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/Optional.h"
54 #include "llvm/ADT/SCCIterator.h"
55 #include "llvm/ADT/SetVector.h"
56 #include "llvm/ADT/SmallPtrSet.h"
57 #include "llvm/ADT/SmallSet.h"
58 #include "llvm/ADT/SmallVector.h"
59 #include "llvm/ADT/Statistic.h"
60 #include "llvm/ADT/StringExtras.h"
61 #include "llvm/Analysis/CodeMetrics.h"
62 #include "llvm/Analysis/GlobalsModRef.h"
63 #include "llvm/Analysis/LoopInfo.h"
64 #include "llvm/Analysis/LoopIterator.h"
65 #include "llvm/Analysis/LoopPass.h"
66 #include "llvm/Analysis/ScalarEvolutionExpander.h"
67 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
68 #include "llvm/Analysis/ValueTracking.h"
69 #include "llvm/Analysis/VectorUtils.h"
70 #include "llvm/IR/Constants.h"
71 #include "llvm/IR/DataLayout.h"
72 #include "llvm/IR/DebugInfo.h"
73 #include "llvm/IR/DerivedTypes.h"
74 #include "llvm/IR/DiagnosticInfo.h"
75 #include "llvm/IR/Dominators.h"
76 #include "llvm/IR/Function.h"
77 #include "llvm/IR/IRBuilder.h"
78 #include "llvm/IR/Instructions.h"
79 #include "llvm/IR/IntrinsicInst.h"
80 #include "llvm/IR/LLVMContext.h"
81 #include "llvm/IR/Module.h"
82 #include "llvm/IR/PatternMatch.h"
83 #include "llvm/IR/Type.h"
84 #include "llvm/IR/User.h"
85 #include "llvm/IR/Value.h"
86 #include "llvm/IR/ValueHandle.h"
87 #include "llvm/IR/Verifier.h"
88 #include "llvm/Pass.h"
89 #include "llvm/Support/BranchProbability.h"
90 #include "llvm/Support/CommandLine.h"
91 #include "llvm/Support/Debug.h"
92 #include "llvm/Support/raw_ostream.h"
93 #include "llvm/Transforms/Scalar.h"
94 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
95 #include "llvm/Transforms/Utils/Local.h"
96 #include "llvm/Transforms/Utils/LoopSimplify.h"
97 #include "llvm/Transforms/Utils/LoopUtils.h"
98 #include "llvm/Transforms/Utils/LoopVersioning.h"
99 #include "llvm/Transforms/Vectorize.h"
104 using namespace llvm;
105 using namespace llvm::PatternMatch;
107 #define LV_NAME "loop-vectorize"
108 #define DEBUG_TYPE LV_NAME
110 STATISTIC(LoopsVectorized, "Number of loops vectorized");
111 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
114 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
115 cl::desc("Enable if-conversion during vectorization."));
117 /// We don't vectorize loops with a known constant trip count below this number.
118 static cl::opt<unsigned> TinyTripCountVectorThreshold(
119 "vectorizer-min-trip-count", cl::init(16), cl::Hidden,
120 cl::desc("Don't vectorize loops with a constant "
121 "trip count that is smaller than this "
124 static cl::opt<bool> MaximizeBandwidth(
125 "vectorizer-maximize-bandwidth", cl::init(false), cl::Hidden,
126 cl::desc("Maximize bandwidth when selecting vectorization factor which "
127 "will be determined by the smallest type in loop."));
129 static cl::opt<bool> EnableInterleavedMemAccesses(
130 "enable-interleaved-mem-accesses", cl::init(false), cl::Hidden,
131 cl::desc("Enable vectorization on interleaved memory accesses in a loop"));
133 /// Maximum factor for an interleaved memory access.
134 static cl::opt<unsigned> MaxInterleaveGroupFactor(
135 "max-interleave-group-factor", cl::Hidden,
136 cl::desc("Maximum factor for an interleaved access group (default = 8)"),
139 /// We don't interleave loops with a known constant trip count below this
141 static const unsigned TinyTripCountInterleaveThreshold = 128;
143 static cl::opt<unsigned> ForceTargetNumScalarRegs(
144 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
145 cl::desc("A flag that overrides the target's number of scalar registers."));
147 static cl::opt<unsigned> ForceTargetNumVectorRegs(
148 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
149 cl::desc("A flag that overrides the target's number of vector registers."));
151 /// Maximum vectorization interleave count.
152 static const unsigned MaxInterleaveFactor = 16;
154 static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
155 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
156 cl::desc("A flag that overrides the target's max interleave factor for "
159 static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
160 "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
161 cl::desc("A flag that overrides the target's max interleave factor for "
162 "vectorized loops."));
164 static cl::opt<unsigned> ForceTargetInstructionCost(
165 "force-target-instruction-cost", cl::init(0), cl::Hidden,
166 cl::desc("A flag that overrides the target's expected cost for "
167 "an instruction to a single constant value. Mostly "
168 "useful for getting consistent testing."));
170 static cl::opt<unsigned> SmallLoopCost(
171 "small-loop-cost", cl::init(20), cl::Hidden,
173 "The cost of a loop that is considered 'small' by the interleaver."));
175 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
176 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
177 cl::desc("Enable the use of the block frequency analysis to access PGO "
178 "heuristics minimizing code growth in cold regions and being more "
179 "aggressive in hot regions."));
181 // Runtime interleave loops for load/store throughput.
182 static cl::opt<bool> EnableLoadStoreRuntimeInterleave(
183 "enable-loadstore-runtime-interleave", cl::init(true), cl::Hidden,
185 "Enable runtime interleaving until load/store ports are saturated"));
187 /// The number of stores in a loop that are allowed to need predication.
188 static cl::opt<unsigned> NumberOfStoresToPredicate(
189 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
190 cl::desc("Max number of stores to be predicated behind an if."));
192 static cl::opt<bool> EnableIndVarRegisterHeur(
193 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
194 cl::desc("Count the induction variable only once when interleaving"));
196 static cl::opt<bool> EnableCondStoresVectorization(
197 "enable-cond-stores-vec", cl::init(true), cl::Hidden,
198 cl::desc("Enable if predication of stores during vectorization."));
200 static cl::opt<unsigned> MaxNestedScalarReductionIC(
201 "max-nested-scalar-reduction-interleave", cl::init(2), cl::Hidden,
202 cl::desc("The maximum interleave count to use when interleaving a scalar "
203 "reduction in a nested loop."));
205 static cl::opt<unsigned> PragmaVectorizeMemoryCheckThreshold(
206 "pragma-vectorize-memory-check-threshold", cl::init(128), cl::Hidden,
207 cl::desc("The maximum allowed number of runtime memory checks with a "
208 "vectorize(enable) pragma."));
210 static cl::opt<unsigned> VectorizeSCEVCheckThreshold(
211 "vectorize-scev-check-threshold", cl::init(16), cl::Hidden,
212 cl::desc("The maximum number of SCEV checks allowed."));
214 static cl::opt<unsigned> PragmaVectorizeSCEVCheckThreshold(
215 "pragma-vectorize-scev-check-threshold", cl::init(128), cl::Hidden,
216 cl::desc("The maximum number of SCEV checks allowed with a "
217 "vectorize(enable) pragma"));
219 /// Create an analysis remark that explains why vectorization failed
221 /// \p PassName is the name of the pass (e.g. can be AlwaysPrint). \p
222 /// RemarkName is the identifier for the remark. If \p I is passed it is an
223 /// instruction that prevents vectorization. Otherwise \p TheLoop is used for
224 /// the location of the remark. \return the remark object that can be
226 static OptimizationRemarkAnalysis
227 createMissedAnalysis(const char *PassName, StringRef RemarkName, Loop *TheLoop,
228 Instruction *I = nullptr) {
229 Value *CodeRegion = TheLoop->getHeader();
230 DebugLoc DL = TheLoop->getStartLoc();
233 CodeRegion = I->getParent();
234 // If there is no debug location attached to the instruction, revert back to
236 if (I->getDebugLoc())
237 DL = I->getDebugLoc();
240 OptimizationRemarkAnalysis R(PassName, RemarkName, DL, CodeRegion);
241 R << "loop not vectorized: ";
247 // Forward declarations.
248 class LoopVectorizeHints;
249 class LoopVectorizationLegality;
250 class LoopVectorizationCostModel;
251 class LoopVectorizationRequirements;
253 /// Returns true if the given loop body has a cycle, excluding the loop
255 static bool hasCyclesInLoopBody(const Loop &L) {
259 for (const auto &SCC :
260 make_range(scc_iterator<Loop, LoopBodyTraits>::begin(L),
261 scc_iterator<Loop, LoopBodyTraits>::end(L))) {
262 if (SCC.size() > 1) {
263 DEBUG(dbgs() << "LVL: Detected a cycle in the loop body:\n");
271 /// A helper function for converting Scalar types to vector types.
272 /// If the incoming type is void, we return void. If the VF is 1, we return
274 static Type *ToVectorTy(Type *Scalar, unsigned VF) {
275 if (Scalar->isVoidTy() || VF == 1)
277 return VectorType::get(Scalar, VF);
280 // FIXME: The following helper functions have multiple implementations
281 // in the project. They can be effectively organized in a common Load/Store
284 /// A helper function that returns the pointer operand of a load or store
286 static Value *getPointerOperand(Value *I) {
287 if (auto *LI = dyn_cast<LoadInst>(I))
288 return LI->getPointerOperand();
289 if (auto *SI = dyn_cast<StoreInst>(I))
290 return SI->getPointerOperand();
294 /// A helper function that returns the type of loaded or stored value.
295 static Type *getMemInstValueType(Value *I) {
296 assert((isa<LoadInst>(I) || isa<StoreInst>(I)) &&
297 "Expected Load or Store instruction");
298 if (auto *LI = dyn_cast<LoadInst>(I))
299 return LI->getType();
300 return cast<StoreInst>(I)->getValueOperand()->getType();
303 /// A helper function that returns the alignment of load or store instruction.
304 static unsigned getMemInstAlignment(Value *I) {
305 assert((isa<LoadInst>(I) || isa<StoreInst>(I)) &&
306 "Expected Load or Store instruction");
307 if (auto *LI = dyn_cast<LoadInst>(I))
308 return LI->getAlignment();
309 return cast<StoreInst>(I)->getAlignment();
312 /// A helper function that returns the address space of the pointer operand of
313 /// load or store instruction.
314 static unsigned getMemInstAddressSpace(Value *I) {
315 assert((isa<LoadInst>(I) || isa<StoreInst>(I)) &&
316 "Expected Load or Store instruction");
317 if (auto *LI = dyn_cast<LoadInst>(I))
318 return LI->getPointerAddressSpace();
319 return cast<StoreInst>(I)->getPointerAddressSpace();
322 /// A helper function that returns true if the given type is irregular. The
323 /// type is irregular if its allocated size doesn't equal the store size of an
324 /// element of the corresponding vector type at the given vectorization factor.
325 static bool hasIrregularType(Type *Ty, const DataLayout &DL, unsigned VF) {
327 // Determine if an array of VF elements of type Ty is "bitcast compatible"
328 // with a <VF x Ty> vector.
330 auto *VectorTy = VectorType::get(Ty, VF);
331 return VF * DL.getTypeAllocSize(Ty) != DL.getTypeStoreSize(VectorTy);
334 // If the vectorization factor is one, we just check if an array of type Ty
335 // requires padding between elements.
336 return DL.getTypeAllocSizeInBits(Ty) != DL.getTypeSizeInBits(Ty);
339 /// A helper function that returns the reciprocal of the block probability of
340 /// predicated blocks. If we return X, we are assuming the predicated block
341 /// will execute once for for every X iterations of the loop header.
343 /// TODO: We should use actual block probability here, if available. Currently,
344 /// we always assume predicated blocks have a 50% chance of executing.
345 static unsigned getReciprocalPredBlockProb() { return 2; }
347 /// A helper function that adds a 'fast' flag to floating-point operations.
348 static Value *addFastMathFlag(Value *V) {
349 if (isa<FPMathOperator>(V)) {
351 Flags.setUnsafeAlgebra();
352 cast<Instruction>(V)->setFastMathFlags(Flags);
357 /// A helper function that returns an integer or floating-point constant with
359 static Constant *getSignedIntOrFpConstant(Type *Ty, int64_t C) {
360 return Ty->isIntegerTy() ? ConstantInt::getSigned(Ty, C)
361 : ConstantFP::get(Ty, C);
364 /// InnerLoopVectorizer vectorizes loops which contain only one basic
365 /// block to a specified vectorization factor (VF).
366 /// This class performs the widening of scalars into vectors, or multiple
367 /// scalars. This class also implements the following features:
368 /// * It inserts an epilogue loop for handling loops that don't have iteration
369 /// counts that are known to be a multiple of the vectorization factor.
370 /// * It handles the code generation for reduction variables.
371 /// * Scalarization (implementation using scalars) of un-vectorizable
373 /// InnerLoopVectorizer does not perform any vectorization-legality
374 /// checks, and relies on the caller to check for the different legality
375 /// aspects. The InnerLoopVectorizer relies on the
376 /// LoopVectorizationLegality class to provide information about the induction
377 /// and reduction variables that were found to a given vectorization factor.
378 class InnerLoopVectorizer {
380 InnerLoopVectorizer(Loop *OrigLoop, PredicatedScalarEvolution &PSE,
381 LoopInfo *LI, DominatorTree *DT,
382 const TargetLibraryInfo *TLI,
383 const TargetTransformInfo *TTI, AssumptionCache *AC,
384 OptimizationRemarkEmitter *ORE, unsigned VecWidth,
385 unsigned UnrollFactor, LoopVectorizationLegality *LVL,
386 LoopVectorizationCostModel *CM)
387 : OrigLoop(OrigLoop), PSE(PSE), LI(LI), DT(DT), TLI(TLI), TTI(TTI),
388 AC(AC), ORE(ORE), VF(VecWidth), UF(UnrollFactor),
389 Builder(PSE.getSE()->getContext()), Induction(nullptr),
390 OldInduction(nullptr), VectorLoopValueMap(UnrollFactor, VecWidth),
391 TripCount(nullptr), VectorTripCount(nullptr), Legal(LVL), Cost(CM),
392 AddedSafetyChecks(false) {}
394 /// Create a new empty loop. Unlink the old loop and connect the new one.
395 void createVectorizedLoopSkeleton();
397 /// Vectorize a single instruction within the innermost loop.
398 void vectorizeInstruction(Instruction &I);
400 /// Fix the vectorized code, taking care of header phi's, live-outs, and more.
401 void fixVectorizedLoop();
403 // Return true if any runtime check is added.
404 bool areSafetyChecksAdded() { return AddedSafetyChecks; }
406 virtual ~InnerLoopVectorizer() {}
409 /// A small list of PHINodes.
410 typedef SmallVector<PHINode *, 4> PhiVector;
412 /// A type for vectorized values in the new loop. Each value from the
413 /// original loop, when vectorized, is represented by UF vector values in the
414 /// new unrolled loop, where UF is the unroll factor.
415 typedef SmallVector<Value *, 2> VectorParts;
417 /// A type for scalarized values in the new loop. Each value from the
418 /// original loop, when scalarized, is represented by UF x VF scalar values
419 /// in the new unrolled loop, where UF is the unroll factor and VF is the
420 /// vectorization factor.
421 typedef SmallVector<SmallVector<Value *, 4>, 2> ScalarParts;
423 // When we if-convert we need to create edge masks. We have to cache values
424 // so that we don't end up with exponential recursion/IR.
425 typedef DenseMap<std::pair<BasicBlock *, BasicBlock *>, VectorParts>
427 typedef DenseMap<BasicBlock *, VectorParts> BlockMaskCacheTy;
429 /// Set up the values of the IVs correctly when exiting the vector loop.
430 void fixupIVUsers(PHINode *OrigPhi, const InductionDescriptor &II,
431 Value *CountRoundDown, Value *EndValue,
432 BasicBlock *MiddleBlock);
434 /// Create a new induction variable inside L.
435 PHINode *createInductionVariable(Loop *L, Value *Start, Value *End,
436 Value *Step, Instruction *DL);
438 /// Handle all cross-iteration phis in the header.
439 void fixCrossIterationPHIs();
441 /// Fix a first-order recurrence. This is the second phase of vectorizing
443 void fixFirstOrderRecurrence(PHINode *Phi);
445 /// Fix a reduction cross-iteration phi. This is the second phase of
446 /// vectorizing this phi node.
447 void fixReduction(PHINode *Phi);
449 /// \brief The Loop exit block may have single value PHI nodes with some
450 /// incoming value. While vectorizing we only handled real values
451 /// that were defined inside the loop and we should have one value for
452 /// each predecessor of its parent basic block. See PR14725.
455 /// Iteratively sink the scalarized operands of a predicated instruction into
456 /// the block that was created for it.
457 void sinkScalarOperands(Instruction *PredInst);
459 /// Predicate conditional instructions that require predication on their
460 /// respective conditions.
461 void predicateInstructions();
463 /// Shrinks vector element sizes to the smallest bitwidth they can be legally
465 void truncateToMinimalBitwidths();
467 /// A helper function that computes the predicate of the block BB, assuming
468 /// that the header block of the loop is set to True. It returns the *entry*
469 /// mask for the block BB.
470 VectorParts createBlockInMask(BasicBlock *BB);
471 /// A helper function that computes the predicate of the edge between SRC
473 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
475 /// Vectorize a single PHINode in a block. This method handles the induction
476 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
477 /// arbitrary length vectors.
478 void widenPHIInstruction(Instruction *PN, unsigned UF, unsigned VF);
480 /// Insert the new loop to the loop hierarchy and pass manager
481 /// and update the analysis passes.
482 void updateAnalysis();
484 /// This instruction is un-vectorizable. Implement it as a sequence
485 /// of scalars. If \p IfPredicateInstr is true we need to 'hide' each
486 /// scalarized instruction behind an if block predicated on the control
487 /// dependence of the instruction.
488 virtual void scalarizeInstruction(Instruction *Instr,
489 bool IfPredicateInstr = false);
491 /// Vectorize Load and Store instructions,
492 virtual void vectorizeMemoryInstruction(Instruction *Instr);
494 /// Create a broadcast instruction. This method generates a broadcast
495 /// instruction (shuffle) for loop invariant values and for the induction
496 /// value. If this is the induction variable then we extend it to N, N+1, ...
497 /// this is needed because each iteration in the loop corresponds to a SIMD
499 virtual Value *getBroadcastInstrs(Value *V);
501 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
502 /// to each vector element of Val. The sequence starts at StartIndex.
503 /// \p Opcode is relevant for FP induction variable.
504 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step,
505 Instruction::BinaryOps Opcode =
506 Instruction::BinaryOpsEnd);
508 /// Compute scalar induction steps. \p ScalarIV is the scalar induction
509 /// variable on which to base the steps, \p Step is the size of the step, and
510 /// \p EntryVal is the value from the original loop that maps to the steps.
511 /// Note that \p EntryVal doesn't have to be an induction variable (e.g., it
512 /// can be a truncate instruction).
513 void buildScalarSteps(Value *ScalarIV, Value *Step, Value *EntryVal,
514 const InductionDescriptor &ID);
516 /// Create a vector induction phi node based on an existing scalar one. \p
517 /// EntryVal is the value from the original loop that maps to the vector phi
518 /// node, and \p Step is the loop-invariant step. If \p EntryVal is a
519 /// truncate instruction, instead of widening the original IV, we widen a
520 /// version of the IV truncated to \p EntryVal's type.
521 void createVectorIntOrFpInductionPHI(const InductionDescriptor &II,
522 Value *Step, Instruction *EntryVal);
524 /// Widen an integer or floating-point induction variable \p IV. If \p Trunc
525 /// is provided, the integer induction variable will first be truncated to
526 /// the corresponding type.
527 void widenIntOrFpInduction(PHINode *IV, TruncInst *Trunc = nullptr);
529 /// Returns true if an instruction \p I should be scalarized instead of
530 /// vectorized for the chosen vectorization factor.
531 bool shouldScalarizeInstruction(Instruction *I) const;
533 /// Returns true if we should generate a scalar version of \p IV.
534 bool needsScalarInduction(Instruction *IV) const;
536 /// Return a constant reference to the VectorParts corresponding to \p V from
537 /// the original loop. If the value has already been vectorized, the
538 /// corresponding vector entry in VectorLoopValueMap is returned. If,
539 /// however, the value has a scalar entry in VectorLoopValueMap, we construct
540 /// new vector values on-demand by inserting the scalar values into vectors
541 /// with an insertelement sequence. If the value has been neither vectorized
542 /// nor scalarized, it must be loop invariant, so we simply broadcast the
543 /// value into vectors.
544 const VectorParts &getVectorValue(Value *V);
546 /// Return a value in the new loop corresponding to \p V from the original
547 /// loop at unroll index \p Part and vector index \p Lane. If the value has
548 /// been vectorized but not scalarized, the necessary extractelement
549 /// instruction will be generated.
550 Value *getScalarValue(Value *V, unsigned Part, unsigned Lane);
552 /// Try to vectorize the interleaved access group that \p Instr belongs to.
553 void vectorizeInterleaveGroup(Instruction *Instr);
555 /// Generate a shuffle sequence that will reverse the vector Vec.
556 virtual Value *reverseVector(Value *Vec);
558 /// Returns (and creates if needed) the original loop trip count.
559 Value *getOrCreateTripCount(Loop *NewLoop);
561 /// Returns (and creates if needed) the trip count of the widened loop.
