(Type *DataType) {
1661 return TTI->isLegalMaskedScatter(DataType);
1663 /// Returns true if the target machine supports masked gather operation
1664 /// for the given \p DataType.
1665 bool isLegalMaskedGather(Type *DataType) {
1666 return TTI->isLegalMaskedGather(DataType);
1668 /// Returns true if the target machine can represent \p V as a masked gather
1669 /// or scatter operation.
1670 bool isLegalGatherOrScatter(Value *V) {
1671 auto *LI = dyn_cast<LoadInst>(V);
1672 auto *SI = dyn_cast<StoreInst>(V);
1675 auto *Ptr = getPointerOperand(V);
1676 auto *Ty = cast<PointerType>(Ptr->getType())->getElementType();
1677 return (LI && isLegalMaskedGather(Ty)) || (SI && isLegalMaskedScatter(Ty));
1680 /// Returns true if vector representation of the instruction \p I
1682 bool isMaskRequired(const Instruction *I) { return (MaskedOp.count(I) != 0); }
1683 unsigned getNumStores() const { return LAI->getNumStores(); }
1684 unsigned getNumLoads() const { return LAI->getNumLoads(); }
1685 unsigned getNumPredStores() const { return NumPredStores; }
1687 /// Returns true if \p I is an instruction that will be scalarized with
1688 /// predication. Such instructions include conditional stores and
1689 /// instructions that may divide by zero.
1690 bool isScalarWithPredication(Instruction *I);
1692 /// Returns true if \p I is a memory instruction that has a consecutive or
1693 /// consecutive-like pointer operand. Consecutive-like pointers are pointers
1694 /// that are treated like consecutive pointers during vectorization. The
1695 /// pointer operands of interleaved accesses are an example.
1696 bool hasConsecutiveLikePtrOperand(Instruction *I);
1698 /// Returns true if \p I is a memory instruction that must be scalarized
1699 /// during vectorization.
1700 bool memoryInstructionMustBeScalarized(Instruction *I, unsigned VF = 1);
1703 /// Check if a single basic block loop is vectorizable.
1704 /// At this point we know that this is a loop with a constant trip count
1705 /// and we only need to check individual instructions.
1706 bool canVectorizeInstrs();
1708 /// When we vectorize loops we may change the order in which
1709 /// we read and write from memory. This method checks if it is
1710 /// legal to vectorize the code, considering only memory constrains.
1711 /// Returns true if the loop is vectorizable
1712 bool canVectorizeMemory();
1714 /// Return true if we can vectorize this loop using the IF-conversion
1716 bool canVectorizeWithIfConvert();
1718 /// Collect the instructions that are uniform after vectorization. An
1719 /// instruction is uniform if we represent it with a single scalar value in
1720 /// the vectorized loop corresponding to each vector iteration. Examples of
1721 /// uniform instructions include pointer operands of consecutive or
1722 /// interleaved memory accesses. Note that although uniformity implies an
1723 /// instruction will be scalar, the reverse is not true. In general, a
1724 /// scalarized instruction will be represented by VF scalar values in the
1725 /// vectorized loop, each corresponding to an iteration of the original
1727 void collectLoopUniforms();
1729 /// Collect the instructions that are scalar after vectorization. An
1730 /// instruction is scalar if it is known to be uniform or will be scalarized
1731 /// during vectorization. Non-uniform scalarized instructions will be
1732 /// represented by VF values in the vectorized loop, each corresponding to an
1733 /// iteration of the original scalar loop.
1734 void collectLoopScalars();
1736 /// Return true if all of the instructions in the block can be speculatively
1737 /// executed. \p SafePtrs is a list of addresses that are known to be legal
1738 /// and we know that we can read from them without segfault.
1739 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
1741 /// Updates the vectorization state by adding \p Phi to the inductions list.
1742 /// This can set \p Phi as the main induction of the loop if \p Phi is a
1743 /// better choice for the main induction than the existing one.
1744 void addInductionPhi(PHINode *Phi, const InductionDescriptor &ID,
1745 SmallPtrSetImpl<Value *> &AllowedExit);
1747 /// Report an analysis message to assist the user in diagnosing loops that are
1748 /// not vectorized. These are handled as LoopAccessReport rather than
1749 /// VectorizationReport because the << operator of VectorizationReport returns
1750 /// LoopAccessReport.
1751 void emitAnalysis(const LoopAccessReport &Message) const {
1752 emitAnalysisDiag(TheLoop, *Hints, *ORE, Message);
1755 /// Create an analysis remark that explains why vectorization failed
1757 /// \p RemarkName is the identifier for the remark. If \p I is passed it is
1758 /// an instruction that prevents vectorization. Otherwise the loop is used
1759 /// for the location of the remark. \return the remark object that can be
1761 OptimizationRemarkAnalysis
1762 createMissedAnalysis(StringRef RemarkName, Instruction *I = nullptr) const {
1763 return ::createMissedAnalysis(Hints->vectorizeAnalysisPassName(),
1764 RemarkName, TheLoop, I);
1767 /// \brief If an access has a symbolic strides, this maps the pointer value to
1768 /// the stride symbol.
1769 const ValueToValueMap *getSymbolicStrides() {
1770 // FIXME: Currently, the set of symbolic strides is sometimes queried before
1771 // it's collected. This happens from canVectorizeWithIfConvert, when the
1772 // pointer is checked to reference consecutive elements suitable for a
1774 return LAI ? &LAI->getSymbolicStrides() : nullptr;
1777 unsigned NumPredStores;
1779 /// The loop that we evaluate.
1781 /// A wrapper around ScalarEvolution used to add runtime SCEV checks.
1782 /// Applies dynamic knowledge to simplify SCEV expressions in the context
1783 /// of existing SCEV assumptions. The analysis will also add a minimal set
1784 /// of new predicates if this is required to enable vectorization and
1786 PredicatedScalarEvolution &PSE;
1787 /// Target Library Info.
1788 TargetLibraryInfo *TLI;
1789 /// Target Transform Info
1790 const TargetTransformInfo *TTI;
1793 // LoopAccess analysis.
1794 std::function<const LoopAccessInfo &(Loop &)> *GetLAA;
1795 // And the loop-accesses info corresponding to this loop. This pointer is
1796 // null until canVectorizeMemory sets it up.
1797 const LoopAccessInfo *LAI;
1798 /// Interface to emit optimization remarks.
1799 OptimizationRemarkEmitter *ORE;
1801 /// The interleave access information contains groups of interleaved accesses
1802 /// with the same stride and close to each other.
1803 InterleavedAccessInfo InterleaveInfo;
1805 // --- vectorization state --- //
1807 /// Holds the integer induction variable. This is the counter of the
1810 /// Holds the reduction variables.
1811 ReductionList Reductions;
1812 /// Holds all of the induction variables that we found in the loop.
1813 /// Notice that inductions don't need to start at zero and that induction
1814 /// variables can be pointers.
1815 InductionList Inductions;
1816 /// Holds the phi nodes that are first-order recurrences.
1817 RecurrenceSet FirstOrderRecurrences;
1818 /// Holds the widest induction type encountered.
1821 /// Allowed outside users. This holds the induction and reduction
1822 /// vars which can be accessed from outside the loop.
1823 SmallPtrSet<Value *, 4> AllowedExit;
1825 /// Holds the instructions known to be uniform after vectorization.
1826 SmallPtrSet<Instruction *, 4> Uniforms;
1828 /// Holds the instructions known to be scalar after vectorization.
1829 SmallPtrSet<Instruction *, 4> Scalars;
1831 /// Can we assume the absence of NaNs.
1832 bool HasFunNoNaNAttr;
1834 /// Vectorization requirements that will go through late-evaluation.
1835 LoopVectorizationRequirements *Requirements;
1837 /// Used to emit an analysis of any legality issues.
1838 LoopVectorizeHints *Hints;
1840 /// While vectorizing these instructions we have to generate a
1841 /// call to the appropriate masked intrinsic
1842 SmallPtrSet<const Instruction *, 8> MaskedOp;
1845 /// LoopVectorizationCostModel - estimates the expected speedups due to
1847 /// In many cases vectorization is not profitable. This can happen because of
1848 /// a number of reasons. In this class we mainly attempt to predict the
1849 /// expected speedup/slowdowns due to the supported instruction set. We use the
1850 /// TargetTransformInfo to query the different backends for the cost of
1851 /// different operations.
1852 class LoopVectorizationCostModel {
1854 LoopVectorizationCostModel(Loop *L, PredicatedScalarEvolution &PSE,
1855 LoopInfo *LI, LoopVectorizationLegality *Legal,
1856 const TargetTransformInfo &TTI,
1857 const TargetLibraryInfo *TLI, DemandedBits *DB,
1858 AssumptionCache *AC,
1859 OptimizationRemarkEmitter *ORE, const Function *F,
1860 const LoopVectorizeHints *Hints)
1861 : TheLoop(L), PSE(PSE), LI(LI), Legal(Legal), TTI(TTI), TLI(TLI), DB(DB),
1862 AC(AC), ORE(ORE), TheFunction(F), Hints(Hints) {}
1864 /// Information about vectorization costs
1865 struct VectorizationFactor {
1866 unsigned Width; // Vector width with best cost
1867 unsigned Cost; // Cost of the loop with that width
1869 /// \return The most profitable vectorization factor and the cost of that VF.
1870 /// This method checks every power of two up to VF. If UserVF is not ZERO
1871 /// then this vectorization factor will be selected if vectorization is
1873 VectorizationFactor selectVectorizationFactor(bool OptForSize);
1875 /// \return The size (in bits) of the smallest and widest types in the code
1876 /// that needs to be vectorized. We ignore values that remain scalar such as
1877 /// 64 bit loop indices.
1878 std::pair<unsigned, unsigned> getSmallestAndWidestTypes();
1880 /// \return The desired interleave count.
1881 /// If interleave count has been specified by metadata it will be returned.
1882 /// Otherwise, the interleave count is computed and returned. VF and LoopCost
1883 /// are the selected vectorization factor and the cost of the selected VF.
1884 unsigned selectInterleaveCount(bool OptForSize, unsigned VF,
1887 /// \brief A struct that represents some properties of the register usage
1889 struct RegisterUsage {
1890 /// Holds the number of loop invariant values that are used in the loop.
1891 unsigned LoopInvariantRegs;
1892 /// Holds the maximum number of concurrent live intervals in the loop.
1893 unsigned MaxLocalUsers;
1894 /// Holds the number of instructions in the loop.
1895 unsigned NumInstructions;
1898 /// \return Returns information about the register usages of the loop for the
1899 /// given vectorization factors.
1900 SmallVector<RegisterUsage, 8> calculateRegisterUsage(ArrayRef<unsigned> VFs);
1902 /// Collect values we want to ignore in the cost model.
1903 void collectValuesToIgnore();
1905 /// \returns The smallest bitwidth each instruction can be represented with.
1906 /// The vector equivalents of these instructions should be truncated to this
1908 const MapVector<Instruction *, uint64_t> &getMinimalBitwidths() const {
1912 /// \returns True if it is more profitable to scalarize instruction \p I for
1913 /// vectorization factor \p VF.
1914 bool isProfitableToScalarize(Instruction *I, unsigned VF) const {
1915 auto Scalars = InstsToScalarize.find(VF);
1916 assert(Scalars != InstsToScalarize.end() &&
1917 "VF not yet analyzed for scalarization profitability");
1918 return Scalars->second.count(I);
1921 /// \returns True if instruction \p I can be truncated to a smaller bitwidth
1922 /// for vectorization factor \p VF.
1923 bool canTruncateToMinimalBitwidth(Instruction *I, unsigned VF) const {
1924 return VF > 1 && MinBWs.count(I) && !isProfitableToScalarize(I, VF) &&
1925 !Legal->isScalarAfterVectorization(I);
1929 /// The vectorization cost is a combination of the cost itself and a boolean
1930 /// indicating whether any of the contributing operations will actually
1932 /// vector values after type legalization in the backend. If this latter value
1934 /// false, then all operations will be scalarized (i.e. no vectorization has
1935 /// actually taken place).
1936 typedef std::pair<unsigned, bool> VectorizationCostTy;
1938 /// Returns the expected execution cost. The unit of the cost does
1939 /// not matter because we use the 'cost' units to compare different
1940 /// vector widths. The cost that is returned is *not* normalized by
1941 /// the factor width.
1942 VectorizationCostTy expectedCost(unsigned VF);
1944 /// Returns the execution time cost of an instruction for a given vector
1945 /// width. Vector width of one means scalar.
1946 VectorizationCostTy getInstructionCost(Instruction *I, unsigned VF);
1948 /// The cost-computation logic from getInstructionCost which provides
1949 /// the vector type as an output parameter.
1950 unsigned getInstructionCost(Instruction *I, unsigned VF, Type *&VectorTy);
1952 /// Returns whether the instruction is a load or store and will be a emitted
1953 /// as a vector operation.
1954 bool isConsecutiveLoadOrStore(Instruction *I);
1956 /// Create an analysis remark that explains why vectorization failed
1958 /// \p RemarkName is the identifier for the remark. \return the remark object
1959 /// that can be streamed to.
1960 OptimizationRemarkAnalysis createMissedAnalysis(StringRef RemarkName) {
1961 return ::createMissedAnalysis(Hints->vectorizeAnalysisPassName(),
1962 RemarkName, TheLoop);
1965 /// Map of scalar integer values to the smallest bitwidth they can be legally
1966 /// represented as. The vector equivalents of these values should be truncated
1968 MapVector<Instruction *, uint64_t> MinBWs;
1970 /// A type representing the costs for instructions if they were to be
1971 /// scalarized rather than vectorized. The entries are Instruction-Cost
1973 typedef DenseMap<Instruction *, unsigned> ScalarCostsTy;
1975 /// A map holding scalar costs for different vectorization factors. The
1976 /// presence of a cost for an instruction in the mapping indicates that the
1977 /// instruction will be scalarized when vectorizing with the associated
1978 /// vectorization factor. The entries are VF-ScalarCostTy pairs.
1979 DenseMap<unsigned, ScalarCostsTy> InstsToScalarize;
1981 /// Returns the expected difference in cost from scalarizing the expression
1982 /// feeding a predicated instruction \p PredInst. The instructions to
1983 /// scalarize and their scalar costs are collected in \p ScalarCosts. A
1984 /// non-negative return value implies the expression will be scalarized.
1985 /// Currently, only single-use chains are considered for scalarization.
1986 int computePredInstDiscount(Instruction *PredInst, ScalarCostsTy &ScalarCosts,
1989 /// Collects the instructions to scalarize for each predicated instruction in
1991 void collectInstsToScalarize(unsigned VF);
1994 /// The loop that we evaluate.
1996 /// Predicated scalar evolution analysis.
1997 PredicatedScalarEvolution &PSE;
1998 /// Loop Info analysis.
2000 /// Vectorization legality.
2001 LoopVectorizationLegality *Legal;
2002 /// Vector target information.
2003 const TargetTransformInfo &TTI;
2004 /// Target Library Info.
2005 const TargetLibraryInfo *TLI;
2006 /// Demanded bits analysis.
2008 /// Assumption cache.
2009 AssumptionCache *AC;
2010 /// Interface to emit optimization remarks.
2011 OptimizationRemarkEmitter *ORE;
2013 const Function *TheFunction;
2014 /// Loop Vectorize Hint.
2015 const LoopVectorizeHints *Hints;
2016 /// Values to ignore in the cost model.
2017 SmallPtrSet<const Value *, 16> ValuesToIgnore;
2018 /// Values to ignore in the cost model when VF > 1.
2019 SmallPtrSet<const Value *, 16> VecValuesToIgnore;
2022 /// \brief This holds vectorization requirements that must be verified late in
2023 /// the process. The requirements are set by legalize and costmodel. Once
2024 /// vectorization has been determined to be possible and profitable the
2025 /// requirements can be verified by looking for metadata or compiler options.
2026 /// For example, some loops require FP commutativity which is only allowed if
2027 /// vectorization is explicitly specified or if the fast-math compiler option
2028 /// has been provided.
2029 /// Late evaluation of these requirements allows helpful diagnostics to be
2030 /// composed that tells the user what need to be done to vectorize the loop. For
2031 /// example, by specifying #pragma clang loop vectorize or -ffast-math. Late
2032 /// evaluation should be used only when diagnostics can generated that can be
2033 /// followed by a non-expert user.
2034 class LoopVectorizationRequirements {
2036 LoopVectorizationRequirements(OptimizationRemarkEmitter &ORE)
2037 : NumRuntimePointerChecks(0), UnsafeAlgebraInst(nullptr), ORE(ORE) {}
2039 void addUnsafeAlgebraInst(Instruction *I) {
2040 // First unsafe algebra instruction.
2041 if (!UnsafeAlgebraInst)
2042 UnsafeAlgebraInst = I;
2045 void addRuntimePointerChecks(unsigned Num) { NumRuntimePointerChecks = Num; }
2047 bool doesNotMeet(Function *F, Loop *L, const LoopVectorizeHints &Hints) {
2048 const char *PassName = Hints.vectorizeAnalysisPassName();
2049 bool Failed = false;
2050 if (UnsafeAlgebraInst && !Hints.allowReordering()) {
2052 OptimizationRemarkAnalysisFPCommute(PassName, "CantReorderFPOps",
2053 UnsafeAlgebraInst->getDebugLoc(),
2054 UnsafeAlgebraInst->getParent())
2055 << "loop not vectorized: cannot prove it is safe to reorder "
2056 "floating-point operations");
2060 // Test if runtime memcheck thresholds are exceeded.
2061 bool PragmaThresholdReached =
2062 NumRuntimePointerChecks > PragmaVectorizeMemoryCheckThreshold;
2063 bool ThresholdReached =
2064 NumRuntimePointerChecks > VectorizerParams::RuntimeMemoryCheckThreshold;
2065 if ((ThresholdReached && !Hints.allowReordering()) ||
2066 PragmaThresholdReached) {
2067 ORE.emit(OptimizationRemarkAnalysisAliasing(PassName, "CantReorderMemOps",
2070 << "loop not vectorized: cannot prove it is safe to reorder "
2071 "memory operations");
2072 DEBUG(dbgs() << "LV: Too many memory checks needed.\n");
2080 unsigned NumRuntimePointerChecks;
2081 Instruction *UnsafeAlgebraInst;
2083 /// Interface to emit optimization remarks.
2084 OptimizationRemarkEmitter &ORE;
2087 static void addAcyclicInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
2089 if (!hasCyclesInLoopBody(L))
2093 for (Loop *InnerL : L)
2094 addAcyclicInnerLoop(*InnerL, V);
2097 /// The LoopVectorize Pass.
2098 struct LoopVectorize : public FunctionPass {
2099 /// Pass identification, replacement for typeid
2102 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
2103 : FunctionPass(ID) {
2104 Impl.DisableUnrolling = NoUnrolling;
2105 Impl.AlwaysVectorize = AlwaysVectorize;
2106 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
2109 LoopVectorizePass Impl;
2111 bool runOnFunction(Function &F) override {
2112 if (skipFunction(F))
2115 auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
2116 auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
2117 auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2118 auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2119 auto *BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI();
2120 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
2121 auto *TLI = TLIP ? &TLIP->getTLI() : nullptr;
2122 auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
2123 auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
2124 auto *LAA = &getAnalysis<LoopAccessLegacyAnalysis>();
2125 auto *DB = &getAnalysis<DemandedBitsWrapperPass>().getDemandedBits();
2126 auto *ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
2128 std::function<const LoopAccessInfo &(Loop &)> GetLAA =
2129 [&](Loop &L) -> const LoopAccessInfo & { return LAA->getInfo(&L); };
2131 return Impl.runImpl(F, *SE, *LI, *TTI, *DT, *BFI, TLI, *DB, *AA, *AC,
2135 void getAnalysisUsage(AnalysisUsage &AU) const override {
2136 AU.addRequired<AssumptionCacheTracker>();
2137 AU.addRequiredID(LoopSimplifyID);
2138 AU.addRequiredID(LCSSAID);
2139 AU.addRequired<BlockFrequencyInfoWrapperPass>();
2140 AU.addRequired<DominatorTreeWrapperPass>();
2141 AU.addRequired<LoopInfoWrapperPass>();
2142 AU.addRequired<ScalarEvolutionWrapperPass>();
2143 AU.addRequired<TargetTransformInfoWrapperPass>();
2144 AU.addRequired<AAResultsWrapperPass>();
2145 AU.addRequired<LoopAccessLegacyAnalysis>();
2146 AU.addRequired<DemandedBitsWrapperPass>();
2147 AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
2148 AU.addPreserved<LoopInfoWrapperPass>();
2149 AU.addPreserved<DominatorTreeWrapperPass>();
2150 AU.addPreserved<BasicAAWrapperPass>();
2151 AU.addPreserved<GlobalsAAWrapperPass>();
2155 } // end anonymous namespace
2157 //===----------------------------------------------------------------------===//
2158 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
2159 // LoopVectorizationCostModel.
2160 //===----------------------------------------------------------------------===//
2162 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
2163 // We need to place the broadcast of invariant variables outside the loop.
2164 Instruction *Instr = dyn_cast<Instruction>(V);
2165 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
2166 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
2168 // Place the code for broadcasting invariant variables in the new preheader.
2169 IRBuilder<>::InsertPointGuard Guard(Builder);
2171 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
2173 // Broadcast the scalar into all locations in the vector.
2174 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
2179 void InnerLoopVectorizer::createVectorIntInductionPHI(
2180 const InductionDescriptor &II, Instruction *EntryVal) {
2181 Value *Start = II.getStartValue();
2182 ConstantInt *Step = II.getConstIntStepValue();
2183 assert(Step && "Can not widen an IV with a non-constant step");
2185 // Construct the initial value of the vector IV in the vector loop preheader
2186 auto CurrIP = Builder.saveIP();
2187 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
2188 if (isa<TruncInst>(EntryVal)) {
2189 auto *TruncType = cast<IntegerType>(EntryVal->getType());
2190 Step = ConstantInt::getSigned(TruncType, Step->getSExtValue());
2191 Start = Builder.CreateCast(Instruction::Trunc, Start, TruncType);
2193 Value *SplatStart = Builder.CreateVectorSplat(VF, Start);
2194 Value *SteppedStart = getStepVector(SplatStart, 0, Step);
2195 Builder.restoreIP(CurrIP);
2198 ConstantVector::getSplat(VF, ConstantInt::getSigned(Start->getType(),
2199 VF * Step->getSExtValue()));
2200 // We may need to add the step a number of times, depending on the unroll
2201 // factor. The last of those goes into the PHI.
2202 PHINode *VecInd = PHINode::Create(SteppedStart->getType(), 2, "vec.ind",
2203 &*LoopVectorBody->getFirstInsertionPt());
2204 Instruction *LastInduction = VecInd;
2205 VectorParts Entry(UF);
2206 for (unsigned Part = 0; Part < UF; ++Part) {
2207 Entry[Part] = LastInduction;
2208 LastInduction = cast<Instruction>(
2209 Builder.CreateAdd(LastInduction, SplatVF, "step.add"));
2211 VectorLoopValueMap.initVector(EntryVal, Entry);
2212 if (isa<TruncInst>(EntryVal))
2213 addMetadata(Entry, EntryVal);
2215 // Move the last step to the end of the latch block. This ensures consistent
2216 // placement of all induction updates.
2217 auto *LoopVectorLatch = LI->getLoopFor(LoopVectorBody)->getLoopLatch();
2218 auto *Br = cast<BranchInst>(LoopVectorLatch->getTerminator());
2219 auto *ICmp = cast<Instruction>(Br->getCondition());
2220 LastInduction->moveBefore(ICmp);
2221 LastInduction->setName("vec.ind.next");
2223 VecInd->addIncoming(SteppedStart, LoopVectorPreHeader);
2224 VecInd->addIncoming(LastInduction, LoopVectorLatch);
2227 bool InnerLoopVectorizer::shouldScalarizeInstruction(Instruction *I) const {
2228 return Legal->isScalarAfterVectorization(I) ||
2229 Cost->isProfitableToScalarize(I, VF);
2232 bool InnerLoopVectorizer::needsScalarInduction(Instruction *IV) const {
2233 if (shouldScalarizeInstruction(IV))
2235 auto isScalarInst = [&](User *U) -> bool {
2236 auto *I = cast<Instruction>(U);
2237 return (OrigLoop->contains(I) && shouldScalarizeInstruction(I));
2239 return any_of(IV->users(), isScalarInst);
2242 void InnerLoopVectorizer::widenIntInduction(PHINode *IV, TruncInst *Trunc) {
2244 auto II = Legal->getInductionVars()->find(IV);
2245 assert(II != Legal->getInductionVars()->end() && "IV is not an induction");
2247 auto ID = II->second;
2248 assert(IV->getType() == ID.getStartValue()->getType() && "Types must match");
2250 // The scalar value to broadcast. This will be derived from the canonical
2251 // induction variable.
2252 Value *ScalarIV = nullptr;
2254 // The step of the induction.
2255 Value *Step = nullptr;
2257 // The value from the original loop to which we are mapping the new induction
2259 Instruction *EntryVal = Trunc ? cast<Instruction>(Trunc) : IV;
2261 // True if we have vectorized the induction variable.
2262 auto VectorizedIV = false;
2264 // Determine if we want a scalar version of the induction variable. This is
2265 // true if the induction variable itself is not widened, or if it has at
2266 // least one user in the loop that is not widened.
2267 auto NeedsScalarIV = VF > 1 && needsScalarInduction(EntryVal);
2269 // If the induction variable has a constant integer step value, go ahead and
2271 if (ID.getConstIntStepValue())
2272 Step = ID.getConstIntStepValue();
2274 // Try to create a new independent vector induction variable. If we can't
2275 // create the phi node, we will splat the scalar induction variable in each
2277 if (VF > 1 && IV->getType() == Induction->getType() && Step &&
2278 !shouldScalarizeInstruction(EntryVal)) {
2279 createVectorIntInductionPHI(ID, EntryVal);
2280 VectorizedIV = true;
2283 // If we haven't yet vectorized the induction variable, or if we will create
2284 // a scalar one, we need to define the scalar induction variable and step
2285 // values. If we were given a truncation type, truncate the canonical
2286 // induction variable and constant step. Otherwise, derive these values from
2287 // the induction descriptor.
2288 if (!VectorizedIV || NeedsScalarIV) {
2290 auto *TruncType = cast<IntegerType>(Trunc->getType());
2291 assert(Step && "Truncation requires constant integer step");
2292 auto StepInt = cast<ConstantInt>(Step)->getSExtValue();
2293 ScalarIV = Builder.CreateCast(Instruction::Trunc, Induction, TruncType);
2294 Step = ConstantInt::getSigned(TruncType, StepInt);
2296 ScalarIV = Induction;
2297 auto &DL = OrigLoop->getHeader()->getModule()->getDataLayout();
2298 if (IV != OldInduction) {
2299 ScalarIV = Builder.CreateSExtOrTrunc(ScalarIV, IV->getType());
2300 ScalarIV = ID.transform(Builder, ScalarIV, PSE.getSE(), DL);
2301 ScalarIV->setName("offset.idx");
2304 SCEVExpander Exp(*PSE.getSE(), DL, "induction");
2305 Step = Exp.expandCodeFor(ID.getStep(), ID.getStep()->getType(),
2306 &*Builder.GetInsertPoint());
2311 // If we haven't yet vectorized the induction variable, splat the scalar
2312 // induction variable, and build the necessary step vectors.
2313 if (!VectorizedIV) {
2314 Value *Broadcasted = getBroadcastInstrs(ScalarIV);
2315 VectorParts Entry(UF);
2316 for (unsigned Part = 0; Part < UF; ++Part)
2317 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
2318 VectorLoopValueMap.initVector(EntryVal, Entry);
2320 addMetadata(Entry, Trunc);
2323 // If an induction variable is only used for counting loop iterations or
2324 // calculating addresses, it doesn't need to be widened. Create scalar steps
2325 // that can be used by instructions we will later scalarize. Note that the
2326 // addition of the scalar steps will not increase the number of instructions
2327 // in the loop in the common case prior to InstCombine. We will be trading
2328 // one vector extract for each scalar step.
2330 buildScalarSteps(ScalarIV, Step, EntryVal);
2333 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx, Value *Step,
2334 Instruction::BinaryOps BinOp) {
2335 // Create and check the types.
2336 assert(Val->getType()->isVectorTy() && "Must be a vector");
2337 int VLen = Val->getType()->getVectorNumElements();
2339 Type *STy = Val->getType()->getScalarType();
2340 assert((STy->isIntegerTy() || STy->isFloatingPointTy()) &&
2341 "Induction Step must be an integer or FP");
2342 assert(Step->getType() == STy && "Step has wrong type");
2344 SmallVector<Constant *, 8> Indices;
2346 if (STy->isIntegerTy()) {
2347 // Create a vector of consecutive numbers from zero to VF.
2348 for (int i = 0; i < VLen; ++i)
2349 Indices.push_back(ConstantInt::get(STy, StartIdx + i));
2351 // Add the consecutive indices to the vector value.