562 Value *getOrCreateVectorTripCount(Loop *NewLoop);
564 /// Emit a bypass check to see if the trip count would overflow, or we
565 /// wouldn't have enough iterations to execute one vector loop.
566 void emitMinimumIterationCountCheck(Loop *L, BasicBlock *Bypass);
567 /// Emit a bypass check to see if the vector trip count is nonzero.
568 void emitVectorLoopEnteredCheck(Loop *L, BasicBlock *Bypass);
569 /// Emit a bypass check to see if all of the SCEV assumptions we've
570 /// had to make are correct.
571 void emitSCEVChecks(Loop *L, BasicBlock *Bypass);
572 /// Emit bypass checks to check any memory assumptions we may have made.
573 void emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass);
575 /// Add additional metadata to \p To that was not present on \p Orig.
577 /// Currently this is used to add the noalias annotations based on the
578 /// inserted memchecks. Use this for instructions that are *cloned* into the
580 void addNewMetadata(Instruction *To, const Instruction *Orig);
582 /// Add metadata from one instruction to another.
584 /// This includes both the original MDs from \p From and additional ones (\see
585 /// addNewMetadata). Use this for *newly created* instructions in the vector
587 void addMetadata(Instruction *To, Instruction *From);
589 /// \brief Similar to the previous function but it adds the metadata to a
590 /// vector of instructions.
591 void addMetadata(ArrayRef<Value *> To, Instruction *From);
593 /// \brief Set the debug location in the builder using the debug location in
595 void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr);
597 /// This is a helper class for maintaining vectorization state. It's used for
598 /// mapping values from the original loop to their corresponding values in
599 /// the new loop. Two mappings are maintained: one for vectorized values and
600 /// one for scalarized values. Vectorized values are represented with UF
601 /// vector values in the new loop, and scalarized values are represented with
602 /// UF x VF scalar values in the new loop. UF and VF are the unroll and
603 /// vectorization factors, respectively.
605 /// Entries can be added to either map with initVector and initScalar, which
606 /// initialize and return a constant reference to the new entry. If a
607 /// non-constant reference to a vector entry is required, getVector can be
608 /// used to retrieve a mutable entry. We currently directly modify the mapped
609 /// values during "fix-up" operations that occur once the first phase of
610 /// widening is complete. These operations include type truncation and the
611 /// second phase of recurrence widening.
613 /// Otherwise, entries from either map should be accessed using the
614 /// getVectorValue or getScalarValue functions from InnerLoopVectorizer.
615 /// getVectorValue and getScalarValue coordinate to generate a vector or
616 /// scalar value on-demand if one is not yet available. When vectorizing a
617 /// loop, we visit the definition of an instruction before its uses. When
618 /// visiting the definition, we either vectorize or scalarize the
619 /// instruction, creating an entry for it in the corresponding map. (In some
620 /// cases, such as induction variables, we will create both vector and scalar
621 /// entries.) Then, as we encounter uses of the definition, we derive values
622 /// for each scalar or vector use unless such a value is already available.
623 /// For example, if we scalarize a definition and one of its uses is vector,
624 /// we build the required vector on-demand with an insertelement sequence
625 /// when visiting the use. Otherwise, if the use is scalar, we can use the
626 /// existing scalar definition.
629 /// Construct an empty map with the given unroll and vectorization factors.
630 ValueMap(unsigned UnrollFactor, unsigned VecWidth)
631 : UF(UnrollFactor), VF(VecWidth) {
632 // The unroll and vectorization factors are only used in asserts builds
633 // to verify map entries are sized appropriately.
638 /// \return True if the map has a vector entry for \p Key.
639 bool hasVector(Value *Key) const { return VectorMapStorage.count(Key); }
641 /// \return True if the map has a scalar entry for \p Key.
642 bool hasScalar(Value *Key) const { return ScalarMapStorage.count(Key); }
644 /// \brief Map \p Key to the given VectorParts \p Entry, and return a
645 /// constant reference to the new vector map entry. The given key should
646 /// not already be in the map, and the given VectorParts should be
647 /// correctly sized for the current unroll factor.
648 const VectorParts &initVector(Value *Key, const VectorParts &Entry) {
649 assert(!hasVector(Key) && "Vector entry already initialized");
650 assert(Entry.size() == UF && "VectorParts has wrong dimensions");
651 VectorMapStorage[Key] = Entry;
652 return VectorMapStorage[Key];
655 /// \brief Map \p Key to the given ScalarParts \p Entry, and return a
656 /// constant reference to the new scalar map entry. The given key should
657 /// not already be in the map, and the given ScalarParts should be
658 /// correctly sized for the current unroll and vectorization factors.
659 const ScalarParts &initScalar(Value *Key, const ScalarParts &Entry) {
660 assert(!hasScalar(Key) && "Scalar entry already initialized");
661 assert(Entry.size() == UF &&
662 all_of(make_range(Entry.begin(), Entry.end()),
663 [&](const SmallVectorImpl<Value *> &Values) -> bool {
664 return Values.size() == VF;
666 "ScalarParts has wrong dimensions");
667 ScalarMapStorage[Key] = Entry;
668 return ScalarMapStorage[Key];
671 /// \return A reference to the vector map entry corresponding to \p Key.
672 /// The key should already be in the map. This function should only be used
673 /// when it's necessary to update values that have already been vectorized.
674 /// This is the case for "fix-up" operations including type truncation and
675 /// the second phase of recurrence vectorization. If a non-const reference
676 /// isn't required, getVectorValue should be used instead.
677 VectorParts &getVector(Value *Key) {
678 assert(hasVector(Key) && "Vector entry not initialized");
679 return VectorMapStorage.find(Key)->second;
682 /// Retrieve an entry from the vector or scalar maps. The preferred way to
683 /// access an existing mapped entry is with getVectorValue or
684 /// getScalarValue from InnerLoopVectorizer. Until those functions can be
685 /// moved inside ValueMap, we have to declare them as friends.
686 friend const VectorParts &InnerLoopVectorizer::getVectorValue(Value *V);
687 friend Value *InnerLoopVectorizer::getScalarValue(Value *V, unsigned Part,
691 /// The unroll factor. Each entry in the vector map contains UF vector
695 /// The vectorization factor. Each entry in the scalar map contains UF x VF
699 /// The vector and scalar map storage. We use std::map and not DenseMap
700 /// because insertions to DenseMap invalidate its iterators.
701 std::map<Value *, VectorParts> VectorMapStorage;
702 std::map<Value *, ScalarParts> ScalarMapStorage;
705 /// The original loop.
707 /// A wrapper around ScalarEvolution used to add runtime SCEV checks. Applies
708 /// dynamic knowledge to simplify SCEV expressions and converts them to a
709 /// more usable form.
710 PredicatedScalarEvolution &PSE;
717 /// Target Library Info.
718 const TargetLibraryInfo *TLI;
719 /// Target Transform Info.
720 const TargetTransformInfo *TTI;
721 /// Assumption Cache.
723 /// Interface to emit optimization remarks.
724 OptimizationRemarkEmitter *ORE;
726 /// \brief LoopVersioning. It's only set up (non-null) if memchecks were
729 /// This is currently only used to add no-alias metadata based on the
730 /// memchecks. The actually versioning is performed manually.
731 std::unique_ptr<LoopVersioning> LVer;
733 /// The vectorization SIMD factor to use. Each vector will have this many
738 /// The vectorization unroll factor to use. Each scalar is vectorized to this
739 /// many different vector instructions.
742 /// The builder that we use
745 // --- Vectorization state ---
747 /// The vector-loop preheader.
748 BasicBlock *LoopVectorPreHeader;
749 /// The scalar-loop preheader.
750 BasicBlock *LoopScalarPreHeader;
751 /// Middle Block between the vector and the scalar.
752 BasicBlock *LoopMiddleBlock;
753 /// The ExitBlock of the scalar loop.
754 BasicBlock *LoopExitBlock;
755 /// The vector loop body.
756 BasicBlock *LoopVectorBody;
757 /// The scalar loop body.
758 BasicBlock *LoopScalarBody;
759 /// A list of all bypass blocks. The first block is the entry of the loop.
760 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
762 /// The new Induction variable which was added to the new block.
764 /// The induction variable of the old basic block.
765 PHINode *OldInduction;
767 /// Maps values from the original loop to their corresponding values in the
768 /// vectorized loop. A key value can map to either vector values, scalar
769 /// values or both kinds of values, depending on whether the key was
770 /// vectorized and scalarized.
771 ValueMap VectorLoopValueMap;
773 /// Store instructions that should be predicated, as a pair
774 /// <StoreInst, Predicate>
775 SmallVector<std::pair<Instruction *, Value *>, 4> PredicatedInstructions;
776 EdgeMaskCacheTy EdgeMaskCache;
777 BlockMaskCacheTy BlockMaskCache;
778 /// Trip count of the original loop.
780 /// Trip count of the widened loop (TripCount - TripCount % (VF*UF))
781 Value *VectorTripCount;
783 /// The legality analysis.
784 LoopVectorizationLegality *Legal;
786 /// The profitablity analysis.
787 LoopVectorizationCostModel *Cost;
789 // Record whether runtime checks are added.
790 bool AddedSafetyChecks;
792 // Holds the end values for each induction variable. We save the end values
793 // so we can later fix-up the external users of the induction variables.
794 DenseMap<PHINode *, Value *> IVEndValues;
797 class InnerLoopUnroller : public InnerLoopVectorizer {
799 InnerLoopUnroller(Loop *OrigLoop, PredicatedScalarEvolution &PSE,
800 LoopInfo *LI, DominatorTree *DT,
801 const TargetLibraryInfo *TLI,
802 const TargetTransformInfo *TTI, AssumptionCache *AC,
803 OptimizationRemarkEmitter *ORE, unsigned UnrollFactor,
804 LoopVectorizationLegality *LVL,
805 LoopVectorizationCostModel *CM)
806 : InnerLoopVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, 1,
807 UnrollFactor, LVL, CM) {}
810 void vectorizeMemoryInstruction(Instruction *Instr) override;
811 Value *getBroadcastInstrs(Value *V) override;
812 Value *getStepVector(Value *Val, int StartIdx, Value *Step,
813 Instruction::BinaryOps Opcode =
814 Instruction::BinaryOpsEnd) override;
815 Value *reverseVector(Value *Vec) override;
818 /// \brief Look for a meaningful debug location on the instruction or it's
820 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
825 if (I->getDebugLoc() != Empty)
828 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
829 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
830 if (OpInst->getDebugLoc() != Empty)
837 void InnerLoopVectorizer::setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
838 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr)) {
839 const DILocation *DIL = Inst->getDebugLoc();
840 if (DIL && Inst->getFunction()->isDebugInfoForProfiling())
841 B.SetCurrentDebugLocation(DIL->cloneWithDuplicationFactor(UF * VF));
843 B.SetCurrentDebugLocation(DIL);
845 B.SetCurrentDebugLocation(DebugLoc());
849 /// \return string containing a file name and a line # for the given loop.
850 static std::string getDebugLocString(const Loop *L) {
853 raw_string_ostream OS(Result);
854 if (const DebugLoc LoopDbgLoc = L->getStartLoc())
855 LoopDbgLoc.print(OS);
857 // Just print the module name.
858 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
865 void InnerLoopVectorizer::addNewMetadata(Instruction *To,
866 const Instruction *Orig) {
867 // If the loop was versioned with memchecks, add the corresponding no-alias
869 if (LVer && (isa<LoadInst>(Orig) || isa<StoreInst>(Orig)))
870 LVer->annotateInstWithNoAlias(To, Orig);
873 void InnerLoopVectorizer::addMetadata(Instruction *To,
875 propagateMetadata(To, From);
876 addNewMetadata(To, From);
879 void InnerLoopVectorizer::addMetadata(ArrayRef<Value *> To,
881 for (Value *V : To) {
882 if (Instruction *I = dyn_cast<Instruction>(V))
883 addMetadata(I, From);
887 /// \brief The group of interleaved loads/stores sharing the same stride and
888 /// close to each other.
890 /// Each member in this group has an index starting from 0, and the largest
891 /// index should be less than interleaved factor, which is equal to the absolute
892 /// value of the access's stride.
894 /// E.g. An interleaved load group of factor 4:
895 /// for (unsigned i = 0; i < 1024; i+=4) {
896 /// a = A[i]; // Member of index 0
897 /// b = A[i+1]; // Member of index 1
898 /// d = A[i+3]; // Member of index 3
902 /// An interleaved store group of factor 4:
903 /// for (unsigned i = 0; i < 1024; i+=4) {
905 /// A[i] = a; // Member of index 0
906 /// A[i+1] = b; // Member of index 1
907 /// A[i+2] = c; // Member of index 2
908 /// A[i+3] = d; // Member of index 3
911 /// Note: the interleaved load group could have gaps (missing members), but
912 /// the interleaved store group doesn't allow gaps.
913 class InterleaveGroup {
915 InterleaveGroup(Instruction *Instr, int Stride, unsigned Align)
916 : Align(Align), SmallestKey(0), LargestKey(0), InsertPos(Instr) {
917 assert(Align && "The alignment should be non-zero");
919 Factor = std::abs(Stride);
920 assert(Factor > 1 && "Invalid interleave factor");
922 Reverse = Stride < 0;
926 bool isReverse() const { return Reverse; }
927 unsigned getFactor() const { return Factor; }
928 unsigned getAlignment() const { return Align; }
929 unsigned getNumMembers() const { return Members.size(); }
931 /// \brief Try to insert a new member \p Instr with index \p Index and
932 /// alignment \p NewAlign. The index is related to the leader and it could be
933 /// negative if it is the new leader.
935 /// \returns false if the instruction doesn't belong to the group.
936 bool insertMember(Instruction *Instr, int Index, unsigned NewAlign) {
937 assert(NewAlign && "The new member's alignment should be non-zero");
939 int Key = Index + SmallestKey;
941 // Skip if there is already a member with the same index.
942 if (Members.count(Key))
945 if (Key > LargestKey) {
946 // The largest index is always less than the interleave factor.
947 if (Index >= static_cast<int>(Factor))
951 } else if (Key < SmallestKey) {
952 // The largest index is always less than the interleave factor.
953 if (LargestKey - Key >= static_cast<int>(Factor))
959 // It's always safe to select the minimum alignment.
960 Align = std::min(Align, NewAlign);
961 Members[Key] = Instr;
965 /// \brief Get the member with the given index \p Index
967 /// \returns nullptr if contains no such member.
968 Instruction *getMember(unsigned Index) const {
969 int Key = SmallestKey + Index;
970 if (!Members.count(Key))
973 return Members.find(Key)->second;
976 /// \brief Get the index for the given member. Unlike the key in the member
977 /// map, the index starts from 0.
978 unsigned getIndex(Instruction *Instr) const {
979 for (auto I : Members)
980 if (I.second == Instr)
981 return I.first - SmallestKey;
983 llvm_unreachable("InterleaveGroup contains no such member");
986 Instruction *getInsertPos() const { return InsertPos; }
987 void setInsertPos(Instruction *Inst) { InsertPos = Inst; }
990 unsigned Factor; // Interleave Factor.
993 DenseMap<int, Instruction *> Members;
997 // To avoid breaking dependences, vectorized instructions of an interleave
998 // group should be inserted at either the first load or the last store in
1001 // E.g. %even = load i32 // Insert Position
1002 // %add = add i32 %even // Use of %even
1006 // %odd = add i32 // Def of %odd
1007 // store i32 %odd // Insert Position
1008 Instruction *InsertPos;
1011 /// \brief Drive the analysis of interleaved memory accesses in the loop.
1013 /// Use this class to analyze interleaved accesses only when we can vectorize
1014 /// a loop. Otherwise it's meaningless to do analysis as the vectorization
1015 /// on interleaved accesses is unsafe.
1017 /// The analysis collects interleave groups and records the relationships
1018 /// between the member and the group in a map.
1019 class InterleavedAccessInfo {
1021 InterleavedAccessInfo(PredicatedScalarEvolution &PSE, Loop *L,
1022 DominatorTree *DT, LoopInfo *LI)
1023 : PSE(PSE), TheLoop(L), DT(DT), LI(LI), LAI(nullptr),
1024 RequiresScalarEpilogue(false) {}
1026 ~InterleavedAccessInfo() {
1027 SmallSet<InterleaveGroup *, 4> DelSet;
1028 // Avoid releasing a pointer twice.
1029 for (auto &I : InterleaveGroupMap)
1030 DelSet.insert(I.second);
1031 for (auto *Ptr : DelSet)
1035 /// \brief Analyze the interleaved accesses and collect them in interleave
1036 /// groups. Substitute symbolic strides using \p Strides.
1037 void analyzeInterleaving(const ValueToValueMap &Strides);
1039 /// \brief Check if \p Instr belongs to any interleave group.
1040 bool isInterleaved(Instruction *Instr) const {
1041 return InterleaveGroupMap.count(Instr);
1044 /// \brief Return the maximum interleave factor of all interleaved groups.
1045 unsigned getMaxInterleaveFactor() const {
1046 unsigned MaxFactor = 1;
1047 for (auto &Entry : InterleaveGroupMap)
1048 MaxFactor = std::max(MaxFactor, Entry.second->getFactor());
1052 /// \brief Get the interleave group that \p Instr belongs to.
1054 /// \returns nullptr if doesn't have such group.
1055 InterleaveGroup *getInterleaveGroup(Instruction *Instr) const {
1056 if (InterleaveGroupMap.count(Instr))
1057 return InterleaveGroupMap.find(Instr)->second;
1061 /// \brief Returns true if an interleaved group that may access memory
1062 /// out-of-bounds requires a scalar epilogue iteration for correctness.
1063 bool requiresScalarEpilogue() const { return RequiresScalarEpilogue; }
1065 /// \brief Initialize the LoopAccessInfo used for dependence checking.
1066 void setLAI(const LoopAccessInfo *Info) { LAI = Info; }
1069 /// A wrapper around ScalarEvolution, used to add runtime SCEV checks.
1070 /// Simplifies SCEV expressions in the context of existing SCEV assumptions.
1071 /// The interleaved access analysis can also add new predicates (for example
1072 /// by versioning strides of pointers).
1073 PredicatedScalarEvolution &PSE;
1077 const LoopAccessInfo *LAI;
1079 /// True if the loop may contain non-reversed interleaved groups with
1080 /// out-of-bounds accesses. We ensure we don't speculatively access memory
1081 /// out-of-bounds by executing at least one scalar epilogue iteration.
1082 bool RequiresScalarEpilogue;
1084 /// Holds the relationships between the members and the interleave group.
1085 DenseMap<Instruction *, InterleaveGroup *> InterleaveGroupMap;
1087 /// Holds dependences among the memory accesses in the loop. It maps a source
1088 /// access to a set of dependent sink accesses.
1089 DenseMap<Instruction *, SmallPtrSet<Instruction *, 2>> Dependences;
1091 /// \brief The descriptor for a strided memory access.
1092 struct StrideDescriptor {
1093 StrideDescriptor(int64_t Stride, const SCEV *Scev, uint64_t Size,
1095 : Stride(Stride), Scev(Scev), Size(Size), Align(Align) {}
1097 StrideDescriptor() = default;
1099 // The access's stride. It is negative for a reverse access.
1101 const SCEV *Scev = nullptr; // The scalar expression of this access
1102 uint64_t Size = 0; // The size of the memory object.
1103 unsigned Align = 0; // The alignment of this access.
1106 /// \brief A type for holding instructions and their stride descriptors.
1107 typedef std::pair<Instruction *, StrideDescriptor> StrideEntry;
1109 /// \brief Create a new interleave group with the given instruction \p Instr,
1110 /// stride \p Stride and alignment \p Align.
1112 /// \returns the newly created interleave group.
1113 InterleaveGroup *createInterleaveGroup(Instruction *Instr, int Stride,
1115 assert(!InterleaveGroupMap.count(Instr) &&
1116 "Already in an interleaved access group");
1117 InterleaveGroupMap[Instr] = new InterleaveGroup(Instr, Stride, Align);
1118 return InterleaveGroupMap[Instr];
1121 /// \brief Release the group and remove all the relationships.
1122 void releaseGroup(InterleaveGroup *Group) {
1123 for (unsigned i = 0; i < Group->getFactor(); i++)
1124 if (Instruction *Member = Group->getMember(i))
1125 InterleaveGroupMap.erase(Member);
1130 /// \brief Collect all the accesses with a constant stride in program order.
1131 void collectConstStrideAccesses(
1132 MapVector<Instruction *, StrideDescriptor> &AccessStrideInfo,
1133 const ValueToValueMap &Strides);
1135 /// \brief Returns true if \p Stride is allowed in an interleaved group.
1136 static bool isStrided(int Stride) {
1137 unsigned Factor = std::abs(Stride);
1138 return Factor >= 2 && Factor <= MaxInterleaveGroupFactor;
1141 /// \brief Returns true if \p BB is a predicated block.
1142 bool isPredicated(BasicBlock *BB) const {
1143 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
1146 /// \brief Returns true if LoopAccessInfo can be used for dependence queries.
1147 bool areDependencesValid() const {
1148 return LAI && LAI->getDepChecker().getDependences();
1151 /// \brief Returns true if memory accesses \p A and \p B can be reordered, if
1152 /// necessary, when constructing interleaved groups.
1154 /// \p A must precede \p B in program order. We return false if reordering is
1155 /// not necessary or is prevented because \p A and \p B may be dependent.
1156 bool canReorderMemAccessesForInterleavedGroups(StrideEntry *A,
1157 StrideEntry *B) const {
1159 // Code motion for interleaved accesses can potentially hoist strided loads
1160 // and sink strided stores. The code below checks the legality of the
1161 // following two conditions:
1163 // 1. Potentially moving a strided load (B) before any store (A) that
1166 // 2. Potentially moving a strided store (A) after any load or store (B)
1169 // It's legal to reorder A and B if we know there isn't a dependence from A
1170 // to B. Note that this determination is conservative since some
1171 // dependences could potentially be reordered safely.
1173 // A is potentially the source of a dependence.
1174 auto *Src = A->first;
1175 auto SrcDes = A->second;
1177 // B is potentially the sink of a dependence.
1178 auto *Sink = B->first;
1179 auto SinkDes = B->second;
1181 // Code motion for interleaved accesses can't violate WAR dependences.
1182 // Thus, reordering is legal if the source isn't a write.
1183 if (!Src->mayWriteToMemory())
1186 // At least one of the accesses must be strided.
1187 if (!isStrided(SrcDes.Stride) && !isStrided(SinkDes.Stride))
1190 // If dependence information is not available from LoopAccessInfo,
1191 // conservatively assume the instructions can't be reordered.
1192 if (!areDependencesValid())
1195 // If we know there is a dependence from source to sink, assume the
1196 // instructions can't be reordered. Otherwise, reordering is legal.
1197 return !Dependences.count(Src) || !Dependences.lookup(Src).count(Sink);
1200 /// \brief Collect the dependences from LoopAccessInfo.
1202 /// We process the dependences once during the interleaved access analysis to
1203 /// enable constant-time dependence queries.
1204 void collectDependences() {
1205 if (!areDependencesValid())
1207 auto *Deps = LAI->getDepChecker().getDependences();
1208 for (auto Dep : *Deps)
1209 Dependences[Dep.getSource(*LAI)].insert(Dep.getDestination(*LAI));
1213 /// Utility class for getting and setting loop vectorizer hints in the form
1214 /// of loop metadata.
1215 /// This class keeps a number of loop annotations locally (as member variables)
1216 /// and can, upon request, write them back as metadata on the loop. It will
1217 /// initially scan the loop for existing metadata, and will update the local
1218 /// values based on information in the loop.
1219 /// We cannot write all values to metadata, as the mere presence of some info,
1220 /// for example 'force', means a decision has been made. So, we need to be
1221 /// careful NOT to add them if the user hasn't specifically asked so.
1222 class LoopVectorizeHints {
1223 enum HintKind { HK_WIDTH, HK_UNROLL, HK_FORCE };
1225 /// Hint - associates name and validation with the hint value.
1228 unsigned Value; // This may have to change for non-numeric values.
1231 Hint(const char *Name, unsigned Value, HintKind Kind)
1232 : Name(Name), Value(Value), Kind(Kind) {}
1234 bool validate(unsigned Val) {
1237 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
1239 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1247 /// Vectorization width.
1249 /// Vectorization interleave factor.
1251 /// Vectorization forced
1254 /// Return the loop metadata prefix.
1255 static StringRef Prefix() { return "llvm.loop."; }
1257 /// True if there is any unsafe math in the loop.
1258 bool PotentiallyUnsafe;
1262 FK_Undefined = -1, ///< Not selected.
1263 FK_Disabled = 0, ///< Forcing disabled.
1264 FK_Enabled = 1, ///< Forcing enabled.
1267 LoopVectorizeHints(const Loop *L, bool DisableInterleaving,
1268 OptimizationRemarkEmitter &ORE)
1269 : Width("vectorize.width", VectorizerParams::VectorizationFactor,
1271 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1272 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1273 PotentiallyUnsafe(false), TheLoop(L), ORE(ORE) {
1274 // Populate values with existing loop metadata.
1275 getHintsFromMetadata();
1277 // force-vector-interleave overrides DisableInterleaving.
1278 if (VectorizerParams::isInterleaveForced())
1279 Interleave.Value = VectorizerParams::VectorizationInterleave;
1281 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
1282 << "LV: Interleaving disabled by the pass manager\n");
1285 /// Mark the loop L as already vectorized by setting the width to 1.
1286 void setAlreadyVectorized() {
1287 Width.Value = Interleave.Value = 1;
1288 Hint Hints[] = {Width, Interleave};
1289 writeHintsToMetadata(Hints);
1292 bool allowVectorization(Function *F, Loop *L, bool AlwaysVectorize) const {
1293 if (getForce() == LoopVectorizeHints::FK_Disabled) {
1294 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1295 emitRemarkWithHints();
1299 if (!AlwaysVectorize && getForce() != LoopVectorizeHints::FK_Enabled) {
1300 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1301 emitRemarkWithHints();
1305 if (getWidth() == 1 && getInterleave() == 1) {
1306 // FIXME: Add a separate metadata to indicate when the loop has already
1307 // been vectorized instead of setting width and count to 1.
1308 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1309 // FIXME: Add interleave.disable metadata. This will allow
1310 // vectorize.disable to be used without disabling the pass and errors
1311 // to differentiate between disabled vectorization and a width of 1.
1312 ORE.emit(OptimizationRemarkAnalysis(vectorizeAnalysisPassName(),
1313 "AllDisabled", L->getStartLoc(),
1315 << "loop not vectorized: vectorization and interleaving are "
1316 "explicitly disabled, or vectorize width and interleave "
1317 "count are both set to 1");
1324 /// Dumps all the hint information.
1325 void emitRemarkWithHints() const {
1326 using namespace ore;
1327 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1328 ORE.emit(OptimizationRemarkMissed(LV_NAME, "MissedExplicitlyDisabled",
1329 TheLoop->getStartLoc(),
1330 TheLoop->getHeader())
1331 << "loop not vectorized: vectorization is explicitly disabled");
1333 OptimizationRemarkMissed R(LV_NAME, "MissedDetails",
1334 TheLoop->getStartLoc(), TheLoop->getHeader());
1335 R << "loop not vectorized";
1336 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1337 R << " (Force=" << NV("Force", true);
1338 if (Width.Value != 0)
1339 R << ", Vector Width=" << NV("VectorWidth", Width.Value);
1340 if (Interleave.Value != 0)
1341 R << ", Interleave Count=" << NV("InterleaveCount", Interleave.Value);
1348 unsigned getWidth() const { return Width.Value; }
1349 unsigned getInterleave() const { return Interleave.Value; }
1350 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1352 /// \brief If hints are provided that force vectorization, use the AlwaysPrint
1353 /// pass name to force the frontend to print the diagnostic.
1354 const char *vectorizeAnalysisPassName() const {
1355 if (getWidth() == 1)
1357 if (getForce() == LoopVectorizeHints::FK_Disabled)
1359 if (getForce() == LoopVectorizeHints::FK_Undefined && getWidth() == 0)
1361 return OptimizationRemarkAnalysis::AlwaysPrint;
1364 bool allowReordering() const {
1365 // When enabling loop hints are provided we allow the vectorizer to change
1366 // the order of operations that is given by the scalar loop. This is not
1367 // enabled by default because can be unsafe or inefficient. For example,
1368 // reordering floating-point operations will change the way round-off
1369 // error accumulates in the loop.
1370 return getForce() == LoopVectorizeHints::FK_Enabled || getWidth() > 1;
1373 bool isPotentiallyUnsafe() const {
1374 // Avoid FP vectorization if the target is unsure about proper support.
1375 // This may be related to the SIMD unit in the target not handling
1376 // IEEE 754 FP ops properly, or bad single-to-double promotions.
1377 // Otherwise, a sequence of vectorized loops, even without reduction,
1378 // could lead to different end results on the destination vectors.
1379 return getForce() != LoopVectorizeHints::FK_Enabled && PotentiallyUnsafe;
1382 void setPotentiallyUnsafe() { PotentiallyUnsafe = true; }
1385 /// Find hints specified in the loop metadata and update local values.
1386 void getHintsFromMetadata() {
1387 MDNode *LoopID = TheLoop->getLoopID();
1391 // First operand should refer to the loop id itself.
1392 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1393 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1395 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1396 const MDString *S = nullptr;
1397 SmallVector<Metadata *, 4> Args;
1399 // The expected hint is either a MDString or a MDNode with the first
1400 // operand a MDString.
1401 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1402 if (!MD || MD->getNumOperands() == 0)
1404 S = dyn_cast<MDString>(MD->getOperand(0));
1405 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1406 Args.push_back(MD->getOperand(i));
1408 S = dyn_cast<MDString>(LoopID->getOperand(i));
1409 assert(Args.size() == 0 && "too many arguments for MDString");
1415 // Check if the hint starts with the loop metadata prefix.
1416 StringRef Name = S->getString();
1417 if (Args.size() == 1)
1418 setHint(Name, Args[0]);
1422 /// Checks string hint with one operand and set value if valid.
1423 void setHint(StringRef Name, Metadata *Arg) {
1424 if (!Name.startswith(Prefix()))
1426 Name = Name.substr(Prefix().size(), StringRef::npos);
1428 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1431 unsigned Val = C->getZExtValue();
1433 Hint *Hints[] = {&Width, &Interleave, &Force};
1434 for (auto H : Hints) {
1435 if (Name == H->Name) {
1436 if (H->validate(Val))
1439 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1445 /// Create a new hint from name / value pair.
1446 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1447 LLVMContext &Context = TheLoop->getHeader()->getContext();
1448 Metadata *MDs[] = {MDString::get(Context, Name),
1449 ConstantAsMetadata::get(
1450 ConstantInt::get(Type::getInt32Ty(Context), V))};
1451 return MDNode::get(Context, MDs);
1454 /// Matches metadata with hint name.
1455 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1456 MDString *Name = dyn_cast<MDString>(Node->getOperand(0));
1460 for (auto H : HintTypes)
1461 if (Name->getString().endswith(H.Name))
1466 /// Sets current hints into loop metadata, keeping other values intact.
1467 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1468 if (HintTypes.size() == 0)
1471 // Reserve the first element to LoopID (see below).
1472 SmallVector<Metadata *, 4> MDs(1);
1473 // If the loop already has metadata, then ignore the existing operands.
1474 MDNode *LoopID = TheLoop->getLoopID();
1476 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1477 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1478 // If node in update list, ignore old value.
1479 if (!matchesHintMetadataName(Node, HintTypes))
1480 MDs.push_back(Node);
1484 // Now, add the missing hints.
1485 for (auto H : HintTypes)
1486 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1488 // Replace current metadata node with new one.
1489 LLVMContext &Context = TheLoop->getHeader()->getContext();
1490 MDNode *NewLoopID = MDNode::get(Context, MDs);
1491 // Set operand 0 to refer to the loop id itself.
1492 NewLoopID->replaceOperandWith(0, NewLoopID);
1494 TheLoop->setLoopID(NewLoopID);
1497 /// The loop these hints belong to.
1498 const Loop *TheLoop;
1500 /// Interface to emit optimization remarks.
1501 OptimizationRemarkEmitter &ORE;
1504 static void emitMissedWarning(Function *F, Loop *L,
1505 const LoopVectorizeHints &LH,
1506 OptimizationRemarkEmitter *ORE) {
1507 LH.emitRemarkWithHints();
1509 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1510 if (LH.getWidth() != 1)
1511 ORE->emit(DiagnosticInfoOptimizationFailure(
1512 DEBUG_TYPE, "FailedRequestedVectorization",
1513 L->getStartLoc(), L->getHeader())
1514 << "loop not vectorized: "
1515 << "failed explicitly specified loop vectorization");
1516 else if (LH.getInterleave() != 1)
1517 ORE->emit(DiagnosticInfoOptimizationFailure(
1518 DEBUG_TYPE, "FailedRequestedInterleaving", L->getStartLoc(),
1520 << "loop not interleaved: "
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 PrimaryInduction(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 primary induction variable.
1570 PHINode *getPrimaryInduction() { return PrimaryInduction; }
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 the information that we collected about runtime memory check.
1611 const RuntimePointerChecking *getRuntimePointerChecking() const {
1612 return LAI->getRuntimePointerChecking();
1615 const LoopAccessInfo *getLAI() const { return LAI; }
1617 /// \brief Check if \p Instr belongs to any interleaved access group.
1618 bool isAccessInterleaved(Instruction *Instr) {
1619 return InterleaveInfo.isInterleaved(Instr);
1622 /// \brief Return the maximum interleave factor of all interleaved groups.
1623 unsigned getMaxInterleaveFactor() const {
1624 return InterleaveInfo.getMaxInterleaveFactor();
1627 /// \brief Get the interleaved access group that \p Instr belongs to.
1628 const InterleaveGroup *getInterleavedAccessGroup(Instruction *Instr) {
1629 return InterleaveInfo.getInterleaveGroup(Instr);
1632 /// \brief Returns true if an interleaved group requires a scalar iteration
1633 /// to handle accesses with gaps.
1634 bool requiresScalarEpilogue() const {
1635 return InterleaveInfo.requiresScalarEpilogue();
1638 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
1640 bool hasStride(Value *V) { return LAI->hasStride(V); }
1642 /// Returns true if the target machine supports masked store operation
1643 /// for the given \p DataType and kind of access to \p Ptr.
1644 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
1645 return isConsecutivePtr(Ptr) && TTI->isLegalMaskedStore(DataType);
1647 /// Returns true if the target machine supports masked load operation
1648 /// for the given \p DataType and kind of access to \p Ptr.
1649 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
1650 return isConsecutivePtr(Ptr) && TTI->isLegalMaskedLoad(DataType);
1652 /// Returns true if the target machine supports masked scatter operation
1653 /// for the given \p DataType.
1654 bool isLegalMaskedScatter(Type *DataType) {
1655 return TTI->isLegalMaskedScatter(DataType);
1657 /// Returns true if the target machine supports masked gather operation
1658 /// for the given \p DataType.
1659 bool isLegalMaskedGather(Type *DataType) {
1660 return TTI->isLegalMaskedGather(DataType);
1662 /// Returns true if the target machine can represent \p V as a masked gather
1663 /// or scatter operation.
1664 bool isLegalGatherOrScatter(Value *V) {
1665 auto *LI = dyn_cast<LoadInst>(V);
1666 auto *SI = dyn_cast<StoreInst>(V);
1669 auto *Ptr = getPointerOperand(V);
1670 auto *Ty = cast<PointerType>(Ptr->getType())->getElementType();
1671 return (LI && isLegalMaskedGather(Ty)) || (SI && isLegalMaskedScatter(Ty));
1674 /// Returns true if vector representation of the instruction \p I
1676 bool isMaskRequired(const Instruction *I) { return (MaskedOp.count(I) != 0); }
1677 unsigned getNumStores() const { return LAI->getNumStores(); }
1678 unsigned getNumLoads() const { return LAI->getNumLoads(); }
1679 unsigned getNumPredStores() const { return NumPredStores; }
1681 /// Returns true if \p I is an instruction that will be scalarized with
1682 /// predication. Such instructions include conditional stores and
1683 /// instructions that may divide by zero.
1684 bool isScalarWithPredication(Instruction *I);
1686 /// Returns true if \p I is a memory instruction with consecutive memory
1687 /// access that can be widened.
1688 bool memoryInstructionCanBeWidened(Instruction *I, unsigned VF = 1);
1690 // Returns true if the NoNaN attribute is set on the function.
1691 bool hasFunNoNaNAttr() const { return HasFunNoNaNAttr; }
1694 /// Check if a single basic block loop is vectorizable.
1695 /// At this point we know that this is a loop with a constant trip count
1696 /// and we only need to check individual instructions.
1697 bool canVectorizeInstrs();
1699 /// When we vectorize loops we may change the order in which
1700 /// we read and write from memory. This method checks if it is
1701 /// legal to vectorize the code, considering only memory constrains.
1702 /// Returns true if the loop is vectorizable
1703 bool canVectorizeMemory();
1705 /// Return true if we can vectorize this loop using the IF-conversion
1707 bool canVectorizeWithIfConvert();
1709 /// Return true if all of the instructions in the block can be speculatively
1710 /// executed. \p SafePtrs is a list of addresses that are known to be legal
1711 /// and we know that we can read from them without segfault.
1712 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
1714 /// Updates the vectorization state by adding \p Phi to the inductions list.
1715 /// This can set \p Phi as the main induction of the loop if \p Phi is a
1716 /// better choice for the main induction than the existing one.
1717 void addInductionPhi(PHINode *Phi, const InductionDescriptor &ID,
1718 SmallPtrSetImpl<Value *> &AllowedExit);
1720 /// Create an analysis remark that explains why vectorization failed
1722 /// \p RemarkName is the identifier for the remark. If \p I is passed it is
1723 /// an instruction that prevents vectorization. Otherwise the loop is used
1724 /// for the location of the remark. \return the remark object that can be
1726 OptimizationRemarkAnalysis
1727 createMissedAnalysis(StringRef RemarkName, Instruction *I = nullptr) const {
1728 return ::createMissedAnalysis(Hints->vectorizeAnalysisPassName(),
1729 RemarkName, TheLoop, I);
1732 /// \brief If an access has a symbolic strides, this maps the pointer value to
1733 /// the stride symbol.
1734 const ValueToValueMap *getSymbolicStrides() {
1735 // FIXME: Currently, the set of symbolic strides is sometimes queried before
1736 // it's collected. This happens from canVectorizeWithIfConvert, when the
1737 // pointer is checked to reference consecutive elements suitable for a
1739 return LAI ? &LAI->getSymbolicStrides() : nullptr;
1742 unsigned NumPredStores;
1744 /// The loop that we evaluate.
1746 /// A wrapper around ScalarEvolution used to add runtime SCEV checks.
1747 /// Applies dynamic knowledge to simplify SCEV expressions in the context
1748 /// of existing SCEV assumptions. The analysis will also add a minimal set
1749 /// of new predicates if this is required to enable vectorization and
1751 PredicatedScalarEvolution &PSE;
1752 /// Target Library Info.
1753 TargetLibraryInfo *TLI;
1754 /// Target Transform Info
1755 const TargetTransformInfo *TTI;
1758 // LoopAccess analysis.
1759 std::function<const LoopAccessInfo &(Loop &)> *GetLAA;
1760 // And the loop-accesses info corresponding to this loop. This pointer is
1761 // null until canVectorizeMemory sets it up.
1762 const LoopAccessInfo *LAI;
1763 /// Interface to emit optimization remarks.
1764 OptimizationRemarkEmitter *ORE;
1766 /// The interleave access information contains groups of interleaved accesses
1767 /// with the same stride and close to each other.
1768 InterleavedAccessInfo InterleaveInfo;
1770 // --- vectorization state --- //
1772 /// Holds the primary induction variable. This is the counter of the
1774 PHINode *PrimaryInduction;
1775 /// Holds the reduction variables.
1776 ReductionList Reductions;
1777 /// Holds all of the induction variables that we found in the loop.
1778 /// Notice that inductions don't need to start at zero and that induction
1779 /// variables can be pointers.
1780 InductionList Inductions;
1781 /// Holds the phi nodes that are first-order recurrences.
1782 RecurrenceSet FirstOrderRecurrences;
1783 /// Holds the widest induction type encountered.
1786 /// Allowed outside users. This holds the induction and reduction
1787 /// vars which can be accessed from outside the loop.
1788 SmallPtrSet<Value *, 4> AllowedExit;
1790 /// Can we assume the absence of NaNs.
1791 bool HasFunNoNaNAttr;
1793 /// Vectorization requirements that will go through late-evaluation.
1794 LoopVectorizationRequirements *Requirements;
1796 /// Used to emit an analysis of any legality issues.
1797 LoopVectorizeHints *Hints;
1799 /// While vectorizing these instructions we have to generate a
1800 /// call to the appropriate masked intrinsic
1801 SmallPtrSet<const Instruction *, 8> MaskedOp;
1804 /// LoopVectorizationCostModel - estimates the expected speedups due to
1806 /// In many cases vectorization is not profitable. This can happen because of
1807 /// a number of reasons. In this class we mainly attempt to predict the
1808 /// expected speedup/slowdowns due to the supported instruction set. We use the
1809 /// TargetTransformInfo to query the different backends for the cost of
1810 /// different operations.
1811 class LoopVectorizationCostModel {
1813 LoopVectorizationCostModel(Loop *L, PredicatedScalarEvolution &PSE,
1814 LoopInfo *LI, LoopVectorizationLegality *Legal,
1815 const TargetTransformInfo &TTI,
1816 const TargetLibraryInfo *TLI, DemandedBits *DB,
1817 AssumptionCache *AC,
1818 OptimizationRemarkEmitter *ORE, const Function *F,
1819 const LoopVectorizeHints *Hints)
1820 : TheLoop(L), PSE(PSE), LI(LI), Legal(Legal), TTI(TTI), TLI(TLI), DB(DB),
1821 AC(AC), ORE(ORE), TheFunction(F), Hints(Hints) {}
1823 /// \return An upper bound for the vectorization factor, or None if
1824 /// vectorization should be avoided up front.
1825 Optional<unsigned> computeMaxVF(bool OptForSize);
1827 /// Information about vectorization costs
1828 struct VectorizationFactor {
1829 unsigned Width; // Vector width with best cost
1830 unsigned Cost; // Cost of the loop with that width
1832 /// \return The most profitable vectorization factor and the cost of that VF.