2352 Constant *Cv = ConstantVector::get(Indices);
2353 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
2354 Step = Builder.CreateVectorSplat(VLen, Step);
2355 assert(Step->getType() == Val->getType() && "Invalid step vec");
2356 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
2357 // which can be found from the original scalar operations.
2358 Step = Builder.CreateMul(Cv, Step);
2359 return Builder.CreateAdd(Val, Step, "induction");
2362 // Floating point induction.
2363 assert((BinOp == Instruction::FAdd || BinOp == Instruction::FSub) &&
2364 "Binary Opcode should be specified for FP induction");
2365 // Create a vector of consecutive numbers from zero to VF.
2366 for (int i = 0; i < VLen; ++i)
2367 Indices.push_back(ConstantFP::get(STy, (double)(StartIdx + i)));
2369 // Add the consecutive indices to the vector value.
2370 Constant *Cv = ConstantVector::get(Indices);
2372 Step = Builder.CreateVectorSplat(VLen, Step);
2374 // Floating point operations had to be 'fast' to enable the induction.
2375 FastMathFlags Flags;
2376 Flags.setUnsafeAlgebra();
2378 Value *MulOp = Builder.CreateFMul(Cv, Step);
2379 if (isa<Instruction>(MulOp))
2380 // Have to check, MulOp may be a constant
2381 cast<Instruction>(MulOp)->setFastMathFlags(Flags);
2383 Value *BOp = Builder.CreateBinOp(BinOp, Val, MulOp, "induction");
2384 if (isa<Instruction>(BOp))
2385 cast<Instruction>(BOp)->setFastMathFlags(Flags);
2389 void InnerLoopVectorizer::buildScalarSteps(Value *ScalarIV, Value *Step,
2392 // We shouldn't have to build scalar steps if we aren't vectorizing.
2393 assert(VF > 1 && "VF should be greater than one");
2395 // Get the value type and ensure it and the step have the same integer type.
2396 Type *ScalarIVTy = ScalarIV->getType()->getScalarType();
2397 assert(ScalarIVTy->isIntegerTy() && ScalarIVTy == Step->getType() &&
2398 "Val and Step should have the same integer type");
2400 // Determine the number of scalars we need to generate for each unroll
2401 // iteration. If EntryVal is uniform, we only need to generate the first
2402 // lane. Otherwise, we generate all VF values.
2404 Legal->isUniformAfterVectorization(cast<Instruction>(EntryVal)) ? 1 : VF;
2406 // Compute the scalar steps and save the results in VectorLoopValueMap.
2407 ScalarParts Entry(UF);
2408 for (unsigned Part = 0; Part < UF; ++Part) {
2409 Entry[Part].resize(VF);
2410 for (unsigned Lane = 0; Lane < Lanes; ++Lane) {
2411 auto *StartIdx = ConstantInt::get(ScalarIVTy, VF * Part + Lane);
2412 auto *Mul = Builder.CreateMul(StartIdx, Step);
2413 auto *Add = Builder.CreateAdd(ScalarIV, Mul);
2414 Entry[Part][Lane] = Add;
2417 VectorLoopValueMap.initScalar(EntryVal, Entry);
2420 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
2422 const ValueToValueMap &Strides = getSymbolicStrides() ? *getSymbolicStrides() :
2425 int Stride = getPtrStride(PSE, Ptr, TheLoop, Strides, true, false);
2426 if (Stride == 1 || Stride == -1)
2431 bool LoopVectorizationLegality::isUniform(Value *V) {
2432 return LAI->isUniform(V);
2435 const InnerLoopVectorizer::VectorParts &
2436 InnerLoopVectorizer::getVectorValue(Value *V) {
2437 assert(V != Induction && "The new induction variable should not be used.");
2438 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
2439 assert(!V->getType()->isVoidTy() && "Type does not produce a value");
2441 // If we have a stride that is replaced by one, do it here.
2442 if (Legal->hasStride(V))
2443 V = ConstantInt::get(V->getType(), 1);
2445 // If we have this scalar in the map, return it.
2446 if (VectorLoopValueMap.hasVector(V))
2447 return VectorLoopValueMap.VectorMapStorage[V];
2449 // If the value has not been vectorized, check if it has been scalarized
2450 // instead. If it has been scalarized, and we actually need the value in
2451 // vector form, we will construct the vector values on demand.
2452 if (VectorLoopValueMap.hasScalar(V)) {
2454 // Initialize a new vector map entry.
2455 VectorParts Entry(UF);
2457 // If we've scalarized a value, that value should be an instruction.
2458 auto *I = cast<Instruction>(V);
2460 // If we aren't vectorizing, we can just copy the scalar map values over to
2463 for (unsigned Part = 0; Part < UF; ++Part)
2464 Entry[Part] = getScalarValue(V, Part, 0);
2465 return VectorLoopValueMap.initVector(V, Entry);
2468 // Get the last scalar instruction we generated for V. If the value is
2469 // known to be uniform after vectorization, this corresponds to lane zero
2470 // of the last unroll iteration. Otherwise, the last instruction is the one
2471 // we created for the last vector lane of the last unroll iteration.
2472 unsigned LastLane = Legal->isUniformAfterVectorization(I) ? 0 : VF - 1;
2473 auto *LastInst = cast<Instruction>(getScalarValue(V, UF - 1, LastLane));
2475 // Set the insert point after the last scalarized instruction. This ensures
2476 // the insertelement sequence will directly follow the scalar definitions.
2477 auto OldIP = Builder.saveIP();
2478 auto NewIP = std::next(BasicBlock::iterator(LastInst));
2479 Builder.SetInsertPoint(&*NewIP);
2481 // However, if we are vectorizing, we need to construct the vector values.
2482 // If the value is known to be uniform after vectorization, we can just
2483 // broadcast the scalar value corresponding to lane zero for each unroll
2484 // iteration. Otherwise, we construct the vector values using insertelement
2485 // instructions. Since the resulting vectors are stored in
2486 // VectorLoopValueMap, we will only generate the insertelements once.
2487 for (unsigned Part = 0; Part < UF; ++Part) {
2488 Value *VectorValue = nullptr;
2489 if (Legal->isUniformAfterVectorization(I)) {
2490 VectorValue = getBroadcastInstrs(getScalarValue(V, Part, 0));
2492 VectorValue = UndefValue::get(VectorType::get(V->getType(), VF));
2493 for (unsigned Lane = 0; Lane < VF; ++Lane)
2494 VectorValue = Builder.CreateInsertElement(
2495 VectorValue, getScalarValue(V, Part, Lane),
2496 Builder.getInt32(Lane));
2498 Entry[Part] = VectorValue;
2500 Builder.restoreIP(OldIP);
2501 return VectorLoopValueMap.initVector(V, Entry);
2504 // If this scalar is unknown, assume that it is a constant or that it is
2505 // loop invariant. Broadcast V and save the value for future uses.
2506 Value *B = getBroadcastInstrs(V);
2507 return VectorLoopValueMap.initVector(V, VectorParts(UF, B));
2510 Value *InnerLoopVectorizer::getScalarValue(Value *V, unsigned Part,
2513 // If the value is not an instruction contained in the loop, it should
2514 // already be scalar.
2515 if (OrigLoop->isLoopInvariant(V))
2518 assert(Lane > 0 ? !Legal->isUniformAfterVectorization(cast<Instruction>(V))
2519 : true && "Uniform values only have lane zero");
2521 // If the value from the original loop has not been vectorized, it is
2522 // represented by UF x VF scalar values in the new loop. Return the requested
2524 if (VectorLoopValueMap.hasScalar(V))
2525 return VectorLoopValueMap.ScalarMapStorage[V][Part][Lane];
2527 // If the value has not been scalarized, get its entry in VectorLoopValueMap
2528 // for the given unroll part. If this entry is not a vector type (i.e., the
2529 // vectorization factor is one), there is no need to generate an
2530 // extractelement instruction.
2531 auto *U = getVectorValue(V)[Part];
2532 if (!U->getType()->isVectorTy()) {
2533 assert(VF == 1 && "Value not scalarized has non-vector type");
2537 // Otherwise, the value from the original loop has been vectorized and is
2538 // represented by UF vector values. Extract and return the requested scalar
2539 // value from the appropriate vector lane.
2540 return Builder.CreateExtractElement(U, Builder.getInt32(Lane));
2543 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
2544 assert(Vec->getType()->isVectorTy() && "Invalid type");
2545 SmallVector<Constant *, 8> ShuffleMask;
2546 for (unsigned i = 0; i < VF; ++i)
2547 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
2549 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
2550 ConstantVector::get(ShuffleMask),
2554 // Get a mask to interleave \p NumVec vectors into a wide vector.
2555 // I.e. <0, VF, VF*2, ..., VF*(NumVec-1), 1, VF+1, VF*2+1, ...>
2556 // E.g. For 2 interleaved vectors, if VF is 4, the mask is:
2557 // <0, 4, 1, 5, 2, 6, 3, 7>
2558 static Constant *getInterleavedMask(IRBuilder<> &Builder, unsigned VF,
2560 SmallVector<Constant *, 16> Mask;
2561 for (unsigned i = 0; i < VF; i++)
2562 for (unsigned j = 0; j < NumVec; j++)
2563 Mask.push_back(Builder.getInt32(j * VF + i));
2565 return ConstantVector::get(Mask);
2568 // Get the strided mask starting from index \p Start.
2569 // I.e. <Start, Start + Stride, ..., Start + Stride*(VF-1)>
2570 static Constant *getStridedMask(IRBuilder<> &Builder, unsigned Start,
2571 unsigned Stride, unsigned VF) {
2572 SmallVector<Constant *, 16> Mask;
2573 for (unsigned i = 0; i < VF; i++)
2574 Mask.push_back(Builder.getInt32(Start + i * Stride));
2576 return ConstantVector::get(Mask);
2579 // Get a mask of two parts: The first part consists of sequential integers
2580 // starting from 0, The second part consists of UNDEFs.
2581 // I.e. <0, 1, 2, ..., NumInt - 1, undef, ..., undef>
2582 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned NumInt,
2583 unsigned NumUndef) {
2584 SmallVector<Constant *, 16> Mask;
2585 for (unsigned i = 0; i < NumInt; i++)
2586 Mask.push_back(Builder.getInt32(i));
2588 Constant *Undef = UndefValue::get(Builder.getInt32Ty());
2589 for (unsigned i = 0; i < NumUndef; i++)
2590 Mask.push_back(Undef);
2592 return ConstantVector::get(Mask);
2595 // Concatenate two vectors with the same element type. The 2nd vector should
2596 // not have more elements than the 1st vector. If the 2nd vector has less
2597 // elements, extend it with UNDEFs.
2598 static Value *ConcatenateTwoVectors(IRBuilder<> &Builder, Value *V1,
2600 VectorType *VecTy1 = dyn_cast<VectorType>(V1->getType());
2601 VectorType *VecTy2 = dyn_cast<VectorType>(V2->getType());
2602 assert(VecTy1 && VecTy2 &&
2603 VecTy1->getScalarType() == VecTy2->getScalarType() &&
2604 "Expect two vectors with the same element type");
2606 unsigned NumElts1 = VecTy1->getNumElements();
2607 unsigned NumElts2 = VecTy2->getNumElements();
2608 assert(NumElts1 >= NumElts2 && "Unexpect the first vector has less elements");
2610 if (NumElts1 > NumElts2) {
2611 // Extend with UNDEFs.
2613 getSequentialMask(Builder, NumElts2, NumElts1 - NumElts2);
2614 V2 = Builder.CreateShuffleVector(V2, UndefValue::get(VecTy2), ExtMask);
2617 Constant *Mask = getSequentialMask(Builder, NumElts1 + NumElts2, 0);
2618 return Builder.CreateShuffleVector(V1, V2, Mask);
2621 // Concatenate vectors in the given list. All vectors have the same type.
2622 static Value *ConcatenateVectors(IRBuilder<> &Builder,
2623 ArrayRef<Value *> InputList) {
2624 unsigned NumVec = InputList.size();
2625 assert(NumVec > 1 && "Should be at least two vectors");
2627 SmallVector<Value *, 8> ResList;
2628 ResList.append(InputList.begin(), InputList.end());
2630 SmallVector<Value *, 8> TmpList;
2631 for (unsigned i = 0; i < NumVec - 1; i += 2) {
2632 Value *V0 = ResList[i], *V1 = ResList[i + 1];
2633 assert((V0->getType() == V1->getType() || i == NumVec - 2) &&
2634 "Only the last vector may have a different type");
2636 TmpList.push_back(ConcatenateTwoVectors(Builder, V0, V1));
2639 // Push the last vector if the total number of vectors is odd.
2640 if (NumVec % 2 != 0)
2641 TmpList.push_back(ResList[NumVec - 1]);
2644 NumVec = ResList.size();
2645 } while (NumVec > 1);
2650 // Try to vectorize the interleave group that \p Instr belongs to.
2652 // E.g. Translate following interleaved load group (factor = 3):
2653 // for (i = 0; i < N; i+=3) {
2654 // R = Pic[i]; // Member of index 0
2655 // G = Pic[i+1]; // Member of index 1
2656 // B = Pic[i+2]; // Member of index 2
2657 // ... // do something to R, G, B
2660 // %wide.vec = load <12 x i32> ; Read 4 tuples of R,G,B
2661 // %R.vec = shuffle %wide.vec, undef, <0, 3, 6, 9> ; R elements
2662 // %G.vec = shuffle %wide.vec, undef, <1, 4, 7, 10> ; G elements
2663 // %B.vec = shuffle %wide.vec, undef, <2, 5, 8, 11> ; B elements
2665 // Or translate following interleaved store group (factor = 3):
2666 // for (i = 0; i < N; i+=3) {
2667 // ... do something to R, G, B
2668 // Pic[i] = R; // Member of index 0
2669 // Pic[i+1] = G; // Member of index 1
2670 // Pic[i+2] = B; // Member of index 2
2673 // %R_G.vec = shuffle %R.vec, %G.vec, <0, 1, 2, ..., 7>
2674 // %B_U.vec = shuffle %B.vec, undef, <0, 1, 2, 3, u, u, u, u>
2675 // %interleaved.vec = shuffle %R_G.vec, %B_U.vec,
2676 // <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> ; Interleave R,G,B elements
2677 // store <12 x i32> %interleaved.vec ; Write 4 tuples of R,G,B
2678 void InnerLoopVectorizer::vectorizeInterleaveGroup(Instruction *Instr) {
2679 const InterleaveGroup *Group = Legal->getInterleavedAccessGroup(Instr);
2680 assert(Group && "Fail to get an interleaved access group.");
2682 // Skip if current instruction is not the insert position.
2683 if (Instr != Group->getInsertPos())
2686 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2687 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2688 Value *Ptr = getPointerOperand(Instr);
2690 // Prepare for the vector type of the interleaved load/store.
2691 Type *ScalarTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2692 unsigned InterleaveFactor = Group->getFactor();
2693 Type *VecTy = VectorType::get(ScalarTy, InterleaveFactor * VF);
2694 Type *PtrTy = VecTy->getPointerTo(Ptr->getType()->getPointerAddressSpace());
2696 // Prepare for the new pointers.
2697 setDebugLocFromInst(Builder, Ptr);
2698 SmallVector<Value *, 2> NewPtrs;
2699 unsigned Index = Group->getIndex(Instr);
2701 // If the group is reverse, adjust the index to refer to the last vector lane
2702 // instead of the first. We adjust the index from the first vector lane,
2703 // rather than directly getting the pointer for lane VF - 1, because the
2704 // pointer operand of the interleaved access is supposed to be uniform. For
2705 // uniform instructions, we're only required to generate a value for the
2706 // first vector lane in each unroll iteration.
2707 if (Group->isReverse())
2708 Index += (VF - 1) * Group->getFactor();
2710 for (unsigned Part = 0; Part < UF; Part++) {
2711 Value *NewPtr = getScalarValue(Ptr, Part, 0);
2713 // Notice current instruction could be any index. Need to adjust the address
2714 // to the member of index 0.
2716 // E.g. a = A[i+1]; // Member of index 1 (Current instruction)
2717 // b = A[i]; // Member of index 0
2718 // Current pointer is pointed to A[i+1], adjust it to A[i].
2720 // E.g. A[i+1] = a; // Member of index 1
2721 // A[i] = b; // Member of index 0
2722 // A[i+2] = c; // Member of index 2 (Current instruction)
2723 // Current pointer is pointed to A[i+2], adjust it to A[i].
2724 NewPtr = Builder.CreateGEP(NewPtr, Builder.getInt32(-Index));
2726 // Cast to the vector pointer type.
2727 NewPtrs.push_back(Builder.CreateBitCast(NewPtr, PtrTy));
2730 setDebugLocFromInst(Builder, Instr);
2731 Value *UndefVec = UndefValue::get(VecTy);
2733 // Vectorize the interleaved load group.
2736 // For each unroll part, create a wide load for the group.
2737 SmallVector<Value *, 2> NewLoads;
2738 for (unsigned Part = 0; Part < UF; Part++) {
2739 auto *NewLoad = Builder.CreateAlignedLoad(
2740 NewPtrs[Part], Group->getAlignment(), "wide.vec");
2741 addMetadata(NewLoad, Instr);
2742 NewLoads.push_back(NewLoad);
2745 // For each member in the group, shuffle out the appropriate data from the
2747 for (unsigned I = 0; I < InterleaveFactor; ++I) {
2748 Instruction *Member = Group->getMember(I);
2750 // Skip the gaps in the group.
2754 VectorParts Entry(UF);
2755 Constant *StrideMask = getStridedMask(Builder, I, InterleaveFactor, VF);
2756 for (unsigned Part = 0; Part < UF; Part++) {
2757 Value *StridedVec = Builder.CreateShuffleVector(
2758 NewLoads[Part], UndefVec, StrideMask, "strided.vec");
2760 // If this member has different type, cast the result type.
2761 if (Member->getType() != ScalarTy) {
2762 VectorType *OtherVTy = VectorType::get(Member->getType(), VF);
2763 StridedVec = Builder.CreateBitOrPointerCast(StridedVec, OtherVTy);
2767 Group->isReverse() ? reverseVector(StridedVec) : StridedVec;
2769 VectorLoopValueMap.initVector(Member, Entry);
2774 // The sub vector type for current instruction.
2775 VectorType *SubVT = VectorType::get(ScalarTy, VF);
2777 // Vectorize the interleaved store group.
2778 for (unsigned Part = 0; Part < UF; Part++) {
2779 // Collect the stored vector from each member.
2780 SmallVector<Value *, 4> StoredVecs;
2781 for (unsigned i = 0; i < InterleaveFactor; i++) {
2782 // Interleaved store group doesn't allow a gap, so each index has a member
2783 Instruction *Member = Group->getMember(i);
2784 assert(Member && "Fail to get a member from an interleaved store group");
2787 getVectorValue(cast<StoreInst>(Member)->getValueOperand())[Part];
2788 if (Group->isReverse())
2789 StoredVec = reverseVector(StoredVec);
2791 // If this member has different type, cast it to an unified type.
2792 if (StoredVec->getType() != SubVT)
2793 StoredVec = Builder.CreateBitOrPointerCast(StoredVec, SubVT);
2795 StoredVecs.push_back(StoredVec);
2798 // Concatenate all vectors into a wide vector.
2799 Value *WideVec = ConcatenateVectors(Builder, StoredVecs);
2801 // Interleave the elements in the wide vector.
2802 Constant *IMask = getInterleavedMask(Builder, VF, InterleaveFactor);
2803 Value *IVec = Builder.CreateShuffleVector(WideVec, UndefVec, IMask,
2806 Instruction *NewStoreInstr =
2807 Builder.CreateAlignedStore(IVec, NewPtrs[Part], Group->getAlignment());
2808 addMetadata(NewStoreInstr, Instr);
2812 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
2813 // Attempt to issue a wide load.
2814 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2815 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2817 assert((LI || SI) && "Invalid Load/Store instruction");
2819 // Try to vectorize the interleave group if this access is interleaved.
2820 if (Legal->isAccessInterleaved(Instr))
2821 return vectorizeInterleaveGroup(Instr);
2823 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2824 Type *DataTy = VectorType::get(ScalarDataTy, VF);
2825 Value *Ptr = getPointerOperand(Instr);
2826 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
2827 // An alignment of 0 means target abi alignment. We need to use the scalar's
2828 // target abi alignment in such a case.
2829 const DataLayout &DL = Instr->getModule()->getDataLayout();
2831 Alignment = DL.getABITypeAlignment(ScalarDataTy);
2832 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
2834 // Scalarize the memory instruction if necessary.
2835 if (Legal->memoryInstructionMustBeScalarized(Instr, VF))
2836 return scalarizeInstruction(Instr, Legal->isScalarWithPredication(Instr));
2838 // Determine if the pointer operand of the access is either consecutive or
2839 // reverse consecutive.
2840 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
2841 bool Reverse = ConsecutiveStride < 0;
2843 // Determine if either a gather or scatter operation is legal.
2844 bool CreateGatherScatter =
2845 !ConsecutiveStride && Legal->isLegalGatherOrScatter(Instr);
2847 VectorParts VectorGep;
2849 // Handle consecutive loads/stores.
2850 GetElementPtrInst *Gep = getGEPInstruction(Ptr);
2851 if (ConsecutiveStride) {
2853 unsigned NumOperands = Gep->getNumOperands();
2855 // The original GEP that identified as a consecutive memory access
2856 // should have only one loop-variant operand.
2857 unsigned NumOfLoopVariantOps = 0;
2858 for (unsigned i = 0; i < NumOperands; ++i)
2859 if (!PSE.getSE()->isLoopInvariant(PSE.getSCEV(Gep->getOperand(i)),
2861 NumOfLoopVariantOps++;
2862 assert(NumOfLoopVariantOps == 1 &&
2863 "Consecutive GEP should have only one loop-variant operand");
2865 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2866 Gep2->setName("gep.indvar");
2868 // A new GEP is created for a 0-lane value of the first unroll iteration.
2869 // The GEPs for the rest of the unroll iterations are computed below as an
2870 // offset from this GEP.
2871 for (unsigned i = 0; i < NumOperands; ++i)
2872 // We can apply getScalarValue() for all GEP indices. It returns an
2873 // original value for loop-invariant operand and 0-lane for consecutive
2875 Gep2->setOperand(i, getScalarValue(Gep->getOperand(i),
2876 0, /* First unroll iteration */
2877 0 /* 0-lane of the vector */ ));
2878 setDebugLocFromInst(Builder, Gep);
2879 Ptr = Builder.Insert(Gep2);
2882 setDebugLocFromInst(Builder, Ptr);
2883 Ptr = getScalarValue(Ptr, 0, 0);
2886 // At this point we should vector version of GEP for Gather or Scatter
2887 assert(CreateGatherScatter && "The instruction should be scalarized");
2889 // Vectorizing GEP, across UF parts. We want to get a vector value for base
2890 // and each index that's defined inside the loop, even if it is
2891 // loop-invariant but wasn't hoisted out. Otherwise we want to keep them
2893 SmallVector<VectorParts, 4> OpsV;
2894 for (Value *Op : Gep->operands()) {
2895 Instruction *SrcInst = dyn_cast<Instruction>(Op);
2896 if (SrcInst && OrigLoop->contains(SrcInst))
2897 OpsV.push_back(getVectorValue(Op));
2899 OpsV.push_back(VectorParts(UF, Op));
2901 for (unsigned Part = 0; Part < UF; ++Part) {
2902 SmallVector<Value *, 4> Ops;
2903 Value *GEPBasePtr = OpsV[0][Part];
2904 for (unsigned i = 1; i < Gep->getNumOperands(); i++)
2905 Ops.push_back(OpsV[i][Part]);
2906 Value *NewGep = Builder.CreateGEP(GEPBasePtr, Ops, "VectorGep");
2907 cast<GetElementPtrInst>(NewGep)->setIsInBounds(Gep->isInBounds());
2908 assert(NewGep->getType()->isVectorTy() && "Expected vector GEP");
2911 Builder.CreateBitCast(NewGep, VectorType::get(Ptr->getType(), VF));
2912 VectorGep.push_back(NewGep);
2915 VectorGep = getVectorValue(Ptr);
2918 VectorParts Mask = createBlockInMask(Instr->getParent());
2921 assert(!Legal->isUniform(SI->getPointerOperand()) &&
2922 "We do not allow storing to uniform addresses");
2923 setDebugLocFromInst(Builder, SI);
2924 // We don't want to update the value in the map as it might be used in
2925 // another expression. So don't use a reference type for "StoredVal".
2926 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
2928 for (unsigned Part = 0; Part < UF; ++Part) {
2929 Instruction *NewSI = nullptr;
2930 if (CreateGatherScatter) {
2931 Value *MaskPart = Legal->isMaskRequired(SI) ? Mask[Part] : nullptr;
2932 NewSI = Builder.CreateMaskedScatter(StoredVal[Part], VectorGep[Part],
2933 Alignment, MaskPart);
2935 // Calculate the pointer for the specific unroll-part.
2937 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2940 // If we store to reverse consecutive memory locations, then we need
2941 // to reverse the order of elements in the stored value.
2942 StoredVal[Part] = reverseVector(StoredVal[Part]);
2943 // If the address is consecutive but reversed, then the
2944 // wide store needs to start at the last vector element.
2946 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2948 Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2949 Mask[Part] = reverseVector(Mask[Part]);
2953 Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace));
2955 if (Legal->isMaskRequired(SI))
2956 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
2960 Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
2962 addMetadata(NewSI, SI);
2968 assert(LI && "Must have a load instruction");
2969 setDebugLocFromInst(Builder, LI);
2970 VectorParts Entry(UF);
2971 for (unsigned Part = 0; Part < UF; ++Part) {
2973 if (CreateGatherScatter) {
2974 Value *MaskPart = Legal->isMaskRequired(LI) ? Mask[Part] : nullptr;
2975 NewLI = Builder.CreateMaskedGather(VectorGep[Part], Alignment, MaskPart,
2976 0, "wide.masked.gather");
2977 Entry[Part] = NewLI;
2979 // Calculate the pointer for the specific unroll-part.
2981 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2984 // If the address is consecutive but reversed, then the
2985 // wide load needs to start at the last vector element.
2986 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2987 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2988 Mask[Part] = reverseVector(Mask[Part]);
2992 Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace));
2993 if (Legal->isMaskRequired(LI))
2994 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
2995 UndefValue::get(DataTy),
2996 "wide.masked.load");
2998 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
2999 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
3001 addMetadata(NewLI, LI);
3003 VectorLoopValueMap.initVector(Instr, Entry);
3006 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr,
3007 bool IfPredicateInstr) {
3008 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
3009 DEBUG(dbgs() << "LV: Scalarizing"
3010 << (IfPredicateInstr ? " and predicating:" : ":") << *Instr
3012 // Holds vector parameters or scalars, in case of uniform vals.
3013 SmallVector<VectorParts, 4> Params;
3015 setDebugLocFromInst(Builder, Instr);
3017 // Does this instruction return a value ?
3018 bool IsVoidRetTy = Instr->getType()->isVoidTy();
3020 // Initialize a new scalar map entry.
3021 ScalarParts Entry(UF);
3024 if (IfPredicateInstr)
3025 Cond = createBlockInMask(Instr->getParent());
3027 // Determine the number of scalars we need to generate for each unroll
3028 // iteration. If the instruction is uniform, we only need to generate the
3029 // first lane. Otherwise, we generate all VF values.
3030 unsigned Lanes = Legal->isUniformAfterVectorization(Instr) ? 1 : VF;
3032 // For each vector unroll 'part':
3033 for (unsigned Part = 0; Part < UF; ++Part) {
3034 Entry[Part].resize(VF);
3035 // For each scalar that we create:
3036 for (unsigned Lane = 0; Lane < Lanes; ++Lane) {
3039 Value *Cmp = nullptr;
3040 if (IfPredicateInstr) {
3041 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Lane));
3042 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp,
3043 ConstantInt::get(Cmp->getType(), 1));
3046 Instruction *Cloned = Instr->clone();
3048 Cloned->setName(Instr->getName() + ".cloned");
3050 // Replace the operands of the cloned instructions with their scalar
3051 // equivalents in the new loop.
3052 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
3053 auto *NewOp = getScalarValue(Instr->getOperand(op), Part, Lane);
3054 Cloned->setOperand(op, NewOp);
3056 addNewMetadata(Cloned, Instr);
3058 // Place the cloned scalar in the new loop.
3059 Builder.Insert(Cloned);
3061 // Add the cloned scalar to the scalar map entry.
3062 Entry[Part][Lane] = Cloned;
3064 // If we just cloned a new assumption, add it the assumption cache.
3065 if (auto *II = dyn_cast<IntrinsicInst>(Cloned))
3066 if (II->getIntrinsicID() == Intrinsic::assume)
3067 AC->registerAssumption(II);
3070 if (IfPredicateInstr)
3071 PredicatedInstructions.push_back(std::make_pair(Cloned, Cmp));
3074 VectorLoopValueMap.initScalar(Instr, Entry);
3077 PHINode *InnerLoopVectorizer::createInductionVariable(Loop *L, Value *Start,
3078 Value *End, Value *Step,
3080 BasicBlock *Header = L->getHeader();
3081 BasicBlock *Latch = L->getLoopLatch();
3082 // As we're just creating this loop, it's possible no latch exists
3083 // yet. If so, use the header as this will be a single block loop.
3087 IRBuilder<> Builder(&*Header->getFirstInsertionPt());
3088 Instruction *OldInst = getDebugLocFromInstOrOperands(OldInduction);
3089 setDebugLocFromInst(Builder, OldInst);
3090 auto *Induction = Builder.CreatePHI(Start->getType(), 2, "index");
3092 Builder.SetInsertPoint(Latch->getTerminator());
3093 setDebugLocFromInst(Builder, OldInst);
3095 // Create i+1 and fill the PHINode.
3096 Value *Next = Builder.CreateAdd(Induction, Step, "index.next");
3097 Induction->addIncoming(Start, L->getLoopPreheader());
3098 Induction->addIncoming(Next, Latch);
3099 // Create the compare.
3100 Value *ICmp = Builder.CreateICmpEQ(Next, End);
3101 Builder.CreateCondBr(ICmp, L->getExitBlock(), Header);
3103 // Now we have two terminators. Remove the old one from the block.
3104 Latch->getTerminator()->eraseFromParent();
3109 Value *InnerLoopVectorizer::getOrCreateTripCount(Loop *L) {
3113 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
3114 // Find the loop boundaries.
3115 ScalarEvolution *SE = PSE.getSE();
3116 const SCEV *BackedgeTakenCount = PSE.getBackedgeTakenCount();
3117 assert(BackedgeTakenCount != SE->getCouldNotCompute() &&
3118 "Invalid loop count");
3120 Type *IdxTy = Legal->getWidestInductionType();
3122 // The exit count might have the type of i64 while the phi is i32. This can
3123 // happen if we have an induction variable that is sign extended before the
3124 // compare. The only way that we get a backedge taken count is that the
3125 // induction variable was signed and as such will not overflow. In such a case
3126 // truncation is legal.
3127 if (BackedgeTakenCount->getType()->getPrimitiveSizeInBits() >
3128 IdxTy->getPrimitiveSizeInBits())
3129 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount, IdxTy);
3130 BackedgeTakenCount = SE->getNoopOrZeroExtend(BackedgeTakenCount, IdxTy);
3132 // Get the total trip count from the count by adding 1.
3133 const SCEV *ExitCount = SE->getAddExpr(
3134 BackedgeTakenCount, SE->getOne(BackedgeTakenCount->getType()));
3136 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
3138 // Expand the trip count and place the new instructions in the preheader.
3139 // Notice that the pre-header does not change, only the loop body.
3140 SCEVExpander Exp(*SE, DL, "induction");
3142 // Count holds the overall loop count (N).
3143 TripCount = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
3144 L->getLoopPreheader()->getTerminator());
3146 if (TripCount->getType()->isPointerTy())
3148 CastInst::CreatePointerCast(TripCount, IdxTy, "exitcount.ptrcnt.to.int",
3149 L->getLoopPreheader()->getTerminator());
3154 Value *InnerLoopVectorizer::getOrCreateVectorTripCount(Loop *L) {
3155 if (VectorTripCount)
3156 return VectorTripCount;
3158 Value *TC = getOrCreateTripCount(L);
3159 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
3161 // Now we need to generate the expression for the part of the loop that the
3162 // vectorized body will execute. This is equal to N - (N % Step) if scalar
3163 // iterations are not required for correctness, or N - Step, otherwise. Step
3164 // is equal to the vectorization factor (number of SIMD elements) times the
3165 // unroll factor (number of SIMD instructions).
3166 Constant *Step = ConstantInt::get(TC->getType(), VF * UF);
3167 Value *R = Builder.CreateURem(TC, Step, "n.mod.vf");
3169 // If there is a non-reversed interleaved group that may speculatively access
3170 // memory out-of-bounds, we need to ensure that there will be at least one
3171 // iteration of the scalar epilogue loop. Thus, if the step evenly divides
3172 // the trip count, we set the remainder to be equal to the step. If the step
3173 // does not evenly divide the trip count, no adjustment is necessary since
3174 // there will already be scalar iterations. Note that the minimum iterations
3175 // check ensures that N >= Step.
3176 if (VF > 1 && Legal->requiresScalarEpilogue()) {
3177 auto *IsZero = Builder.CreateICmpEQ(R, ConstantInt::get(R->getType(), 0));
3178 R = Builder.CreateSelect(IsZero, Step, R);
3181 VectorTripCount = Builder.CreateSub(TC, R, "n.vec");
3183 return VectorTripCount;
3186 void InnerLoopVectorizer::emitMinimumIterationCountCheck(Loop *L,
3187 BasicBlock *Bypass) {
3188 Value *Count = getOrCreateTripCount(L);
3189 BasicBlock *BB = L->getLoopPreheader();
3190 IRBuilder<> Builder(BB->getTerminator());
3192 // Generate code to check that the loop's trip count that we computed by
3193 // adding one to the backedge-taken count will not overflow.
3194 Value *CheckMinIters = Builder.CreateICmpULT(
3195 Count, ConstantInt::get(Count->getType(), VF * UF), "min.iters.check");
3198 BB->splitBasicBlock(BB->getTerminator(), "min.iters.checked");
3199 // Update dominator tree immediately if the generated block is a
3200 // LoopBypassBlock because SCEV expansions to generate loop bypass
3201 // checks may query it before the current function is finished.
3202 DT->addNewBlock(NewBB, BB);
3203 if (L->getParentLoop())
3204 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
3205 ReplaceInstWithInst(BB->getTerminator(),
3206 BranchInst::Create(Bypass, NewBB, CheckMinIters));
3207 LoopBypassBlocks.push_back(BB);
3210 void InnerLoopVectorizer::emitVectorLoopEnteredCheck(Loop *L,
3211 BasicBlock *Bypass) {
3212 Value *TC = getOrCreateVectorTripCount(L);
3213 BasicBlock *BB = L->getLoopPreheader();
3214 IRBuilder<> Builder(BB->getTerminator());
3216 // Now, compare the new count to zero. If it is zero skip the vector loop and
3217 // jump to the scalar loop.
3218 Value *Cmp = Builder.CreateICmpEQ(TC, Constant::getNullValue(TC->getType()),
3221 // Generate code to check that the loop's trip count that we computed by
3222 // adding one to the backedge-taken count will not overflow.
3223 BasicBlock *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
3224 // Update dominator tree immediately if the generated block is a
3225 // LoopBypassBlock because SCEV expansions to generate loop bypass
3226 // checks may query it before the current function is finished.
3227 DT->addNewBlock(NewBB, BB);
3228 if (L->getParentLoop())
3229 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
3230 ReplaceInstWithInst(BB->getTerminator(),
3231 BranchInst::Create(Bypass, NewBB, Cmp));
3232 LoopBypassBlocks.push_back(BB);
3235 void InnerLoopVectorizer::emitSCEVChecks(Loop *L, BasicBlock *Bypass) {
3236 BasicBlock *BB = L->getLoopPreheader();
3238 // Generate the code to check that the SCEV assumptions that we made.
3239 // We want the new basic block to start at the first instruction in a
3240 // sequence of instructions that form a check.
3241 SCEVExpander Exp(*PSE.getSE(), Bypass->getModule()->getDataLayout(),
3244 Exp.expandCodeForPredicate(&PSE.getUnionPredicate(), BB->getTerminator());
3246 if (auto *C = dyn_cast<ConstantInt>(SCEVCheck))
3250 // Create a new block containing the stride check.
3251 BB->setName("vector.scevcheck");
3252 auto *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
3253 // Update dominator tree immediately if the generated block is a
3254 // LoopBypassBlock because SCEV expansions to generate loop bypass
3255 // checks may query it before the current function is finished.
3256 DT->addNewBlock(NewBB, BB);
3257 if (L->getParentLoop())
3258 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
3259 ReplaceInstWithInst(BB->getTerminator(),
3260 BranchInst::Create(Bypass, NewBB, SCEVCheck));
3261 LoopBypassBlocks.push_back(BB);
3262 AddedSafetyChecks = true;
3265 void InnerLoopVectorizer::emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass) {
3266 BasicBlock *BB = L->getLoopPreheader();
3268 // Generate the code that checks in runtime if arrays overlap. We put the
3269 // checks into a separate block to make the more common case of few elements
3271 Instruction *FirstCheckInst;
3272 Instruction *MemRuntimeCheck;
3273 std::tie(FirstCheckInst, MemRuntimeCheck) =
3274 Legal->getLAI()->addRuntimeChecks(BB->getTerminator());
3275 if (!MemRuntimeCheck)
3278 // Create a new block containing the memory check.
3279 BB->setName("vector.memcheck");
3280 auto *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
3281 // Update dominator tree immediately if the generated block is a
3282 // LoopBypassBlock because SCEV expansions to generate loop bypass
3283 // checks may query it before the current function is finished.
3284 DT->addNewBlock(NewBB, BB);
3285 if (L->getParentLoop())
3286 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
3287 ReplaceInstWithInst(BB->getTerminator(),
3288 BranchInst::Create(Bypass, NewBB, MemRuntimeCheck));
3289 LoopBypassBlocks.push_back(BB);
3290 AddedSafetyChecks = true;
3292 // We currently don't use LoopVersioning for the actual loop cloning but we
3293 // still use it to add the noalias metadata.
3294 LVer = llvm::make_unique<LoopVersioning>(*Legal->getLAI(), OrigLoop, LI, DT,
3296 LVer->prepareNoAliasMetadata();
3299 void InnerLoopVectorizer::createEmptyLoop() {
3301 In this function we generate a new loop. The new loop will contain
3302 the vectorized instructions while the old loop will continue to run the
3305 [ ] <-- loop iteration number check.
3308 | [ ] <-- vector loop bypass (may consist of multiple blocks).
3311 || [ ] <-- vector pre header.
3315 | [ ]_| <-- vector loop.
3318 | -[ ] <--- middle-block.
3321 -|- >[ ] <--- new preheader.
3325 | [ ]_| <-- old scalar loop to handle remainder.
3328 >[ ] <-- exit block.
3332 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
3333 BasicBlock *VectorPH = OrigLoop->getLoopPreheader();
3334 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
3335 assert(VectorPH && "Invalid loop structure");
3336 assert(ExitBlock && "Must have an exit block");
3338 // Some loops have a single integer induction variable, while other loops
3339 // don't. One example is c++ iterators that often have multiple pointer
3340 // induction variables. In the code below we also support a case where we
3341 // don't have a single induction variable.
3343 // We try to obtain an induction variable from the original loop as hard
3344 // as possible. However if we don't find one that:
3346 // - counts from zero, stepping by one
3347 // - is the size of the widest induction variable type
3348 // then we create a new one.
3349 OldInduction = Legal->getInduction();
3350 Type *IdxTy = Legal->getWidestInductionType();
3352 // Split the single block loop into the two loop structure described above.
3353 BasicBlock *VecBody =
3354 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
3355 BasicBlock *MiddleBlock =
3356 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
3357 BasicBlock *ScalarPH =
3358 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
3360 // Create and register the new vector loop.
3361 Loop *Lp = new Loop();
3362 Loop *ParentLoop = OrigLoop->getParentLoop();
3364 // Insert the new loop into the loop nest and register the new basic blocks
3365 // before calling any utilities such as SCEV that require valid LoopInfo.
3367 ParentLoop->addChildLoop(Lp);
3368 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
3369 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
3371 LI->addTopLevelLoop(Lp);
3373 Lp->addBasicBlockToLoop(VecBody, *LI);
3375 // Find the loop boundaries.
3376 Value *Count = getOrCreateTripCount(Lp);
3378 Value *StartIdx = ConstantInt::get(IdxTy, 0);
3380 // We need to test whether the backedge-taken count is uint##_max. Adding one
3381 // to it will cause overflow and an incorrect loop trip count in the vector
3382 // body. In case of overflow we want to directly jump to the scalar remainder
3384 emitMinimumIterationCountCheck(Lp, ScalarPH);
3385 // Now, compare the new count to zero. If it is zero skip the vector loop and
3386 // jump to the scalar loop.
3387 emitVectorLoopEnteredCheck(Lp, ScalarPH);
3388 // Generate the code to check any assumptions that we've made for SCEV
3390 emitSCEVChecks(Lp, ScalarPH);
3392 // Generate the code that checks in runtime if arrays overlap. We put the
3393 // checks into a separate block to make the more common case of few elements
3395 emitMemRuntimeChecks(Lp, ScalarPH);
3397 // Generate the induction variable.
3398 // The loop step is equal to the vectorization factor (num of SIMD elements)
3399 // times the unroll factor (num of SIMD instructions).
3400 Value *CountRoundDown = getOrCreateVectorTripCount(Lp);
3401 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
3403 createInductionVariable(Lp, StartIdx, CountRoundDown, Step,
3404 getDebugLocFromInstOrOperands(OldInduction));
3406 // We are going to resume the execution of the scalar loop.
3407 // Go over all of the induction variables that we found and fix the
3408 // PHIs that are left in the scalar version of the loop.
3409 // The starting values of PHI nodes depend on the counter of the last
3410 // iteration in the vectorized loop.
3411 // If we come from a bypass edge then we need to start from the original
3414 // This variable saves the new starting index for the scalar loop. It is used
3415 // to test if there are any tail iterations left once the vector loop has
3417 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
3418 for (auto &InductionEntry : *List) {
3419 PHINode *OrigPhi = InductionEntry.first;
3420 InductionDescriptor II = InductionEntry.second;
3422 // Create phi nodes to merge from the backedge-taken check block.
3423 PHINode *BCResumeVal = PHINode::Create(
3424 OrigPhi->getType(), 3, "bc.resume.val", ScalarPH->getTerminator());
3425 Value *&EndValue = IVEndValues[OrigPhi];
3426 if (OrigPhi == OldInduction) {
3427 // We know what the end value is.
3428 EndValue = CountRoundDown;
3430 IRBuilder<> B(LoopBypassBlocks.back()->getTerminator());
3431 Type *StepType = II.getStep()->getType();
3432 Instruction::CastOps CastOp =
3433 CastInst::getCastOpcode(CountRoundDown, true, StepType, true);
3434 Value *CRD = B.CreateCast(CastOp, CountRoundDown, StepType, "cast.crd");
3435 const DataLayout &DL = OrigLoop->getHeader()->getModule()->getDataLayout();
3436 EndValue = II.transform(B, CRD, PSE.getSE(), DL);
3437 EndValue->setName("ind.end");
3440 // The new PHI merges the original incoming value, in case of a bypass,
3441 // or the value at the end of the vectorized loop.
3442 BCResumeVal->addIncoming(EndValue, MiddleBlock);
3444 // Fix the scalar body counter (PHI node).
3445 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
3447 // The old induction's phi node in the scalar body needs the truncated
3449 for (BasicBlock *BB : LoopBypassBlocks)
3450 BCResumeVal->addIncoming(II.getStartValue(), BB);
3451 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
3454 // Add a check in the middle block to see if we have completed
3455 // all of the iterations in the first vector loop.
3456 // If (N - N%VF) == N, then we *don't* need to run the remainder.
3458 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, Count,
3459 CountRoundDown, "cmp.n", MiddleBlock->getTerminator());
3460 ReplaceInstWithInst(MiddleBlock->getTerminator(),
3461 BranchInst::Create(ExitBlock, ScalarPH, CmpN));
3463 // Get ready to start creating new instructions into the vectorized body.
3464 Builder.SetInsertPoint(&*VecBody->getFirstInsertionPt());
3467 LoopVectorPreHeader = Lp->getLoopPreheader();
3468 LoopScalarPreHeader = ScalarPH;
3469 LoopMiddleBlock = MiddleBlock;
3470 LoopExitBlock = ExitBlock;
3471 LoopVectorBody = VecBody;
3472 LoopScalarBody = OldBasicBlock;
3474 // Keep all loop hints from the original loop on the vector loop (we'll
3475 // replace the vectorizer-specific hints below).
3476 if (MDNode *LID = OrigLoop->getLoopID())
3479 LoopVectorizeHints Hints(Lp, true, *ORE);
3480 Hints.setAlreadyVectorized();
3483 // Fix up external users of the induction variable. At this point, we are
3484 // in LCSSA form, with all external PHIs that use the IV having one input value,
3485 // coming from the remainder loop. We need those PHIs to also have a correct
3486 // value for the IV when arriving directly from the middle block.
3487 void InnerLoopVectorizer::fixupIVUsers(PHINode *OrigPhi,
3488 const InductionDescriptor &II,
3489 Value *CountRoundDown, Value *EndValue,
3490 BasicBlock *MiddleBlock) {
3491 // There are two kinds of external IV usages - those that use the value
3492 // computed in the last iteration (the PHI) and those that use the penultimate
3493 // value (the value that feeds into the phi from the loop latch).
3494 // We allow both, but they, obviously, have different values.
3496 assert(OrigLoop->getExitBlock() && "Expected a single exit block");
3498 DenseMap<Value *, Value *> MissingVals;
3500 // An external user of the last iteration's value should see the value that
3501 // the remainder loop uses to initialize its own IV.
3502 Value *PostInc = OrigPhi->getIncomingValueForBlock(OrigLoop->getLoopLatch());
3503 for (User *U : PostInc->users()) {
3504 Instruction *UI = cast<Instruction>(U);
3505 if (!OrigLoop->contains(UI)) {
3506 assert(isa<PHINode>(UI) && "Expected LCSSA form");
3507 MissingVals[UI] = EndValue;
3511 // An external user of the penultimate value need to see EndValue - Step.
3512 // The simplest way to get this is to recompute it from the constituent SCEVs,
3513 // that is Start + (Step * (CRD - 1)).
3514 for (User *U : OrigPhi->users()) {
3515 auto *UI = cast<Instruction>(U);
3516 if (!OrigLoop->contains(UI)) {
3517 const DataLayout &DL =
3518 OrigLoop->getHeader()->getModule()->getDataLayout();
3519 assert(isa<PHINode>(UI) && "Expected LCSSA form");
3521 IRBuilder<> B(MiddleBlock->getTerminator());
3522 Value *CountMinusOne = B.CreateSub(
3523 CountRoundDown, ConstantInt::get(CountRoundDown->getType(), 1));
3524 Value *CMO = B.CreateSExtOrTrunc(CountMinusOne, II.getStep()->getType(),
3526 Value *Escape = II.transform(B, CMO, PSE.getSE(), DL);
3527 Escape->setName("ind.escape");
3528 MissingVals[UI] = Escape;
3532 for (auto &I : MissingVals) {
3533 PHINode *PHI = cast<PHINode>(I.first);
3534 // One corner case we have to handle is two IVs "chasing" each-other,
3535 // that is %IV2 = phi [...], [ %IV1, %latch ]
3536 // In this case, if IV1 has an external use, we need to avoid adding both
3537 // "last value of IV1" and "penultimate value of IV2". So, verify that we
3538 // don't already have an incoming value for the middle block.
3539 if (PHI->getBasicBlockIndex(MiddleBlock) == -1)
3540 PHI->addIncoming(I.second, MiddleBlock);
3545 struct CSEDenseMapInfo {
3546 static bool canHandle(Instruction *I) {
3547 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
3548 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
3550 static inline Instruction *getEmptyKey() {
3551 return DenseMapInfo<Instruction *>::getEmptyKey();
3553 static inline Instruction *getTombstoneKey() {
3554 return DenseMapInfo<Instruction *>::getTombstoneKey();
3556 static unsigned getHashValue(Instruction *I) {
3557 assert(canHandle(I) && "Unknown instruction!");
3558 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
3559 I->value_op_end()));
3561 static bool isEqual(Instruction *LHS, Instruction *RHS) {
3562 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
3563 LHS == getTombstoneKey() || RHS == getTombstoneKey())
3565 return LHS->isIdenticalTo(RHS);
3570 ///\brief Perform cse of induction variable instructions.
3571 static void cse(BasicBlock *BB) {
3572 // Perform simple cse.
3573 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
3574 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
3575 Instruction *In = &*I++;
3577 if (!CSEDenseMapInfo::canHandle(In))
3580 // Check if we can replace this instruction with any of the
3581 // visited instructions.
3582 if (Instruction *V = CSEMap.lookup(In)) {
3583 In->replaceAllUsesWith(V);
3584 In->eraseFromParent();
3592 /// \brief Adds a 'fast' flag to floating point operations.
3593 static Value *addFastMathFlag(Value *V) {
3594 if (isa<FPMathOperator>(V)) {
3595 FastMathFlags Flags;
3596 Flags.setUnsafeAlgebra();
3597 cast<Instruction>(V)->setFastMathFlags(Flags);
3602 /// \brief Estimate the overhead of scalarizing a value based on its type.
3603 /// Insert and Extract are set if the result needs to be inserted and/or
3604 /// extracted from vectors.
3605 static unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract,
3606 const TargetTransformInfo &TTI) {
3610 assert(Ty->isVectorTy() && "Can only scalarize vectors");
3613 for (unsigned I = 0, E = Ty->getVectorNumElements(); I < E; ++I) {
3615 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, Ty, I);
3617 Cost += TTI.getVectorInstrCost(Instruction::InsertElement, Ty, I);
3623 /// \brief Estimate the overhead of scalarizing an Instruction based on the
3624 /// types of its operands and return value.
3625 static unsigned getScalarizationOverhead(SmallVectorImpl<Type *> &OpTys,
3627 const TargetTransformInfo &TTI) {
3628 unsigned ScalarizationCost =
3629 getScalarizationOverhead(RetTy, true, false, TTI);
3631 for (Type *Ty : OpTys)
3632 ScalarizationCost += getScalarizationOverhead(Ty, false, true, TTI);
3634 return ScalarizationCost;
3637 /// \brief Estimate the overhead of scalarizing an instruction. This is a
3638 /// convenience wrapper for the type-based getScalarizationOverhead API.
3639 static unsigned getScalarizationOverhead(Instruction *I, unsigned VF,
3640 const TargetTransformInfo &TTI) {
3644 Type *RetTy = ToVectorTy(I->getType(), VF);
3646 SmallVector<Type *, 4> OpTys;
3647 unsigned OperandsNum = I->getNumOperands();
3648 for (unsigned OpInd = 0; OpInd < OperandsNum; ++OpInd)
3649 OpTys.push_back(ToVectorTy(I->getOperand(OpInd)->getType(), VF));
3651 return getScalarizationOverhead(OpTys, RetTy, TTI);
3654 // Estimate cost of a call instruction CI if it were vectorized with factor VF.
3655 // Return the cost of the instruction, including scalarization overhead if it's
3656 // needed. The flag NeedToScalarize shows if the call needs to be scalarized -
3657 // i.e. either vector version isn't available, or is too expensive.
3658 static unsigned getVectorCallCost(CallInst *CI, unsigned VF,
3659 const TargetTransformInfo &TTI,
3660 const TargetLibraryInfo *TLI,
3661 bool &NeedToScalarize) {
3662 Function *F = CI->getCalledFunction();
3663 StringRef FnName = CI->getCalledFunction()->getName();
3664 Type *ScalarRetTy = CI->getType();
3665 SmallVector<Type *, 4> Tys, ScalarTys;
3666 for (auto &ArgOp : CI->arg_operands())
3667 ScalarTys.push_back(ArgOp->getType());
3669 // Estimate cost of scalarized vector call. The source operands are assumed
3670 // to be vectors, so we need to extract individual elements from there,
3671 // execute VF scalar calls, and then gather the result into the vector return
3673 unsigned ScalarCallCost = TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys);
3675 return ScalarCallCost;
3677 // Compute corresponding vector type for return value and arguments.
3678 Type *RetTy = ToVectorTy(ScalarRetTy, VF);
3679 for (Type *ScalarTy : ScalarTys)
3680 Tys.push_back(ToVectorTy(ScalarTy, VF));
3682 // Compute costs of unpacking argument values for the scalar calls and
3683 // packing the return values to a vector.
3684 unsigned ScalarizationCost = getScalarizationOverhead(Tys, RetTy, TTI);
3686 unsigned Cost = ScalarCallCost * VF + ScalarizationCost;
3688 // If we can't emit a vector call for this function, then the currently found
3689 // cost is the cost we need to return.
3690 NeedToScalarize = true;
3691 if (!TLI || !TLI->isFunctionVectorizable(FnName, VF) || CI->isNoBuiltin())
3694 // If the corresponding vector cost is cheaper, return its cost.
3695 unsigned VectorCallCost = TTI.getCallInstrCost(nullptr, RetTy, Tys);
3696 if (VectorCallCost < Cost) {
3697 NeedToScalarize = false;
3698 return VectorCallCost;
3703 // Estimate cost of an intrinsic call instruction CI if it were vectorized with
3704 // factor VF. Return the cost of the instruction, including scalarization
3705 // overhead if it's needed.
3706 static unsigned getVectorIntrinsicCost(CallInst *CI, unsigned VF,
3707 const TargetTransformInfo &TTI,
3708 const TargetLibraryInfo *TLI) {
3709 Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
3710 assert(ID && "Expected intrinsic call!");
3712 Type *RetTy = ToVectorTy(CI->getType(), VF);
3713 SmallVector<Type *, 4> Tys;
3714 for (Value *ArgOperand : CI->arg_operands())
3715 Tys.push_back(ToVectorTy(ArgOperand->getType(), VF));
3718 if (auto *FPMO = dyn_cast<FPMathOperator>(CI))
3719 FMF = FPMO->getFastMathFlags();
3721 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys, FMF);
3724 static Type *smallestIntegerVectorType(Type *T1, Type *T2) {
3725 auto *I1 = cast<IntegerType>(T1->getVectorElementType());
3726 auto *I2 = cast<IntegerType>(T2->getVectorElementType());
3727 return I1->getBitWidth() < I2->getBitWidth() ? T1 : T2;
3729 static Type *largestIntegerVectorType(Type *T1, Type *T2) {
3730 auto *I1 = cast<IntegerType>(T1->getVectorElementType());
3731 auto *I2 = cast<IntegerType>(T2->getVectorElementType());
3732 return I1->getBitWidth() > I2->getBitWidth() ? T1 : T2;
3735 void InnerLoopVectorizer::truncateToMinimalBitwidths() {
3736 // For every instruction `I` in MinBWs, truncate the operands, create a
3737 // truncated version of `I` and reextend its result. InstCombine runs
3738 // later and will remove any ext/trunc pairs.
3740 SmallPtrSet<Value *, 4> Erased;
3741 for (const auto &KV : Cost->getMinimalBitwidths()) {
3742 // If the value wasn't vectorized, we must maintain the original scalar
3743 // type. The absence of the value from VectorLoopValueMap indicates that it
3744 // wasn't vectorized.
3745 if (!VectorLoopValueMap.hasVector(KV.first))
3747 VectorParts &Parts = VectorLoopValueMap.getVector(KV.first);
3748 for (Value *&I : Parts) {
3749 if (Erased.count(I) || I->use_empty() || !isa<Instruction>(I))
3751 Type *OriginalTy = I->getType();
3752 Type *ScalarTruncatedTy =
3753 IntegerType::get(OriginalTy->getContext(), KV.second);
3754 Type *TruncatedTy = VectorType::get(ScalarTruncatedTy,
3755 OriginalTy->getVectorNumElements());
3756 if (TruncatedTy == OriginalTy)
3759 IRBuilder<> B(cast<Instruction>(I));
3760 auto ShrinkOperand = [&](Value *V) -> Value * {
3761 if (auto *ZI = dyn_cast<ZExtInst>(V))
3762 if (ZI->getSrcTy() == TruncatedTy)
3763 return ZI->getOperand(0);
3764 return B.CreateZExtOrTrunc(V, TruncatedTy);
3767 // The actual instruction modification depends on the instruction type,
3769 Value *NewI = nullptr;
3770 if (auto *BO = dyn_cast<BinaryOperator>(I)) {
3771 NewI = B.CreateBinOp(BO->getOpcode(), ShrinkOperand(BO->getOperand(0)),
3772 ShrinkOperand(BO->getOperand(1)));
3773 cast<BinaryOperator>(NewI)->copyIRFlags(I);
3774 } else if (auto *CI = dyn_cast<ICmpInst>(I)) {
3776 B.CreateICmp(CI->getPredicate(), ShrinkOperand(CI->getOperand(0)),
3777 ShrinkOperand(CI->getOperand(1)));
3778 } else if (auto *SI = dyn_cast<SelectInst>(I)) {
3779 NewI = B.CreateSelect(SI->getCondition(),
3780 ShrinkOperand(SI->getTrueValue()),
3781 ShrinkOperand(SI->getFalseValue()));
3782 } else if (auto *CI = dyn_cast<CastInst>(I)) {
3783 switch (CI->getOpcode()) {
3785 llvm_unreachable("Unhandled cast!");
3786 case Instruction::Trunc:
3787 NewI = ShrinkOperand(CI->getOperand(0));
3789 case Instruction::SExt:
3790 NewI = B.CreateSExtOrTrunc(
3792 smallestIntegerVectorType(OriginalTy, TruncatedTy));
3794 case Instruction::ZExt:
3795 NewI = B.CreateZExtOrTrunc(
3797 smallestIntegerVectorType(OriginalTy, TruncatedTy));
3800 } else if (auto *SI = dyn_cast<ShuffleVectorInst>(I)) {
3801 auto Elements0 = SI->getOperand(0)->getType()->getVectorNumElements();
3802 auto *O0 = B.CreateZExtOrTrunc(
3803 SI->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements0));
3804 auto Elements1 = SI->getOperand(1)->getType()->getVectorNumElements();
3805 auto *O1 = B.CreateZExtOrTrunc(
3806 SI->getOperand(1), VectorType::get(ScalarTruncatedTy, Elements1));
3808 NewI = B.CreateShuffleVector(O0, O1, SI->getMask());
3809 } else if (isa<LoadInst>(I)) {
3810 // Don't do anything with the operands, just extend the result.