1833 /// This method checks every power of two up to MaxVF. If UserVF is not ZERO
1834 /// then this vectorization factor will be selected if vectorization is
1836 VectorizationFactor selectVectorizationFactor(unsigned MaxVF);
1838 /// Setup cost-based decisions for user vectorization factor.
1839 void selectUserVectorizationFactor(unsigned UserVF) {
1840 collectUniformsAndScalars(UserVF);
1841 collectInstsToScalarize(UserVF);
1844 /// \return The size (in bits) of the smallest and widest types in the code
1845 /// that needs to be vectorized. We ignore values that remain scalar such as
1846 /// 64 bit loop indices.
1847 std::pair<unsigned, unsigned> getSmallestAndWidestTypes();
1849 /// \return The desired interleave count.
1850 /// If interleave count has been specified by metadata it will be returned.
1851 /// Otherwise, the interleave count is computed and returned. VF and LoopCost
1852 /// are the selected vectorization factor and the cost of the selected VF.
1853 unsigned selectInterleaveCount(bool OptForSize, unsigned VF,
1856 /// Memory access instruction may be vectorized in more than one way.
1857 /// Form of instruction after vectorization depends on cost.
1858 /// This function takes cost-based decisions for Load/Store instructions
1859 /// and collects them in a map. This decisions map is used for building
1860 /// the lists of loop-uniform and loop-scalar instructions.
1861 /// The calculated cost is saved with widening decision in order to
1862 /// avoid redundant calculations.
1863 void setCostBasedWideningDecision(unsigned VF);
1865 /// \brief A struct that represents some properties of the register usage
1867 struct RegisterUsage {
1868 /// Holds the number of loop invariant values that are used in the loop.
1869 unsigned LoopInvariantRegs;
1870 /// Holds the maximum number of concurrent live intervals in the loop.
1871 unsigned MaxLocalUsers;
1872 /// Holds the number of instructions in the loop.
1873 unsigned NumInstructions;
1876 /// \return Returns information about the register usages of the loop for the
1877 /// given vectorization factors.
1878 SmallVector<RegisterUsage, 8> calculateRegisterUsage(ArrayRef<unsigned> VFs);
1880 /// Collect values we want to ignore in the cost model.
1881 void collectValuesToIgnore();
1883 /// \returns The smallest bitwidth each instruction can be represented with.
1884 /// The vector equivalents of these instructions should be truncated to this
1886 const MapVector<Instruction *, uint64_t> &getMinimalBitwidths() const {
1890 /// \returns True if it is more profitable to scalarize instruction \p I for
1891 /// vectorization factor \p VF.
1892 bool isProfitableToScalarize(Instruction *I, unsigned VF) const {
1893 auto Scalars = InstsToScalarize.find(VF);
1894 assert(Scalars != InstsToScalarize.end() &&
1895 "VF not yet analyzed for scalarization profitability");
1896 return Scalars->second.count(I);
1899 /// Returns true if \p I is known to be uniform after vectorization.
1900 bool isUniformAfterVectorization(Instruction *I, unsigned VF) const {
1903 assert(Uniforms.count(VF) && "VF not yet analyzed for uniformity");
1904 auto UniformsPerVF = Uniforms.find(VF);
1905 return UniformsPerVF->second.count(I);
1908 /// Returns true if \p I is known to be scalar after vectorization.
1909 bool isScalarAfterVectorization(Instruction *I, unsigned VF) const {
1912 assert(Scalars.count(VF) && "Scalar values are not calculated for VF");
1913 auto ScalarsPerVF = Scalars.find(VF);
1914 return ScalarsPerVF->second.count(I);
1917 /// \returns True if instruction \p I can be truncated to a smaller bitwidth
1918 /// for vectorization factor \p VF.
1919 bool canTruncateToMinimalBitwidth(Instruction *I, unsigned VF) const {
1920 return VF > 1 && MinBWs.count(I) && !isProfitableToScalarize(I, VF) &&
1921 !isScalarAfterVectorization(I, VF);
1924 /// Decision that was taken during cost calculation for memory instruction.
1933 /// Save vectorization decision \p W and \p Cost taken by the cost model for
1934 /// instruction \p I and vector width \p VF.
1935 void setWideningDecision(Instruction *I, unsigned VF, InstWidening W,
1937 assert(VF >= 2 && "Expected VF >=2");
1938 WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, Cost);
1941 /// Save vectorization decision \p W and \p Cost taken by the cost model for
1942 /// interleaving group \p Grp and vector width \p VF.
1943 void setWideningDecision(const InterleaveGroup *Grp, unsigned VF,
1944 InstWidening W, unsigned Cost) {
1945 assert(VF >= 2 && "Expected VF >=2");
1946 /// Broadcast this decicion to all instructions inside the group.
1947 /// But the cost will be assigned to one instruction only.
1948 for (unsigned i = 0; i < Grp->getFactor(); ++i) {
1949 if (auto *I = Grp->getMember(i)) {
1950 if (Grp->getInsertPos() == I)
1951 WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, Cost);
1953 WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, 0);
1958 /// Return the cost model decision for the given instruction \p I and vector
1959 /// width \p VF. Return CM_Unknown if this instruction did not pass
1960 /// through the cost modeling.
1961 InstWidening getWideningDecision(Instruction *I, unsigned VF) {
1962 assert(VF >= 2 && "Expected VF >=2");
1963 std::pair<Instruction *, unsigned> InstOnVF = std::make_pair(I, VF);
1964 auto Itr = WideningDecisions.find(InstOnVF);
1965 if (Itr == WideningDecisions.end())
1967 return Itr->second.first;
1970 /// Return the vectorization cost for the given instruction \p I and vector
1972 unsigned getWideningCost(Instruction *I, unsigned VF) {
1973 assert(VF >= 2 && "Expected VF >=2");
1974 std::pair<Instruction *, unsigned> InstOnVF = std::make_pair(I, VF);
1975 assert(WideningDecisions.count(InstOnVF) && "The cost is not calculated");
1976 return WideningDecisions[InstOnVF].second;
1979 /// Return True if instruction \p I is an optimizable truncate whose operand
1980 /// is an induction variable. Such a truncate will be removed by adding a new
1981 /// induction variable with the destination type.
1982 bool isOptimizableIVTruncate(Instruction *I, unsigned VF) {
1984 // If the instruction is not a truncate, return false.
1985 auto *Trunc = dyn_cast<TruncInst>(I);
1989 // Get the source and destination types of the truncate.
1990 Type *SrcTy = ToVectorTy(cast<CastInst>(I)->getSrcTy(), VF);
1991 Type *DestTy = ToVectorTy(cast<CastInst>(I)->getDestTy(), VF);
1993 // If the truncate is free for the given types, return false. Replacing a
1994 // free truncate with an induction variable would add an induction variable
1995 // update instruction to each iteration of the loop. We exclude from this
1996 // check the primary induction variable since it will need an update
1997 // instruction regardless.
1998 Value *Op = Trunc->getOperand(0);
1999 if (Op != Legal->getPrimaryInduction() && TTI.isTruncateFree(SrcTy, DestTy))
2002 // If the truncated value is not an induction variable, return false.
2003 return Legal->isInductionVariable(Op);
2007 /// \return An upper bound for the vectorization factor, larger than zero.
2008 /// One is returned if vectorization should best be avoided due to cost.
2009 unsigned computeFeasibleMaxVF(bool OptForSize);
2011 /// The vectorization cost is a combination of the cost itself and a boolean
2012 /// indicating whether any of the contributing operations will actually
2014 /// vector values after type legalization in the backend. If this latter value
2016 /// false, then all operations will be scalarized (i.e. no vectorization has
2017 /// actually taken place).
2018 typedef std::pair<unsigned, bool> VectorizationCostTy;
2020 /// Returns the expected execution cost. The unit of the cost does
2021 /// not matter because we use the 'cost' units to compare different
2022 /// vector widths. The cost that is returned is *not* normalized by
2023 /// the factor width.
2024 VectorizationCostTy expectedCost(unsigned VF);
2026 /// Returns the execution time cost of an instruction for a given vector
2027 /// width. Vector width of one means scalar.
2028 VectorizationCostTy getInstructionCost(Instruction *I, unsigned VF);
2030 /// The cost-computation logic from getInstructionCost which provides
2031 /// the vector type as an output parameter.
2032 unsigned getInstructionCost(Instruction *I, unsigned VF, Type *&VectorTy);
2034 /// Calculate vectorization cost of memory instruction \p I.
2035 unsigned getMemoryInstructionCost(Instruction *I, unsigned VF);
2037 /// The cost computation for scalarized memory instruction.
2038 unsigned getMemInstScalarizationCost(Instruction *I, unsigned VF);
2040 /// The cost computation for interleaving group of memory instructions.
2041 unsigned getInterleaveGroupCost(Instruction *I, unsigned VF);
2043 /// The cost computation for Gather/Scatter instruction.
2044 unsigned getGatherScatterCost(Instruction *I, unsigned VF);
2046 /// The cost computation for widening instruction \p I with consecutive
2048 unsigned getConsecutiveMemOpCost(Instruction *I, unsigned VF);
2050 /// The cost calculation for Load instruction \p I with uniform pointer -
2051 /// scalar load + broadcast.
2052 unsigned getUniformMemOpCost(Instruction *I, unsigned VF);
2054 /// Returns whether the instruction is a load or store and will be a emitted
2055 /// as a vector operation.
2056 bool isConsecutiveLoadOrStore(Instruction *I);
2058 /// Create an analysis remark that explains why vectorization failed
2060 /// \p RemarkName is the identifier for the remark. \return the remark object
2061 /// that can be streamed to.
2062 OptimizationRemarkAnalysis createMissedAnalysis(StringRef RemarkName) {
2063 return ::createMissedAnalysis(Hints->vectorizeAnalysisPassName(),
2064 RemarkName, TheLoop);
2067 /// Map of scalar integer values to the smallest bitwidth they can be legally
2068 /// represented as. The vector equivalents of these values should be truncated
2070 MapVector<Instruction *, uint64_t> MinBWs;
2072 /// A type representing the costs for instructions if they were to be
2073 /// scalarized rather than vectorized. The entries are Instruction-Cost
2075 typedef DenseMap<Instruction *, unsigned> ScalarCostsTy;
2077 /// A set containing all BasicBlocks that are known to present after
2078 /// vectorization as a predicated block.
2079 SmallPtrSet<BasicBlock *, 4> PredicatedBBsAfterVectorization;
2081 /// A map holding scalar costs for different vectorization factors. The
2082 /// presence of a cost for an instruction in the mapping indicates that the
2083 /// instruction will be scalarized when vectorizing with the associated
2084 /// vectorization factor. The entries are VF-ScalarCostTy pairs.
2085 DenseMap<unsigned, ScalarCostsTy> InstsToScalarize;
2087 /// Holds the instructions known to be uniform after vectorization.
2088 /// The data is collected per VF.
2089 DenseMap<unsigned, SmallPtrSet<Instruction *, 4>> Uniforms;
2091 /// Holds the instructions known to be scalar after vectorization.
2092 /// The data is collected per VF.
2093 DenseMap<unsigned, SmallPtrSet<Instruction *, 4>> Scalars;
2095 /// Returns the expected difference in cost from scalarizing the expression
2096 /// feeding a predicated instruction \p PredInst. The instructions to
2097 /// scalarize and their scalar costs are collected in \p ScalarCosts. A
2098 /// non-negative return value implies the expression will be scalarized.
2099 /// Currently, only single-use chains are considered for scalarization.
2100 int computePredInstDiscount(Instruction *PredInst, ScalarCostsTy &ScalarCosts,
2103 /// Collects the instructions to scalarize for each predicated instruction in
2105 void collectInstsToScalarize(unsigned VF);
2107 /// Collect the instructions that are uniform after vectorization. An
2108 /// instruction is uniform if we represent it with a single scalar value in
2109 /// the vectorized loop corresponding to each vector iteration. Examples of
2110 /// uniform instructions include pointer operands of consecutive or
2111 /// interleaved memory accesses. Note that although uniformity implies an
2112 /// instruction will be scalar, the reverse is not true. In general, a
2113 /// scalarized instruction will be represented by VF scalar values in the
2114 /// vectorized loop, each corresponding to an iteration of the original
2116 void collectLoopUniforms(unsigned VF);
2118 /// Collect the instructions that are scalar after vectorization. An
2119 /// instruction is scalar if it is known to be uniform or will be scalarized
2120 /// during vectorization. Non-uniform scalarized instructions will be
2121 /// represented by VF values in the vectorized loop, each corresponding to an
2122 /// iteration of the original scalar loop.
2123 void collectLoopScalars(unsigned VF);
2125 /// Collect Uniform and Scalar values for the given \p VF.
2126 /// The sets depend on CM decision for Load/Store instructions
2127 /// that may be vectorized as interleave, gather-scatter or scalarized.
2128 void collectUniformsAndScalars(unsigned VF) {
2129 // Do the analysis once.
2130 if (VF == 1 || Uniforms.count(VF))
2132 setCostBasedWideningDecision(VF);
2133 collectLoopUniforms(VF);
2134 collectLoopScalars(VF);
2137 /// Keeps cost model vectorization decision and cost for instructions.
2138 /// Right now it is used for memory instructions only.
2139 typedef DenseMap<std::pair<Instruction *, unsigned>,
2140 std::pair<InstWidening, unsigned>>
2143 DecisionList WideningDecisions;
2146 /// The loop that we evaluate.
2148 /// Predicated scalar evolution analysis.
2149 PredicatedScalarEvolution &PSE;
2150 /// Loop Info analysis.
2152 /// Vectorization legality.
2153 LoopVectorizationLegality *Legal;
2154 /// Vector target information.
2155 const TargetTransformInfo &TTI;
2156 /// Target Library Info.
2157 const TargetLibraryInfo *TLI;
2158 /// Demanded bits analysis.
2160 /// Assumption cache.
2161 AssumptionCache *AC;
2162 /// Interface to emit optimization remarks.
2163 OptimizationRemarkEmitter *ORE;
2165 const Function *TheFunction;
2166 /// Loop Vectorize Hint.
2167 const LoopVectorizeHints *Hints;
2168 /// Values to ignore in the cost model.
2169 SmallPtrSet<const Value *, 16> ValuesToIgnore;
2170 /// Values to ignore in the cost model when VF > 1.
2171 SmallPtrSet<const Value *, 16> VecValuesToIgnore;
2174 /// LoopVectorizationPlanner - drives the vectorization process after having
2175 /// passed Legality checks.
2176 class LoopVectorizationPlanner {
2178 LoopVectorizationPlanner(Loop *OrigLoop, LoopInfo *LI,
2179 LoopVectorizationLegality *Legal,
2180 LoopVectorizationCostModel &CM)
2181 : OrigLoop(OrigLoop), LI(LI), Legal(Legal), CM(CM) {}
2183 ~LoopVectorizationPlanner() {}
2185 /// Plan how to best vectorize, return the best VF and its cost.
2186 LoopVectorizationCostModel::VectorizationFactor plan(bool OptForSize,
2189 /// Generate the IR code for the vectorized loop.
2190 void executePlan(InnerLoopVectorizer &ILV);
2193 /// Collect the instructions from the original loop that would be trivially
2194 /// dead in the vectorized loop if generated.
2195 void collectTriviallyDeadInstructions(
2196 SmallPtrSetImpl<Instruction *> &DeadInstructions);
2199 /// The loop that we evaluate.
2202 /// Loop Info analysis.
2205 /// The legality analysis.
2206 LoopVectorizationLegality *Legal;
2208 /// The profitablity analysis.
2209 LoopVectorizationCostModel &CM;
2212 /// \brief This holds vectorization requirements that must be verified late in
2213 /// the process. The requirements are set by legalize and costmodel. Once
2214 /// vectorization has been determined to be possible and profitable the
2215 /// requirements can be verified by looking for metadata or compiler options.
2216 /// For example, some loops require FP commutativity which is only allowed if
2217 /// vectorization is explicitly specified or if the fast-math compiler option
2218 /// has been provided.
2219 /// Late evaluation of these requirements allows helpful diagnostics to be
2220 /// composed that tells the user what need to be done to vectorize the loop. For
2221 /// example, by specifying #pragma clang loop vectorize or -ffast-math. Late
2222 /// evaluation should be used only when diagnostics can generated that can be
2223 /// followed by a non-expert user.
2224 class LoopVectorizationRequirements {
2226 LoopVectorizationRequirements(OptimizationRemarkEmitter &ORE)
2227 : NumRuntimePointerChecks(0), UnsafeAlgebraInst(nullptr), ORE(ORE) {}
2229 void addUnsafeAlgebraInst(Instruction *I) {
2230 // First unsafe algebra instruction.
2231 if (!UnsafeAlgebraInst)
2232 UnsafeAlgebraInst = I;
2235 void addRuntimePointerChecks(unsigned Num) { NumRuntimePointerChecks = Num; }
2237 bool doesNotMeet(Function *F, Loop *L, const LoopVectorizeHints &Hints) {
2238 const char *PassName = Hints.vectorizeAnalysisPassName();
2239 bool Failed = false;
2240 if (UnsafeAlgebraInst && !Hints.allowReordering()) {
2242 OptimizationRemarkAnalysisFPCommute(PassName, "CantReorderFPOps",
2243 UnsafeAlgebraInst->getDebugLoc(),
2244 UnsafeAlgebraInst->getParent())
2245 << "loop not vectorized: cannot prove it is safe to reorder "
2246 "floating-point operations");
2250 // Test if runtime memcheck thresholds are exceeded.
2251 bool PragmaThresholdReached =
2252 NumRuntimePointerChecks > PragmaVectorizeMemoryCheckThreshold;
2253 bool ThresholdReached =
2254 NumRuntimePointerChecks > VectorizerParams::RuntimeMemoryCheckThreshold;
2255 if ((ThresholdReached && !Hints.allowReordering()) ||
2256 PragmaThresholdReached) {
2257 ORE.emit(OptimizationRemarkAnalysisAliasing(PassName, "CantReorderMemOps",
2260 << "loop not vectorized: cannot prove it is safe to reorder "
2261 "memory operations");
2262 DEBUG(dbgs() << "LV: Too many memory checks needed.\n");
2270 unsigned NumRuntimePointerChecks;
2271 Instruction *UnsafeAlgebraInst;
2273 /// Interface to emit optimization remarks.
2274 OptimizationRemarkEmitter &ORE;
2277 static void addAcyclicInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
2279 if (!hasCyclesInLoopBody(L))
2283 for (Loop *InnerL : L)
2284 addAcyclicInnerLoop(*InnerL, V);
2287 /// The LoopVectorize Pass.
2288 struct LoopVectorize : public FunctionPass {
2289 /// Pass identification, replacement for typeid
2292 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
2293 : FunctionPass(ID) {
2294 Impl.DisableUnrolling = NoUnrolling;
2295 Impl.AlwaysVectorize = AlwaysVectorize;
2296 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
2299 LoopVectorizePass Impl;
2301 bool runOnFunction(Function &F) override {
2302 if (skipFunction(F))
2305 auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
2306 auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
2307 auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2308 auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2309 auto *BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI();
2310 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
2311 auto *TLI = TLIP ? &TLIP->getTLI() : nullptr;
2312 auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
2313 auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
2314 auto *LAA = &getAnalysis<LoopAccessLegacyAnalysis>();
2315 auto *DB = &getAnalysis<DemandedBitsWrapperPass>().getDemandedBits();
2316 auto *ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
2318 std::function<const LoopAccessInfo &(Loop &)> GetLAA =
2319 [&](Loop &L) -> const LoopAccessInfo & { return LAA->getInfo(&L); };
2321 return Impl.runImpl(F, *SE, *LI, *TTI, *DT, *BFI, TLI, *DB, *AA, *AC,
2325 void getAnalysisUsage(AnalysisUsage &AU) const override {
2326 AU.addRequired<AssumptionCacheTracker>();
2327 AU.addRequired<BlockFrequencyInfoWrapperPass>();
2328 AU.addRequired<DominatorTreeWrapperPass>();
2329 AU.addRequired<LoopInfoWrapperPass>();
2330 AU.addRequired<ScalarEvolutionWrapperPass>();
2331 AU.addRequired<TargetTransformInfoWrapperPass>();
2332 AU.addRequired<AAResultsWrapperPass>();
2333 AU.addRequired<LoopAccessLegacyAnalysis>();
2334 AU.addRequired<DemandedBitsWrapperPass>();
2335 AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
2336 AU.addPreserved<LoopInfoWrapperPass>();
2337 AU.addPreserved<DominatorTreeWrapperPass>();
2338 AU.addPreserved<BasicAAWrapperPass>();
2339 AU.addPreserved<GlobalsAAWrapperPass>();
2343 } // end anonymous namespace
2345 //===----------------------------------------------------------------------===//
2346 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
2347 // LoopVectorizationCostModel and LoopVectorizationPlanner.
2348 //===----------------------------------------------------------------------===//
2350 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
2351 // We need to place the broadcast of invariant variables outside the loop.
2352 Instruction *Instr = dyn_cast<Instruction>(V);
2353 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
2354 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
2356 // Place the code for broadcasting invariant variables in the new preheader.
2357 IRBuilder<>::InsertPointGuard Guard(Builder);
2359 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
2361 // Broadcast the scalar into all locations in the vector.