3812 } else if (auto *IE = dyn_cast<InsertElementInst>(I)) {
3813 auto Elements = IE->getOperand(0)->getType()->getVectorNumElements();
3814 auto *O0 = B.CreateZExtOrTrunc(
3815 IE->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements));
3816 auto *O1 = B.CreateZExtOrTrunc(IE->getOperand(1), ScalarTruncatedTy);
3817 NewI = B.CreateInsertElement(O0, O1, IE->getOperand(2));
3818 } else if (auto *EE = dyn_cast<ExtractElementInst>(I)) {
3819 auto Elements = EE->getOperand(0)->getType()->getVectorNumElements();
3820 auto *O0 = B.CreateZExtOrTrunc(
3821 EE->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements));
3822 NewI = B.CreateExtractElement(O0, EE->getOperand(2));
3824 llvm_unreachable("Unhandled instruction type!");
3827 // Lastly, extend the result.
3828 NewI->takeName(cast<Instruction>(I));
3829 Value *Res = B.CreateZExtOrTrunc(NewI, OriginalTy);
3830 I->replaceAllUsesWith(Res);
3831 cast<Instruction>(I)->eraseFromParent();
3837 // We'll have created a bunch of ZExts that are now parentless. Clean up.
3838 for (const auto &KV : Cost->getMinimalBitwidths()) {
3839 // If the value wasn't vectorized, we must maintain the original scalar
3840 // type. The absence of the value from VectorLoopValueMap indicates that it
3841 // wasn't vectorized.
3842 if (!VectorLoopValueMap.hasVector(KV.first))
3844 VectorParts &Parts = VectorLoopValueMap.getVector(KV.first);
3845 for (Value *&I : Parts) {
3846 ZExtInst *Inst = dyn_cast<ZExtInst>(I);
3847 if (Inst && Inst->use_empty()) {
3848 Value *NewI = Inst->getOperand(0);
3849 Inst->eraseFromParent();
3856 void InnerLoopVectorizer::vectorizeLoop() {
3857 //===------------------------------------------------===//
3859 // Notice: any optimization or new instruction that go
3860 // into the code below should be also be implemented in
3863 //===------------------------------------------------===//
3864 Constant *Zero = Builder.getInt32(0);
3866 // In order to support recurrences we need to be able to vectorize Phi nodes.
3867 // Phi nodes have cycles, so we need to vectorize them in two stages. First,
3868 // we create a new vector PHI node with no incoming edges. We use this value
3869 // when we vectorize all of the instructions that use the PHI. Next, after
3870 // all of the instructions in the block are complete we add the new incoming
3871 // edges to the PHI. At this point all of the instructions in the basic block
3872 // are vectorized, so we can use them to construct the PHI.
3873 PhiVector PHIsToFix;
3875 // Collect instructions from the original loop that will become trivially
3876 // dead in the vectorized loop. We don't need to vectorize these
3878 collectTriviallyDeadInstructions();
3880 // Scan the loop in a topological order to ensure that defs are vectorized
3882 LoopBlocksDFS DFS(OrigLoop);
3885 // Vectorize all of the blocks in the original loop.
3886 for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO()))
3887 vectorizeBlockInLoop(BB, &PHIsToFix);
3889 // Insert truncates and extends for any truncated instructions as hints to
3892 truncateToMinimalBitwidths();
3894 // At this point every instruction in the original loop is widened to a
3895 // vector form. Now we need to fix the recurrences in PHIsToFix. These PHI
3896 // nodes are currently empty because we did not want to introduce cycles.
3897 // This is the second stage of vectorizing recurrences.
3898 for (PHINode *Phi : PHIsToFix) {
3899 assert(Phi && "Unable to recover vectorized PHI");
3901 // Handle first-order recurrences that need to be fixed.
3902 if (Legal->isFirstOrderRecurrence(Phi)) {
3903 fixFirstOrderRecurrence(Phi);
3907 // If the phi node is not a first-order recurrence, it must be a reduction.
3908 // Get it's reduction variable descriptor.
3909 assert(Legal->isReductionVariable(Phi) &&
3910 "Unable to find the reduction variable");
3911 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[Phi];
3913 RecurrenceDescriptor::RecurrenceKind RK = RdxDesc.getRecurrenceKind();
3914 TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue();
3915 Instruction *LoopExitInst = RdxDesc.getLoopExitInstr();
3916 RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind =
3917 RdxDesc.getMinMaxRecurrenceKind();
3918 setDebugLocFromInst(Builder, ReductionStartValue);
3920 // We need to generate a reduction vector from the incoming scalar.
3921 // To do so, we need to generate the 'identity' vector and override
3922 // one of the elements with the incoming scalar reduction. We need
3923 // to do it in the vector-loop preheader.
3924 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
3926 // This is the vector-clone of the value that leaves the loop.
3927 const VectorParts &VectorExit = getVectorValue(LoopExitInst);
3928 Type *VecTy = VectorExit[0]->getType();
3930 // Find the reduction identity variable. Zero for addition, or, xor,
3931 // one for multiplication, -1 for And.
3934 if (RK == RecurrenceDescriptor::RK_IntegerMinMax ||
3935 RK == RecurrenceDescriptor::RK_FloatMinMax) {
3936 // MinMax reduction have the start value as their identify.
3938 VectorStart = Identity = ReductionStartValue;
3940 VectorStart = Identity =
3941 Builder.CreateVectorSplat(VF, ReductionStartValue, "minmax.ident");
3944 // Handle other reduction kinds:
3945 Constant *Iden = RecurrenceDescriptor::getRecurrenceIdentity(
3946 RK, VecTy->getScalarType());
3949 // This vector is the Identity vector where the first element is the
3950 // incoming scalar reduction.
3951 VectorStart = ReductionStartValue;
3953 Identity = ConstantVector::getSplat(VF, Iden);
3955 // This vector is the Identity vector where the first element is the
3956 // incoming scalar reduction.
3958 Builder.CreateInsertElement(Identity, ReductionStartValue, Zero);
3962 // Fix the vector-loop phi.
3964 // Reductions do not have to start at zero. They can start with
3965 // any loop invariant values.
3966 const VectorParts &VecRdxPhi = getVectorValue(Phi);
3967 BasicBlock *Latch = OrigLoop->getLoopLatch();
3968 Value *LoopVal = Phi->getIncomingValueForBlock(Latch);
3969 const VectorParts &Val = getVectorValue(LoopVal);
3970 for (unsigned part = 0; part < UF; ++part) {
3971 // Make sure to add the reduction stat value only to the
3972 // first unroll part.
3973 Value *StartVal = (part == 0) ? VectorStart : Identity;
3974 cast<PHINode>(VecRdxPhi[part])
3975 ->addIncoming(StartVal, LoopVectorPreHeader);
3976 cast<PHINode>(VecRdxPhi[part])
3977 ->addIncoming(Val[part], LoopVectorBody);
3980 // Before each round, move the insertion point right between
3981 // the PHIs and the values we are going to write.
3982 // This allows us to write both PHINodes and the extractelement
3984 Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt());
3986 VectorParts &RdxParts = VectorLoopValueMap.getVector(LoopExitInst);
3987 setDebugLocFromInst(Builder, LoopExitInst);
3989 // If the vector reduction can be performed in a smaller type, we truncate
3990 // then extend the loop exit value to enable InstCombine to evaluate the
3991 // entire expression in the smaller type.
3992 if (VF > 1 && Phi->getType() != RdxDesc.getRecurrenceType()) {
3993 Type *RdxVecTy = VectorType::get(RdxDesc.getRecurrenceType(), VF);
3994 Builder.SetInsertPoint(LoopVectorBody->getTerminator());
3995 for (unsigned part = 0; part < UF; ++part) {
3996 Value *Trunc = Builder.CreateTrunc(RdxParts[part], RdxVecTy);
3997 Value *Extnd = RdxDesc.isSigned() ? Builder.CreateSExt(Trunc, VecTy)
3998 : Builder.CreateZExt(Trunc, VecTy);
3999 for (Value::user_iterator UI = RdxParts[part]->user_begin();
4000 UI != RdxParts[part]->user_end();)
4002 (*UI++)->replaceUsesOfWith(RdxParts[part], Extnd);
4003 RdxParts[part] = Extnd;
4008 Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt());
4009 for (unsigned part = 0; part < UF; ++part)
4010 RdxParts[part] = Builder.CreateTrunc(RdxParts[part], RdxVecTy);
4013 // Reduce all of the unrolled parts into a single vector.
4014 Value *ReducedPartRdx = RdxParts[0];
4015 unsigned Op = RecurrenceDescriptor::getRecurrenceBinOp(RK);
4016 setDebugLocFromInst(Builder, ReducedPartRdx);
4017 for (unsigned part = 1; part < UF; ++part) {
4018 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
4019 // Floating point operations had to be 'fast' to enable the reduction.
4020 ReducedPartRdx = addFastMathFlag(
4021 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
4022 ReducedPartRdx, "bin.rdx"));
4024 ReducedPartRdx = RecurrenceDescriptor::createMinMaxOp(
4025 Builder, MinMaxKind, ReducedPartRdx, RdxParts[part]);
4029 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
4030 // and vector ops, reducing the set of values being computed by half each
4032 assert(isPowerOf2_32(VF) &&
4033 "Reduction emission only supported for pow2 vectors!");
4034 Value *TmpVec = ReducedPartRdx;
4035 SmallVector<Constant *, 32> ShuffleMask(VF, nullptr);
4036 for (unsigned i = VF; i != 1; i >>= 1) {
4037 // Move the upper half of the vector to the lower half.
4038 for (unsigned j = 0; j != i / 2; ++j)
4039 ShuffleMask[j] = Builder.getInt32(i / 2 + j);
4041 // Fill the rest of the mask with undef.
4042 std::fill(&ShuffleMask[i / 2], ShuffleMask.end(),
4043 UndefValue::get(Builder.getInt32Ty()));
4045 Value *Shuf = Builder.CreateShuffleVector(
4046 TmpVec, UndefValue::get(TmpVec->getType()),
4047 ConstantVector::get(ShuffleMask), "rdx.shuf");
4049 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
4050 // Floating point operations had to be 'fast' to enable the reduction.
4051 TmpVec = addFastMathFlag(Builder.CreateBinOp(
4052 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
4054 TmpVec = RecurrenceDescriptor::createMinMaxOp(Builder, MinMaxKind,
4058 // The result is in the first element of the vector.
4060 Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
4062 // If the reduction can be performed in a smaller type, we need to extend
4063 // the reduction to the wider type before we branch to the original loop.
4064 if (Phi->getType() != RdxDesc.getRecurrenceType())
4067 ? Builder.CreateSExt(ReducedPartRdx, Phi->getType())
4068 : Builder.CreateZExt(ReducedPartRdx, Phi->getType());
4071 // Create a phi node that merges control-flow from the backedge-taken check
4072 // block and the middle block.
4073 PHINode *BCBlockPhi = PHINode::Create(Phi->getType(), 2, "bc.merge.rdx",
4074 LoopScalarPreHeader->getTerminator());
4075 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
4076 BCBlockPhi->addIncoming(ReductionStartValue, LoopBypassBlocks[I]);
4077 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
4079 // Now, we need to fix the users of the reduction variable
4080 // inside and outside of the scalar remainder loop.
4081 // We know that the loop is in LCSSA form. We need to update the
4082 // PHI nodes in the exit blocks.
4083 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
4084 LEE = LoopExitBlock->end();
4085 LEI != LEE; ++LEI) {
4086 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
4090 // All PHINodes need to have a single entry edge, or two if
4091 // we already fixed them.
4092 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
4094 // We found our reduction value exit-PHI. Update it with the
4095 // incoming bypass edge.
4096 if (LCSSAPhi->getIncomingValue(0) == LoopExitInst) {
4097 // Add an edge coming from the bypass.
4098 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
4101 } // end of the LCSSA phi scan.
4103 // Fix the scalar loop reduction variable with the incoming reduction sum
4104 // from the vector body and from the backedge value.
4105 int IncomingEdgeBlockIdx =
4106 Phi->getBasicBlockIndex(OrigLoop->getLoopLatch());
4107 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
4108 // Pick the other block.
4109 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
4110 Phi->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
4111 Phi->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst);
4112 } // end of for each Phi in PHIsToFix.
4114 // Update the dominator tree.
4116 // FIXME: After creating the structure of the new loop, the dominator tree is
4117 // no longer up-to-date, and it remains that way until we update it
4118 // here. An out-of-date dominator tree is problematic for SCEV,
4119 // because SCEVExpander uses it to guide code generation. The
4120 // vectorizer use SCEVExpanders in several places. Instead, we should
4121 // keep the dominator tree up-to-date as we go.
4124 // Fix-up external users of the induction variables.
4125 for (auto &Entry : *Legal->getInductionVars())
4126 fixupIVUsers(Entry.first, Entry.second,
4127 getOrCreateVectorTripCount(LI->getLoopFor(LoopVectorBody)),
4128 IVEndValues[Entry.first], LoopMiddleBlock);
4131 predicateInstructions();
4133 // Remove redundant induction instructions.
4134 cse(LoopVectorBody);
4137 void InnerLoopVectorizer::fixFirstOrderRecurrence(PHINode *Phi) {
4139 // This is the second phase of vectorizing first-order recurrences. An
4140 // overview of the transformation is described below. Suppose we have the
4143 // for (int i = 0; i < n; ++i)
4144 // b[i] = a[i] - a[i - 1];
4146 // There is a first-order recurrence on "a". For this loop, the shorthand
4147 // scalar IR looks like:
4154 // i = phi [0, scalar.ph], [i+1, scalar.body]
4155 // s1 = phi [s_init, scalar.ph], [s2, scalar.body]
4158 // br cond, scalar.body, ...
4160 // In this example, s1 is a recurrence because it's value depends on the
4161 // previous iteration. In the first phase of vectorization, we created a
4162 // temporary value for s1. We now complete the vectorization and produce the
4163 // shorthand vector IR shown below (for VF = 4, UF = 1).
4166 // v_init = vector(..., ..., ..., a[-1])
4170 // i = phi [0, vector.ph], [i+4, vector.body]
4171 // v1 = phi [v_init, vector.ph], [v2, vector.body]
4172 // v2 = a[i, i+1, i+2, i+3];
4173 // v3 = vector(v1(3), v2(0, 1, 2))
4174 // b[i, i+1, i+2, i+3] = v2 - v3
4175 // br cond, vector.body, middle.block
4182 // s_init = phi [x, middle.block], [a[-1], otherwise]
4185 // After execution completes the vector loop, we extract the next value of
4186 // the recurrence (x) to use as the initial value in the scalar loop.
4188 // Get the original loop preheader and single loop latch.
4189 auto *Preheader = OrigLoop->getLoopPreheader();
4190 auto *Latch = OrigLoop->getLoopLatch();
4192 // Get the initial and previous values of the scalar recurrence.
4193 auto *ScalarInit = Phi->getIncomingValueForBlock(Preheader);
4194 auto *Previous = Phi->getIncomingValueForBlock(Latch);
4196 // Create a vector from the initial value.
4197 auto *VectorInit = ScalarInit;
4199 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
4200 VectorInit = Builder.CreateInsertElement(
4201 UndefValue::get(VectorType::get(VectorInit->getType(), VF)), VectorInit,
4202 Builder.getInt32(VF - 1), "vector.recur.init");
4205 // We constructed a temporary phi node in the first phase of vectorization.
4206 // This phi node will eventually be deleted.
4207 VectorParts &PhiParts = VectorLoopValueMap.getVector(Phi);
4208 Builder.SetInsertPoint(cast<Instruction>(PhiParts[0]));
4210 // Create a phi node for the new recurrence. The current value will either be
4211 // the initial value inserted into a vector or loop-varying vector value.
4212 auto *VecPhi = Builder.CreatePHI(VectorInit->getType(), 2, "vector.recur");
4213 VecPhi->addIncoming(VectorInit, LoopVectorPreHeader);
4215 // Get the vectorized previous value. We ensured the previous values was an
4216 // instruction when detecting the recurrence.
4217 auto &PreviousParts = getVectorValue(Previous);
4219 // Set the insertion point to be after this instruction. We ensured the
4220 // previous value dominated all uses of the phi when detecting the
4222 Builder.SetInsertPoint(
4223 &*++BasicBlock::iterator(cast<Instruction>(PreviousParts[UF - 1])));
4225 // We will construct a vector for the recurrence by combining the values for
4226 // the current and previous iterations. This is the required shuffle mask.
4227 SmallVector<Constant *, 8> ShuffleMask(VF);
4228 ShuffleMask[0] = Builder.getInt32(VF - 1);
4229 for (unsigned I = 1; I < VF; ++I)
4230 ShuffleMask[I] = Builder.getInt32(I + VF - 1);
4232 // The vector from which to take the initial value for the current iteration
4233 // (actual or unrolled). Initially, this is the vector phi node.
4234 Value *Incoming = VecPhi;
4236 // Shuffle the current and previous vector and update the vector parts.
4237 for (unsigned Part = 0; Part < UF; ++Part) {
4240 ? Builder.CreateShuffleVector(Incoming, PreviousParts[Part],
4241 ConstantVector::get(ShuffleMask))
4243 PhiParts[Part]->replaceAllUsesWith(Shuffle);
4244 cast<Instruction>(PhiParts[Part])->eraseFromParent();
4245 PhiParts[Part] = Shuffle;
4246 Incoming = PreviousParts[Part];
4249 // Fix the latch value of the new recurrence in the vector loop.
4250 VecPhi->addIncoming(Incoming, LI->getLoopFor(LoopVectorBody)->getLoopLatch());
4252 // Extract the last vector element in the middle block. This will be the
4253 // initial value for the recurrence when jumping to the scalar loop.
4254 auto *Extract = Incoming;
4256 Builder.SetInsertPoint(LoopMiddleBlock->getTerminator());
4257 Extract = Builder.CreateExtractElement(Extract, Builder.getInt32(VF - 1),
4258 "vector.recur.extract");
4261 // Fix the initial value of the original recurrence in the scalar loop.
4262 Builder.SetInsertPoint(&*LoopScalarPreHeader->begin());
4263 auto *Start = Builder.CreatePHI(Phi->getType(), 2, "scalar.recur.init");
4264 for (auto *BB : predecessors(LoopScalarPreHeader)) {
4265 auto *Incoming = BB == LoopMiddleBlock ? Extract : ScalarInit;
4266 Start->addIncoming(Incoming, BB);
4269 Phi->setIncomingValue(Phi->getBasicBlockIndex(LoopScalarPreHeader), Start);
4270 Phi->setName("scalar.recur");
4272 // Finally, fix users of the recurrence outside the loop. The users will need
4273 // either the last value of the scalar recurrence or the last value of the
4274 // vector recurrence we extracted in the middle block. Since the loop is in
4275 // LCSSA form, we just need to find the phi node for the original scalar
4276 // recurrence in the exit block, and then add an edge for the middle block.
4277 for (auto &I : *LoopExitBlock) {
4278 auto *LCSSAPhi = dyn_cast<PHINode>(&I);
4281 if (LCSSAPhi->getIncomingValue(0) == Phi) {
4282 LCSSAPhi->addIncoming(Extract, LoopMiddleBlock);
4288 void InnerLoopVectorizer::fixLCSSAPHIs() {
4289 for (Instruction &LEI : *LoopExitBlock) {
4290 auto *LCSSAPhi = dyn_cast<PHINode>(&LEI);
4293 if (LCSSAPhi->getNumIncomingValues() == 1)
4294 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
4299 void InnerLoopVectorizer::collectTriviallyDeadInstructions() {
4300 BasicBlock *Latch = OrigLoop->getLoopLatch();
4302 // We create new control-flow for the vectorized loop, so the original
4303 // condition will be dead after vectorization if it's only used by the
4305 auto *Cmp = dyn_cast<Instruction>(Latch->getTerminator()->getOperand(0));
4306 if (Cmp && Cmp->hasOneUse())
4307 DeadInstructions.insert(Cmp);
4309 // We create new "steps" for induction variable updates to which the original
4310 // induction variables map. An original update instruction will be dead if
4311 // all its users except the induction variable are dead.
4312 for (auto &Induction : *Legal->getInductionVars()) {
4313 PHINode *Ind = Induction.first;
4314 auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch));
4315 if (all_of(IndUpdate->users(), [&](User *U) -> bool {
4316 return U == Ind || DeadInstructions.count(cast<Instruction>(U));
4318 DeadInstructions.insert(IndUpdate);
4322 void InnerLoopVectorizer::sinkScalarOperands(Instruction *PredInst) {
4324 // The basic block and loop containing the predicated instruction.
4325 auto *PredBB = PredInst->getParent();
4326 auto *VectorLoop = LI->getLoopFor(PredBB);
4328 // Initialize a worklist with the operands of the predicated instruction.
4329 SetVector<Value *> Worklist(PredInst->op_begin(), PredInst->op_end());
4331 // Holds instructions that we need to analyze again. An instruction may be
4332 // reanalyzed if we don't yet know if we can sink it or not.
4333 SmallVector<Instruction *, 8> InstsToReanalyze;
4335 // Returns true if a given use occurs in the predicated block. Phi nodes use
4336 // their operands in their corresponding predecessor blocks.
4337 auto isBlockOfUsePredicated = [&](Use &U) -> bool {
4338 auto *I = cast<Instruction>(U.getUser());
4339 BasicBlock *BB = I->getParent();
4340 if (auto *Phi = dyn_cast<PHINode>(I))
4341 BB = Phi->getIncomingBlock(
4342 PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
4343 return BB == PredBB;
4346 // Iteratively sink the scalarized operands of the predicated instruction
4347 // into the block we created for it. When an instruction is sunk, it's
4348 // operands are then added to the worklist. The algorithm ends after one pass
4349 // through the worklist doesn't sink a single instruction.
4353 // Add the instructions that need to be reanalyzed to the worklist, and
4354 // reset the changed indicator.
4355 Worklist.insert(InstsToReanalyze.begin(), InstsToReanalyze.end());
4356 InstsToReanalyze.clear();
4359 while (!Worklist.empty()) {
4360 auto *I = dyn_cast<Instruction>(Worklist.pop_back_val());
4362 // We can't sink an instruction if it is a phi node, is already in the
4363 // predicated block, is not in the loop, or may have side effects.
4364 if (!I || isa<PHINode>(I) || I->getParent() == PredBB ||
4365 !VectorLoop->contains(I) || I->mayHaveSideEffects())
4368 // It's legal to sink the instruction if all its uses occur in the
4369 // predicated block. Otherwise, there's nothing to do yet, and we may
4370 // need to reanalyze the instruction.
4371 if (!all_of(I->uses(), isBlockOfUsePredicated)) {
4372 InstsToReanalyze.push_back(I);
4376 // Move the instruction to the beginning of the predicated block, and add
4377 // it's operands to the worklist.
4378 I->moveBefore(&*PredBB->getFirstInsertionPt());
4379 Worklist.insert(I->op_begin(), I->op_end());
4381 // The sinking may have enabled other instructions to be sunk, so we will
4388 void InnerLoopVectorizer::predicateInstructions() {
4390 // For each instruction I marked for predication on value C, split I into its
4391 // own basic block to form an if-then construct over C. Since I may be fed by
4392 // an extractelement instruction or other scalar operand, we try to
4393 // iteratively sink its scalar operands into the predicated block. If I feeds
4394 // an insertelement instruction, we try to move this instruction into the
4395 // predicated block as well. For non-void types, a phi node will be created
4396 // for the resulting value (either vector or scalar).
4398 // So for some predicated instruction, e.g. the conditional sdiv in:
4402 // %add = add nsw i32 %mul, %0
4403 // %cmp5 = icmp sgt i32 %2, 7
4404 // br i1 %cmp5, label %if.then, label %if.end
4407 // %div = sdiv i32 %0, %1
4411 // %x.0 = phi i32 [ %div, %if.then ], [ %add, %for.body ]
4413 // the sdiv at this point is scalarized and if-converted using a select.
4414 // The inactive elements in the vector are not used, but the predicated
4415 // instruction is still executed for all vector elements, essentially:
4419 // %17 = add nsw <2 x i32> %16, %wide.load
4420 // %29 = extractelement <2 x i32> %wide.load, i32 0
4421 // %30 = extractelement <2 x i32> %wide.load51, i32 0
4422 // %31 = sdiv i32 %29, %30
4423 // %32 = insertelement <2 x i32> undef, i32 %31, i32 0
4424 // %35 = extractelement <2 x i32> %wide.load, i32 1
4425 // %36 = extractelement <2 x i32> %wide.load51, i32 1
4426 // %37 = sdiv i32 %35, %36
4427 // %38 = insertelement <2 x i32> %32, i32 %37, i32 1
4428 // %predphi = select <2 x i1> %26, <2 x i32> %38, <2 x i32> %17
4430 // Predication will now re-introduce the original control flow to avoid false
4431 // side-effects by the sdiv instructions on the inactive elements, yielding
4436 // %5 = add nsw <2 x i32> %4, %wide.load
4437 // %8 = icmp sgt <2 x i32> %wide.load52, <i32 7, i32 7>
4438 // %9 = extractelement <2 x i1> %8, i32 0
4439 // br i1 %9, label %pred.sdiv.if, label %pred.sdiv.continue
4442 // %10 = extractelement <2 x i32> %wide.load, i32 0
4443 // %11 = extractelement <2 x i32> %wide.load51, i32 0
4444 // %12 = sdiv i32 %10, %11
4445 // %13 = insertelement <2 x i32> undef, i32 %12, i32 0
4446 // br label %pred.sdiv.continue
4448 // pred.sdiv.continue:
4449 // %14 = phi <2 x i32> [ undef, %vector.body ], [ %13, %pred.sdiv.if ]
4450 // %15 = extractelement <2 x i1> %8, i32 1
4451 // br i1 %15, label %pred.sdiv.if54, label %pred.sdiv.continue55
4454 // %16 = extractelement <2 x i32> %wide.load, i32 1
4455 // %17 = extractelement <2 x i32> %wide.load51, i32 1
4456 // %18 = sdiv i32 %16, %17
4457 // %19 = insertelement <2 x i32> %14, i32 %18, i32 1
4458 // br label %pred.sdiv.continue55
4460 // pred.sdiv.continue55:
4461 // %20 = phi <2 x i32> [ %14, %pred.sdiv.continue ], [ %19, %pred.sdiv.if54 ]
4462 // %predphi = select <2 x i1> %8, <2 x i32> %20, <2 x i32> %5
4464 for (auto KV : PredicatedInstructions) {
4465 BasicBlock::iterator I(KV.first);
4466 BasicBlock *Head = I->getParent();
4467 auto *BB = SplitBlock(Head, &*std::next(I), DT, LI);
4468 auto *T = SplitBlockAndInsertIfThen(KV.second, &*I, /*Unreachable=*/false,
4469 /*BranchWeights=*/nullptr, DT, LI);
4471 sinkScalarOperands(&*I);
4473 I->getParent()->setName(Twine("pred.") + I->getOpcodeName() + ".if");
4474 BB->setName(Twine("pred.") + I->getOpcodeName() + ".continue");
4476 // If the instruction is non-void create a Phi node at reconvergence point.
4477 if (!I->getType()->isVoidTy()) {
4478 Value *IncomingTrue = nullptr;
4479 Value *IncomingFalse = nullptr;
4481 if (I->hasOneUse() && isa<InsertElementInst>(*I->user_begin())) {
4482 // If the predicated instruction is feeding an insert-element, move it
4483 // into the Then block; Phi node will be created for the vector.
4484 InsertElementInst *IEI = cast<InsertElementInst>(*I->user_begin());
4486 IncomingTrue = IEI; // the new vector with the inserted element.
4487 IncomingFalse = IEI->getOperand(0); // the unmodified vector
4489 // Phi node will be created for the scalar predicated instruction.
4491 IncomingFalse = UndefValue::get(I->getType());
4494 BasicBlock *PostDom = I->getParent()->getSingleSuccessor();
4495 assert(PostDom && "Then block has multiple successors");
4497 PHINode::Create(IncomingTrue->getType(), 2, "", &PostDom->front());
4498 IncomingTrue->replaceAllUsesWith(Phi);
4499 Phi->addIncoming(IncomingFalse, Head);
4500 Phi->addIncoming(IncomingTrue, I->getParent());
4504 DEBUG(DT->verifyDomTree());
4507 InnerLoopVectorizer::VectorParts
4508 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
4509 assert(is_contained(predecessors(Dst), Src) && "Invalid edge");
4511 // Look for cached value.
4512 std::pair<BasicBlock *, BasicBlock *> Edge(Src, Dst);
4513 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
4514 if (ECEntryIt != MaskCache.end())
4515 return ECEntryIt->second;
4517 VectorParts SrcMask = createBlockInMask(Src);
4519 // The terminator has to be a branch inst!
4520 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
4521 assert(BI && "Unexpected terminator found");
4523 if (BI->isConditional()) {
4524 VectorParts EdgeMask = getVectorValue(BI->getCondition());
4526 if (BI->getSuccessor(0) != Dst)
4527 for (unsigned part = 0; part < UF; ++part)
4528 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
4530 for (unsigned part = 0; part < UF; ++part)
4531 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
4533 MaskCache[Edge] = EdgeMask;
4537 MaskCache[Edge] = SrcMask;
4541 InnerLoopVectorizer::VectorParts
4542 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
4543 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
4545 // Loop incoming mask is all-one.
4546 if (OrigLoop->getHeader() == BB) {
4547 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
4548 return getVectorValue(C);
4551 // This is the block mask. We OR all incoming edges, and with zero.
4552 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
4553 VectorParts BlockMask = getVectorValue(Zero);
4556 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
4557 VectorParts EM = createEdgeMask(*it, BB);
4558 for (unsigned part = 0; part < UF; ++part)
4559 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
4565 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN, unsigned UF,
4566 unsigned VF, PhiVector *PV) {
4567 PHINode *P = cast<PHINode>(PN);
4568 // Handle recurrences.
4569 if (Legal->isReductionVariable(P) || Legal->isFirstOrderRecurrence(P)) {
4570 VectorParts Entry(UF);
4571 for (unsigned part = 0; part < UF; ++part) {
4572 // This is phase one of vectorizing PHIs.
4574 (VF == 1) ? PN->getType() : VectorType::get(PN->getType(), VF);
4575 Entry[part] = PHINode::Create(
4576 VecTy, 2, "vec.phi", &*LoopVectorBody->getFirstInsertionPt());
4578 VectorLoopValueMap.initVector(P, Entry);
4583 setDebugLocFromInst(Builder, P);
4584 // Check for PHI nodes that are lowered to vector selects.
4585 if (P->getParent() != OrigLoop->getHeader()) {
4586 // We know that all PHIs in non-header blocks are converted into
4587 // selects, so we don't have to worry about the insertion order and we
4588 // can just use the builder.
4589 // At this point we generate the predication tree. There may be
4590 // duplications since this is a simple recursive scan, but future
4591 // optimizations will clean it up.
4593 unsigned NumIncoming = P->getNumIncomingValues();
4595 // Generate a sequence of selects of the form:
4596 // SELECT(Mask3, In3,
4597 // SELECT(Mask2, In2,
4599 VectorParts Entry(UF);
4600 for (unsigned In = 0; In < NumIncoming; In++) {
4602 createEdgeMask(P->getIncomingBlock(In), P->getParent());
4603 const VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
4605 for (unsigned part = 0; part < UF; ++part) {
4606 // We might have single edge PHIs (blocks) - use an identity
4607 // 'select' for the first PHI operand.
4609 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], In0[part]);
4611 // Select between the current value and the previous incoming edge
4612 // based on the incoming mask.
4613 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], Entry[part],
4617 VectorLoopValueMap.initVector(P, Entry);
4621 // This PHINode must be an induction variable.
4622 // Make sure that we know about it.
4623 assert(Legal->getInductionVars()->count(P) && "Not an induction variable");
4625 InductionDescriptor II = Legal->getInductionVars()->lookup(P);
4626 const DataLayout &DL = OrigLoop->getHeader()->getModule()->getDataLayout();
4628 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
4629 // which can be found from the original scalar operations.
4630 switch (II.getKind()) {
4631 case InductionDescriptor::IK_NoInduction:
4632 llvm_unreachable("Unknown induction");
4633 case InductionDescriptor::IK_IntInduction:
4634 return widenIntInduction(P);
4635 case InductionDescriptor::IK_PtrInduction: {
4636 // Handle the pointer induction variable case.
4637 assert(P->getType()->isPointerTy() && "Unexpected type.");
4638 // This is the normalized GEP that starts counting at zero.
4639 Value *PtrInd = Induction;
4640 PtrInd = Builder.CreateSExtOrTrunc(PtrInd, II.getStep()->getType());
4641 // Determine the number of scalars we need to generate for each unroll
4642 // iteration. If the instruction is uniform, we only need to generate the
4643 // first lane. Otherwise, we generate all VF values.
4644 unsigned Lanes = Legal->isUniformAfterVectorization(P) ? 1 : VF;
4645 // These are the scalar results. Notice that we don't generate vector GEPs
4646 // because scalar GEPs result in better code.
4647 ScalarParts Entry(UF);
4648 for (unsigned Part = 0; Part < UF; ++Part) {
4649 Entry[Part].resize(VF);
4650 for (unsigned Lane = 0; Lane < Lanes; ++Lane) {
4651 Constant *Idx = ConstantInt::get(PtrInd->getType(), Lane + Part * VF);
4652 Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx);
4653 Value *SclrGep = II.transform(Builder, GlobalIdx, PSE.getSE(), DL);
4654 SclrGep->setName("next.gep");
4655 Entry[Part][Lane] = SclrGep;
4658 VectorLoopValueMap.initScalar(P, Entry);
4661 case InductionDescriptor::IK_FpInduction: {
4662 assert(P->getType() == II.getStartValue()->getType() &&
4663 "Types must match");
4664 // Handle other induction variables that are now based on the
4666 assert(P != OldInduction && "Primary induction can be integer only");
4668 Value *V = Builder.CreateCast(Instruction::SIToFP, Induction, P->getType());
4669 V = II.transform(Builder, V, PSE.getSE(), DL);
4670 V->setName("fp.offset.idx");
4672 // Now we have scalar op: %fp.offset.idx = StartVal +/- Induction*StepVal
4674 Value *Broadcasted = getBroadcastInstrs(V);
4675 // After broadcasting the induction variable we need to make the vector
4676 // consecutive by adding StepVal*0, StepVal*1, StepVal*2, etc.
4677 Value *StepVal = cast<SCEVUnknown>(II.getStep())->getValue();
4678 VectorParts Entry(UF);
4679 for (unsigned part = 0; part < UF; ++part)
4680 Entry[part] = getStepVector(Broadcasted, VF * part, StepVal,
4681 II.getInductionOpcode());
4682 VectorLoopValueMap.initVector(P, Entry);
4688 /// A helper function for checking whether an integer division-related
4689 /// instruction may divide by zero (in which case it must be predicated if
4690 /// executed conditionally in the scalar code).
4691 /// TODO: It may be worthwhile to generalize and check isKnownNonZero().
4692 /// Non-zero divisors that are non compile-time constants will not be
4693 /// converted into multiplication, so we will still end up scalarizing
4694 /// the division, but can do so w/o predication.
4695 static bool mayDivideByZero(Instruction &I) {
4696 assert((I.getOpcode() == Instruction::UDiv ||
4697 I.getOpcode() == Instruction::SDiv ||
4698 I.getOpcode() == Instruction::URem ||
4699 I.getOpcode() == Instruction::SRem) &&
4700 "Unexpected instruction");
4701 Value *Divisor = I.getOperand(1);
4702 auto *CInt = dyn_cast<ConstantInt>(Divisor);
4703 return !CInt || CInt->isZero();
4706 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
4707 // For each instruction in the old loop.
4708 for (Instruction &I : *BB) {
4710 // If the instruction will become trivially dead when vectorized, we don't
4711 // need to generate it.
4712 if (DeadInstructions.count(&I))
4715 // Scalarize instructions that should remain scalar after vectorization.
4717 !(isa<BranchInst>(&I) || isa<PHINode>(&I) ||
4718 isa<DbgInfoIntrinsic>(&I)) &&
4719 shouldScalarizeInstruction(&I)) {
4720 scalarizeInstruction(&I, Legal->isScalarWithPredication(&I));
4724 switch (I.getOpcode()) {
4725 case Instruction::Br:
4726 // Nothing to do for PHIs and BR, since we already took care of the
4727 // loop control flow instructions.
4729 case Instruction::PHI: {
4730 // Vectorize PHINodes.
4731 widenPHIInstruction(&I, UF, VF, PV);
4735 case Instruction::UDiv:
4736 case Instruction::SDiv:
4737 case Instruction::SRem:
4738 case Instruction::URem:
4739 // Scalarize with predication if this instruction may divide by zero and
4740 // block execution is conditional, otherwise fallthrough.
4741 if (Legal->isScalarWithPredication(&I)) {
4742 scalarizeInstruction(&I, true);
4745 case Instruction::Add:
4746 case Instruction::FAdd:
4747 case Instruction::Sub:
4748 case Instruction::FSub:
4749 case Instruction::Mul:
4750 case Instruction::FMul:
4751 case Instruction::FDiv:
4752 case Instruction::FRem:
4753 case Instruction::Shl:
4754 case Instruction::LShr:
4755 case Instruction::AShr:
4756 case Instruction::And:
4757 case Instruction::Or:
4758 case Instruction::Xor: {
4759 // Just widen binops.
4760 auto *BinOp = cast<BinaryOperator>(&I);
4761 setDebugLocFromInst(Builder, BinOp);
4762 const VectorParts &A = getVectorValue(BinOp->getOperand(0));
4763 const VectorParts &B = getVectorValue(BinOp->getOperand(1));
4765 // Use this vector value for all users of the original instruction.
4766 VectorParts Entry(UF);
4767 for (unsigned Part = 0; Part < UF; ++Part) {
4768 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
4770 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
4771 VecOp->copyIRFlags(BinOp);
4776 VectorLoopValueMap.initVector(&I, Entry);
4777 addMetadata(Entry, BinOp);
4780 case Instruction::Select: {
4782 // If the selector is loop invariant we can create a select
4783 // instruction with a scalar condition. Otherwise, use vector-select.
4784 auto *SE = PSE.getSE();
4785 bool InvariantCond =
4786 SE->isLoopInvariant(PSE.getSCEV(I.getOperand(0)), OrigLoop);
4787 setDebugLocFromInst(Builder, &I);
4789 // The condition can be loop invariant but still defined inside the
4790 // loop. This means that we can't just use the original 'cond' value.
4791 // We have to take the 'vectorized' value and pick the first lane.
4792 // Instcombine will make this a no-op.
4793 const VectorParts &Cond = getVectorValue(I.getOperand(0));
4794 const VectorParts &Op0 = getVectorValue(I.getOperand(1));
4795 const VectorParts &Op1 = getVectorValue(I.getOperand(2));
4797 auto *ScalarCond = getScalarValue(I.getOperand(0), 0, 0);
4799 VectorParts Entry(UF);
4800 for (unsigned Part = 0; Part < UF; ++Part) {
4801 Entry[Part] = Builder.CreateSelect(
4802 InvariantCond ? ScalarCond : Cond[Part], Op0[Part], Op1[Part]);
4805 VectorLoopValueMap.initVector(&I, Entry);
4806 addMetadata(Entry, &I);
4810 case Instruction::ICmp:
4811 case Instruction::FCmp: {
4812 // Widen compares. Generate vector compares.
4813 bool FCmp = (I.getOpcode() == Instruction::FCmp);
4814 auto *Cmp = dyn_cast<CmpInst>(&I);
4815 setDebugLocFromInst(Builder, Cmp);
4816 const VectorParts &A = getVectorValue(Cmp->getOperand(0));
4817 const VectorParts &B = getVectorValue(Cmp->getOperand(1));
4818 VectorParts Entry(UF);
4819 for (unsigned Part = 0; Part < UF; ++Part) {
4822 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
4823 cast<FCmpInst>(C)->copyFastMathFlags(Cmp);
4825 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
4830 VectorLoopValueMap.initVector(&I, Entry);
4831 addMetadata(Entry, &I);
4835 case Instruction::Store:
4836 case Instruction::Load:
4837 vectorizeMemoryInstruction(&I);
4839 case Instruction::ZExt:
4840 case Instruction::SExt:
4841 case Instruction::FPToUI:
4842 case Instruction::FPToSI:
4843 case Instruction::FPExt:
4844 case Instruction::PtrToInt:
4845 case Instruction::IntToPtr:
4846 case Instruction::SIToFP:
4847 case Instruction::UIToFP:
4848 case Instruction::Trunc:
4849 case Instruction::FPTrunc:
4850 case Instruction::BitCast: {
4851 auto *CI = dyn_cast<CastInst>(&I);
4852 setDebugLocFromInst(Builder, CI);
4854 // Optimize the special case where the source is a constant integer
4855 // induction variable. Notice that we can only optimize the 'trunc' case
4856 // because (a) FP conversions lose precision, (b) sext/zext may wrap, and
4857 // (c) other casts depend on pointer size.
4858 auto ID = Legal->getInductionVars()->lookup(OldInduction);
4859 if (isa<TruncInst>(CI) && CI->getOperand(0) == OldInduction &&
4860 ID.getConstIntStepValue()) {
4861 widenIntInduction(OldInduction, cast<TruncInst>(CI));
4865 /// Vectorize casts.
4867 (VF == 1) ? CI->getType() : VectorType::get(CI->getType(), VF);
4869 const VectorParts &A = getVectorValue(CI->getOperand(0));
4870 VectorParts Entry(UF);
4871 for (unsigned Part = 0; Part < UF; ++Part)
4872 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
4873 VectorLoopValueMap.initVector(&I, Entry);
4874 addMetadata(Entry, &I);
4878 case Instruction::Call: {
4879 // Ignore dbg intrinsics.
4880 if (isa<DbgInfoIntrinsic>(I))
4882 setDebugLocFromInst(Builder, &I);
4884 Module *M = BB->getParent()->getParent();
4885 auto *CI = cast<CallInst>(&I);
4887 StringRef FnName = CI->getCalledFunction()->getName();
4888 Function *F = CI->getCalledFunction();
4889 Type *RetTy = ToVectorTy(CI->getType(), VF);
4890 SmallVector<Type *, 4> Tys;
4891 for (Value *ArgOperand : CI->arg_operands())
4892 Tys.push_back(ToVectorTy(ArgOperand->getType(), VF));
4894 Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
4895 if (ID && (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end ||
4896 ID == Intrinsic::lifetime_start)) {
4897 scalarizeInstruction(&I);
4900 // The flag shows whether we use Intrinsic or a usual Call for vectorized
4901 // version of the instruction.
4902 // Is it beneficial to perform intrinsic call compared to lib call?
4903 bool NeedToScalarize;
4904 unsigned CallCost = getVectorCallCost(CI, VF, *TTI, TLI, NeedToScalarize);
4905 bool UseVectorIntrinsic =
4906 ID && getVectorIntrinsicCost(CI, VF, *TTI, TLI) <= CallCost;
4907 if (!UseVectorIntrinsic && NeedToScalarize) {
4908 scalarizeInstruction(&I);
4912 VectorParts Entry(UF);
4913 for (unsigned Part = 0; Part < UF; ++Part) {
4914 SmallVector<Value *, 4> Args;
4915 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
4916 Value *Arg = CI->getArgOperand(i);
4917 // Some intrinsics have a scalar argument - don't replace it with a
4919 if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, i)) {
4920 const VectorParts &VectorArg = getVectorValue(CI->getArgOperand(i));
4921 Arg = VectorArg[Part];
4923 Args.push_back(Arg);
4927 if (UseVectorIntrinsic) {
4928 // Use vector version of the intrinsic.
4929 Type *TysForDecl[] = {CI->getType()};
4931 TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF);
4932 VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl);
4934 // Use vector version of the library call.
4935 StringRef VFnName = TLI->getVectorizedFunction(FnName, VF);
4936 assert(!VFnName.empty() && "Vector function name is empty.");
4937 VectorF = M->getFunction(VFnName);
4939 // Generate a declaration
4940 FunctionType *FTy = FunctionType::get(RetTy, Tys, false);
4942 Function::Create(FTy, Function::ExternalLinkage, VFnName, M);
4943 VectorF->copyAttributesFrom(F);
4946 assert(VectorF && "Can't create vector function.");
4948 SmallVector<OperandBundleDef, 1> OpBundles;
4949 CI->getOperandBundlesAsDefs(OpBundles);
4950 CallInst *V = Builder.CreateCall(VectorF, Args, OpBundles);
4952 if (isa<FPMathOperator>(V))
4953 V->copyFastMathFlags(CI);
4958 VectorLoopValueMap.initVector(&I, Entry);
4959 addMetadata(Entry, &I);
4964 // All other instructions are unsupported. Scalarize them.
4965 scalarizeInstruction(&I);
4968 } // end of for_each instr.
4971 void InnerLoopVectorizer::updateAnalysis() {
4972 // Forget the original basic block.
4973 PSE.getSE()->forgetLoop(OrigLoop);
4975 // Update the dominator tree information.
4976 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
4977 "Entry does not dominate exit.");
4979 // We don't predicate stores by this point, so the vector body should be a
4981 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
4983 DT->addNewBlock(LoopMiddleBlock, LoopVectorBody);
4984 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
4985 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
4986 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
4988 DEBUG(DT->verifyDomTree());
4991 /// \brief Check whether it is safe to if-convert this phi node.
4993 /// Phi nodes with constant expressions that can trap are not safe to if
4995 static bool canIfConvertPHINodes(BasicBlock *BB) {
4996 for (Instruction &I : *BB) {
4997 auto *Phi = dyn_cast<PHINode>(&I);
5000 for (Value *V : Phi->incoming_values())
5001 if (auto *C = dyn_cast<Constant>(V))
5008 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
5009 if (!EnableIfConversion) {
5010 ORE->emit(createMissedAnalysis("IfConversionDisabled")
5011 << "if-conversion is disabled");
5015 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
5017 // A list of pointers that we can safely read and write to.
5018 SmallPtrSet<Value *, 8> SafePointes;
5020 // Collect safe addresses.
5021 for (BasicBlock *BB : TheLoop->blocks()) {
5022 if (blockNeedsPredication(BB))
5025 for (Instruction &I : *BB)
5026 if (auto *Ptr = getPointerOperand(&I))
5027 SafePointes.insert(Ptr);
5030 // Collect the blocks that need predication.
5031 BasicBlock *Header = TheLoop->getHeader();
5032 for (BasicBlock *BB : TheLoop->blocks()) {
5033 // We don't support switch statements inside loops.
5034 if (!isa<BranchInst>(BB->getTerminator())) {
5035 ORE->emit(createMissedAnalysis("LoopContainsSwitch", BB->getTerminator())
5036 << "loop contains a switch statement");
5040 // We must be able to predicate all blocks that need to be predicated.
5041 if (blockNeedsPredication(BB)) {
5042 if (!blockCanBePredicated(BB, SafePointes)) {
5043 ORE->emit(createMissedAnalysis("NoCFGForSelect", BB->getTerminator())
5044 << "control flow cannot be substituted for a select");
5047 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
5048 ORE->emit(createMissedAnalysis("NoCFGForSelect", BB->getTerminator())
5049 << "control flow cannot be substituted for a select");
5054 // We can if-convert this loop.
5058 bool LoopVectorizationLegality::canVectorize() {
5059 // We must have a loop in canonical form. Loops with indirectbr in them cannot
5060 // be canonicalized.
5061 if (!TheLoop->getLoopPreheader()) {
5062 ORE->emit(createMissedAnalysis("CFGNotUnderstood")
5063 << "loop control flow is not understood by vectorizer");
5067 // FIXME: The code is currently dead, since the loop gets sent to
5068 // LoopVectorizationLegality is already an innermost loop.
5070 // We can only vectorize innermost loops.
5071 if (!TheLoop->empty()) {
5072 ORE->emit(createMissedAnalysis("NotInnermostLoop")
5073 << "loop is not the innermost loop");
5077 // We must have a single backedge.
5078 if (TheLoop->getNumBackEdges() != 1) {
5079 ORE->emit(createMissedAnalysis("CFGNotUnderstood")
5080 << "loop control flow is not understood by vectorizer");
5084 // We must have a single exiting block.
5085 if (!TheLoop->getExitingBlock()) {
5086 ORE->emit(createMissedAnalysis("CFGNotUnderstood")
5087 << "loop control flow is not understood by vectorizer");
5091 // We only handle bottom-tested loops, i.e. loop in which the condition is
5092 // checked at the end of each iteration. With that we can assume that all
5093 // instructions in the loop are executed the same number of times.
5094 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
5095 ORE->emit(createMissedAnalysis("CFGNotUnderstood")
5096 << "loop control flow is not understood by vectorizer");
5100 // We need to have a loop header.
5101 DEBUG(dbgs() << "LV: Found a loop: " << TheLoop->getHeader()->getName()
5104 // Check if we can if-convert non-single-bb loops.
5105 unsigned NumBlocks = TheLoop->getNumBlocks();
5106 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
5107 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
5111 // ScalarEvolution needs to be able to find the exit count.
5112 const SCEV *ExitCount = PSE.getBackedgeTakenCount();
5113 if (ExitCount == PSE.getSE()->getCouldNotCompute()) {
5114 ORE->emit(createMissedAnalysis("CantComputeNumberOfIterations")
5115 << "could not determine number of loop iterations");
5116 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
5120 // Check if we can vectorize the instructions and CFG in this loop.
5121 if (!canVectorizeInstrs()) {
5122 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
5126 // Go over each instruction and look at memory deps.
5127 if (!canVectorizeMemory()) {
5128 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
5132 DEBUG(dbgs() << "LV: We can vectorize this loop"
5133 << (LAI->getRuntimePointerChecking()->Need
5134 ? " (with a runtime bound check)"
5138 bool UseInterleaved = TTI->enableInterleavedAccessVectorization();
5140 // If an override option has been passed in for interleaved accesses, use it.
5141 if (EnableInterleavedMemAccesses.getNumOccurrences() > 0)
5142 UseInterleaved = EnableInterleavedMemAccesses;
5144 // Analyze interleaved memory accesses.
5146 InterleaveInfo.analyzeInterleaving(*getSymbolicStrides());
5148 // Collect all instructions that are known to be uniform after vectorization.
5149 collectLoopUniforms();
5151 // Collect all instructions that are known to be scalar after vectorization.
5152 collectLoopScalars();
5154 unsigned SCEVThreshold = VectorizeSCEVCheckThreshold;
5155 if (Hints->getForce() == LoopVectorizeHints::FK_Enabled)
5156 SCEVThreshold = PragmaVectorizeSCEVCheckThreshold;
5158 if (PSE.getUnionPredicate().getComplexity() > SCEVThreshold) {
5159 ORE->emit(createMissedAnalysis("TooManySCEVRunTimeChecks")
5160 << "Too many SCEV assumptions need to be made and checked "
5162 DEBUG(dbgs() << "LV: Too many SCEV checks needed.\n");
5166 // Okay! We can vectorize. At this point we don't have any other mem analysis
5167 // which may limit our maximum vectorization factor, so just return true with
5172 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
5173 if (Ty->isPointerTy())
5174 return DL.getIntPtrType(Ty);
5176 // It is possible that char's or short's overflow when we ask for the loop's
5177 // trip count, work around this by changing the type size.
5178 if (Ty->getScalarSizeInBits() < 32)
5179 return Type::getInt32Ty(Ty->getContext());
5184 static Type *getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
5185 Ty0 = convertPointerToIntegerType(DL, Ty0);
5186 Ty1 = convertPointerToIntegerType(DL, Ty1);
5187 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
5192 /// \brief Check that the instruction has outside loop users and is not an
5193 /// identified reduction variable.
5194 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
5195 SmallPtrSetImpl<Value *> &AllowedExit) {
5196 // Reduction and Induction instructions are allowed to have exit users. All
5197 // other instructions must not have external users.
5198 if (!AllowedExit.count(Inst))
5199 // Check that all of the users of the loop are inside the BB.
5200 for (User *U : Inst->users()) {
5201 Instruction *UI = cast<Instruction>(U);
5202 // This user may be a reduction exit value.
5203 if (!TheLoop->contains(UI)) {
5204 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
5211 void LoopVectorizationLegality::addInductionPhi(
5212 PHINode *Phi, const InductionDescriptor &ID,
5213 SmallPtrSetImpl<Value *> &AllowedExit) {
5214 Inductions[Phi] = ID;
5215 Type *PhiTy = Phi->getType();
5216 const DataLayout &DL = Phi->getModule()->getDataLayout();
5218 // Get the widest type.
5219 if (!PhiTy->isFloatingPointTy()) {
5221 WidestIndTy = convertPointerToIntegerType(DL, PhiTy);
5223 WidestIndTy = getWiderType(DL, PhiTy, WidestIndTy);
5226 // Int inductions are special because we only allow one IV.
5227 if (ID.getKind() == InductionDescriptor::IK_IntInduction &&
5228 ID.getConstIntStepValue() &&
5229 ID.getConstIntStepValue()->isOne() &&
5230 isa<Constant>(ID.getStartValue()) &&
5231 cast<Constant>(ID.getStartValue())->isNullValue()) {
5233 // Use the phi node with the widest type as induction. Use the last
5234 // one if there are multiple (no good reason for doing this other
5235 // than it is expedient). We've checked that it begins at zero and
5236 // steps by one, so this is a canonical induction variable.
5237 if (!Induction || PhiTy == WidestIndTy)
5241 // Both the PHI node itself, and the "post-increment" value feeding
5242 // back into the PHI node may have external users.
5243 AllowedExit.insert(Phi);
5244 AllowedExit.insert(Phi->getIncomingValueForBlock(TheLoop->getLoopLatch()));
5246 DEBUG(dbgs() << "LV: Found an induction variable.\n");
5250 bool LoopVectorizationLegality::canVectorizeInstrs() {
5251 BasicBlock *Header = TheLoop->getHeader();
5253 // Look for the attribute signaling the absence of NaNs.
5254 Function &F = *Header->getParent();
5256 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
5258 // For each block in the loop.
5259 for (BasicBlock *BB : TheLoop->blocks()) {
5260 // Scan the instructions in the block and look for hazards.
5261 for (Instruction &I : *BB) {
5262 if (auto *Phi = dyn_cast<PHINode>(&I)) {
5263 Type *PhiTy = Phi->getType();
5264 // Check that this PHI type is allowed.
5265 if (!PhiTy->isIntegerTy() && !PhiTy->isFloatingPointTy() &&
5266 !PhiTy->isPointerTy()) {
5267 ORE->emit(createMissedAnalysis("CFGNotUnderstood", Phi)
5268 << "loop control flow is not understood by vectorizer");
5269 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
5273 // If this PHINode is not in the header block, then we know that we
5274 // can convert it to select during if-conversion. No need to check if
5275 // the PHIs in this block are induction or reduction variables.
5277 // Check that this instruction has no outside users or is an
5278 // identified reduction value with an outside user.
5279 if (!hasOutsideLoopUser(TheLoop, Phi, AllowedExit))
5281 ORE->emit(createMissedAnalysis("NeitherInductionNorReduction", Phi)
5282 << "value could not be identified as "
5283 "an induction or reduction variable");
5287 // We only allow if-converted PHIs with exactly two incoming values.
5288 if (Phi->getNumIncomingValues() != 2) {
5289 ORE->emit(createMissedAnalysis("CFGNotUnderstood", Phi)
5290 << "control flow not understood by vectorizer");
5291 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
5295 RecurrenceDescriptor RedDes;
5296 if (RecurrenceDescriptor::isReductionPHI(Phi, TheLoop, RedDes)) {
5297 if (RedDes.hasUnsafeAlgebra())
5298 Requirements->addUnsafeAlgebraInst(RedDes.getUnsafeAlgebraInst());
5299 AllowedExit.insert(RedDes.getLoopExitInstr());
5300 Reductions[Phi] = RedDes;
5304 InductionDescriptor ID;
5305 if (InductionDescriptor::isInductionPHI(Phi, TheLoop, PSE, ID)) {
5306 addInductionPhi(Phi, ID, AllowedExit);
5307 if (ID.hasUnsafeAlgebra() && !HasFunNoNaNAttr)
5308 Requirements->addUnsafeAlgebraInst(ID.getUnsafeAlgebraInst());
5312 if (RecurrenceDescriptor::isFirstOrderRecurrence(Phi, TheLoop, DT)) {
5313 FirstOrderRecurrences.insert(Phi);
5317 // As a last resort, coerce the PHI to a AddRec expression
5318 // and re-try classifying it a an induction PHI.