2362 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
2367 void InnerLoopVectorizer::createVectorIntOrFpInductionPHI(
2368 const InductionDescriptor &II, Value *Step, Instruction *EntryVal) {
2369 Value *Start = II.getStartValue();
2371 // Construct the initial value of the vector IV in the vector loop preheader
2372 auto CurrIP = Builder.saveIP();
2373 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
2374 if (isa<TruncInst>(EntryVal)) {
2375 assert(Start->getType()->isIntegerTy() &&
2376 "Truncation requires an integer type");
2377 auto *TruncType = cast<IntegerType>(EntryVal->getType());
2378 Step = Builder.CreateTrunc(Step, TruncType);
2379 Start = Builder.CreateCast(Instruction::Trunc, Start, TruncType);
2381 Value *SplatStart = Builder.CreateVectorSplat(VF, Start);
2382 Value *SteppedStart =
2383 getStepVector(SplatStart, 0, Step, II.getInductionOpcode());
2385 // We create vector phi nodes for both integer and floating-point induction
2386 // variables. Here, we determine the kind of arithmetic we will perform.
2387 Instruction::BinaryOps AddOp;
2388 Instruction::BinaryOps MulOp;
2389 if (Step->getType()->isIntegerTy()) {
2390 AddOp = Instruction::Add;
2391 MulOp = Instruction::Mul;
2393 AddOp = II.getInductionOpcode();
2394 MulOp = Instruction::FMul;
2397 // Multiply the vectorization factor by the step using integer or
2398 // floating-point arithmetic as appropriate.
2399 Value *ConstVF = getSignedIntOrFpConstant(Step->getType(), VF);
2400 Value *Mul = addFastMathFlag(Builder.CreateBinOp(MulOp, Step, ConstVF));
2402 // Create a vector splat to use in the induction update.
2404 // FIXME: If the step is non-constant, we create the vector splat with
2405 // IRBuilder. IRBuilder can constant-fold the multiply, but it doesn't
2406 // handle a constant vector splat.
2407 Value *SplatVF = isa<Constant>(Mul)
2408 ? ConstantVector::getSplat(VF, cast<Constant>(Mul))
2409 : Builder.CreateVectorSplat(VF, Mul);
2410 Builder.restoreIP(CurrIP);
2412 // We may need to add the step a number of times, depending on the unroll
2413 // factor. The last of those goes into the PHI.
2414 PHINode *VecInd = PHINode::Create(SteppedStart->getType(), 2, "vec.ind",
2415 &*LoopVectorBody->getFirstInsertionPt());
2416 Instruction *LastInduction = VecInd;
2417 VectorParts Entry(UF);
2418 for (unsigned Part = 0; Part < UF; ++Part) {
2419 Entry[Part] = LastInduction;
2420 LastInduction = cast<Instruction>(addFastMathFlag(
2421 Builder.CreateBinOp(AddOp, LastInduction, SplatVF, "step.add")));
2423 VectorLoopValueMap.initVector(EntryVal, Entry);
2424 if (isa<TruncInst>(EntryVal))
2425 addMetadata(Entry, EntryVal);
2427 // Move the last step to the end of the latch block. This ensures consistent
2428 // placement of all induction updates.
2429 auto *LoopVectorLatch = LI->getLoopFor(LoopVectorBody)->getLoopLatch();
2430 auto *Br = cast<BranchInst>(LoopVectorLatch->getTerminator());
2431 auto *ICmp = cast<Instruction>(Br->getCondition());
2432 LastInduction->moveBefore(ICmp);
2433 LastInduction->setName("vec.ind.next");
2435 VecInd->addIncoming(SteppedStart, LoopVectorPreHeader);
2436 VecInd->addIncoming(LastInduction, LoopVectorLatch);
2439 bool InnerLoopVectorizer::shouldScalarizeInstruction(Instruction *I) const {
2440 return Cost->isScalarAfterVectorization(I, VF) ||
2441 Cost->isProfitableToScalarize(I, VF);
2444 bool InnerLoopVectorizer::needsScalarInduction(Instruction *IV) const {
2445 if (shouldScalarizeInstruction(IV))
2447 auto isScalarInst = [&](User *U) -> bool {
2448 auto *I = cast<Instruction>(U);
2449 return (OrigLoop->contains(I) && shouldScalarizeInstruction(I));
2451 return any_of(IV->users(), isScalarInst);
2454 void InnerLoopVectorizer::widenIntOrFpInduction(PHINode *IV, TruncInst *Trunc) {
2456 assert((IV->getType()->isIntegerTy() || IV != OldInduction) &&
2457 "Primary induction variable must have an integer type");
2459 auto II = Legal->getInductionVars()->find(IV);
2460 assert(II != Legal->getInductionVars()->end() && "IV is not an induction");
2462 auto ID = II->second;
2463 assert(IV->getType() == ID.getStartValue()->getType() && "Types must match");
2465 // The scalar value to broadcast. This will be derived from the canonical
2466 // induction variable.
2467 Value *ScalarIV = nullptr;
2469 // The value from the original loop to which we are mapping the new induction
2471 Instruction *EntryVal = Trunc ? cast<Instruction>(Trunc) : IV;
2473 // True if we have vectorized the induction variable.
2474 auto VectorizedIV = false;
2476 // Determine if we want a scalar version of the induction variable. This is
2477 // true if the induction variable itself is not widened, or if it has at
2478 // least one user in the loop that is not widened.
2479 auto NeedsScalarIV = VF > 1 && needsScalarInduction(EntryVal);
2481 // Generate code for the induction step. Note that induction steps are
2482 // required to be loop-invariant
2483 assert(PSE.getSE()->isLoopInvariant(ID.getStep(), OrigLoop) &&
2484 "Induction step should be loop invariant");
2485 auto &DL = OrigLoop->getHeader()->getModule()->getDataLayout();
2486 Value *Step = nullptr;
2487 if (PSE.getSE()->isSCEVable(IV->getType())) {
2488 SCEVExpander Exp(*PSE.getSE(), DL, "induction");
2489 Step = Exp.expandCodeFor(ID.getStep(), ID.getStep()->getType(),
2490 LoopVectorPreHeader->getTerminator());
2492 Step = cast<SCEVUnknown>(ID.getStep())->getValue();
2495 // Try to create a new independent vector induction variable. If we can't
2496 // create the phi node, we will splat the scalar induction variable in each
2498 if (VF > 1 && !shouldScalarizeInstruction(EntryVal)) {
2499 createVectorIntOrFpInductionPHI(ID, Step, EntryVal);
2500 VectorizedIV = true;
2503 // If we haven't yet vectorized the induction variable, or if we will create
2504 // a scalar one, we need to define the scalar induction variable and step
2505 // values. If we were given a truncation type, truncate the canonical
2506 // induction variable and step. Otherwise, derive these values from the
2507 // induction descriptor.
2508 if (!VectorizedIV || NeedsScalarIV) {
2509 ScalarIV = Induction;
2510 if (IV != OldInduction) {
2511 ScalarIV = IV->getType()->isIntegerTy()
2512 ? Builder.CreateSExtOrTrunc(Induction, IV->getType())
2513 : Builder.CreateCast(Instruction::SIToFP, Induction,
2515 ScalarIV = ID.transform(Builder, ScalarIV, PSE.getSE(), DL);
2516 ScalarIV->setName("offset.idx");
2519 auto *TruncType = cast<IntegerType>(Trunc->getType());
2520 assert(Step->getType()->isIntegerTy() &&
2521 "Truncation requires an integer step");
2522 ScalarIV = Builder.CreateTrunc(ScalarIV, TruncType);
2523 Step = Builder.CreateTrunc(Step, TruncType);
2527 // If we haven't yet vectorized the induction variable, splat the scalar
2528 // induction variable, and build the necessary step vectors.
2529 if (!VectorizedIV) {
2530 Value *Broadcasted = getBroadcastInstrs(ScalarIV);
2531 VectorParts Entry(UF);
2532 for (unsigned Part = 0; Part < UF; ++Part)
2534 getStepVector(Broadcasted, VF * Part, Step, ID.getInductionOpcode());
2535 VectorLoopValueMap.initVector(EntryVal, Entry);
2537 addMetadata(Entry, Trunc);
2540 // If an induction variable is only used for counting loop iterations or
2541 // calculating addresses, it doesn't need to be widened. Create scalar steps
2542 // that can be used by instructions we will later scalarize. Note that the
2543 // addition of the scalar steps will not increase the number of instructions
2544 // in the loop in the common case prior to InstCombine. We will be trading
2545 // one vector extract for each scalar step.
2547 buildScalarSteps(ScalarIV, Step, EntryVal, ID);
2550 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx, Value *Step,
2551 Instruction::BinaryOps BinOp) {
2552 // Create and check the types.
2553 assert(Val->getType()->isVectorTy() && "Must be a vector");
2554 int VLen = Val->getType()->getVectorNumElements();
2556 Type *STy = Val->getType()->getScalarType();
2557 assert((STy->isIntegerTy() || STy->isFloatingPointTy()) &&
2558 "Induction Step must be an integer or FP");
2559 assert(Step->getType() == STy && "Step has wrong type");
2561 SmallVector<Constant *, 8> Indices;
2563 if (STy->isIntegerTy()) {
2564 // Create a vector of consecutive numbers from zero to VF.
2565 for (int i = 0; i < VLen; ++i)
2566 Indices.push_back(ConstantInt::get(STy, StartIdx + i));
2568 // Add the consecutive indices to the vector value.
2569 Constant *Cv = ConstantVector::get(Indices);
2570 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
2571 Step = Builder.CreateVectorSplat(VLen, Step);
2572 assert(Step->getType() == Val->getType() && "Invalid step vec");
2573 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
2574 // which can be found from the original scalar operations.
2575 Step = Builder.CreateMul(Cv, Step);
2576 return Builder.CreateAdd(Val, Step, "induction");
2579 // Floating point induction.
2580 assert((BinOp == Instruction::FAdd || BinOp == Instruction::FSub) &&
2581 "Binary Opcode should be specified for FP induction");
2582 // Create a vector of consecutive numbers from zero to VF.
2583 for (int i = 0; i < VLen; ++i)
2584 Indices.push_back(ConstantFP::get(STy, (double)(StartIdx + i)));
2586 // Add the consecutive indices to the vector value.
2587 Constant *Cv = ConstantVector::get(Indices);
2589 Step = Builder.CreateVectorSplat(VLen, Step);
2591 // Floating point operations had to be 'fast' to enable the induction.
2592 FastMathFlags Flags;
2593 Flags.setUnsafeAlgebra();
2595 Value *MulOp = Builder.CreateFMul(Cv, Step);
2596 if (isa<Instruction>(MulOp))
2597 // Have to check, MulOp may be a constant
2598 cast<Instruction>(MulOp)->setFastMathFlags(Flags);
2600 Value *BOp = Builder.CreateBinOp(BinOp, Val, MulOp, "induction");
2601 if (isa<Instruction>(BOp))
2602 cast<Instruction>(BOp)->setFastMathFlags(Flags);
2606 void InnerLoopVectorizer::buildScalarSteps(Value *ScalarIV, Value *Step,
2608 const InductionDescriptor &ID) {
2610 // We shouldn't have to build scalar steps if we aren't vectorizing.
2611 assert(VF > 1 && "VF should be greater than one");
2613 // Get the value type and ensure it and the step have the same integer type.
2614 Type *ScalarIVTy = ScalarIV->getType()->getScalarType();
2615 assert(ScalarIVTy == Step->getType() &&
2616 "Val and Step should have the same type");
2618 // We build scalar steps for both integer and floating-point induction
2619 // variables. Here, we determine the kind of arithmetic we will perform.
2620 Instruction::BinaryOps AddOp;
2621 Instruction::BinaryOps MulOp;
2622 if (ScalarIVTy->isIntegerTy()) {
2623 AddOp = Instruction::Add;
2624 MulOp = Instruction::Mul;
2626 AddOp = ID.getInductionOpcode();
2627 MulOp = Instruction::FMul;
2630 // Determine the number of scalars we need to generate for each unroll
2631 // iteration. If EntryVal is uniform, we only need to generate the first
2632 // lane. Otherwise, we generate all VF values.
2634 Cost->isUniformAfterVectorization(cast<Instruction>(EntryVal), VF) ? 1 : VF;
2636 // Compute the scalar steps and save the results in VectorLoopValueMap.
2637 ScalarParts Entry(UF);
2638 for (unsigned Part = 0; Part < UF; ++Part) {
2639 Entry[Part].resize(VF);
2640 for (unsigned Lane = 0; Lane < Lanes; ++Lane) {
2641 auto *StartIdx = getSignedIntOrFpConstant(ScalarIVTy, VF * Part + Lane);
2642 auto *Mul = addFastMathFlag(Builder.CreateBinOp(MulOp, StartIdx, Step));
2643 auto *Add = addFastMathFlag(Builder.CreateBinOp(AddOp, ScalarIV, Mul));
2644 Entry[Part][Lane] = Add;
2647 VectorLoopValueMap.initScalar(EntryVal, Entry);
2650 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
2652 const ValueToValueMap &Strides = getSymbolicStrides() ? *getSymbolicStrides() :
2655 int Stride = getPtrStride(PSE, Ptr, TheLoop, Strides, true, false);
2656 if (Stride == 1 || Stride == -1)
2661 bool LoopVectorizationLegality::isUniform(Value *V) {
2662 return LAI->isUniform(V);
2665 const InnerLoopVectorizer::VectorParts &
2666 InnerLoopVectorizer::getVectorValue(Value *V) {
2667 assert(V != Induction && "The new induction variable should not be used.");
2668 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
2669 assert(!V->getType()->isVoidTy() && "Type does not produce a value");
2671 // If we have a stride that is replaced by one, do it here.
2672 if (Legal->hasStride(V))
2673 V = ConstantInt::get(V->getType(), 1);
2675 // If we have this scalar in the map, return it.
2676 if (VectorLoopValueMap.hasVector(V))
2677 return VectorLoopValueMap.VectorMapStorage[V];
2679 // If the value has not been vectorized, check if it has been scalarized
2680 // instead. If it has been scalarized, and we actually need the value in
2681 // vector form, we will construct the vector values on demand.
2682 if (VectorLoopValueMap.hasScalar(V)) {
2684 // Initialize a new vector map entry.
2685 VectorParts Entry(UF);
2687 // If we've scalarized a value, that value should be an instruction.
2688 auto *I = cast<Instruction>(V);
2690 // If we aren't vectorizing, we can just copy the scalar map values over to
2693 for (unsigned Part = 0; Part < UF; ++Part)
2694 Entry[Part] = getScalarValue(V, Part, 0);
2695 return VectorLoopValueMap.initVector(V, Entry);
2698 // Get the last scalar instruction we generated for V. If the value is
2699 // known to be uniform after vectorization, this corresponds to lane zero
2700 // of the last unroll iteration. Otherwise, the last instruction is the one
2701 // we created for the last vector lane of the last unroll iteration.
2702 unsigned LastLane = Cost->isUniformAfterVectorization(I, VF) ? 0 : VF - 1;
2703 auto *LastInst = cast<Instruction>(getScalarValue(V, UF - 1, LastLane));
2705 // Set the insert point after the last scalarized instruction. This ensures
2706 // the insertelement sequence will directly follow the scalar definitions.
2707 auto OldIP = Builder.saveIP();
2708 auto NewIP = std::next(BasicBlock::iterator(LastInst));
2709 Builder.SetInsertPoint(&*NewIP);
2711 // However, if we are vectorizing, we need to construct the vector values.
2712 // If the value is known to be uniform after vectorization, we can just
2713 // broadcast the scalar value corresponding to lane zero for each unroll
2714 // iteration. Otherwise, we construct the vector values using insertelement
2715 // instructions. Since the resulting vectors are stored in
2716 // VectorLoopValueMap, we will only generate the insertelements once.
2717 for (unsigned Part = 0; Part < UF; ++Part) {
2718 Value *VectorValue = nullptr;
2719 if (Cost->isUniformAfterVectorization(I, VF)) {
2720 VectorValue = getBroadcastInstrs(getScalarValue(V, Part, 0));
2722 VectorValue = UndefValue::get(VectorType::get(V->getType(), VF));
2723 for (unsigned Lane = 0; Lane < VF; ++Lane)
2724 VectorValue = Builder.CreateInsertElement(
2725 VectorValue, getScalarValue(V, Part, Lane),
2726 Builder.getInt32(Lane));
2728 Entry[Part] = VectorValue;
2730 Builder.restoreIP(OldIP);
2731 return VectorLoopValueMap.initVector(V, Entry);
2734 // If this scalar is unknown, assume that it is a constant or that it is
2735 // loop invariant. Broadcast V and save the value for future uses.
2736 Value *B = getBroadcastInstrs(V);
2737 return VectorLoopValueMap.initVector(V, VectorParts(UF, B));
2740 Value *InnerLoopVectorizer::getScalarValue(Value *V, unsigned Part,
2743 // If the value is not an instruction contained in the loop, it should
2744 // already be scalar.
2745 if (OrigLoop->isLoopInvariant(V))
2749 !Cost->isUniformAfterVectorization(cast<Instruction>(V), VF)
2750 : true && "Uniform values only have lane zero");
2752 // If the value from the original loop has not been vectorized, it is
2753 // represented by UF x VF scalar values in the new loop. Return the requested
2755 if (VectorLoopValueMap.hasScalar(V))
2756 return VectorLoopValueMap.ScalarMapStorage[V][Part][Lane];
2758 // If the value has not been scalarized, get its entry in VectorLoopValueMap
2759 // for the given unroll part. If this entry is not a vector type (i.e., the
2760 // vectorization factor is one), there is no need to generate an
2761 // extractelement instruction.
2762 auto *U = getVectorValue(V)[Part];
2763 if (!U->getType()->isVectorTy()) {
2764 assert(VF == 1 && "Value not scalarized has non-vector type");
2768 // Otherwise, the value from the original loop has been vectorized and is
2769 // represented by UF vector values. Extract and return the requested scalar
2770 // value from the appropriate vector lane.
2771 return Builder.CreateExtractElement(U, Builder.getInt32(Lane));
2774 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
2775 assert(Vec->getType()->isVectorTy() && "Invalid type");
2776 SmallVector<Constant *, 8> ShuffleMask;
2777 for (unsigned i = 0; i < VF; ++i)
2778 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
2780 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
2781 ConstantVector::get(ShuffleMask),
2785 // Try to vectorize the interleave group that \p Instr belongs to.
2787 // E.g. Translate following interleaved load group (factor = 3):
2788 // for (i = 0; i < N; i+=3) {
2789 // R = Pic[i]; // Member of index 0
2790 // G = Pic[i+1]; // Member of index 1
2791 // B = Pic[i+2]; // Member of index 2
2792 // ... // do something to R, G, B
2795 // %wide.vec = load <12 x i32> ; Read 4 tuples of R,G,B
2796 // %R.vec = shuffle %wide.vec, undef, <0, 3, 6, 9> ; R elements
2797 // %G.vec = shuffle %wide.vec, undef, <1, 4, 7, 10> ; G elements
2798 // %B.vec = shuffle %wide.vec, undef, <2, 5, 8, 11> ; B elements
2800 // Or translate following interleaved store group (factor = 3):
2801 // for (i = 0; i < N; i+=3) {
2802 // ... do something to R, G, B
2803 // Pic[i] = R; // Member of index 0
2804 // Pic[i+1] = G; // Member of index 1
2805 // Pic[i+2] = B; // Member of index 2
2808 // %R_G.vec = shuffle %R.vec, %G.vec, <0, 1, 2, ..., 7>
2809 // %B_U.vec = shuffle %B.vec, undef, <0, 1, 2, 3, u, u, u, u>
2810 // %interleaved.vec = shuffle %R_G.vec, %B_U.vec,
2811 // <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> ; Interleave R,G,B elements
2812 // store <12 x i32> %interleaved.vec ; Write 4 tuples of R,G,B
2813 void InnerLoopVectorizer::vectorizeInterleaveGroup(Instruction *Instr) {
2814 const InterleaveGroup *Group = Legal->getInterleavedAccessGroup(Instr);
2815 assert(Group && "Fail to get an interleaved access group.");
2817 // Skip if current instruction is not the insert position.
2818 if (Instr != Group->getInsertPos())
2821 Value *Ptr = getPointerOperand(Instr);
2823 // Prepare for the vector type of the interleaved load/store.
2824 Type *ScalarTy = getMemInstValueType(Instr);
2825 unsigned InterleaveFactor = Group->getFactor();
2826 Type *VecTy = VectorType::get(ScalarTy, InterleaveFactor * VF);
2827 Type *PtrTy = VecTy->getPointerTo(getMemInstAddressSpace(Instr));
2829 // Prepare for the new pointers.
2830 setDebugLocFromInst(Builder, Ptr);
2831 SmallVector<Value *, 2> NewPtrs;
2832 unsigned Index = Group->getIndex(Instr);
2834 // If the group is reverse, adjust the index to refer to the last vector lane
2835 // instead of the first. We adjust the index from the first vector lane,
2836 // rather than directly getting the pointer for lane VF - 1, because the
2837 // pointer operand of the interleaved access is supposed to be uniform. For
2838 // uniform instructions, we're only required to generate a value for the
2839 // first vector lane in each unroll iteration.
2840 if (Group->isReverse())
2841 Index += (VF - 1) * Group->getFactor();
2843 for (unsigned Part = 0; Part < UF; Part++) {
2844 Value *NewPtr = getScalarValue(Ptr, Part, 0);
2846 // Notice current instruction could be any index. Need to adjust the address
2847 // to the member of index 0.