5319 if (InductionDescriptor::isInductionPHI(Phi, TheLoop, PSE, ID, true)) {
5320 addInductionPhi(Phi, ID, AllowedExit);
5324 ORE->emit(createMissedAnalysis("NonReductionValueUsedOutsideLoop", Phi)
5325 << "value that could not be identified as "
5326 "reduction is used outside the loop");
5327 DEBUG(dbgs() << "LV: Found an unidentified PHI." << *Phi << "\n");
5329 } // end of PHI handling
5331 // We handle calls that:
5332 // * Are debug info intrinsics.
5333 // * Have a mapping to an IR intrinsic.
5334 // * Have a vector version available.
5335 auto *CI = dyn_cast<CallInst>(&I);
5336 if (CI && !getVectorIntrinsicIDForCall(CI, TLI) &&
5337 !isa<DbgInfoIntrinsic>(CI) &&
5338 !(CI->getCalledFunction() && TLI &&
5339 TLI->isFunctionVectorizable(CI->getCalledFunction()->getName()))) {
5340 ORE->emit(createMissedAnalysis("CantVectorizeCall", CI)
5341 << "call instruction cannot be vectorized");
5342 DEBUG(dbgs() << "LV: Found a non-intrinsic, non-libfunc callsite.\n");
5346 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
5347 // second argument is the same (i.e. loop invariant)
5348 if (CI && hasVectorInstrinsicScalarOpd(
5349 getVectorIntrinsicIDForCall(CI, TLI), 1)) {
5350 auto *SE = PSE.getSE();
5351 if (!SE->isLoopInvariant(PSE.getSCEV(CI->getOperand(1)), TheLoop)) {
5352 ORE->emit(createMissedAnalysis("CantVectorizeIntrinsic", CI)
5353 << "intrinsic instruction cannot be vectorized");
5354 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
5359 // Check that the instruction return type is vectorizable.
5360 // Also, we can't vectorize extractelement instructions.
5361 if ((!VectorType::isValidElementType(I.getType()) &&
5362 !I.getType()->isVoidTy()) ||
5363 isa<ExtractElementInst>(I)) {
5364 ORE->emit(createMissedAnalysis("CantVectorizeInstructionReturnType", &I)
5365 << "instruction return type cannot be vectorized");
5366 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
5370 // Check that the stored type is vectorizable.
5371 if (auto *ST = dyn_cast<StoreInst>(&I)) {
5372 Type *T = ST->getValueOperand()->getType();
5373 if (!VectorType::isValidElementType(T)) {
5374 ORE->emit(createMissedAnalysis("CantVectorizeStore", ST)
5375 << "store instruction cannot be vectorized");
5379 // FP instructions can allow unsafe algebra, thus vectorizable by
5380 // non-IEEE-754 compliant SIMD units.
5381 // This applies to floating-point math operations and calls, not memory
5382 // operations, shuffles, or casts, as they don't change precision or
5384 } else if (I.getType()->isFloatingPointTy() && (CI || I.isBinaryOp()) &&
5385 !I.hasUnsafeAlgebra()) {
5386 DEBUG(dbgs() << "LV: Found FP op with unsafe algebra.\n");
5387 Hints->setPotentiallyUnsafe();
5390 // Reduction instructions are allowed to have exit users.
5391 // All other instructions must not have external users.
5392 if (hasOutsideLoopUser(TheLoop, &I, AllowedExit)) {
5393 ORE->emit(createMissedAnalysis("ValueUsedOutsideLoop", &I)
5394 << "value cannot be used outside the loop");
5402 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
5403 if (Inductions.empty()) {
5404 ORE->emit(createMissedAnalysis("NoInductionVariable")
5405 << "loop induction variable could not be identified");
5410 // Now we know the widest induction type, check if our found induction
5411 // is the same size. If it's not, unset it here and InnerLoopVectorizer
5412 // will create another.
5413 if (Induction && WidestIndTy != Induction->getType())
5414 Induction = nullptr;
5419 void LoopVectorizationLegality::collectLoopScalars() {
5421 // If an instruction is uniform after vectorization, it will remain scalar.
5422 Scalars.insert(Uniforms.begin(), Uniforms.end());
5424 // Collect the getelementptr instructions that will not be vectorized. A
5425 // getelementptr instruction is only vectorized if it is used for a legal
5426 // gather or scatter operation.
5427 for (auto *BB : TheLoop->blocks())
5428 for (auto &I : *BB) {
5429 if (auto *GEP = dyn_cast<GetElementPtrInst>(&I)) {
5430 Scalars.insert(GEP);
5433 auto *Ptr = getPointerOperand(&I);
5436 auto *GEP = getGEPInstruction(Ptr);
5437 if (GEP && isLegalGatherOrScatter(&I))
5441 // An induction variable will remain scalar if all users of the induction
5442 // variable and induction variable update remain scalar.
5443 auto *Latch = TheLoop->getLoopLatch();
5444 for (auto &Induction : *getInductionVars()) {
5445 auto *Ind = Induction.first;
5446 auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch));
5448 // Determine if all users of the induction variable are scalar after
5450 auto ScalarInd = all_of(Ind->users(), [&](User *U) -> bool {
5451 auto *I = cast<Instruction>(U);
5452 return I == IndUpdate || !TheLoop->contains(I) || Scalars.count(I);
5457 // Determine if all users of the induction variable update instruction are
5458 // scalar after vectorization.
5459 auto ScalarIndUpdate = all_of(IndUpdate->users(), [&](User *U) -> bool {
5460 auto *I = cast<Instruction>(U);
5461 return I == Ind || !TheLoop->contains(I) || Scalars.count(I);
5463 if (!ScalarIndUpdate)
5466 // The induction variable and its update instruction will remain scalar.
5467 Scalars.insert(Ind);
5468 Scalars.insert(IndUpdate);
5472 bool LoopVectorizationLegality::hasConsecutiveLikePtrOperand(Instruction *I) {
5473 if (isAccessInterleaved(I))
5475 if (auto *Ptr = getPointerOperand(I))
5476 return isConsecutivePtr(Ptr);
5480 bool LoopVectorizationLegality::isScalarWithPredication(Instruction *I) {
5481 if (!blockNeedsPredication(I->getParent()))
5483 switch(I->getOpcode()) {
5486 case Instruction::Store:
5487 return !isMaskRequired(I);
5488 case Instruction::UDiv:
5489 case Instruction::SDiv:
5490 case Instruction::SRem:
5491 case Instruction::URem:
5492 return mayDivideByZero(*I);
5497 bool LoopVectorizationLegality::memoryInstructionMustBeScalarized(
5498 Instruction *I, unsigned VF) {
5500 // If the memory instruction is in an interleaved group, it will be
5501 // vectorized and its pointer will remain uniform.
5502 if (isAccessInterleaved(I))
5505 // Get and ensure we have a valid memory instruction.
5506 LoadInst *LI = dyn_cast<LoadInst>(I);
5507 StoreInst *SI = dyn_cast<StoreInst>(I);
5508 assert((LI || SI) && "Invalid memory instruction");
5510 // If the pointer operand is uniform (loop invariant), the memory instruction
5511 // will be scalarized.
5512 auto *Ptr = getPointerOperand(I);
5513 if (LI && isUniform(Ptr))
5516 // If the pointer operand is non-consecutive and neither a gather nor a
5517 // scatter operation is legal, the memory instruction will be scalarized.
5518 if (!isConsecutivePtr(Ptr) && !isLegalGatherOrScatter(I))
5521 // If the instruction is a store located in a predicated block, it will be
5523 if (isScalarWithPredication(I))
5526 // If the instruction's allocated size doesn't equal it's type size, it
5527 // requires padding and will be scalarized.
5528 auto &DL = I->getModule()->getDataLayout();
5529 auto *ScalarTy = LI ? LI->getType() : SI->getValueOperand()->getType();
5530 if (hasIrregularType(ScalarTy, DL, VF))
5533 // Otherwise, the memory instruction should be vectorized if the rest of the
5538 void LoopVectorizationLegality::collectLoopUniforms() {
5539 // We now know that the loop is vectorizable!
5540 // Collect instructions inside the loop that will remain uniform after
5543 // Global values, params and instructions outside of current loop are out of
5545 auto isOutOfScope = [&](Value *V) -> bool {
5546 Instruction *I = dyn_cast<Instruction>(V);
5547 return (!I || !TheLoop->contains(I));
5550 SetVector<Instruction *> Worklist;
5551 BasicBlock *Latch = TheLoop->getLoopLatch();
5553 // Start with the conditional branch. If the branch condition is an
5554 // instruction contained in the loop that is only used by the branch, it is
5556 auto *Cmp = dyn_cast<Instruction>(Latch->getTerminator()->getOperand(0));
5557 if (Cmp && TheLoop->contains(Cmp) && Cmp->hasOneUse()) {
5558 Worklist.insert(Cmp);
5559 DEBUG(dbgs() << "LV: Found uniform instruction: " << *Cmp << "\n");
5562 // Holds consecutive and consecutive-like pointers. Consecutive-like pointers
5563 // are pointers that are treated like consecutive pointers during
5564 // vectorization. The pointer operands of interleaved accesses are an
5566 SmallSetVector<Instruction *, 8> ConsecutiveLikePtrs;
5568 // Holds pointer operands of instructions that are possibly non-uniform.
5569 SmallPtrSet<Instruction *, 8> PossibleNonUniformPtrs;
5571 // Iterate over the instructions in the loop, and collect all
5572 // consecutive-like pointer operands in ConsecutiveLikePtrs. If it's possible
5573 // that a consecutive-like pointer operand will be scalarized, we collect it
5574 // in PossibleNonUniformPtrs instead. We use two sets here because a single
5575 // getelementptr instruction can be used by both vectorized and scalarized
5576 // memory instructions. For example, if a loop loads and stores from the same
5577 // location, but the store is conditional, the store will be scalarized, and
5578 // the getelementptr won't remain uniform.
5579 for (auto *BB : TheLoop->blocks())
5580 for (auto &I : *BB) {
5582 // If there's no pointer operand, there's nothing to do.
5583 auto *Ptr = dyn_cast_or_null<Instruction>(getPointerOperand(&I));
5587 // True if all users of Ptr are memory accesses that have Ptr as their
5589 auto UsersAreMemAccesses = all_of(Ptr->users(), [&](User *U) -> bool {
5590 return getPointerOperand(U) == Ptr;
5593 // Ensure the memory instruction will not be scalarized, making its
5594 // pointer operand non-uniform. If the pointer operand is used by some
5595 // instruction other than a memory access, we're not going to check if
5596 // that other instruction may be scalarized here. Thus, conservatively
5597 // assume the pointer operand may be non-uniform.
5598 if (!UsersAreMemAccesses || memoryInstructionMustBeScalarized(&I))
5599 PossibleNonUniformPtrs.insert(Ptr);
5601 // If the memory instruction will be vectorized and its pointer operand
5602 // is consecutive-like, the pointer operand should remain uniform.
5603 else if (hasConsecutiveLikePtrOperand(&I))
5604 ConsecutiveLikePtrs.insert(Ptr);
5607 // Add to the Worklist all consecutive and consecutive-like pointers that
5608 // aren't also identified as possibly non-uniform.
5609 for (auto *V : ConsecutiveLikePtrs)
5610 if (!PossibleNonUniformPtrs.count(V)) {
5611 DEBUG(dbgs() << "LV: Found uniform instruction: " << *V << "\n");
5615 // Expand Worklist in topological order: whenever a new instruction
5616 // is added , its users should be either already inside Worklist, or
5617 // out of scope. It ensures a uniform instruction will only be used
5618 // by uniform instructions or out of scope instructions.
5620 while (idx != Worklist.size()) {
5621 Instruction *I = Worklist[idx++];
5623 for (auto OV : I->operand_values()) {
5624 if (isOutOfScope(OV))
5626 auto *OI = cast<Instruction>(OV);
5627 if (all_of(OI->users(), [&](User *U) -> bool {
5628 return isOutOfScope(U) || Worklist.count(cast<Instruction>(U));
5630 Worklist.insert(OI);
5631 DEBUG(dbgs() << "LV: Found uniform instruction: " << *OI << "\n");
5636 // Returns true if Ptr is the pointer operand of a memory access instruction
5637 // I, and I is known to not require scalarization.
5638 auto isVectorizedMemAccessUse = [&](Instruction *I, Value *Ptr) -> bool {
5639 return getPointerOperand(I) == Ptr && !memoryInstructionMustBeScalarized(I);
5642 // For an instruction to be added into Worklist above, all its users inside
5643 // the loop should also be in Worklist. However, this condition cannot be
5644 // true for phi nodes that form a cyclic dependence. We must process phi
5645 // nodes separately. An induction variable will remain uniform if all users
5646 // of the induction variable and induction variable update remain uniform.
5647 // The code below handles both pointer and non-pointer induction variables.
5648 for (auto &Induction : Inductions) {
5649 auto *Ind = Induction.first;
5650 auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch));
5652 // Determine if all users of the induction variable are uniform after
5654 auto UniformInd = all_of(Ind->users(), [&](User *U) -> bool {
5655 auto *I = cast<Instruction>(U);
5656 return I == IndUpdate || !TheLoop->contains(I) || Worklist.count(I) ||
5657 isVectorizedMemAccessUse(I, Ind);
5662 // Determine if all users of the induction variable update instruction are
5663 // uniform after vectorization.
5664 auto UniformIndUpdate = all_of(IndUpdate->users(), [&](User *U) -> bool {
5665 auto *I = cast<Instruction>(U);
5666 return I == Ind || !TheLoop->contains(I) || Worklist.count(I) ||
5667 isVectorizedMemAccessUse(I, IndUpdate);
5669 if (!UniformIndUpdate)
5672 // The induction variable and its update instruction will remain uniform.
5673 Worklist.insert(Ind);
5674 Worklist.insert(IndUpdate);
5675 DEBUG(dbgs() << "LV: Found uniform instruction: " << *Ind << "\n");
5676 DEBUG(dbgs() << "LV: Found uniform instruction: " << *IndUpdate << "\n");
5679 Uniforms.insert(Worklist.begin(), Worklist.end());
5682 bool LoopVectorizationLegality::canVectorizeMemory() {
5683 LAI = &(*GetLAA)(*TheLoop);
5684 InterleaveInfo.setLAI(LAI);
5685 const OptimizationRemarkAnalysis *LAR = LAI->getReport();
5687 OptimizationRemarkAnalysis VR(Hints->vectorizeAnalysisPassName(),
5688 "loop not vectorized: ", *LAR);
5691 if (!LAI->canVectorizeMemory())
5694 if (LAI->hasStoreToLoopInvariantAddress()) {
5695 ORE->emit(createMissedAnalysis("CantVectorizeStoreToLoopInvariantAddress")
5696 << "write to a loop invariant address could not be vectorized");
5697 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
5701 Requirements->addRuntimePointerChecks(LAI->getNumRuntimePointerChecks());
5702 PSE.addPredicate(LAI->getPSE().getUnionPredicate());
5707 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
5708 Value *In0 = const_cast<Value *>(V);
5709 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
5713 return Inductions.count(PN);
5716 bool LoopVectorizationLegality::isFirstOrderRecurrence(const PHINode *Phi) {
5717 return FirstOrderRecurrences.count(Phi);
5720 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
5721 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
5724 bool LoopVectorizationLegality::blockCanBePredicated(
5725 BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs) {
5726 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
5728 for (Instruction &I : *BB) {
5729 // Check that we don't have a constant expression that can trap as operand.
5730 for (Value *Operand : I.operands()) {
5731 if (auto *C = dyn_cast<Constant>(Operand))
5735 // We might be able to hoist the load.
5736 if (I.mayReadFromMemory()) {
5737 auto *LI = dyn_cast<LoadInst>(&I);
5740 if (!SafePtrs.count(LI->getPointerOperand())) {
5741 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand()) ||
5742 isLegalMaskedGather(LI->getType())) {
5743 MaskedOp.insert(LI);
5746 // !llvm.mem.parallel_loop_access implies if-conversion safety.
5747 if (IsAnnotatedParallel)
5753 if (I.mayWriteToMemory()) {
5754 auto *SI = dyn_cast<StoreInst>(&I);
5755 // We only support predication of stores in basic blocks with one
5760 // Build a masked store if it is legal for the target.
5761 if (isLegalMaskedStore(SI->getValueOperand()->getType(),
5762 SI->getPointerOperand()) ||
5763 isLegalMaskedScatter(SI->getValueOperand()->getType())) {
5764 MaskedOp.insert(SI);
5768 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
5769 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
5771 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
5772 !isSinglePredecessor)
5782 void InterleavedAccessInfo::collectConstStrideAccesses(
5783 MapVector<Instruction *, StrideDescriptor> &AccessStrideInfo,
5784 const ValueToValueMap &Strides) {
5786 auto &DL = TheLoop->getHeader()->getModule()->getDataLayout();
5788 // Since it's desired that the load/store instructions be maintained in
5789 // "program order" for the interleaved access analysis, we have to visit the
5790 // blocks in the loop in reverse postorder (i.e., in a topological order).
5791 // Such an ordering will ensure that any load/store that may be executed
5792 // before a second load/store will precede the second load/store in
5793 // AccessStrideInfo.
5794 LoopBlocksDFS DFS(TheLoop);
5796 for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO()))
5797 for (auto &I : *BB) {
5798 auto *LI = dyn_cast<LoadInst>(&I);
5799 auto *SI = dyn_cast<StoreInst>(&I);
5803 Value *Ptr = getPointerOperand(&I);
5804 // We don't check wrapping here because we don't know yet if Ptr will be
5805 // part of a full group or a group with gaps. Checking wrapping for all
5806 // pointers (even those that end up in groups with no gaps) will be overly
5807 // conservative. For full groups, wrapping should be ok since if we would
5808 // wrap around the address space we would do a memory access at nullptr
5809 // even without the transformation. The wrapping checks are therefore
5810 // deferred until after we've formed the interleaved groups.
5811 int64_t Stride = getPtrStride(PSE, Ptr, TheLoop, Strides,
5812 /*Assume=*/true, /*ShouldCheckWrap=*/false);
5814 const SCEV *Scev = replaceSymbolicStrideSCEV(PSE, Strides, Ptr);
5815 PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
5816 uint64_t Size = DL.getTypeAllocSize(PtrTy->getElementType());
5818 // An alignment of 0 means target ABI alignment.
5819 unsigned Align = LI ? LI->getAlignment() : SI->getAlignment();
5821 Align = DL.getABITypeAlignment(PtrTy->getElementType());
5823 AccessStrideInfo[&I] = StrideDescriptor(Stride, Scev, Size, Align);
5827 // Analyze interleaved accesses and collect them into interleaved load and
5830 // When generating code for an interleaved load group, we effectively hoist all
5831 // loads in the group to the location of the first load in program order. When
5832 // generating code for an interleaved store group, we sink all stores to the
5833 // location of the last store. This code motion can change the order of load
5834 // and store instructions and may break dependences.
5836 // The code generation strategy mentioned above ensures that we won't violate
5837 // any write-after-read (WAR) dependences.
5839 // E.g., for the WAR dependence: a = A[i]; // (1)
5842 // The store group of (2) is always inserted at or below (2), and the load
5843 // group of (1) is always inserted at or above (1). Thus, the instructions will
5844 // never be reordered. All other dependences are checked to ensure the
5845 // correctness of the instruction reordering.
5847 // The algorithm visits all memory accesses in the loop in bottom-up program
5848 // order. Program order is established by traversing the blocks in the loop in
5849 // reverse postorder when collecting the accesses.
5851 // We visit the memory accesses in bottom-up order because it can simplify the
5852 // construction of store groups in the presence of write-after-write (WAW)
5855 // E.g., for the WAW dependence: A[i] = a; // (1)
5857 // A[i + 1] = c; // (3)
5859 // We will first create a store group with (3) and (2). (1) can't be added to
5860 // this group because it and (2) are dependent. However, (1) can be grouped
5861 // with other accesses that may precede it in program order. Note that a
5862 // bottom-up order does not imply that WAW dependences should not be checked.
5863 void InterleavedAccessInfo::analyzeInterleaving(
5864 const ValueToValueMap &Strides) {
5865 DEBUG(dbgs() << "LV: Analyzing interleaved accesses...\n");
5867 // Holds all accesses with a constant stride.
5868 MapVector<Instruction *, StrideDescriptor> AccessStrideInfo;
5869 collectConstStrideAccesses(AccessStrideInfo, Strides);
5871 if (AccessStrideInfo.empty())
5874 // Collect the dependences in the loop.
5875 collectDependences();
5877 // Holds all interleaved store groups temporarily.
5878 SmallSetVector<InterleaveGroup *, 4> StoreGroups;
5879 // Holds all interleaved load groups temporarily.
5880 SmallSetVector<InterleaveGroup *, 4> LoadGroups;
5882 // Search in bottom-up program order for pairs of accesses (A and B) that can
5883 // form interleaved load or store groups. In the algorithm below, access A
5884 // precedes access B in program order. We initialize a group for B in the
5885 // outer loop of the algorithm, and then in the inner loop, we attempt to
5886 // insert each A into B's group if:
5888 // 1. A and B have the same stride,
5889 // 2. A and B have the same memory object size, and
5890 // 3. A belongs in B's group according to its distance from B.
5892 // Special care is taken to ensure group formation will not break any
5894 for (auto BI = AccessStrideInfo.rbegin(), E = AccessStrideInfo.rend();
5896 Instruction *B = BI->first;
5897 StrideDescriptor DesB = BI->second;
5899 // Initialize a group for B if it has an allowable stride. Even if we don't
5900 // create a group for B, we continue with the bottom-up algorithm to ensure
5901 // we don't break any of B's dependences.
5902 InterleaveGroup *Group = nullptr;
5903 if (isStrided(DesB.Stride)) {
5904 Group = getInterleaveGroup(B);
5906 DEBUG(dbgs() << "LV: Creating an interleave group with:" << *B << '\n');
5907 Group = createInterleaveGroup(B, DesB.Stride, DesB.Align);
5909 if (B->mayWriteToMemory())
5910 StoreGroups.insert(Group);
5912 LoadGroups.insert(Group);
5915 for (auto AI = std::next(BI); AI != E; ++AI) {
5916 Instruction *A = AI->first;
5917 StrideDescriptor DesA = AI->second;
5919 // Our code motion strategy implies that we can't have dependences
5920 // between accesses in an interleaved group and other accesses located
5921 // between the first and last member of the group. Note that this also
5922 // means that a group can't have more than one member at a given offset.
5923 // The accesses in a group can have dependences with other accesses, but
5924 // we must ensure we don't extend the boundaries of the group such that
5925 // we encompass those dependent accesses.
5927 // For example, assume we have the sequence of accesses shown below in a
5930 // (1, 2) is a group | A[i] = a; // (1)
5931 // | A[i-1] = b; // (2) |
5932 // A[i-3] = c; // (3)
5933 // A[i] = d; // (4) | (2, 4) is not a group
5935 // Because accesses (2) and (3) are dependent, we can group (2) with (1)
5936 // but not with (4). If we did, the dependent access (3) would be within
5937 // the boundaries of the (2, 4) group.
5938 if (!canReorderMemAccessesForInterleavedGroups(&*AI, &*BI)) {
5940 // If a dependence exists and A is already in a group, we know that A
5941 // must be a store since A precedes B and WAR dependences are allowed.
5942 // Thus, A would be sunk below B. We release A's group to prevent this
5943 // illegal code motion. A will then be free to form another group with
5944 // instructions that precede it.
5945 if (isInterleaved(A)) {
5946 InterleaveGroup *StoreGroup = getInterleaveGroup(A);
5947 StoreGroups.remove(StoreGroup);
5948 releaseGroup(StoreGroup);
5951 // If a dependence exists and A is not already in a group (or it was
5952 // and we just released it), B might be hoisted above A (if B is a
5953 // load) or another store might be sunk below A (if B is a store). In
5954 // either case, we can't add additional instructions to B's group. B
5955 // will only form a group with instructions that it precedes.
5959 // At this point, we've checked for illegal code motion. If either A or B
5960 // isn't strided, there's nothing left to do.
5961 if (!isStrided(DesA.Stride) || !isStrided(DesB.Stride))
5964 // Ignore A if it's already in a group or isn't the same kind of memory
5966 if (isInterleaved(A) || A->mayReadFromMemory() != B->mayReadFromMemory())
5969 // Check rules 1 and 2. Ignore A if its stride or size is different from
5971 if (DesA.Stride != DesB.Stride || DesA.Size != DesB.Size)
5974 // Calculate the distance from A to B.
5975 const SCEVConstant *DistToB = dyn_cast<SCEVConstant>(
5976 PSE.getSE()->getMinusSCEV(DesA.Scev, DesB.Scev));
5979 int64_t DistanceToB = DistToB->getAPInt().getSExtValue();
5981 // Check rule 3. Ignore A if its distance to B is not a multiple of the
5983 if (DistanceToB % static_cast<int64_t>(DesB.Size))
5986 // Ignore A if either A or B is in a predicated block. Although we
5987 // currently prevent group formation for predicated accesses, we may be
5988 // able to relax this limitation in the future once we handle more
5989 // complicated blocks.
5990 if (isPredicated(A->getParent()) || isPredicated(B->getParent()))
5993 // The index of A is the index of B plus A's distance to B in multiples
5996 Group->getIndex(B) + DistanceToB / static_cast<int64_t>(DesB.Size);
5998 // Try to insert A into B's group.
5999 if (Group->insertMember(A, IndexA, DesA.Align)) {
6000 DEBUG(dbgs() << "LV: Inserted:" << *A << '\n'
6001 << " into the interleave group with" << *B << '\n');
6002 InterleaveGroupMap[A] = Group;
6004 // Set the first load in program order as the insert position.
6005 if (A->mayReadFromMemory())
6006 Group->setInsertPos(A);
6008 } // Iteration over A accesses.
6009 } // Iteration over B accesses.
6011 // Remove interleaved store groups with gaps.
6012 for (InterleaveGroup *Group : StoreGroups)
6013 if (Group->getNumMembers() != Group->getFactor())
6014 releaseGroup(Group);
6016 // Remove interleaved groups with gaps (currently only loads) whose memory
6017 // accesses may wrap around. We have to revisit the getPtrStride analysis,
6018 // this time with ShouldCheckWrap=true, since collectConstStrideAccesses does
6019 // not check wrapping (see documentation there).
6020 // FORNOW we use Assume=false;
6021 // TODO: Change to Assume=true but making sure we don't exceed the threshold
6022 // of runtime SCEV assumptions checks (thereby potentially failing to
6023 // vectorize altogether).
6024 // Additional optional optimizations:
6025 // TODO: If we are peeling the loop and we know that the first pointer doesn't
6026 // wrap then we can deduce that all pointers in the group don't wrap.
6027 // This means that we can forcefully peel the loop in order to only have to
6028 // check the first pointer for no-wrap. When we'll change to use Assume=true
6029 // we'll only need at most one runtime check per interleaved group.
6031 for (InterleaveGroup *Group : LoadGroups) {
6033 // Case 1: A full group. Can Skip the checks; For full groups, if the wide
6034 // load would wrap around the address space we would do a memory access at
6035 // nullptr even without the transformation.
6036 if (Group->getNumMembers() == Group->getFactor())
6039 // Case 2: If first and last members of the group don't wrap this implies
6040 // that all the pointers in the group don't wrap.
6041 // So we check only group member 0 (which is always guaranteed to exist),
6042 // and group member Factor - 1; If the latter doesn't exist we rely on
6043 // peeling (if it is a non-reveresed accsess -- see Case 3).
6044 Value *FirstMemberPtr = getPointerOperand(Group->getMember(0));
6045 if (!getPtrStride(PSE, FirstMemberPtr, TheLoop, Strides, /*Assume=*/false,
6046 /*ShouldCheckWrap=*/true)) {
6047 DEBUG(dbgs() << "LV: Invalidate candidate interleaved group due to "
6048 "first group member potentially pointer-wrapping.\n");
6049 releaseGroup(Group);
6052 Instruction *LastMember = Group->getMember(Group->getFactor() - 1);
6054 Value *LastMemberPtr = getPointerOperand(LastMember);
6055 if (!getPtrStride(PSE, LastMemberPtr, TheLoop, Strides, /*Assume=*/false,
6056 /*ShouldCheckWrap=*/true)) {
6057 DEBUG(dbgs() << "LV: Invalidate candidate interleaved group due to "
6058 "last group member potentially pointer-wrapping.\n");
6059 releaseGroup(Group);
6063 // Case 3: A non-reversed interleaved load group with gaps: We need
6064 // to execute at least one scalar epilogue iteration. This will ensure
6065 // we don't speculatively access memory out-of-bounds. We only need
6066 // to look for a member at index factor - 1, since every group must have
6067 // a member at index zero.