2849 // E.g. a = A[i+1]; // Member of index 1 (Current instruction)
2850 // b = A[i]; // Member of index 0
2851 // Current pointer is pointed to A[i+1], adjust it to A[i].
2853 // E.g. A[i+1] = a; // Member of index 1
2854 // A[i] = b; // Member of index 0
2855 // A[i+2] = c; // Member of index 2 (Current instruction)
2856 // Current pointer is pointed to A[i+2], adjust it to A[i].
2857 NewPtr = Builder.CreateGEP(NewPtr, Builder.getInt32(-Index));
2859 // Cast to the vector pointer type.
2860 NewPtrs.push_back(Builder.CreateBitCast(NewPtr, PtrTy));
2863 setDebugLocFromInst(Builder, Instr);
2864 Value *UndefVec = UndefValue::get(VecTy);
2866 // Vectorize the interleaved load group.
2867 if (isa<LoadInst>(Instr)) {
2869 // For each unroll part, create a wide load for the group.
2870 SmallVector<Value *, 2> NewLoads;
2871 for (unsigned Part = 0; Part < UF; Part++) {
2872 auto *NewLoad = Builder.CreateAlignedLoad(
2873 NewPtrs[Part], Group->getAlignment(), "wide.vec");
2874 addMetadata(NewLoad, Instr);
2875 NewLoads.push_back(NewLoad);
2878 // For each member in the group, shuffle out the appropriate data from the
2880 for (unsigned I = 0; I < InterleaveFactor; ++I) {
2881 Instruction *Member = Group->getMember(I);
2883 // Skip the gaps in the group.
2887 VectorParts Entry(UF);
2888 Constant *StrideMask = createStrideMask(Builder, I, InterleaveFactor, VF);
2889 for (unsigned Part = 0; Part < UF; Part++) {
2890 Value *StridedVec = Builder.CreateShuffleVector(
2891 NewLoads[Part], UndefVec, StrideMask, "strided.vec");
2893 // If this member has different type, cast the result type.
2894 if (Member->getType() != ScalarTy) {
2895 VectorType *OtherVTy = VectorType::get(Member->getType(), VF);
2896 StridedVec = Builder.CreateBitOrPointerCast(StridedVec, OtherVTy);
2900 Group->isReverse() ? reverseVector(StridedVec) : StridedVec;
2902 VectorLoopValueMap.initVector(Member, Entry);
2907 // The sub vector type for current instruction.
2908 VectorType *SubVT = VectorType::get(ScalarTy, VF);
2910 // Vectorize the interleaved store group.
2911 for (unsigned Part = 0; Part < UF; Part++) {
2912 // Collect the stored vector from each member.
2913 SmallVector<Value *, 4> StoredVecs;
2914 for (unsigned i = 0; i < InterleaveFactor; i++) {
2915 // Interleaved store group doesn't allow a gap, so each index has a member
2916 Instruction *Member = Group->getMember(i);
2917 assert(Member && "Fail to get a member from an interleaved store group");
2920 getVectorValue(cast<StoreInst>(Member)->getValueOperand())[Part];
2921 if (Group->isReverse())
2922 StoredVec = reverseVector(StoredVec);
2924 // If this member has different type, cast it to an unified type.
2925 if (StoredVec->getType() != SubVT)
2926 StoredVec = Builder.CreateBitOrPointerCast(StoredVec, SubVT);
2928 StoredVecs.push_back(StoredVec);
2931 // Concatenate all vectors into a wide vector.
2932 Value *WideVec = concatenateVectors(Builder, StoredVecs);
2934 // Interleave the elements in the wide vector.
2935 Constant *IMask = createInterleaveMask(Builder, VF, InterleaveFactor);
2936 Value *IVec = Builder.CreateShuffleVector(WideVec, UndefVec, IMask,
2939 Instruction *NewStoreInstr =
2940 Builder.CreateAlignedStore(IVec, NewPtrs[Part], Group->getAlignment());
2941 addMetadata(NewStoreInstr, Instr);
2945 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
2946 // Attempt to issue a wide load.
2947 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2948 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2950 assert((LI || SI) && "Invalid Load/Store instruction");
2952 LoopVectorizationCostModel::InstWidening Decision =
2953 Cost->getWideningDecision(Instr, VF);
2954 assert(Decision != LoopVectorizationCostModel::CM_Unknown &&
2955 "CM decision should be taken at this point");
2956 if (Decision == LoopVectorizationCostModel::CM_Interleave)
2957 return vectorizeInterleaveGroup(Instr);
2959 Type *ScalarDataTy = getMemInstValueType(Instr);
2960 Type *DataTy = VectorType::get(ScalarDataTy, VF);
2961 Value *Ptr = getPointerOperand(Instr);
2962 unsigned Alignment = getMemInstAlignment(Instr);
2963 // An alignment of 0 means target abi alignment. We need to use the scalar's
2964 // target abi alignment in such a case.
2965 const DataLayout &DL = Instr->getModule()->getDataLayout();
2967 Alignment = DL.getABITypeAlignment(ScalarDataTy);
2968 unsigned AddressSpace = getMemInstAddressSpace(Instr);
2970 // Scalarize the memory instruction if necessary.
2971 if (Decision == LoopVectorizationCostModel::CM_Scalarize)
2972 return scalarizeInstruction(Instr, Legal->isScalarWithPredication(Instr));
2974 // Determine if the pointer operand of the access is either consecutive or
2975 // reverse consecutive.
2976 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
2977 bool Reverse = ConsecutiveStride < 0;
2978 bool CreateGatherScatter =
2979 (Decision == LoopVectorizationCostModel::CM_GatherScatter);
2981 VectorParts VectorGep;
2983 // Handle consecutive loads/stores.
2984 if (ConsecutiveStride) {
2985 Ptr = getScalarValue(Ptr, 0, 0);
2987 // At this point we should vector version of GEP for Gather or Scatter
2988 assert(CreateGatherScatter && "The instruction should be scalarized");
2989 VectorGep = getVectorValue(Ptr);
2992 VectorParts Mask = createBlockInMask(Instr->getParent());
2995 assert(!Legal->isUniform(SI->getPointerOperand()) &&
2996 "We do not allow storing to uniform addresses");
2997 setDebugLocFromInst(Builder, SI);
2998 // We don't want to update the value in the map as it might be used in
2999 // another expression. So don't use a reference type for "StoredVal".
3000 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
3002 for (unsigned Part = 0; Part < UF; ++Part) {
3003 Instruction *NewSI = nullptr;
3004 if (CreateGatherScatter) {
3005 Value *MaskPart = Legal->isMaskRequired(SI) ? Mask[Part] : nullptr;
3006 NewSI = Builder.CreateMaskedScatter(StoredVal[Part], VectorGep[Part],
3007 Alignment, MaskPart);
3009 // Calculate the pointer for the specific unroll-part.
3011 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
3014 // If we store to reverse consecutive memory locations, then we need
3015 // to reverse the order of elements in the stored value.
3016 StoredVal[Part] = reverseVector(StoredVal[Part]);
3017 // If the address is consecutive but reversed, then the
3018 // wide store needs to start at the last vector element.
3020 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
3022 Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
3023 Mask[Part] = reverseVector(Mask[Part]);
3027 Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace));
3029 if (Legal->isMaskRequired(SI))
3030 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
3034 Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
3036 addMetadata(NewSI, SI);
3042 assert(LI && "Must have a load instruction");
3043 setDebugLocFromInst(Builder, LI);
3044 VectorParts Entry(UF);
3045 for (unsigned Part = 0; Part < UF; ++Part) {
3047 if (CreateGatherScatter) {
3048 Value *MaskPart = Legal->isMaskRequired(LI) ? Mask[Part] : nullptr;
3049 NewLI = Builder.CreateMaskedGather(VectorGep[Part], Alignment, MaskPart,
3050 nullptr, "wide.masked.gather");
3051 Entry[Part] = NewLI;
3053 // Calculate the pointer for the specific unroll-part.
3055 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
3058 // If the address is consecutive but reversed, then the
3059 // wide load needs to start at the last vector element.
3060 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
3061 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
3062 Mask[Part] = reverseVector(Mask[Part]);
3066 Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace));
3067 if (Legal->isMaskRequired(LI))
3068 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
3069 UndefValue::get(DataTy),
3070 "wide.masked.load");
3072 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
3073 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
3075 addMetadata(NewLI, LI);
3077 VectorLoopValueMap.initVector(Instr, Entry);
3080 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr,
3081 bool IfPredicateInstr) {
3082 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
3083 DEBUG(dbgs() << "LV: Scalarizing"
3084 << (IfPredicateInstr ? " and predicating:" : ":") << *Instr
3086 // Holds vector parameters or scalars, in case of uniform vals.
3087 SmallVector<VectorParts, 4> Params;
3089 setDebugLocFromInst(Builder, Instr);
3091 // Does this instruction return a value ?
3092 bool IsVoidRetTy = Instr->getType()->isVoidTy();
3094 // Initialize a new scalar map entry.
3095 ScalarParts Entry(UF);
3098 if (IfPredicateInstr)
3099 Cond = createBlockInMask(Instr->getParent());
3101 // Determine the number of scalars we need to generate for each unroll
3102 // iteration. If the instruction is uniform, we only need to generate the
3103 // first lane. Otherwise, we generate all VF values.
3104 unsigned Lanes = Cost->isUniformAfterVectorization(Instr, VF) ? 1 : VF;
3106 // For each vector unroll 'part':
3107 for (unsigned Part = 0; Part < UF; ++Part) {
3108 Entry[Part].resize(VF);
3109 // For each scalar that we create:
3110 for (unsigned Lane = 0; Lane < Lanes; ++Lane) {
3113 Value *Cmp = nullptr;
3114 if (IfPredicateInstr) {
3116 if (Cmp->getType()->isVectorTy())
3117 Cmp = Builder.CreateExtractElement(Cmp, Builder.getInt32(Lane));
3118 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp,
3119 ConstantInt::get(Cmp->getType(), 1));
3122 Instruction *Cloned = Instr->clone();
3124 Cloned->setName(Instr->getName() + ".cloned");
3126 // Replace the operands of the cloned instructions with their scalar
3127 // equivalents in the new loop.
3128 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
3129 auto *NewOp = getScalarValue(Instr->getOperand(op), Part, Lane);
3130 Cloned->setOperand(op, NewOp);
3132 addNewMetadata(Cloned, Instr);
3134 // Place the cloned scalar in the new loop.
3135 Builder.Insert(Cloned);
3137 // Add the cloned scalar to the scalar map entry.
3138 Entry[Part][Lane] = Cloned;
3140 // If we just cloned a new assumption, add it the assumption cache.
3141 if (auto *II = dyn_cast<IntrinsicInst>(Cloned))
3142 if (II->getIntrinsicID() == Intrinsic::assume)
3143 AC->registerAssumption(II);
3146 if (IfPredicateInstr)
3147 PredicatedInstructions.push_back(std::make_pair(Cloned, Cmp));
3150 VectorLoopValueMap.initScalar(Instr, Entry);
3153 PHINode *InnerLoopVectorizer::createInductionVariable(Loop *L, Value *Start,
3154 Value *End, Value *Step,
3156 BasicBlock *Header = L->getHeader();
3157 BasicBlock *Latch = L->getLoopLatch();
3158 // As we're just creating this loop, it's possible no latch exists
3159 // yet. If so, use the header as this will be a single block loop.
3163 IRBuilder<> Builder(&*Header->getFirstInsertionPt());
3164 Instruction *OldInst = getDebugLocFromInstOrOperands(OldInduction);
3165 setDebugLocFromInst(Builder, OldInst);
3166 auto *Induction = Builder.CreatePHI(Start->getType(), 2, "index");
3168 Builder.SetInsertPoint(Latch->getTerminator());
3169 setDebugLocFromInst(Builder, OldInst);
3171 // Create i+1 and fill the PHINode.
3172 Value *Next = Builder.CreateAdd(Induction, Step, "index.next");
3173 Induction->addIncoming(Start, L->getLoopPreheader());
3174 Induction->addIncoming(Next, Latch);
3175 // Create the compare.
3176 Value *ICmp = Builder.CreateICmpEQ(Next, End);
3177 Builder.CreateCondBr(ICmp, L->getExitBlock(), Header);
3179 // Now we have two terminators. Remove the old one from the block.
3180 Latch->getTerminator()->eraseFromParent();
3185 Value *InnerLoopVectorizer::getOrCreateTripCount(Loop *L) {
3189 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
3190 // Find the loop boundaries.
3191 ScalarEvolution *SE = PSE.getSE();
3192 const SCEV *BackedgeTakenCount = PSE.getBackedgeTakenCount();
3193 assert(BackedgeTakenCount != SE->getCouldNotCompute() &&
3194 "Invalid loop count");
3196 Type *IdxTy = Legal->getWidestInductionType();
3198 // The exit count might have the type of i64 while the phi is i32. This can
3199 // happen if we have an induction variable that is sign extended before the
3200 // compare. The only way that we get a backedge taken count is that the
3201 // induction variable was signed and as such will not overflow. In such a case
3202 // truncation is legal.
3203 if (BackedgeTakenCount->getType()->getPrimitiveSizeInBits() >
3204 IdxTy->getPrimitiveSizeInBits())
3205 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount, IdxTy);
3206 BackedgeTakenCount = SE->getNoopOrZeroExtend(BackedgeTakenCount, IdxTy);
3208 // Get the total trip count from the count by adding 1.
3209 const SCEV *ExitCount = SE->getAddExpr(
3210 BackedgeTakenCount, SE->getOne(BackedgeTakenCount->getType()));
3212 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
3214 // Expand the trip count and place the new instructions in the preheader.
3215 // Notice that the pre-header does not change, only the loop body.
3216 SCEVExpander Exp(*SE, DL, "induction");
3218 // Count holds the overall loop count (N).
3219 TripCount = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
3220 L->getLoopPreheader()->getTerminator());
3222 if (TripCount->getType()->isPointerTy())
3224 CastInst::CreatePointerCast(TripCount, IdxTy, "exitcount.ptrcnt.to.int",
3225 L->getLoopPreheader()->getTerminator());
3230 Value *InnerLoopVectorizer::getOrCreateVectorTripCount(Loop *L) {
3231 if (VectorTripCount)
3232 return VectorTripCount;
3234 Value *TC = getOrCreateTripCount(L);
3235 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
3237 // Now we need to generate the expression for the part of the loop that the
3238 // vectorized body will execute. This is equal to N - (N % Step) if scalar
3239 // iterations are not required for correctness, or N - Step, otherwise. Step
3240 // is equal to the vectorization factor (number of SIMD elements) times the
3241 // unroll factor (number of SIMD instructions).
3242 Constant *Step = ConstantInt::get(TC->getType(), VF * UF);
3243 Value *R = Builder.CreateURem(TC, Step, "n.mod.vf");
3245 // If there is a non-reversed interleaved group that may speculatively access
3246 // memory out-of-bounds, we need to ensure that there will be at least one
3247 // iteration of the scalar epilogue loop. Thus, if the step evenly divides
3248 // the trip count, we set the remainder to be equal to the step. If the step
3249 // does not evenly divide the trip count, no adjustment is necessary since
3250 // there will already be scalar iterations. Note that the minimum iterations
3251 // check ensures that N >= Step.
3252 if (VF > 1 && Legal->requiresScalarEpilogue()) {
3253 auto *IsZero = Builder.CreateICmpEQ(R, ConstantInt::get(R->getType(), 0));
3254 R = Builder.CreateSelect(IsZero, Step, R);
3257 VectorTripCount = Builder.CreateSub(TC, R, "n.vec");
3259 return VectorTripCount;
3262 void InnerLoopVectorizer::emitMinimumIterationCountCheck(Loop *L,
3263 BasicBlock *Bypass) {
3264 Value *Count = getOrCreateTripCount(L);
3265 BasicBlock *BB = L->getLoopPreheader();
3266 IRBuilder<> Builder(BB->getTerminator());
3268 // Generate code to check that the loop's trip count that we computed by
3269 // adding one to the backedge-taken count will not overflow.
3270 Value *CheckMinIters = Builder.CreateICmpULT(
3271 Count, ConstantInt::get(Count->getType(), VF * UF), "min.iters.check");
3274 BB->splitBasicBlock(BB->getTerminator(), "min.iters.checked");
3275 // Update dominator tree immediately if the generated block is a
3276 // LoopBypassBlock because SCEV expansions to generate loop bypass
3277 // checks may query it before the current function is finished.
3278 DT->addNewBlock(NewBB, BB);
3279 if (L->getParentLoop())
3280 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
3281 ReplaceInstWithInst(BB->getTerminator(),
3282 BranchInst::Create(Bypass, NewBB, CheckMinIters));
3283 LoopBypassBlocks.push_back(BB);
3286 void InnerLoopVectorizer::emitVectorLoopEnteredCheck(Loop *L,
3287 BasicBlock *Bypass) {
3288 Value *TC = getOrCreateVectorTripCount(L);
3289 BasicBlock *BB = L->getLoopPreheader();
3290 IRBuilder<> Builder(BB->getTerminator());
3292 // Now, compare the new count to zero. If it is zero skip the vector loop and
3293 // jump to the scalar loop.
3294 Value *Cmp = Builder.CreateICmpEQ(TC, Constant::getNullValue(TC->getType()),
3297 // Generate code to check that the loop's trip count that we computed by
3298 // adding one to the backedge-taken count will not overflow.
3299 BasicBlock *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
3300 // Update dominator tree immediately if the generated block is a
3301 // LoopBypassBlock because SCEV expansions to generate loop bypass
3302 // checks may query it before the current function is finished.
3303 DT->addNewBlock(NewBB, BB);
3304 if (L->getParentLoop())
3305 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
3306 ReplaceInstWithInst(BB->getTerminator(),
3307 BranchInst::Create(Bypass, NewBB, Cmp));
3308 LoopBypassBlocks.push_back(BB);
3311 void InnerLoopVectorizer::emitSCEVChecks(Loop *L, BasicBlock *Bypass) {
3312 BasicBlock *BB = L->getLoopPreheader();
3314 // Generate the code to check that the SCEV assumptions that we made.
3315 // We want the new basic block to start at the first instruction in a
3316 // sequence of instructions that form a check.
3317 SCEVExpander Exp(*PSE.getSE(), Bypass->getModule()->getDataLayout(),
3320 Exp.expandCodeForPredicate(&PSE.getUnionPredicate(), BB->getTerminator());
3322 if (auto *C = dyn_cast<ConstantInt>(SCEVCheck))
3326 // Create a new block containing the stride check.
3327 BB->setName("vector.scevcheck");
3328 auto *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
3329 // Update dominator tree immediately if the generated block is a
3330 // LoopBypassBlock because SCEV expansions to generate loop bypass
3331 // checks may query it before the current function is finished.
3332 DT->addNewBlock(NewBB, BB);
3333 if (L->getParentLoop())
3334 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
3335 ReplaceInstWithInst(BB->getTerminator(),
3336 BranchInst::Create(Bypass, NewBB, SCEVCheck));
3337 LoopBypassBlocks.push_back(BB);
3338 AddedSafetyChecks = true;
3341 void InnerLoopVectorizer::emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass) {
3342 BasicBlock *BB = L->getLoopPreheader();
3344 // Generate the code that checks in runtime if arrays overlap. We put the
3345 // checks into a separate block to make the more common case of few elements
3347 Instruction *FirstCheckInst;
3348 Instruction *MemRuntimeCheck;
3349 std::tie(FirstCheckInst, MemRuntimeCheck) =
3350 Legal->getLAI()->addRuntimeChecks(BB->getTerminator());
3351 if (!MemRuntimeCheck)
3354 // Create a new block containing the memory check.
3355 BB->setName("vector.memcheck");
3356 auto *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
3357 // Update dominator tree immediately if the generated block is a
3358 // LoopBypassBlock because SCEV expansions to generate loop bypass
3359 // checks may query it before the current function is finished.
3360 DT->addNewBlock(NewBB, BB);
3361 if (L->getParentLoop())
3362 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
3363 ReplaceInstWithInst(BB->getTerminator(),
3364 BranchInst::Create(Bypass, NewBB, MemRuntimeCheck));
3365 LoopBypassBlocks.push_back(BB);
3366 AddedSafetyChecks = true;
3368 // We currently don't use LoopVersioning for the actual loop cloning but we
3369 // still use it to add the noalias metadata.
3370 LVer = llvm::make_unique<LoopVersioning>(*Legal->getLAI(), OrigLoop, LI, DT,
3372 LVer->prepareNoAliasMetadata();
3375 void InnerLoopVectorizer::createVectorizedLoopSkeleton() {
3377 In this function we generate a new loop. The new loop will contain
3378 the vectorized instructions while the old loop will continue to run the
3381 [ ] <-- loop iteration number check.
3384 | [ ] <-- vector loop bypass (may consist of multiple blocks).
3387 || [ ] <-- vector pre header.
3391 | [ ]_| <-- vector loop.
3394 | -[ ] <--- middle-block.
3397 -|- >[ ] <--- new preheader.
3401 | [ ]_| <-- old scalar loop to handle remainder.
3404 >[ ] <-- exit block.
3408 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
3409 BasicBlock *VectorPH = OrigLoop->getLoopPreheader();
3410 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
3411 assert(VectorPH && "Invalid loop structure");
3412 assert(ExitBlock && "Must have an exit block");
3414 // Some loops have a single integer induction variable, while other loops
3415 // don't. One example is c++ iterators that often have multiple pointer
3416 // induction variables. In the code below we also support a case where we
3417 // don't have a single induction variable.