6068 if (Group->isReverse()) {
6069 releaseGroup(Group);
6072 DEBUG(dbgs() << "LV: Interleaved group requires epilogue iteration.\n");
6073 RequiresScalarEpilogue = true;
6078 LoopVectorizationCostModel::VectorizationFactor
6079 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
6080 // Width 1 means no vectorize
6081 VectorizationFactor Factor = {1U, 0U};
6082 if (OptForSize && Legal->getRuntimePointerChecking()->Need) {
6083 ORE->emit(createMissedAnalysis("CantVersionLoopWithOptForSize")
6084 << "runtime pointer checks needed. Enable vectorization of this "
6085 "loop with '#pragma clang loop vectorize(enable)' when "
6086 "compiling with -Os/-Oz");
6088 << "LV: Aborting. Runtime ptr check is required with -Os/-Oz.\n");
6092 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
6093 ORE->emit(createMissedAnalysis("ConditionalStore")
6094 << "store that is conditionally executed prevents vectorization");
6095 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
6099 MinBWs = computeMinimumValueSizes(TheLoop->getBlocks(), *DB, &TTI);
6100 unsigned SmallestType, WidestType;
6101 std::tie(SmallestType, WidestType) = getSmallestAndWidestTypes();
6102 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
6103 unsigned MaxSafeDepDist = -1U;
6105 // Get the maximum safe dependence distance in bits computed by LAA. If the
6106 // loop contains any interleaved accesses, we divide the dependence distance
6107 // by the maximum interleave factor of all interleaved groups. Note that
6108 // although the division ensures correctness, this is a fairly conservative
6109 // computation because the maximum distance computed by LAA may not involve
6110 // any of the interleaved accesses.
6111 if (Legal->getMaxSafeDepDistBytes() != -1U)
6113 Legal->getMaxSafeDepDistBytes() * 8 / Legal->getMaxInterleaveFactor();
6116 ((WidestRegister < MaxSafeDepDist) ? WidestRegister : MaxSafeDepDist);
6117 unsigned MaxVectorSize = WidestRegister / WidestType;
6119 DEBUG(dbgs() << "LV: The Smallest and Widest types: " << SmallestType << " / "
6120 << WidestType << " bits.\n");
6121 DEBUG(dbgs() << "LV: The Widest register is: " << WidestRegister
6124 if (MaxVectorSize == 0) {
6125 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
6129 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
6130 " into one vector!");
6132 unsigned VF = MaxVectorSize;
6133 if (MaximizeBandwidth && !OptForSize) {
6134 // Collect all viable vectorization factors.
6135 SmallVector<unsigned, 8> VFs;
6136 unsigned NewMaxVectorSize = WidestRegister / SmallestType;
6137 for (unsigned VS = MaxVectorSize; VS <= NewMaxVectorSize; VS *= 2)
6140 // For each VF calculate its register usage.
6141 auto RUs = calculateRegisterUsage(VFs);
6143 // Select the largest VF which doesn't require more registers than existing
6145 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(true);
6146 for (int i = RUs.size() - 1; i >= 0; --i) {
6147 if (RUs[i].MaxLocalUsers <= TargetNumRegisters) {
6154 // If we optimize the program for size, avoid creating the tail loop.
6156 unsigned TC = PSE.getSE()->getSmallConstantTripCount(TheLoop);
6157 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
6159 // If we don't know the precise trip count, don't try to vectorize.
6162 createMissedAnalysis("UnknownLoopCountComplexCFG")
6163 << "unable to calculate the loop count due to complex control flow");
6164 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n");
6168 // Find the maximum SIMD width that can fit within the trip count.
6169 VF = TC % MaxVectorSize;
6174 // If the trip count that we found modulo the vectorization factor is not
6175 // zero then we require a tail.
6176 ORE->emit(createMissedAnalysis("NoTailLoopWithOptForSize")
6177 << "cannot optimize for size and vectorize at the "
6178 "same time. Enable vectorization of this loop "
6179 "with '#pragma clang loop vectorize(enable)' "
6180 "when compiling with -Os/-Oz");
6181 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n");
6186 int UserVF = Hints->getWidth();
6188 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
6189 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
6191 Factor.Width = UserVF;
6192 collectInstsToScalarize(UserVF);
6196 float Cost = expectedCost(1).first;
6198 const float ScalarCost = Cost;
6201 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
6203 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
6204 // Ignore scalar width, because the user explicitly wants vectorization.
6205 if (ForceVectorization && VF > 1) {
6207 Cost = expectedCost(Width).first / (float)Width;
6210 for (unsigned i = 2; i <= VF; i *= 2) {
6211 // Notice that the vector loop needs to be executed less times, so
6212 // we need to divide the cost of the vector loops by the width of
6213 // the vector elements.
6214 VectorizationCostTy C = expectedCost(i);
6215 float VectorCost = C.first / (float)i;
6216 DEBUG(dbgs() << "LV: Vector loop of width " << i
6217 << " costs: " << (int)VectorCost << ".\n");
6218 if (!C.second && !ForceVectorization) {
6220 dbgs() << "LV: Not considering vector loop of width " << i
6221 << " because it will not generate any vector instructions.\n");
6224 if (VectorCost < Cost) {
6230 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
6231 << "LV: Vectorization seems to be not beneficial, "
6232 << "but was forced by a user.\n");
6233 DEBUG(dbgs() << "LV: Selecting VF: " << Width << ".\n");
6234 Factor.Width = Width;
6235 Factor.Cost = Width * Cost;
6239 std::pair<unsigned, unsigned>
6240 LoopVectorizationCostModel::getSmallestAndWidestTypes() {
6241 unsigned MinWidth = -1U;
6242 unsigned MaxWidth = 8;
6243 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
6246 for (BasicBlock *BB : TheLoop->blocks()) {
6247 // For each instruction in the loop.
6248 for (Instruction &I : *BB) {
6249 Type *T = I.getType();
6251 // Skip ignored values.
6252 if (ValuesToIgnore.count(&I))
6255 // Only examine Loads, Stores and PHINodes.
6256 if (!isa<LoadInst>(I) && !isa<StoreInst>(I) && !isa<PHINode>(I))
6259 // Examine PHI nodes that are reduction variables. Update the type to
6260 // account for the recurrence type.
6261 if (auto *PN = dyn_cast<PHINode>(&I)) {
6262 if (!Legal->isReductionVariable(PN))
6264 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[PN];
6265 T = RdxDesc.getRecurrenceType();
6268 // Examine the stored values.
6269 if (auto *ST = dyn_cast<StoreInst>(&I))
6270 T = ST->getValueOperand()->getType();
6272 // Ignore loaded pointer types and stored pointer types that are not
6273 // consecutive. However, we do want to take consecutive stores/loads of
6274 // pointer vectors into account.
6275 if (T->isPointerTy() && !isConsecutiveLoadOrStore(&I))
6278 MinWidth = std::min(MinWidth,
6279 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
6280 MaxWidth = std::max(MaxWidth,
6281 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
6285 return {MinWidth, MaxWidth};
6288 unsigned LoopVectorizationCostModel::selectInterleaveCount(bool OptForSize,
6290 unsigned LoopCost) {
6292 // -- The interleave heuristics --
6293 // We interleave the loop in order to expose ILP and reduce the loop overhead.
6294 // There are many micro-architectural considerations that we can't predict
6295 // at this level. For example, frontend pressure (on decode or fetch) due to
6296 // code size, or the number and capabilities of the execution ports.
6298 // We use the following heuristics to select the interleave count:
6299 // 1. If the code has reductions, then we interleave to break the cross
6300 // iteration dependency.
6301 // 2. If the loop is really small, then we interleave to reduce the loop
6303 // 3. We don't interleave if we think that we will spill registers to memory
6304 // due to the increased register pressure.
6306 // When we optimize for size, we don't interleave.
6310 // We used the distance for the interleave count.
6311 if (Legal->getMaxSafeDepDistBytes() != -1U)
6314 // Do not interleave loops with a relatively small trip count.
6315 unsigned TC = PSE.getSE()->getSmallConstantTripCount(TheLoop);
6316 if (TC > 1 && TC < TinyTripCountInterleaveThreshold)
6319 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
6320 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters
6324 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
6325 TargetNumRegisters = ForceTargetNumScalarRegs;
6327 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
6328 TargetNumRegisters = ForceTargetNumVectorRegs;
6331 RegisterUsage R = calculateRegisterUsage({VF})[0];
6332 // We divide by these constants so assume that we have at least one
6333 // instruction that uses at least one register.
6334 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
6335 R.NumInstructions = std::max(R.NumInstructions, 1U);
6337 // We calculate the interleave count using the following formula.
6338 // Subtract the number of loop invariants from the number of available
6339 // registers. These registers are used by all of the interleaved instances.
6340 // Next, divide the remaining registers by the number of registers that is
6341 // required by the loop, in order to estimate how many parallel instances
6342 // fit without causing spills. All of this is rounded down if necessary to be
6343 // a power of two. We want power of two interleave count to simplify any
6344 // addressing operations or alignment considerations.
6345 unsigned IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
6348 // Don't count the induction variable as interleaved.
6349 if (EnableIndVarRegisterHeur)
6350 IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
6351 std::max(1U, (R.MaxLocalUsers - 1)));
6353 // Clamp the interleave ranges to reasonable counts.
6354 unsigned MaxInterleaveCount = TTI.getMaxInterleaveFactor(VF);
6356 // Check if the user has overridden the max.
6358 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
6359 MaxInterleaveCount = ForceTargetMaxScalarInterleaveFactor;
6361 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
6362 MaxInterleaveCount = ForceTargetMaxVectorInterleaveFactor;
6365 // If we did not calculate the cost for VF (because the user selected the VF)
6366 // then we calculate the cost of VF here.
6368 LoopCost = expectedCost(VF).first;
6370 // Clamp the calculated IC to be between the 1 and the max interleave count
6371 // that the target allows.
6372 if (IC > MaxInterleaveCount)
6373 IC = MaxInterleaveCount;
6377 // Interleave if we vectorized this loop and there is a reduction that could
6378 // benefit from interleaving.
6379 if (VF > 1 && Legal->getReductionVars()->size()) {
6380 DEBUG(dbgs() << "LV: Interleaving because of reductions.\n");
6384 // Note that if we've already vectorized the loop we will have done the
6385 // runtime check and so interleaving won't require further checks.
6386 bool InterleavingRequiresRuntimePointerCheck =
6387 (VF == 1 && Legal->getRuntimePointerChecking()->Need);
6389 // We want to interleave small loops in order to reduce the loop overhead and
6390 // potentially expose ILP opportunities.
6391 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
6392 if (!InterleavingRequiresRuntimePointerCheck && LoopCost < SmallLoopCost) {
6393 // We assume that the cost overhead is 1 and we use the cost model
6394 // to estimate the cost of the loop and interleave until the cost of the
6395 // loop overhead is about 5% of the cost of the loop.
6397 std::min(IC, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
6399 // Interleave until store/load ports (estimated by max interleave count) are
6401 unsigned NumStores = Legal->getNumStores();
6402 unsigned NumLoads = Legal->getNumLoads();
6403 unsigned StoresIC = IC / (NumStores ? NumStores : 1);
6404 unsigned LoadsIC = IC / (NumLoads ? NumLoads : 1);
6406 // If we have a scalar reduction (vector reductions are already dealt with
6407 // by this point), we can increase the critical path length if the loop
6408 // we're interleaving is inside another loop. Limit, by default to 2, so the
6409 // critical path only gets increased by one reduction operation.
6410 if (Legal->getReductionVars()->size() && TheLoop->getLoopDepth() > 1) {
6411 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionIC);
6412 SmallIC = std::min(SmallIC, F);
6413 StoresIC = std::min(StoresIC, F);
6414 LoadsIC = std::min(LoadsIC, F);
6417 if (EnableLoadStoreRuntimeInterleave &&
6418 std::max(StoresIC, LoadsIC) > SmallIC) {
6419 DEBUG(dbgs() << "LV: Interleaving to saturate store or load ports.\n");
6420 return std::max(StoresIC, LoadsIC);
6423 DEBUG(dbgs() << "LV: Interleaving to reduce branch cost.\n");
6427 // Interleave if this is a large loop (small loops are already dealt with by
6428 // this point) that could benefit from interleaving.
6429 bool HasReductions = (Legal->getReductionVars()->size() > 0);
6430 if (TTI.enableAggressiveInterleaving(HasReductions)) {
6431 DEBUG(dbgs() << "LV: Interleaving to expose ILP.\n");
6435 DEBUG(dbgs() << "LV: Not Interleaving.\n");
6439 SmallVector<LoopVectorizationCostModel::RegisterUsage, 8>
6440 LoopVectorizationCostModel::calculateRegisterUsage(ArrayRef<unsigned> VFs) {
6441 // This function calculates the register usage by measuring the highest number
6442 // of values that are alive at a single location. Obviously, this is a very
6443 // rough estimation. We scan the loop in a topological order in order and
6444 // assign a number to each instruction. We use RPO to ensure that defs are
6445 // met before their users. We assume that each instruction that has in-loop
6446 // users starts an interval. We record every time that an in-loop value is
6447 // used, so we have a list of the first and last occurrences of each
6448 // instruction. Next, we transpose this data structure into a multi map that
6449 // holds the list of intervals that *end* at a specific location. This multi
6450 // map allows us to perform a linear search. We scan the instructions linearly
6451 // and record each time that a new interval starts, by placing it in a set.
6452 // If we find this value in the multi-map then we remove it from the set.
6453 // The max register usage is the maximum size of the set.
6454 // We also search for instructions that are defined outside the loop, but are
6455 // used inside the loop. We need this number separately from the max-interval
6456 // usage number because when we unroll, loop-invariant values do not take
6458 LoopBlocksDFS DFS(TheLoop);
6462 RU.NumInstructions = 0;
6464 // Each 'key' in the map opens a new interval. The values
6465 // of the map are the index of the 'last seen' usage of the
6466 // instruction that is the key.
6467 typedef DenseMap<Instruction *, unsigned> IntervalMap;
6468 // Maps instruction to its index.
6469 DenseMap<unsigned, Instruction *> IdxToInstr;
6470 // Marks the end of each interval.
6471 IntervalMap EndPoint;
6472 // Saves the list of instruction indices that are used in the loop.
6473 SmallSet<Instruction *, 8> Ends;
6474 // Saves the list of values that are used in the loop but are
6475 // defined outside the loop, such as arguments and constants.
6476 SmallPtrSet<Value *, 8> LoopInvariants;
6479 for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO())) {
6480 RU.NumInstructions += BB->size();
6481 for (Instruction &I : *BB) {
6482 IdxToInstr[Index++] = &I;
6484 // Save the end location of each USE.
6485 for (Value *U : I.operands()) {
6486 auto *Instr = dyn_cast<Instruction>(U);
6488 // Ignore non-instruction values such as arguments, constants, etc.
6492 // If this instruction is outside the loop then record it and continue.
6493 if (!TheLoop->contains(Instr)) {
6494 LoopInvariants.insert(Instr);
6498 // Overwrite previous end points.
6499 EndPoint[Instr] = Index;
6505 // Saves the list of intervals that end with the index in 'key'.
6506 typedef SmallVector<Instruction *, 2> InstrList;
6507 DenseMap<unsigned, InstrList> TransposeEnds;
6509 // Transpose the EndPoints to a list of values that end at each index.
6510 for (auto &Interval : EndPoint)
6511 TransposeEnds[Interval.second].push_back(Interval.first);
6513 SmallSet<Instruction *, 8> OpenIntervals;
6515 // Get the size of the widest register.
6516 unsigned MaxSafeDepDist = -1U;
6517 if (Legal->getMaxSafeDepDistBytes() != -1U)
6518 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
6519 unsigned WidestRegister =
6520 std::min(TTI.getRegisterBitWidth(true), MaxSafeDepDist);
6521 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
6523 SmallVector<RegisterUsage, 8> RUs(VFs.size());
6524 SmallVector<unsigned, 8> MaxUsages(VFs.size(), 0);
6526 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
6528 // A lambda that gets the register usage for the given type and VF.
6529 auto GetRegUsage = [&DL, WidestRegister](Type *Ty, unsigned VF) {
6530 if (Ty->isTokenTy())
6532 unsigned TypeSize = DL.getTypeSizeInBits(Ty->getScalarType());
6533 return std::max<unsigned>(1, VF * TypeSize / WidestRegister);
6536 for (unsigned int i = 0; i < Index; ++i) {
6537 Instruction *I = IdxToInstr[i];
6539 // Remove all of the instructions that end at this location.
6540 InstrList &List = TransposeEnds[i];
6541 for (Instruction *ToRemove : List)
6542 OpenIntervals.erase(ToRemove);
6544 // Ignore instructions that are never used within the loop.
6548 // Skip ignored values.
6549 if (ValuesToIgnore.count(I))
6552 // For each VF find the maximum usage of registers.
6553 for (unsigned j = 0, e = VFs.size(); j < e; ++j) {
6555 MaxUsages[j] = std::max(MaxUsages[j], OpenIntervals.size());
6559 // Count the number of live intervals.
6560 unsigned RegUsage = 0;
6561 for (auto Inst : OpenIntervals) {
6562 // Skip ignored values for VF > 1.
6563 if (VecValuesToIgnore.count(Inst))
6565 RegUsage += GetRegUsage(Inst->getType(), VFs[j]);
6567 MaxUsages[j] = std::max(MaxUsages[j], RegUsage);
6570 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # "
6571 << OpenIntervals.size() << '\n');
6573 // Add the current instruction to the list of open intervals.
6574 OpenIntervals.insert(I);
6577 for (unsigned i = 0, e = VFs.size(); i < e; ++i) {
6578 unsigned Invariant = 0;
6580 Invariant = LoopInvariants.size();
6582 for (auto Inst : LoopInvariants)
6583 Invariant += GetRegUsage(Inst->getType(), VFs[i]);
6586 DEBUG(dbgs() << "LV(REG): VF = " << VFs[i] << '\n');
6587 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsages[i] << '\n');
6588 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
6589 DEBUG(dbgs() << "LV(REG): LoopSize: " << RU.NumInstructions << '\n');
6591 RU.LoopInvariantRegs = Invariant;
6592 RU.MaxLocalUsers = MaxUsages[i];
6599 void LoopVectorizationCostModel::collectInstsToScalarize(unsigned VF) {
6601 // If we aren't vectorizing the loop, or if we've already collected the
6602 // instructions to scalarize, there's nothing to do. Collection may already
6603 // have occurred if we have a user-selected VF and are now computing the
6604 // expected cost for interleaving.
6605 if (VF < 2 || InstsToScalarize.count(VF))
6608 // Initialize a mapping for VF in InstsToScalalarize. If we find that it's
6609 // not profitable to scalarize any instructions, the presence of VF in the
6610 // map will indicate that we've analyzed it already.
6611 ScalarCostsTy &ScalarCostsVF = InstsToScalarize[VF];
6613 // Find all the instructions that are scalar with predication in the loop and
6614 // determine if it would be better to not if-convert the blocks they are in.
6615 // If so, we also record the instructions to scalarize.
6616 for (BasicBlock *BB : TheLoop->blocks()) {
6617 if (!Legal->blockNeedsPredication(BB))
6619 for (Instruction &I : *BB)
6620 if (Legal->isScalarWithPredication(&I)) {
6621 ScalarCostsTy ScalarCosts;
6622 if (computePredInstDiscount(&I, ScalarCosts, VF) >= 0)
6623 ScalarCostsVF.insert(ScalarCosts.begin(), ScalarCosts.end());
6628 int LoopVectorizationCostModel::computePredInstDiscount(
6629 Instruction *PredInst, DenseMap<Instruction *, unsigned> &ScalarCosts,
6632 assert(!Legal->isUniformAfterVectorization(PredInst) &&
6633 "Instruction marked uniform-after-vectorization will be predicated");
6635 // Initialize the discount to zero, meaning that the scalar version and the
6636 // vector version cost the same.
6639 // Holds instructions to analyze. The instructions we visit are mapped in
6640 // ScalarCosts. Those instructions are the ones that would be scalarized if
6641 // we find that the scalar version costs less.
6642 SmallVector<Instruction *, 8> Worklist;
6644 // Returns true if the given instruction can be scalarized.
6645 auto canBeScalarized = [&](Instruction *I) -> bool {
6647 // We only attempt to scalarize instructions forming a single-use chain
6648 // from the original predicated block that would otherwise be vectorized.
6649 // Although not strictly necessary, we give up on instructions we know will
6650 // already be scalar to avoid traversing chains that are unlikely to be
6652 if (!I->hasOneUse() || PredInst->getParent() != I->getParent() ||
6653 Legal->isScalarAfterVectorization(I))
6656 // If the instruction is scalar with predication, it will be analyzed
6657 // separately. We ignore it within the context of PredInst.
6658 if (Legal->isScalarWithPredication(I))
6661 // If any of the instruction's operands are uniform after vectorization,
6662 // the instruction cannot be scalarized. This prevents, for example, a
6663 // masked load from being scalarized.
6665 // We assume we will only emit a value for lane zero of an instruction
6666 // marked uniform after vectorization, rather than VF identical values.
6667 // Thus, if we scalarize an instruction that uses a uniform, we would
6668 // create uses of values corresponding to the lanes we aren't emitting code
6669 // for. This behavior can be changed by allowing getScalarValue to clone
6670 // the lane zero values for uniforms rather than asserting.
6671 for (Use &U : I->operands())
6672 if (auto *J = dyn_cast<Instruction>(U.get()))
6673 if (Legal->isUniformAfterVectorization(J))
6676 // Otherwise, we can scalarize the instruction.
6680 // Returns true if an operand that cannot be scalarized must be extracted
6681 // from a vector. We will account for this scalarization overhead below. Note
6682 // that the non-void predicated instructions are placed in their own blocks,
6683 // and their return values are inserted into vectors. Thus, an extract would
6684 // still be required.
6685 auto needsExtract = [&](Instruction *I) -> bool {
6686 return TheLoop->contains(I) && !Legal->isScalarAfterVectorization(I);
6689 // Compute the expected cost discount from scalarizing the entire expression
6690 // feeding the predicated instruction. We currently only consider expressions
6691 // that are single-use instruction chains.
6692 Worklist.push_back(PredInst);
6693 while (!Worklist.empty()) {
6694 Instruction *I = Worklist.pop_back_val();
6696 // If we've already analyzed the instruction, there's nothing to do.
6697 if (ScalarCosts.count(I))
6700 // Compute the cost of the vector instruction. Note that this cost already
6701 // includes the scalarization overhead of the predicated instruction.
6702 unsigned VectorCost = getInstructionCost(I, VF).first;
6704 // Compute the cost of the scalarized instruction. This cost is the cost of
6705 // the instruction as if it wasn't if-converted and instead remained in the
6706 // predicated block. We will scale this cost by block probability after
6707 // computing the scalarization overhead.
6708 unsigned ScalarCost = VF * getInstructionCost(I, 1).first;
6710 // Compute the scalarization overhead of needed insertelement instructions
6712 if (Legal->isScalarWithPredication(I) && !I->getType()->isVoidTy()) {
6713 ScalarCost += getScalarizationOverhead(ToVectorTy(I->getType(), VF), true,
6715 ScalarCost += VF * TTI.getCFInstrCost(Instruction::PHI);
6718 // Compute the scalarization overhead of needed extractelement
6719 // instructions. For each of the instruction's operands, if the operand can
6720 // be scalarized, add it to the worklist; otherwise, account for the
6722 for (Use &U : I->operands())
6723 if (auto *J = dyn_cast<Instruction>(U.get())) {
6724 assert(VectorType::isValidElementType(J->getType()) &&
6725 "Instruction has non-scalar type");
6726 if (canBeScalarized(J))
6727 Worklist.push_back(J);
6728 else if (needsExtract(J))
6729 ScalarCost += getScalarizationOverhead(ToVectorTy(J->getType(), VF),
6733 // Scale the total scalar cost by block probability.
6734 ScalarCost /= getReciprocalPredBlockProb();
6736 // Compute the discount. A non-negative discount means the vector version
6737 // of the instruction costs more, and scalarizing would be beneficial.
6738 Discount += VectorCost - ScalarCost;
6739 ScalarCosts[I] = ScalarCost;
6745 LoopVectorizationCostModel::VectorizationCostTy
6746 LoopVectorizationCostModel::expectedCost(unsigned VF) {
6747 VectorizationCostTy Cost;
6749 // Collect the instructions (and their associated costs) that will be more
6750 // profitable to scalarize.
6751 collectInstsToScalarize(VF);
6754 for (BasicBlock *BB : TheLoop->blocks()) {
6755 VectorizationCostTy BlockCost;
6757 // For each instruction in the old loop.
6758 for (Instruction &I : *BB) {
6759 // Skip dbg intrinsics.
6760 if (isa<DbgInfoIntrinsic>(I))
6763 // Skip ignored values.
6764 if (ValuesToIgnore.count(&I))
6767 VectorizationCostTy C = getInstructionCost(&I, VF);
6769 // Check if we should override the cost.
6770 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
6771 C.first = ForceTargetInstructionCost;
6773 BlockCost.first += C.first;
6774 BlockCost.second |= C.second;
6775 DEBUG(dbgs() << "LV: Found an estimated cost of " << C.first << " for VF "
6776 << VF << " For instruction: " << I << '\n');
6779 // If we are vectorizing a predicated block, it will have been
6780 // if-converted. This means that the block's instructions (aside from
6781 // stores and instructions that may divide by zero) will now be
6782 // unconditionally executed. For the scalar case, we may not always execute
6783 // the predicated block. Thus, scale the block's cost by the probability of
6785 if (VF == 1 && Legal->blockNeedsPredication(BB))
6786 BlockCost.first /= getReciprocalPredBlockProb();
6788 Cost.first += BlockCost.first;
6789 Cost.second |= BlockCost.second;
6795 /// \brief Gets Address Access SCEV after verifying that the access pattern
6796 /// is loop invariant except the induction variable dependence.
6798 /// This SCEV can be sent to the Target in order to estimate the address
6799 /// calculation cost.
6800 static const SCEV *getAddressAccessSCEV(
6802 LoopVectorizationLegality *Legal,
6803 ScalarEvolution *SE,
6804 const Loop *TheLoop) {
6805 auto *Gep = dyn_cast<GetElementPtrInst>(Ptr);
6809 // We are looking for a gep with all loop invariant indices except for one
6810 // which should be an induction variable.
6811 unsigned NumOperands = Gep->getNumOperands();
6812 for (unsigned i = 1; i < NumOperands; ++i) {
6813 Value *Opd = Gep->getOperand(i);
6814 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
6815 !Legal->isInductionVariable(Opd))
6819 // Now we know we have a GEP ptr, %inv, %ind, %inv. return the Ptr SCEV.
6820 return SE->getSCEV(Ptr);
6823 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
6824 return Legal->hasStride(I->getOperand(0)) ||
6825 Legal->hasStride(I->getOperand(1));
6828 LoopVectorizationCostModel::VectorizationCostTy
6829 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
6830 // If we know that this instruction will remain uniform, check the cost of
6831 // the scalar version.
6832 if (Legal->isUniformAfterVectorization(I))
6835 if (VF > 1 && isProfitableToScalarize(I, VF))
6836 return VectorizationCostTy(InstsToScalarize[VF][I], false);
6839 unsigned C = getInstructionCost(I, VF, VectorTy);
6841 bool TypeNotScalarized =
6842 VF > 1 && !VectorTy->isVoidTy() && TTI.getNumberOfParts(VectorTy) < VF;
6843 return VectorizationCostTy(C, TypeNotScalarized);
6846 unsigned LoopVectorizationCostModel::getInstructionCost(Instruction *I,
6849 Type *RetTy = I->getType();
6850 if (canTruncateToMinimalBitwidth(I, VF))
6851 RetTy = IntegerType::get(RetTy->getContext(), MinBWs[I]);
6852 VectorTy = ToVectorTy(RetTy, VF);
6853 auto SE = PSE.getSE();
6855 // TODO: We need to estimate the cost of intrinsic calls.
6856 switch (I->getOpcode()) {
6857 case Instruction::GetElementPtr:
6858 // We mark this instruction as zero-cost because the cost of GEPs in
6859 // vectorized code depends on whether the corresponding memory instruction
6860 // is scalarized or not. Therefore, we handle GEPs with the memory
6861 // instruction cost.
6863 case Instruction::Br: {
6864 return TTI.getCFInstrCost(I->getOpcode());
6866 case Instruction::PHI: {
6867 auto *Phi = cast<PHINode>(I);
6869 // First-order recurrences are replaced by vector shuffles inside the loop.
6870 if (VF > 1 && Legal->isFirstOrderRecurrence(Phi))
6871 return TTI.getShuffleCost(TargetTransformInfo::SK_ExtractSubvector,
6872 VectorTy, VF - 1, VectorTy);
6874 // TODO: IF-converted IFs become selects.
6877 case Instruction::UDiv:
6878 case Instruction::SDiv:
6879 case Instruction::URem:
6880 case Instruction::SRem:
6881 // If we have a predicated instruction, it may not be executed for each
6882 // vector lane. Get the scalarization cost and scale this amount by the
6883 // probability of executing the predicated block. If the instruction is not
6884 // predicated, we fall through to the next case.
6885 if (VF > 1 && Legal->isScalarWithPredication(I)) {
6888 // These instructions have a non-void type, so account for the phi nodes
6889 // that we will create. This cost is likely to be zero. The phi node
6890 // cost, if any, should be scaled by the block probability because it
6891 // models a copy at the end of each predicated block.
6892 Cost += VF * TTI.getCFInstrCost(Instruction::PHI);
6894 // The cost of the non-predicated instruction.
6895 Cost += VF * TTI.getArithmeticInstrCost(I->getOpcode(), RetTy);
6897 // The cost of insertelement and extractelement instructions needed for
6899 Cost += getScalarizationOverhead(I, VF, TTI);
6901 // Scale the cost by the probability of executing the predicated blocks.
6902 // This assumes the predicated block for each vector lane is equally
6904 return Cost / getReciprocalPredBlockProb();
6906 case Instruction::Add:
6907 case Instruction::FAdd:
6908 case Instruction::Sub:
6909 case Instruction::FSub:
6910 case Instruction::Mul:
6911 case Instruction::FMul:
6912 case Instruction::FDiv:
6913 case Instruction::FRem:
6914 case Instruction::Shl:
6915 case Instruction::LShr:
6916 case Instruction::AShr:
6917 case Instruction::And:
6918 case Instruction::Or:
6919 case Instruction::Xor: {
6920 // Since we will replace the stride by 1 the multiplication should go away.
6921 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
6923 // Certain instructions can be cheaper to vectorize if they have a constant
6924 // second vector operand. One example of this are shifts on x86.
6925 TargetTransformInfo::OperandValueKind Op1VK =
6926 TargetTransformInfo::OK_AnyValue;
6927 TargetTransformInfo::OperandValueKind Op2VK =
6928 TargetTransformInfo::OK_AnyValue;
6929 TargetTransformInfo::OperandValueProperties Op1VP =
6930 TargetTransformInfo::OP_None;
6931 TargetTransformInfo::OperandValueProperties Op2VP =
6932 TargetTransformInfo::OP_None;
6933 Value *Op2 = I->getOperand(1);
6935 // Check for a splat or for a non uniform vector of constants.
6936 if (isa<ConstantInt>(Op2)) {
6937 ConstantInt *CInt = cast<ConstantInt>(Op2);
6938 if (CInt && CInt->getValue().isPowerOf2())
6939 Op2VP = TargetTransformInfo::OP_PowerOf2;
6940 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
6941 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
6942 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
6943 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
6945 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
6946 if (CInt && CInt->getValue().isPowerOf2())
6947 Op2VP = TargetTransformInfo::OP_PowerOf2;
6948 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
6950 } else if (Legal->isUniform(Op2)) {
6951 Op2VK = TargetTransformInfo::OK_UniformValue;
6953 SmallVector<const Value *, 4> Operands(I->operand_values());
6954 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK,
6955 Op2VK, Op1VP, Op2VP, Operands);
6957 case Instruction::Select: {
6958 SelectInst *SI = cast<SelectInst>(I);
6959 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
6960 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
6961 Type *CondTy = SI->getCondition()->getType();
6963 CondTy = VectorType::get(CondTy, VF);
6965 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
6967 case Instruction::ICmp:
6968 case Instruction::FCmp: {
6969 Type *ValTy = I->getOperand(0)->getType();
6970 Instruction *Op0AsInstruction = dyn_cast<Instruction>(I->getOperand(0));
6971 if (canTruncateToMinimalBitwidth(Op0AsInstruction, VF))
6972 ValTy = IntegerType::get(ValTy->getContext(), MinBWs[Op0AsInstruction]);
6973 VectorTy = ToVectorTy(ValTy, VF);
6974 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
6976 case Instruction::Store:
6977 case Instruction::Load: {
6978 StoreInst *SI = dyn_cast<StoreInst>(I);
6979 LoadInst *LI = dyn_cast<LoadInst>(I);
6980 Type *ValTy = (SI ? SI->getValueOperand()->getType() : LI->getType());
6981 VectorTy = ToVectorTy(ValTy, VF);
6983 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
6985 SI ? SI->getPointerAddressSpace() : LI->getPointerAddressSpace();
6986 Value *Ptr = getPointerOperand(I);
6987 // We add the cost of address computation here instead of with the gep
6988 // instruction because only here we know whether the operation is
6991 return TTI.getAddressComputationCost(VectorTy) +
6992 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
6994 if (LI && Legal->isUniform(Ptr)) {
6995 // Scalar load + broadcast
6996 unsigned Cost = TTI.getAddressComputationCost(ValTy->getScalarType());
6997 Cost += TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
7000 TTI.getShuffleCost(TargetTransformInfo::SK_Broadcast, ValTy);
7003 // For an interleaved access, calculate the total cost of the whole
7004 // interleave group.
7005 if (Legal->isAccessInterleaved(I)) {
7006 auto Group = Legal->getInterleavedAccessGroup(I);
7007 assert(Group && "Fail to get an interleaved access group.");
7009 // Only calculate the cost once at the insert position.
7010 if (Group->getInsertPos() != I)
7013 unsigned InterleaveFactor = Group->getFactor();
7015 VectorType::get(VectorTy->getVectorElementType(),
7016 VectorTy->getVectorNumElements() * InterleaveFactor);
7018 // Holds the indices of existing members in an interleaved load group.
7019 // An interleaved store group doesn't need this as it doesn't allow gaps.
7020 SmallVector<unsigned, 4> Indices;
7022 for (unsigned i = 0; i < InterleaveFactor; i++)
7023 if (Group->getMember(i))
7024 Indices.push_back(i);
7027 // Calculate the cost of the whole interleaved group.
7028 unsigned Cost = TTI.getInterleavedMemoryOpCost(
7029 I->getOpcode(), WideVecTy, Group->getFactor(), Indices,
7030 Group->getAlignment(), AS);
7032 if (Group->isReverse())
7034 Group->getNumMembers() *
7035 TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
7037 // FIXME: The interleaved load group with a huge gap could be even more
7038 // expensive than scalar operations. Then we could ignore such group and
7039 // use scalar operations instead.
7043 // Check if the memory instruction will be scalarized.
7044 if (Legal->memoryInstructionMustBeScalarized(I, VF)) {
7046 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
7048 // Figure out whether the access is strided and get the stride value
7049 // if it's known in compile time
7050 const SCEV *PtrSCEV = getAddressAccessSCEV(Ptr, Legal, SE, TheLoop);
7052 // Get the cost of the scalar memory instruction and address computation.
7053 Cost += VF * TTI.getAddressComputationCost(PtrTy, SE, PtrSCEV);
7055 TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
7058 // Get the overhead of the extractelement and insertelement instructions
7059 // we might create due to scalarization.
7060 Cost += getScalarizationOverhead(I, VF, TTI);
7062 // If we have a predicated store, it may not be executed for each vector
7063 // lane. Scale the cost by the probability of executing the predicated
7065 if (Legal->isScalarWithPredication(I))
7066 Cost /= getReciprocalPredBlockProb();
7071 // Determine if the pointer operand of the access is either consecutive or
7072 // reverse consecutive.
7073 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
7074 bool Reverse = ConsecutiveStride < 0;
7076 // Determine if either a gather or scatter operation is legal.
7077 bool UseGatherOrScatter =
7078 !ConsecutiveStride && Legal->isLegalGatherOrScatter(I);
7080 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
7081 if (UseGatherOrScatter) {
7082 assert(ConsecutiveStride == 0 &&
7083 "Gather/Scatter are not used for consecutive stride");
7085 TTI.getGatherScatterOpCost(I->getOpcode(), VectorTy, Ptr,
7086 Legal->isMaskRequired(I), Alignment);
7088 // Wide load/stores.
7089 if (Legal->isMaskRequired(I))
7091 TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
7093 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
7096 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
7099 case Instruction::ZExt:
7100 case Instruction::SExt:
7101 case Instruction::FPToUI:
7102 case Instruction::FPToSI:
7103 case Instruction::FPExt:
7104 case Instruction::PtrToInt:
7105 case Instruction::IntToPtr:
7106 case Instruction::SIToFP:
7107 case Instruction::UIToFP:
7108 case Instruction::Trunc:
7109 case Instruction::FPTrunc:
7110 case Instruction::BitCast: {
7111 // We optimize the truncation of induction variable.
7112 // The cost of these is the same as the scalar operation.
7113 if (I->getOpcode() == Instruction::Trunc &&
7114 Legal->isInductionVariable(I->getOperand(0)))
7115 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
7116 I->getOperand(0)->getType());
7118 Type *SrcScalarTy = I->getOperand(0)->getType();
7119 Type *SrcVecTy = ToVectorTy(SrcScalarTy, VF);
7120 if (canTruncateToMinimalBitwidth(I, VF)) {
7121 // This cast is going to be shrunk. This may remove the cast or it might
7122 // turn it into slightly different cast. For example, if MinBW == 16,
7123 // "zext i8 %1 to i32" becomes "zext i8 %1 to i16".
7125 // Calculate the modified src and dest types.
7126 Type *MinVecTy = VectorTy;
7127 if (I->getOpcode() == Instruction::Trunc) {
7128 SrcVecTy = smallestIntegerVectorType(SrcVecTy, MinVecTy);
7130 largestIntegerVectorType(ToVectorTy(I->getType(), VF), MinVecTy);
7131 } else if (I->getOpcode() == Instruction::ZExt ||
7132 I->getOpcode() == Instruction::SExt) {
7133 SrcVecTy = largestIntegerVectorType(SrcVecTy, MinVecTy);
7135 smallestIntegerVectorType(ToVectorTy(I->getType(), VF), MinVecTy);
7139 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
7141 case Instruction::Call: {
7142 bool NeedToScalarize;
7143 CallInst *CI = cast<CallInst>(I);
7144 unsigned CallCost = getVectorCallCost(CI, VF, TTI, TLI, NeedToScalarize);
7145 if (getVectorIntrinsicIDForCall(CI, TLI))
7146 return std::min(CallCost, getVectorIntrinsicCost(CI, VF, TTI, TLI));
7150 // The cost of executing VF copies of the scalar instruction. This opcode
7151 // is unknown. Assume that it is the same as 'mul'.
7152 return VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy) +
7153 getScalarizationOverhead(I, VF, TTI);
7157 char LoopVectorize::ID = 0;
7158 static const char lv_name[] = "Loop Vectorization";
7159 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
7160 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
7161 INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass)
7162 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
7163 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
7164 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
7165 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
7166 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
7167 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
7168 INITIALIZE_PASS_DEPENDENCY(LCSSAWrapperPass)
7169 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
7170 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
7171 INITIALIZE_PASS_DEPENDENCY(LoopAccessLegacyAnalysis)
7172 INITIALIZE_PASS_DEPENDENCY(DemandedBitsWrapperPass)
7173 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
7174 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
7177 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
7178 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
7182 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
7184 // Check if the pointer operand of a load or store instruction is
7186 if (auto *Ptr = getPointerOperand(Inst))
7187 return Legal->isConsecutivePtr(Ptr);
7191 void LoopVectorizationCostModel::collectValuesToIgnore() {
7192 // Ignore ephemeral values.
7193 CodeMetrics::collectEphemeralValues(TheLoop, AC, ValuesToIgnore);
7195 // Ignore type-promoting instructions we identified during reduction
7197 for (auto &Reduction : *Legal->getReductionVars()) {
7198 RecurrenceDescriptor &RedDes = Reduction.second;
7199 SmallPtrSetImpl<Instruction *> &Casts = RedDes.getCastInsts();
7200 VecValuesToIgnore.insert(Casts.begin(), Casts.end());
7203 // Insert values known to be scalar into VecValuesToIgnore. This is a
7204 // conservative estimation of the values that will later be scalarized.
7206 // FIXME: Even though an instruction is not scalar-after-vectoriztion, it may
7207 // still be scalarized. For example, we may find an instruction to be
7208 // more profitable for a given vectorization factor if it were to be
7209 // scalarized. But at this point, we haven't yet computed the
7210 // vectorization factor.
7211 for (auto *BB : TheLoop->getBlocks())
7213 if (Legal->isScalarAfterVectorization(&I))
7214 VecValuesToIgnore.insert(&I);
7217 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
7218 bool IfPredicateInstr) {
7219 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
7220 // Holds vector parameters or scalars, in case of uniform vals.
7221 SmallVector<VectorParts, 4> Params;
7223 setDebugLocFromInst(Builder, Instr);
7225 // Does this instruction return a value ?
7226 bool IsVoidRetTy = Instr->getType()->isVoidTy();
7228 // Initialize a new scalar map entry.
7229 ScalarParts Entry(UF);
7232 if (IfPredicateInstr)
7233 Cond = createBlockInMask(Instr->getParent());
7235 // For each vector unroll 'part':
7236 for (unsigned Part = 0; Part < UF; ++Part) {
7237 Entry[Part].resize(1);
7238 // For each scalar that we create:
7240 // Start an "if (pred) a[i] = ..." block.
7241 Value *Cmp = nullptr;
7242 if (IfPredicateInstr) {
7243 if (Cond[Part]->getType()->isVectorTy())
7245 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
7246 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
7247 ConstantInt::get(Cond[Part]->getType(), 1));
7250 Instruction *Cloned = Instr->clone();
7252 Cloned->setName(Instr->getName() + ".cloned");
7254 // Replace the operands of the cloned instructions with their scalar
7255 // equivalents in the new loop.
7256 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
7257 auto *NewOp = getScalarValue(Instr->getOperand(op), Part, 0);
7258 Cloned->setOperand(op, NewOp);
7261 // Place the cloned scalar in the new loop.
7262 Builder.Insert(Cloned);
7264 // Add the cloned scalar to the scalar map entry.
7265 Entry[Part][0] = Cloned;
7267 // If we just cloned a new assumption, add it the assumption cache.
7268 if (auto *II = dyn_cast<IntrinsicInst>(Cloned))
7269 if (II->getIntrinsicID() == Intrinsic::assume)
7270 AC->registerAssumption(II);
7273 if (IfPredicateInstr)
7274 PredicatedInstructions.push_back(std::make_pair(Cloned, Cmp));
7276 VectorLoopValueMap.initScalar(Instr, Entry);
7279 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
7280 auto *SI = dyn_cast<StoreInst>(Instr);
7281 bool IfPredicateInstr = (SI && Legal->blockNeedsPredication(SI->getParent()));
7283 return scalarizeInstruction(Instr, IfPredicateInstr);
7286 Value *InnerLoopUnroller::reverseVector(Value *Vec) { return Vec; }
7288 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) { return V; }
7290 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step,
7291 Instruction::BinaryOps BinOp) {
7292 // When unrolling and the VF is 1, we only need to add a simple scalar.
7293 Type *Ty = Val->getType();
7294 assert(!Ty->isVectorTy() && "Val must be a scalar");
7296 if (Ty->isFloatingPointTy()) {
7297 Constant *C = ConstantFP::get(Ty, (double)StartIdx);
7299 // Floating point operations had to be 'fast' to enable the unrolling.
7300 Value *MulOp = addFastMathFlag(Builder.CreateFMul(C, Step));
7301 return addFastMathFlag(Builder.CreateBinOp(BinOp, Val, MulOp));
7303 Constant *C = ConstantInt::get(Ty, StartIdx);
7304 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");
7307 static void AddRuntimeUnrollDisableMetaData(Loop *L) {
7308 SmallVector<Metadata *, 4> MDs;
7309 // Reserve first location for self reference to the LoopID metadata node.
7310 MDs.push_back(nullptr);
7311 bool IsUnrollMetadata = false;
7312 MDNode *LoopID = L->getLoopID();
7314 // First find existing loop unrolling disable metadata.
7315 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
7316 auto *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
7318 const auto *S = dyn_cast<MDString>(MD->getOperand(0));
7320 S && S->getString().startswith("llvm.loop.unroll.disable");
7322 MDs.push_back(LoopID->getOperand(i));
7326 if (!IsUnrollMetadata) {
7327 // Add runtime unroll disable metadata.
7328 LLVMContext &Context = L->getHeader()->getContext();
7329 SmallVector<Metadata *, 1> DisableOperands;
7330 DisableOperands.push_back(
7331 MDString::get(Context, "llvm.loop.unroll.runtime.disable"));
7332 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
7333 MDs.push_back(DisableNode);
7334 MDNode *NewLoopID = MDNode::get(Context, MDs);
7335 // Set operand 0 to refer to the loop id itself.
7336 NewLoopID->replaceOperandWith(0, NewLoopID);
7337 L->setLoopID(NewLoopID);
7341 bool LoopVectorizePass::processLoop(Loop *L) {
7342 assert(L->empty() && "Only process inner loops.");
7345 const std::string DebugLocStr = getDebugLocString(L);
7348 DEBUG(dbgs() << "\nLV: Checking a loop in \""
7349 << L->getHeader()->getParent()->getName() << "\" from "
7350 << DebugLocStr << "\n");
7352 LoopVectorizeHints Hints(L, DisableUnrolling, *ORE);
7354 DEBUG(dbgs() << "LV: Loop hints:"
7356 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
7358 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
7361 << " width=" << Hints.getWidth()
7362 << " unroll=" << Hints.getInterleave() << "\n");
7364 // Function containing loop
7365 Function *F = L->getHeader()->getParent();
7367 // Looking at the diagnostic output is the only way to determine if a loop
7368 // was vectorized (other than looking at the IR or machine code), so it
7369 // is important to generate an optimization remark for each loop. Most of
7370 // these messages are generated as OptimizationRemarkAnalysis. Remarks
7371 // generated as OptimizationRemark and OptimizationRemarkMissed are
7372 // less verbose reporting vectorized loops and unvectorized loops that may
7373 // benefit from vectorization, respectively.
7375 if (!Hints.allowVectorization(F, L, AlwaysVectorize)) {
7376 DEBUG(dbgs() << "LV: Loop hints prevent vectorization.\n");
7380 // Check the loop for a trip count threshold:
7381 // do not vectorize loops with a tiny trip count.
7382 const unsigned MaxTC = SE->getSmallConstantMaxTripCount(L);
7383 if (MaxTC > 0u && MaxTC < TinyTripCountVectorThreshold) {
7384 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
7385 << "This loop is not worth vectorizing.");
7386 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
7387 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
7389 DEBUG(dbgs() << "\n");
7390 ORE->emit(createMissedAnalysis(Hints.vectorizeAnalysisPassName(),
7392 << "vectorization is not beneficial "
7393 "and is not explicitly forced");
7398 PredicatedScalarEvolution PSE(*SE, *L);
7400 // Check if it is legal to vectorize the loop.
7401 LoopVectorizationRequirements Requirements(*ORE);
7402 LoopVectorizationLegality LVL(L, PSE, DT, TLI, AA, F, TTI, GetLAA, LI, ORE,
7403 &Requirements, &Hints);
7404 if (!LVL.canVectorize()) {
7405 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
7406 emitMissedWarning(F, L, Hints, ORE);
7410 // Use the cost model.
7411 LoopVectorizationCostModel CM(L, PSE, LI, &LVL, *TTI, TLI, DB, AC, ORE, F,
7413 CM.collectValuesToIgnore();
7415 // Check the function attributes to find out if this function should be
7416 // optimized for size.
7418 Hints.getForce() != LoopVectorizeHints::FK_Enabled && F->optForSize();
7420 // Compute the weighted frequency of this loop being executed and see if it
7421 // is less than 20% of the function entry baseline frequency. Note that we
7422 // always have a canonical loop here because we think we *can* vectorize.
7423 // FIXME: This is hidden behind a flag due to pervasive problems with
7424 // exactly what block frequency models.
7425 if (LoopVectorizeWithBlockFrequency) {
7426 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
7427 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
7428 LoopEntryFreq < ColdEntryFreq)
7432 // Check the function attributes to see if implicit floats are allowed.
7433 // FIXME: This check doesn't seem possibly correct -- what if the loop is
7434 // an integer loop and the vector instructions selected are purely integer
7435 // vector instructions?
7436 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
7437 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
7438 "attribute is used.\n");
7439 ORE->emit(createMissedAnalysis(Hints.vectorizeAnalysisPassName(),
7440 "NoImplicitFloat", L)
7441 << "loop not vectorized due to NoImplicitFloat attribute");
7442 emitMissedWarning(F, L, Hints, ORE);
7446 // Check if the target supports potentially unsafe FP vectorization.
7447 // FIXME: Add a check for the type of safety issue (denormal, signaling)
7448 // for the target we're vectorizing for, to make sure none of the
7449 // additional fp-math flags can help.
7450 if (Hints.isPotentiallyUnsafe() &&
7451 TTI->isFPVectorizationPotentiallyUnsafe()) {
7452 DEBUG(dbgs() << "LV: Potentially unsafe FP op prevents vectorization.\n");
7454 createMissedAnalysis(Hints.vectorizeAnalysisPassName(), "UnsafeFP", L)
7455 << "loop not vectorized due to unsafe FP support.");
7456 emitMissedWarning(F, L, Hints, ORE);
7460 // Select the optimal vectorization factor.
7461 const LoopVectorizationCostModel::VectorizationFactor VF =
7462 CM.selectVectorizationFactor(OptForSize);
7464 // Select the interleave count.
7465 unsigned IC = CM.selectInterleaveCount(OptForSize, VF.Width, VF.Cost);
7467 // Get user interleave count.
7468 unsigned UserIC = Hints.getInterleave();
7470 // Identify the diagnostic messages that should be produced.
7471 std::pair<StringRef, std::string> VecDiagMsg, IntDiagMsg;
7472 bool VectorizeLoop = true, InterleaveLoop = true;
7473 if (Requirements.doesNotMeet(F, L, Hints)) {
7474 DEBUG(dbgs() << "LV: Not vectorizing: loop did not meet vectorization "
7476 emitMissedWarning(F, L, Hints, ORE);
7480 if (VF.Width == 1) {
7481 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
7482 VecDiagMsg = std::make_pair(
7483 "VectorizationNotBeneficial",
7484 "the cost-model indicates that vectorization is not beneficial");
7485 VectorizeLoop = false;
7488 if (IC == 1 && UserIC <= 1) {
7489 // Tell the user interleaving is not beneficial.
7490 DEBUG(dbgs() << "LV: Interleaving is not beneficial.\n");
7491 IntDiagMsg = std::make_pair(
7492 "InterleavingNotBeneficial",
7493 "the cost-model indicates that interleaving is not beneficial");
7494 InterleaveLoop = false;
7496 IntDiagMsg.first = "InterleavingNotBeneficialAndDisabled";
7497 IntDiagMsg.second +=
7498 " and is explicitly disabled or interleave count is set to 1";
7500 } else if (IC > 1 && UserIC == 1) {
7501 // Tell the user interleaving is beneficial, but it explicitly disabled.
7503 << "LV: Interleaving is beneficial but is explicitly disabled.");
7504 IntDiagMsg = std::make_pair(
7505 "InterleavingBeneficialButDisabled",
7506 "the cost-model indicates that interleaving is beneficial "
7507 "but is explicitly disabled or interleave count is set to 1");
7508 InterleaveLoop = false;
7511 // Override IC if user provided an interleave count.
7512 IC = UserIC > 0 ? UserIC : IC;
7514 // Emit diagnostic messages, if any.
7515 const char *VAPassName = Hints.vectorizeAnalysisPassName();
7516 if (!VectorizeLoop && !InterleaveLoop) {
7517 // Do not vectorize or interleaving the loop.
7518 ORE->emit(OptimizationRemarkAnalysis(VAPassName, VecDiagMsg.first,
7519 L->getStartLoc(), L->getHeader())
7520 << VecDiagMsg.second);
7521 ORE->emit(OptimizationRemarkAnalysis(LV_NAME, IntDiagMsg.first,
7522 L->getStartLoc(), L->getHeader())
7523 << IntDiagMsg.second);
7525 } else if (!VectorizeLoop && InterleaveLoop) {
7526 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
7527 ORE->emit(OptimizationRemarkAnalysis(VAPassName, VecDiagMsg.first,
7528 L->getStartLoc(), L->getHeader())
7529 << VecDiagMsg.second);
7530 } else if (VectorizeLoop && !InterleaveLoop) {
7531 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
7532 << DebugLocStr << '\n');
7533 ORE->emit(OptimizationRemarkAnalysis(LV_NAME, IntDiagMsg.first,
7534 L->getStartLoc(), L->getHeader())
7535 << IntDiagMsg.second);
7536 } else if (VectorizeLoop && InterleaveLoop) {
7537 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
7538 << DebugLocStr << '\n');
7539 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
7542 using namespace ore;
7543 if (!VectorizeLoop) {
7544 assert(IC > 1 && "interleave count should not be 1 or 0");
7545 // If we decided that it is not legal to vectorize the loop, then
7547 InnerLoopUnroller Unroller(L, PSE, LI, DT, TLI, TTI, AC, ORE, IC, &LVL,
7549 Unroller.vectorize();
7551 ORE->emit(OptimizationRemark(LV_NAME, "Interleaved", L->getStartLoc(),
7553 << "interleaved loop (interleaved count: "
7554 << NV("InterleaveCount", IC) << ")");
7556 // If we decided that it is *legal* to vectorize the loop, then do it.
7557 InnerLoopVectorizer LB(L, PSE, LI, DT, TLI, TTI, AC, ORE, VF.Width, IC,
7562 // Add metadata to disable runtime unrolling a scalar loop when there are
7563 // no runtime checks about strides and memory. A scalar loop that is
7564 // rarely used is not worth unrolling.
7565 if (!LB.areSafetyChecksAdded())
7566 AddRuntimeUnrollDisableMetaData(L);
7568 // Report the vectorization decision.
7569 ORE->emit(OptimizationRemark(LV_NAME, "Vectorized", L->getStartLoc(),
7571 << "vectorized loop (vectorization width: "
7572 << NV("VectorizationFactor", VF.Width)
7573 << ", interleaved count: " << NV("InterleaveCount", IC) << ")");
7576 // Mark the loop as already vectorized to avoid vectorizing again.
7577 Hints.setAlreadyVectorized();
7579 DEBUG(verifyFunction(*L->getHeader()->getParent()));
7583 bool LoopVectorizePass::runImpl(
7584 Function &F, ScalarEvolution &SE_, LoopInfo &LI_, TargetTransformInfo &TTI_,
7585 DominatorTree &DT_, BlockFrequencyInfo &BFI_, TargetLibraryInfo *TLI_,
7586 DemandedBits &DB_, AliasAnalysis &AA_, AssumptionCache &AC_,
7587 std::function<const LoopAccessInfo &(Loop &)> &GetLAA_,
7588 OptimizationRemarkEmitter &ORE_) {
7602 // Compute some weights outside of the loop over the loops. Compute this
7603 // using a BranchProbability to re-use its scaling math.
7604 const BranchProbability ColdProb(1, 5); // 20%
7605 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
7608 // 1. the target claims to have no vector registers, and
7609 // 2. interleaving won't help ILP.
7611 // The second condition is necessary because, even if the target has no
7612 // vector registers, loop vectorization may still enable scalar
7614 if (!TTI->getNumberOfRegisters(true) && TTI->getMaxInterleaveFactor(1) < 2)
7617 // Build up a worklist of inner-loops to vectorize. This is necessary as
7618 // the act of vectorizing or partially unrolling a loop creates new loops
7619 // and can invalidate iterators across the loops.
7620 SmallVector<Loop *, 8> Worklist;
7623 addAcyclicInnerLoop(*L, Worklist);
7625 LoopsAnalyzed += Worklist.size();
7627 // Now walk the identified inner loops.
7628 bool Changed = false;
7629 while (!Worklist.empty())
7630 Changed |= processLoop(Worklist.pop_back_val());
7632 // Process each loop nest in the function.
7638 PreservedAnalyses LoopVectorizePass::run(Function &F,
7639 FunctionAnalysisManager &AM) {
7640 auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F);
7641 auto &LI = AM.getResult<LoopAnalysis>(F);
7642 auto &TTI = AM.getResult<TargetIRAnalysis>(F);
7643 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
7644 auto &BFI = AM.getResult<BlockFrequencyAnalysis>(F);
7645 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
7646 auto &AA = AM.getResult<AAManager>(F);
7647 auto &AC = AM.getResult<AssumptionAnalysis>(F);
7648 auto &DB = AM.getResult<DemandedBitsAnalysis>(F);
7649 auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
7651 auto &LAM = AM.getResult<LoopAnalysisManagerFunctionProxy>(F).getManager();
7652 std::function<const LoopAccessInfo &(Loop &)> GetLAA =
7653 [&](Loop &L) -> const LoopAccessInfo & {
7654 LoopStandardAnalysisResults AR = {AA, AC, DT, LI, SE, TLI, TTI};
7655 return LAM.getResult<LoopAccessAnalysis>(L, AR);
7658 runImpl(F, SE, LI, TTI, DT, BFI, &TLI, DB, AA, AC, GetLAA, ORE);
7660 return PreservedAnalyses::all();
7661 PreservedAnalyses PA;
7662 PA.preserve<LoopAnalysis>();
7663 PA.preserve<DominatorTreeAnalysis>();
7664 PA.preserve<BasicAA>();
7665 PA.preserve<GlobalsAA>();