3419 // We try to obtain an induction variable from the original loop as hard
3420 // as possible. However if we don't find one that:
3422 // - counts from zero, stepping by one
3423 // - is the size of the widest induction variable type
3424 // then we create a new one.
3425 OldInduction = Legal->getPrimaryInduction();
3426 Type *IdxTy = Legal->getWidestInductionType();
3428 // Split the single block loop into the two loop structure described above.
3429 BasicBlock *VecBody =
3430 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
3431 BasicBlock *MiddleBlock =
3432 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
3433 BasicBlock *ScalarPH =
3434 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
3436 // Create and register the new vector loop.
3437 Loop *Lp = new Loop();
3438 Loop *ParentLoop = OrigLoop->getParentLoop();
3440 // Insert the new loop into the loop nest and register the new basic blocks
3441 // before calling any utilities such as SCEV that require valid LoopInfo.
3443 ParentLoop->addChildLoop(Lp);
3444 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
3445 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
3447 LI->addTopLevelLoop(Lp);
3449 Lp->addBasicBlockToLoop(VecBody, *LI);
3451 // Find the loop boundaries.
3452 Value *Count = getOrCreateTripCount(Lp);
3454 Value *StartIdx = ConstantInt::get(IdxTy, 0);
3456 // We need to test whether the backedge-taken count is uint##_max. Adding one
3457 // to it will cause overflow and an incorrect loop trip count in the vector
3458 // body. In case of overflow we want to directly jump to the scalar remainder
3460 emitMinimumIterationCountCheck(Lp, ScalarPH);
3461 // Now, compare the new count to zero. If it is zero skip the vector loop and
3462 // jump to the scalar loop.
3463 emitVectorLoopEnteredCheck(Lp, ScalarPH);
3464 // Generate the code to check any assumptions that we've made for SCEV
3466 emitSCEVChecks(Lp, ScalarPH);
3468 // Generate the code that checks in runtime if arrays overlap. We put the
3469 // checks into a separate block to make the more common case of few elements
3471 emitMemRuntimeChecks(Lp, ScalarPH);
3473 // Generate the induction variable.
3474 // The loop step is equal to the vectorization factor (num of SIMD elements)
3475 // times the unroll factor (num of SIMD instructions).
3476 Value *CountRoundDown = getOrCreateVectorTripCount(Lp);
3477 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
3479 createInductionVariable(Lp, StartIdx, CountRoundDown, Step,
3480 getDebugLocFromInstOrOperands(OldInduction));
3482 // We are going to resume the execution of the scalar loop.
3483 // Go over all of the induction variables that we found and fix the
3484 // PHIs that are left in the scalar version of the loop.
3485 // The starting values of PHI nodes depend on the counter of the last
3486 // iteration in the vectorized loop.
3487 // If we come from a bypass edge then we need to start from the original
3490 // This variable saves the new starting index for the scalar loop. It is used
3491 // to test if there are any tail iterations left once the vector loop has
3493 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
3494 for (auto &InductionEntry : *List) {
3495 PHINode *OrigPhi = InductionEntry.first;
3496 InductionDescriptor II = InductionEntry.second;
3498 // Create phi nodes to merge from the backedge-taken check block.
3499 PHINode *BCResumeVal = PHINode::Create(
3500 OrigPhi->getType(), 3, "bc.resume.val", ScalarPH->getTerminator());
3501 Value *&EndValue = IVEndValues[OrigPhi];
3502 if (OrigPhi == OldInduction) {
3503 // We know what the end value is.
3504 EndValue = CountRoundDown;
3506 IRBuilder<> B(LoopBypassBlocks.back()->getTerminator());
3507 Type *StepType = II.getStep()->getType();
3508 Instruction::CastOps CastOp =
3509 CastInst::getCastOpcode(CountRoundDown, true, StepType, true);
3510 Value *CRD = B.CreateCast(CastOp, CountRoundDown, StepType, "cast.crd");
3511 const DataLayout &DL = OrigLoop->getHeader()->getModule()->getDataLayout();
3512 EndValue = II.transform(B, CRD, PSE.getSE(), DL);
3513 EndValue->setName("ind.end");
3516 // The new PHI merges the original incoming value, in case of a bypass,
3517 // or the value at the end of the vectorized loop.
3518 BCResumeVal->addIncoming(EndValue, MiddleBlock);
3520 // Fix the scalar body counter (PHI node).
3521 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
3523 // The old induction's phi node in the scalar body needs the truncated
3525 for (BasicBlock *BB : LoopBypassBlocks)
3526 BCResumeVal->addIncoming(II.getStartValue(), BB);
3527 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
3530 // Add a check in the middle block to see if we have completed
3531 // all of the iterations in the first vector loop.
3532 // If (N - N%VF) == N, then we *don't* need to run the remainder.
3534 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, Count,
3535 CountRoundDown, "cmp.n", MiddleBlock->getTerminator());
3536 ReplaceInstWithInst(MiddleBlock->getTerminator(),
3537 BranchInst::Create(ExitBlock, ScalarPH, CmpN));
3539 // Get ready to start creating new instructions into the vectorized body.
3540 Builder.SetInsertPoint(&*VecBody->getFirstInsertionPt());
3543 LoopVectorPreHeader = Lp->getLoopPreheader();
3544 LoopScalarPreHeader = ScalarPH;
3545 LoopMiddleBlock = MiddleBlock;
3546 LoopExitBlock = ExitBlock;
3547 LoopVectorBody = VecBody;
3548 LoopScalarBody = OldBasicBlock;
3550 // Keep all loop hints from the original loop on the vector loop (we'll
3551 // replace the vectorizer-specific hints below).
3552 if (MDNode *LID = OrigLoop->getLoopID())
3555 LoopVectorizeHints Hints(Lp, true, *ORE);
3556 Hints.setAlreadyVectorized();
3559 // Fix up external users of the induction variable. At this point, we are
3560 // in LCSSA form, with all external PHIs that use the IV having one input value,
3561 // coming from the remainder loop. We need those PHIs to also have a correct
3562 // value for the IV when arriving directly from the middle block.
3563 void InnerLoopVectorizer::fixupIVUsers(PHINode *OrigPhi,
3564 const InductionDescriptor &II,
3565 Value *CountRoundDown, Value *EndValue,
3566 BasicBlock *MiddleBlock) {
3567 // There are two kinds of external IV usages - those that use the value
3568 // computed in the last iteration (the PHI) and those that use the penultimate
3569 // value (the value that feeds into the phi from the loop latch).
3570 // We allow both, but they, obviously, have different values.
3572 assert(OrigLoop->getExitBlock() && "Expected a single exit block");
3574 DenseMap<Value *, Value *> MissingVals;
3576 // An external user of the last iteration's value should see the value that
3577 // the remainder loop uses to initialize its own IV.
3578 Value *PostInc = OrigPhi->getIncomingValueForBlock(OrigLoop->getLoopLatch());
3579 for (User *U : PostInc->users()) {
3580 Instruction *UI = cast<Instruction>(U);
3581 if (!OrigLoop->contains(UI)) {
3582 assert(isa<PHINode>(UI) && "Expected LCSSA form");
3583 MissingVals[UI] = EndValue;
3587 // An external user of the penultimate value need to see EndValue - Step.
3588 // The simplest way to get this is to recompute it from the constituent SCEVs,
3589 // that is Start + (Step * (CRD - 1)).
3590 for (User *U : OrigPhi->users()) {
3591 auto *UI = cast<Instruction>(U);
3592 if (!OrigLoop->contains(UI)) {
3593 const DataLayout &DL =
3594 OrigLoop->getHeader()->getModule()->getDataLayout();
3595 assert(isa<PHINode>(UI) && "Expected LCSSA form");
3597 IRBuilder<> B(MiddleBlock->getTerminator());
3598 Value *CountMinusOne = B.CreateSub(
3599 CountRoundDown, ConstantInt::get(CountRoundDown->getType(), 1));
3601 !II.getStep()->getType()->isIntegerTy()
3602 ? B.CreateCast(Instruction::SIToFP, CountMinusOne,
3603 II.getStep()->getType())
3604 : B.CreateSExtOrTrunc(CountMinusOne, II.getStep()->getType());
3605 CMO->setName("cast.cmo");
3606 Value *Escape = II.transform(B, CMO, PSE.getSE(), DL);
3607 Escape->setName("ind.escape");
3608 MissingVals[UI] = Escape;
3612 for (auto &I : MissingVals) {
3613 PHINode *PHI = cast<PHINode>(I.first);
3614 // One corner case we have to handle is two IVs "chasing" each-other,
3615 // that is %IV2 = phi [...], [ %IV1, %latch ]
3616 // In this case, if IV1 has an external use, we need to avoid adding both
3617 // "last value of IV1" and "penultimate value of IV2". So, verify that we
3618 // don't already have an incoming value for the middle block.
3619 if (PHI->getBasicBlockIndex(MiddleBlock) == -1)
3620 PHI->addIncoming(I.second, MiddleBlock);
3625 struct CSEDenseMapInfo {
3626 static bool canHandle(const Instruction *I) {
3627 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
3628 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
3630 static inline Instruction *getEmptyKey() {
3631 return DenseMapInfo<Instruction *>::getEmptyKey();
3633 static inline Instruction *getTombstoneKey() {
3634 return DenseMapInfo<Instruction *>::getTombstoneKey();
3636 static unsigned getHashValue(const Instruction *I) {
3637 assert(canHandle(I) && "Unknown instruction!");
3638 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
3639 I->value_op_end()));
3641 static bool isEqual(const Instruction *LHS, const Instruction *RHS) {
3642 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
3643 LHS == getTombstoneKey() || RHS == getTombstoneKey())
3645 return LHS->isIdenticalTo(RHS);
3650 ///\brief Perform cse of induction variable instructions.
3651 static void cse(BasicBlock *BB) {
3652 // Perform simple cse.
3653 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
3654 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
3655 Instruction *In = &*I++;
3657 if (!CSEDenseMapInfo::canHandle(In))
3660 // Check if we can replace this instruction with any of the
3661 // visited instructions.
3662 if (Instruction *V = CSEMap.lookup(In)) {
3663 In->replaceAllUsesWith(V);
3664 In->eraseFromParent();
3672 /// \brief Estimate the overhead of scalarizing an instruction. This is a
3673 /// convenience wrapper for the type-based getScalarizationOverhead API.
3674 static unsigned getScalarizationOverhead(Instruction *I, unsigned VF,
3675 const TargetTransformInfo &TTI) {
3680 Type *RetTy = ToVectorTy(I->getType(), VF);
3681 if (!RetTy->isVoidTy() &&
3682 (!isa<LoadInst>(I) ||
3683 !TTI.supportsEfficientVectorElementLoadStore()))
3684 Cost += TTI.getScalarizationOverhead(RetTy, true, false);
3686 if (CallInst *CI = dyn_cast<CallInst>(I)) {
3687 SmallVector<const Value *, 4> Operands(CI->arg_operands());
3688 Cost += TTI.getOperandsScalarizationOverhead(Operands, VF);
3690 else if (!isa<StoreInst>(I) ||
3691 !TTI.supportsEfficientVectorElementLoadStore()) {
3692 SmallVector<const Value *, 4> Operands(I->operand_values());
3693 Cost += TTI.getOperandsScalarizationOverhead(Operands, VF);
3699 // Estimate cost of a call instruction CI if it were vectorized with factor VF.
3700 // Return the cost of the instruction, including scalarization overhead if it's
3701 // needed. The flag NeedToScalarize shows if the call needs to be scalarized -
3702 // i.e. either vector version isn't available, or is too expensive.
3703 static unsigned getVectorCallCost(CallInst *CI, unsigned VF,
3704 const TargetTransformInfo &TTI,
3705 const TargetLibraryInfo *TLI,
3706 bool &NeedToScalarize) {
3707 Function *F = CI->getCalledFunction();
3708 StringRef FnName = CI->getCalledFunction()->getName();
3709 Type *ScalarRetTy = CI->getType();
3710 SmallVector<Type *, 4> Tys, ScalarTys;
3711 for (auto &ArgOp : CI->arg_operands())
3712 ScalarTys.push_back(ArgOp->getType());
3714 // Estimate cost of scalarized vector call. The source operands are assumed
3715 // to be vectors, so we need to extract individual elements from there,
3716 // execute VF scalar calls, and then gather the result into the vector return
3718 unsigned ScalarCallCost = TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys);
3720 return ScalarCallCost;
3722 // Compute corresponding vector type for return value and arguments.
3723 Type *RetTy = ToVectorTy(ScalarRetTy, VF);
3724 for (Type *ScalarTy : ScalarTys)
3725 Tys.push_back(ToVectorTy(ScalarTy, VF));
3727 // Compute costs of unpacking argument values for the scalar calls and
3728 // packing the return values to a vector.
3729 unsigned ScalarizationCost = getScalarizationOverhead(CI, VF, TTI);
3731 unsigned Cost = ScalarCallCost * VF + ScalarizationCost;
3733 // If we can't emit a vector call for this function, then the currently found
3734 // cost is the cost we need to return.
3735 NeedToScalarize = true;
3736 if (!TLI || !TLI->isFunctionVectorizable(FnName, VF) || CI->isNoBuiltin())
3739 // If the corresponding vector cost is cheaper, return its cost.
3740 unsigned VectorCallCost = TTI.getCallInstrCost(nullptr, RetTy, Tys);
3741 if (VectorCallCost < Cost) {
3742 NeedToScalarize = false;
3743 return VectorCallCost;
3748 // Estimate cost of an intrinsic call instruction CI if it were vectorized with
3749 // factor VF. Return the cost of the instruction, including scalarization
3750 // overhead if it's needed.
3751 static unsigned getVectorIntrinsicCost(CallInst *CI, unsigned VF,
3752 const TargetTransformInfo &TTI,
3753 const TargetLibraryInfo *TLI) {
3754 Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
3755 assert(ID && "Expected intrinsic call!");
3758 if (auto *FPMO = dyn_cast<FPMathOperator>(CI))
3759 FMF = FPMO->getFastMathFlags();
3761 SmallVector<Value *, 4> Operands(CI->arg_operands());
3762 return TTI.getIntrinsicInstrCost(ID, CI->getType(), Operands, FMF, VF);
3765 static Type *smallestIntegerVectorType(Type *T1, Type *T2) {
3766 auto *I1 = cast<IntegerType>(T1->getVectorElementType());
3767 auto *I2 = cast<IntegerType>(T2->getVectorElementType());
3768 return I1->getBitWidth() < I2->getBitWidth() ? T1 : T2;
3770 static Type *largestIntegerVectorType(Type *T1, Type *T2) {
3771 auto *I1 = cast<IntegerType>(T1->getVectorElementType());
3772 auto *I2 = cast<IntegerType>(T2->getVectorElementType());
3773 return I1->getBitWidth() > I2->getBitWidth() ? T1 : T2;
3776 void InnerLoopVectorizer::truncateToMinimalBitwidths() {
3777 // For every instruction `I` in MinBWs, truncate the operands, create a
3778 // truncated version of `I` and reextend its result. InstCombine runs
3779 // later and will remove any ext/trunc pairs.
3781 SmallPtrSet<Value *, 4> Erased;
3782 for (const auto &KV : Cost->getMinimalBitwidths()) {
3783 // If the value wasn't vectorized, we must maintain the original scalar
3784 // type. The absence of the value from VectorLoopValueMap indicates that it
3785 // wasn't vectorized.
3786 if (!VectorLoopValueMap.hasVector(KV.first))
3788 VectorParts &Parts = VectorLoopValueMap.getVector(KV.first);
3789 for (Value *&I : Parts) {
3790 if (Erased.count(I) || I->use_empty() || !isa<Instruction>(I))
3792 Type *OriginalTy = I->getType();
3793 Type *ScalarTruncatedTy =
3794 IntegerType::get(OriginalTy->getContext(), KV.second);
3795 Type *TruncatedTy = VectorType::get(ScalarTruncatedTy,
3796 OriginalTy->getVectorNumElements());
3797 if (TruncatedTy == OriginalTy)
3800 IRBuilder<> B(cast<Instruction>(I));
3801 auto ShrinkOperand = [&](Value *V) -> Value * {
3802 if (auto *ZI = dyn_cast<ZExtInst>(V))
3803 if (ZI->getSrcTy() == TruncatedTy)
3804 return ZI->getOperand(0);
3805 return B.CreateZExtOrTrunc(V, TruncatedTy);
3808 // The actual instruction modification depends on the instruction type,
3810 Value *NewI = nullptr;
3811 if (auto *BO = dyn_cast<BinaryOperator>(I)) {
3812 NewI = B.CreateBinOp(BO->getOpcode(), ShrinkOperand(BO->getOperand(0)),
3813 ShrinkOperand(BO->getOperand(1)));
3814 cast<BinaryOperator>(NewI)->copyIRFlags(I);
3815 } else if (auto *CI = dyn_cast<ICmpInst>(I)) {
3817 B.CreateICmp(CI->getPredicate(), ShrinkOperand(CI->getOperand(0)),
3818 ShrinkOperand(CI->getOperand(1)));
3819 } else if (auto *SI = dyn_cast<SelectInst>(I)) {
3820 NewI = B.CreateSelect(SI->getCondition(),
3821 ShrinkOperand(SI->getTrueValue()),
3822 ShrinkOperand(SI->getFalseValue()));
3823 } else if (auto *CI = dyn_cast<CastInst>(I)) {
3824 switch (CI->getOpcode()) {
3826 llvm_unreachable("Unhandled cast!");
3827 case Instruction::Trunc:
3828 NewI = ShrinkOperand(CI->getOperand(0));
3830 case Instruction::SExt:
3831 NewI = B.CreateSExtOrTrunc(
3833 smallestIntegerVectorType(OriginalTy, TruncatedTy));
3835 case Instruction::ZExt:
3836 NewI = B.CreateZExtOrTrunc(
3838 smallestIntegerVectorType(OriginalTy, TruncatedTy));
3841 } else if (auto *SI = dyn_cast<ShuffleVectorInst>(I)) {
3842 auto Elements0 = SI->getOperand(0)->getType()->getVectorNumElements();
3843 auto *O0 = B.CreateZExtOrTrunc(
3844 SI->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements0));
3845 auto Elements1 = SI->getOperand(1)->getType()->getVectorNumElements();
3846 auto *O1 = B.CreateZExtOrTrunc(
3847 SI->getOperand(1), VectorType::get(ScalarTruncatedTy, Elements1));
3849 NewI = B.CreateShuffleVector(O0, O1, SI->getMask());
3850 } else if (isa<LoadInst>(I)) {
3851 // Don't do anything with the operands, just extend the result.
3853 } else if (auto *IE = dyn_cast<InsertElementInst>(I)) {
3854 auto Elements = IE->getOperand(0)->getType()->getVectorNumElements();
3855 auto *O0 = B.CreateZExtOrTrunc(
3856 IE->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements));
3857 auto *O1 = B.CreateZExtOrTrunc(IE->getOperand(1), ScalarTruncatedTy);
3858 NewI = B.CreateInsertElement(O0, O1, IE->getOperand(2));
3859 } else if (auto *EE = dyn_cast<ExtractElementInst>(I)) {
3860 auto Elements = EE->getOperand(0)->getType()->getVectorNumElements();
3861 auto *O0 = B.CreateZExtOrTrunc(
3862 EE->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements));
3863 NewI = B.CreateExtractElement(O0, EE->getOperand(2));
3865 llvm_unreachable("Unhandled instruction type!");
3868 // Lastly, extend the result.
3869 NewI->takeName(cast<Instruction>(I));
3870 Value *Res = B.CreateZExtOrTrunc(NewI, OriginalTy);
3871 I->replaceAllUsesWith(Res);
3872 cast<Instruction>(I)->eraseFromParent();
3878 // We'll have created a bunch of ZExts that are now parentless. Clean up.
3879 for (const auto &KV : Cost->getMinimalBitwidths()) {
3880 // If the value wasn't vectorized, we must maintain the original scalar
3881 // type. The absence of the value from VectorLoopValueMap indicates that it
3882 // wasn't vectorized.
3883 if (!VectorLoopValueMap.hasVector(KV.first))
3885 VectorParts &Parts = VectorLoopValueMap.getVector(KV.first);
3886 for (Value *&I : Parts) {
3887 ZExtInst *Inst = dyn_cast<ZExtInst>(I);
3888 if (Inst && Inst->use_empty()) {
3889 Value *NewI = Inst->getOperand(0);
3890 Inst->eraseFromParent();
3897 void InnerLoopVectorizer::fixVectorizedLoop() {
3898 // Insert truncates and extends for any truncated instructions as hints to
3901 truncateToMinimalBitwidths();
3903 // At this point every instruction in the original loop is widened to a
3904 // vector form. Now we need to fix the recurrences in the loop. These PHI
3905 // nodes are currently empty because we did not want to introduce cycles.
3906 // This is the second stage of vectorizing recurrences.
3907 fixCrossIterationPHIs();
3909 // Update the dominator tree.
3911 // FIXME: After creating the structure of the new loop, the dominator tree is
3912 // no longer up-to-date, and it remains that way until we update it
3913 // here. An out-of-date dominator tree is problematic for SCEV,
3914 // because SCEVExpander uses it to guide code generation. The
3915 // vectorizer use SCEVExpanders in several places. Instead, we should
3916 // keep the dominator tree up-to-date as we go.
3919 // Fix-up external users of the induction variables.
3920 for (auto &Entry : *Legal->getInductionVars())
3921 fixupIVUsers(Entry.first, Entry.second,
3922 getOrCreateVectorTripCount(LI->getLoopFor(LoopVectorBody)),
3923 IVEndValues[Entry.first], LoopMiddleBlock);
3926 predicateInstructions();
3928 // Remove redundant induction instructions.
3929 cse(LoopVectorBody);
3932 void InnerLoopVectorizer::fixCrossIterationPHIs() {
3933 // In order to support recurrences we need to be able to vectorize Phi nodes.
3934 // Phi nodes have cycles, so we need to vectorize them in two stages. This is
3935 // stage #2: We now need to fix the recurrences by adding incoming edges to
3936 // the currently empty PHI nodes. At this point every instruction in the
3937 // original loop is widened to a vector form so we can use them to construct
3938 // the incoming edges.
3939 for (Instruction &I : *OrigLoop->getHeader()) {
3940 PHINode *Phi = dyn_cast<PHINode>(&I);
3943 // Handle first-order recurrences and reductions that need to be fixed.
3944 if (Legal->isFirstOrderRecurrence(Phi))
3945 fixFirstOrderRecurrence(Phi);
3946 else if (Legal->isReductionVariable(Phi))
3951 void InnerLoopVectorizer::fixFirstOrderRecurrence(PHINode *Phi) {
3953 // This is the second phase of vectorizing first-order recurrences. An
3954 // overview of the transformation is described below. Suppose we have the
3957 // for (int i = 0; i < n; ++i)
3958 // b[i] = a[i] - a[i - 1];
3960 // There is a first-order recurrence on "a". For this loop, the shorthand
3961 // scalar IR looks like:
3968 // i = phi [0, scalar.ph], [i+1, scalar.body]
3969 // s1 = phi [s_init, scalar.ph], [s2, scalar.body]
3972 // br cond, scalar.body, ...
3974 // In this example, s1 is a recurrence because it's value depends on the
3975 // previous iteration. In the first phase of vectorization, we created a
3976 // temporary value for s1. We now complete the vectorization and produce the
3977 // shorthand vector IR shown below (for VF = 4, UF = 1).
3980 // v_init = vector(..., ..., ..., a[-1])
3984 // i = phi [0, vector.ph], [i+4, vector.body]
3985 // v1 = phi [v_init, vector.ph], [v2, vector.body]
3986 // v2 = a[i, i+1, i+2, i+3];
3987 // v3 = vector(v1(3), v2(0, 1, 2))
3988 // b[i, i+1, i+2, i+3] = v2 - v3
3989 // br cond, vector.body, middle.block
3996 // s_init = phi [x, middle.block], [a[-1], otherwise]
3999 // After execution completes the vector loop, we extract the next value of
4000 // the recurrence (x) to use as the initial value in the scalar loop.
4002 // Get the original loop preheader and single loop latch.
4003 auto *Preheader = OrigLoop->getLoopPreheader();
4004 auto *Latch = OrigLoop->getLoopLatch();
4006 // Get the initial and previous values of the scalar recurrence.
4007 auto *ScalarInit = Phi->getIncomingValueForBlock(Preheader);
4008 auto *Previous = Phi->getIncomingValueForBlock(Latch);
4010 // Create a vector from the initial value.
4011 auto *VectorInit = ScalarInit;
4013 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
4014 VectorInit = Builder.CreateInsertElement(
4015 UndefValue::get(VectorType::get(VectorInit->getType(), VF)), VectorInit,
4016 Builder.getInt32(VF - 1), "vector.recur.init");
4019 // We constructed a temporary phi node in the first phase of vectorization.
4020 // This phi node will eventually be deleted.
4021 VectorParts &PhiParts = VectorLoopValueMap.getVector(Phi);
4022 Builder.SetInsertPoint(cast<Instruction>(PhiParts[0]));
4024 // Create a phi node for the new recurrence. The current value will either be
4025 // the initial value inserted into a vector or loop-varying vector value.
4026 auto *VecPhi = Builder.CreatePHI(VectorInit->getType(), 2, "vector.recur");
4027 VecPhi->addIncoming(VectorInit, LoopVectorPreHeader);
4029 // Get the vectorized previous value.
4030 auto &PreviousParts = getVectorValue(Previous);
4032 // Set the insertion point after the previous value if it is an instruction.
4033 // Note that the previous value may have been constant-folded so it is not
4034 // guaranteed to be an instruction in the vector loop. Also, if the previous
4035 // value is a phi node, we should insert after all the phi nodes to avoid
4036 // breaking basic block verification.
4037 if (LI->getLoopFor(LoopVectorBody)->isLoopInvariant(PreviousParts[UF - 1]) ||
4038 isa<PHINode>(PreviousParts[UF - 1]))
4039 Builder.SetInsertPoint(&*LoopVectorBody->getFirstInsertionPt());
4041 Builder.SetInsertPoint(
4042 &*++BasicBlock::iterator(cast<Instruction>(PreviousParts[UF - 1])));
4044 // We will construct a vector for the recurrence by combining the values for
4045 // the current and previous iterations. This is the required shuffle mask.
4046 SmallVector<Constant *, 8> ShuffleMask(VF);
4047 ShuffleMask[0] = Builder.getInt32(VF - 1);
4048 for (unsigned I = 1; I < VF; ++I)
4049 ShuffleMask[I] = Builder.getInt32(I + VF - 1);
4051 // The vector from which to take the initial value for the current iteration
4052 // (actual or unrolled). Initially, this is the vector phi node.
4053 Value *Incoming = VecPhi;
4055 // Shuffle the current and previous vector and update the vector parts.
4056 for (unsigned Part = 0; Part < UF; ++Part) {
4059 ? Builder.CreateShuffleVector(Incoming, PreviousParts[Part],
4060 ConstantVector::get(ShuffleMask))
4062 PhiParts[Part]->replaceAllUsesWith(Shuffle);
4063 cast<Instruction>(PhiParts[Part])->eraseFromParent();
4064 PhiParts[Part] = Shuffle;
4065 Incoming = PreviousParts[Part];
4068 // Fix the latch value of the new recurrence in the vector loop.
4069 VecPhi->addIncoming(Incoming, LI->getLoopFor(LoopVectorBody)->getLoopLatch());
4071 // Extract the last vector element in the middle block. This will be the
4072 // initial value for the recurrence when jumping to the scalar loop.
4073 auto *ExtractForScalar = Incoming;
4075 Builder.SetInsertPoint(LoopMiddleBlock->getTerminator());
4076 ExtractForScalar = Builder.CreateExtractElement(
4077 ExtractForScalar, Builder.getInt32(VF - 1), "vector.recur.extract");
4079 // Extract the second last element in the middle block if the
4080 // Phi is used outside the loop. We need to extract the phi itself
4081 // and not the last element (the phi update in the current iteration). This
4082 // will be the value when jumping to the exit block from the LoopMiddleBlock,
4083 // when the scalar loop is not run at all.
4084 Value *ExtractForPhiUsedOutsideLoop = nullptr;
4086 ExtractForPhiUsedOutsideLoop = Builder.CreateExtractElement(
4087 Incoming, Builder.getInt32(VF - 2), "vector.recur.extract.for.phi");
4088 // When loop is unrolled without vectorizing, initialize
4089 // ExtractForPhiUsedOutsideLoop with the value just prior to unrolled value of
4090 // `Incoming`. This is analogous to the vectorized case above: extracting the
4091 // second last element when VF > 1.
4093 ExtractForPhiUsedOutsideLoop = PreviousParts[UF - 2];
4095 // Fix the initial value of the original recurrence in the scalar loop.
4096 Builder.SetInsertPoint(&*LoopScalarPreHeader->begin());
4097 auto *Start = Builder.CreatePHI(Phi->getType(), 2, "scalar.recur.init");
4098 for (auto *BB : predecessors(LoopScalarPreHeader)) {
4099 auto *Incoming = BB == LoopMiddleBlock ? ExtractForScalar : ScalarInit;
4100 Start->addIncoming(Incoming, BB);
4103 Phi->setIncomingValue(Phi->getBasicBlockIndex(LoopScalarPreHeader), Start);
4104 Phi->setName("scalar.recur");
4106 // Finally, fix users of the recurrence outside the loop. The users will need
4107 // either the last value of the scalar recurrence or the last value of the
4108 // vector recurrence we extracted in the middle block. Since the loop is in
4109 // LCSSA form, we just need to find the phi node for the original scalar
4110 // recurrence in the exit block, and then add an edge for the middle block.
4111 for (auto &I : *LoopExitBlock) {
4112 auto *LCSSAPhi = dyn_cast<PHINode>(&I);
4115 if (LCSSAPhi->getIncomingValue(0) == Phi) {
4116 LCSSAPhi->addIncoming(ExtractForPhiUsedOutsideLoop, LoopMiddleBlock);
4122 void InnerLoopVectorizer::fixReduction(PHINode *Phi) {
4123 Constant *Zero = Builder.getInt32(0);
4125 // Get it's reduction variable descriptor.
4126 assert(Legal->isReductionVariable(Phi) &&
4127 "Unable to find the reduction variable");
4128 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[Phi];
4130 RecurrenceDescriptor::RecurrenceKind RK = RdxDesc.getRecurrenceKind();
4131 TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue();
4132 Instruction *LoopExitInst = RdxDesc.getLoopExitInstr();
4133 RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind =
4134 RdxDesc.getMinMaxRecurrenceKind();
4135 setDebugLocFromInst(Builder, ReductionStartValue);
4137 // We need to generate a reduction vector from the incoming scalar.
4138 // To do so, we need to generate the 'identity' vector and override
4139 // one of the elements with the incoming scalar reduction. We need
4140 // to do it in the vector-loop preheader.
4141 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
4143 // This is the vector-clone of the value that leaves the loop.
4144 const VectorParts &VectorExit = getVectorValue(LoopExitInst);
4145 Type *VecTy = VectorExit[0]->getType();
4147 // Find the reduction identity variable. Zero for addition, or, xor,
4148 // one for multiplication, -1 for And.
4151 if (RK == RecurrenceDescriptor::RK_IntegerMinMax ||
4152 RK == RecurrenceDescriptor::RK_FloatMinMax) {
4153 // MinMax reduction have the start value as their identify.
4155 VectorStart = Identity = ReductionStartValue;
4157 VectorStart = Identity =
4158 Builder.CreateVectorSplat(VF, ReductionStartValue, "minmax.ident");
4161 // Handle other reduction kinds:
4162 Constant *Iden = RecurrenceDescriptor::getRecurrenceIdentity(
4163 RK, VecTy->getScalarType());
4166 // This vector is the Identity vector where the first element is the
4167 // incoming scalar reduction.
4168 VectorStart = ReductionStartValue;
4170 Identity = ConstantVector::getSplat(VF, Iden);
4172 // This vector is the Identity vector where the first element is the
4173 // incoming scalar reduction.
4175 Builder.CreateInsertElement(Identity, ReductionStartValue, Zero);
4179 // Fix the vector-loop phi.
4181 // Reductions do not have to start at zero. They can start with
4182 // any loop invariant values.
4183 const VectorParts &VecRdxPhi = getVectorValue(Phi);
4184 BasicBlock *Latch = OrigLoop->getLoopLatch();
4185 Value *LoopVal = Phi->getIncomingValueForBlock(Latch);
4186 const VectorParts &Val = getVectorValue(LoopVal);
4187 for (unsigned part = 0; part < UF; ++part) {
4188 // Make sure to add the reduction stat value only to the
4189 // first unroll part.
4190 Value *StartVal = (part == 0) ? VectorStart : Identity;
4191 cast<PHINode>(VecRdxPhi[part])
4192 ->addIncoming(StartVal, LoopVectorPreHeader);
4193 cast<PHINode>(VecRdxPhi[part])
4194 ->addIncoming(Val[part], LI->getLoopFor(LoopVectorBody)->getLoopLatch());
4197 // Before each round, move the insertion point right between
4198 // the PHIs and the values we are going to write.
4199 // This allows us to write both PHINodes and the extractelement
4201 Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt());
4203 VectorParts &RdxParts = VectorLoopValueMap.getVector(LoopExitInst);
4204 setDebugLocFromInst(Builder, LoopExitInst);
4206 // If the vector reduction can be performed in a smaller type, we truncate
4207 // then extend the loop exit value to enable InstCombine to evaluate the
4208 // entire expression in the smaller type.
4209 if (VF > 1 && Phi->getType() != RdxDesc.getRecurrenceType()) {
4210 Type *RdxVecTy = VectorType::get(RdxDesc.getRecurrenceType(), VF);
4211 Builder.SetInsertPoint(LoopVectorBody->getTerminator());
4212 for (unsigned part = 0; part < UF; ++part) {
4213 Value *Trunc = Builder.CreateTrunc(RdxParts[part], RdxVecTy);
4214 Value *Extnd = RdxDesc.isSigned() ? Builder.CreateSExt(Trunc, VecTy)
4215 : Builder.CreateZExt(Trunc, VecTy);
4216 for (Value::user_iterator UI = RdxParts[part]->user_begin();
4217 UI != RdxParts[part]->user_end();)
4219 (*UI++)->replaceUsesOfWith(RdxParts[part], Extnd);
4220 RdxParts[part] = Extnd;
4225 Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt());
4226 for (unsigned part = 0; part < UF; ++part)
4227 RdxParts[part] = Builder.CreateTrunc(RdxParts[part], RdxVecTy);
4230 // Reduce all of the unrolled parts into a single vector.
4231 Value *ReducedPartRdx = RdxParts[0];
4232 unsigned Op = RecurrenceDescriptor::getRecurrenceBinOp(RK);
4233 setDebugLocFromInst(Builder, ReducedPartRdx);
4234 for (unsigned part = 1; part < UF; ++part) {
4235 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
4236 // Floating point operations had to be 'fast' to enable the reduction.
4237 ReducedPartRdx = addFastMathFlag(
4238 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
4239 ReducedPartRdx, "bin.rdx"));
4241 ReducedPartRdx = RecurrenceDescriptor::createMinMaxOp(
4242 Builder, MinMaxKind, ReducedPartRdx, RdxParts[part]);
4246 bool NoNaN = Legal->hasFunNoNaNAttr();
4248 createTargetReduction(Builder, TTI, RdxDesc, ReducedPartRdx, NoNaN);
4249 // If the reduction can be performed in a smaller type, we need to extend
4250 // the reduction to the wider type before we branch to the original loop.
4251 if (Phi->getType() != RdxDesc.getRecurrenceType())
4254 ? Builder.CreateSExt(ReducedPartRdx, Phi->getType())
4255 : Builder.CreateZExt(ReducedPartRdx, Phi->getType());
4258 // Create a phi node that merges control-flow from the backedge-taken check
4259 // block and the middle block.
4260 PHINode *BCBlockPhi = PHINode::Create(Phi->getType(), 2, "bc.merge.rdx",
4261 LoopScalarPreHeader->getTerminator());
4262 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
4263 BCBlockPhi->addIncoming(ReductionStartValue, LoopBypassBlocks[I]);
4264 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
4266 // Now, we need to fix the users of the reduction variable
4267 // inside and outside of the scalar remainder loop.
4268 // We know that the loop is in LCSSA form. We need to update the
4269 // PHI nodes in the exit blocks.
4270 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
4271 LEE = LoopExitBlock->end();
4272 LEI != LEE; ++LEI) {
4273 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
4277 // All PHINodes need to have a single entry edge, or two if
4278 // we already fixed them.
4279 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
4281 // We found a reduction value exit-PHI. Update it with the
4282 // incoming bypass edge.
4283 if (LCSSAPhi->getIncomingValue(0) == LoopExitInst)
4284 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
4285 } // end of the LCSSA phi scan.
4287 // Fix the scalar loop reduction variable with the incoming reduction sum
4288 // from the vector body and from the backedge value.
4289 int IncomingEdgeBlockIdx =
4290 Phi->getBasicBlockIndex(OrigLoop->getLoopLatch());
4291 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
4292 // Pick the other block.
4293 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
4294 Phi->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
4295 Phi->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst);
4298 void InnerLoopVectorizer::fixLCSSAPHIs() {
4299 for (Instruction &LEI : *LoopExitBlock) {
4300 auto *LCSSAPhi = dyn_cast<PHINode>(&LEI);
4303 if (LCSSAPhi->getNumIncomingValues() == 1) {
4304 assert(OrigLoop->isLoopInvariant(LCSSAPhi->getIncomingValue(0)) &&
4305 "Incoming value isn't loop invariant");
4306 LCSSAPhi->addIncoming(LCSSAPhi->getIncomingValue(0), LoopMiddleBlock);
4311 void InnerLoopVectorizer::sinkScalarOperands(Instruction *PredInst) {
4313 // The basic block and loop containing the predicated instruction.
4314 auto *PredBB = PredInst->getParent();
4315 auto *VectorLoop = LI->getLoopFor(PredBB);
4317 // Initialize a worklist with the operands of the predicated instruction.
4318 SetVector<Value *> Worklist(PredInst->op_begin(), PredInst->op_end());
4320 // Holds instructions that we need to analyze again. An instruction may be
4321 // reanalyzed if we don't yet know if we can sink it or not.
4322 SmallVector<Instruction *, 8> InstsToReanalyze;
4324 // Returns true if a given use occurs in the predicated block. Phi nodes use
4325 // their operands in their corresponding predecessor blocks.
4326 auto isBlockOfUsePredicated = [&](Use &U) -> bool {
4327 auto *I = cast<Instruction>(U.getUser());
4328 BasicBlock *BB = I->getParent();
4329 if (auto *Phi = dyn_cast<PHINode>(I))
4330 BB = Phi->getIncomingBlock(
4331 PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
4332 return BB == PredBB;
4335 // Iteratively sink the scalarized operands of the predicated instruction
4336 // into the block we created for it. When an instruction is sunk, it's
4337 // operands are then added to the worklist. The algorithm ends after one pass
4338 // through the worklist doesn't sink a single instruction.
4342 // Add the instructions that need to be reanalyzed to the worklist, and
4343 // reset the changed indicator.
4344 Worklist.insert(InstsToReanalyze.begin(), InstsToReanalyze.end());
4345 InstsToReanalyze.clear();
4348 while (!Worklist.empty()) {
4349 auto *I = dyn_cast<Instruction>(Worklist.pop_back_val());
4351 // We can't sink an instruction if it is a phi node, is already in the
4352 // predicated block, is not in the loop, or may have side effects.
4353 if (!I || isa<PHINode>(I) || I->getParent() == PredBB ||
4354 !VectorLoop->contains(I) || I->mayHaveSideEffects())
4357 // It's legal to sink the instruction if all its uses occur in the
4358 // predicated block. Otherwise, there's nothing to do yet, and we may
4359 // need to reanalyze the instruction.
4360 if (!all_of(I->uses(), isBlockOfUsePredicated)) {
4361 InstsToReanalyze.push_back(I);
4365 // Move the instruction to the beginning of the predicated block, and add
4366 // it's operands to the worklist.
4367 I->moveBefore(&*PredBB->getFirstInsertionPt());
4368 Worklist.insert(I->op_begin(), I->op_end());
4370 // The sinking may have enabled other instructions to be sunk, so we will
4377 void InnerLoopVectorizer::predicateInstructions() {
4379 // For each instruction I marked for predication on value C, split I into its
4380 // own basic block to form an if-then construct over C. Since I may be fed by
4381 // an extractelement instruction or other scalar operand, we try to
4382 // iteratively sink its scalar operands into the predicated block. If I feeds
4383 // an insertelement instruction, we try to move this instruction into the
4384 // predicated block as well. For non-void types, a phi node will be created
4385 // for the resulting value (either vector or scalar).
4387 // So for some predicated instruction, e.g. the conditional sdiv in:
4391 // %add = add nsw i32 %mul, %0
4392 // %cmp5 = icmp sgt i32 %2, 7
4393 // br i1 %cmp5, label %if.then, label %if.end
4396 // %div = sdiv i32 %0, %1
4400 // %x.0 = phi i32 [ %div, %if.then ], [ %add, %for.body ]
4402 // the sdiv at this point is scalarized and if-converted using a select.
4403 // The inactive elements in the vector are not used, but the predicated
4404 // instruction is still executed for all vector elements, essentially:
4408 // %17 = add nsw <2 x i32> %16, %wide.load
4409 // %29 = extractelement <2 x i32> %wide.load, i32 0
4410 // %30 = extractelement <2 x i32> %wide.load51, i32 0
4411 // %31 = sdiv i32 %29, %30
4412 // %32 = insertelement <2 x i32> undef, i32 %31, i32 0
4413 // %35 = extractelement <2 x i32> %wide.load, i32 1
4414 // %36 = extractelement <2 x i32> %wide.load51, i32 1
4415 // %37 = sdiv i32 %35, %36
4416 // %38 = insertelement <2 x i32> %32, i32 %37, i32 1
4417 // %predphi = select <2 x i1> %26, <2 x i32> %38, <2 x i32> %17
4419 // Predication will now re-introduce the original control flow to avoid false
4420 // side-effects by the sdiv instructions on the inactive elements, yielding
4425 // %5 = add nsw <2 x i32> %4, %wide.load
4426 // %8 = icmp sgt <2 x i32> %wide.load52, <i32 7, i32 7>
4427 // %9 = extractelement <2 x i1> %8, i32 0
4428 // br i1 %9, label %pred.sdiv.if, label %pred.sdiv.continue
4431 // %10 = extractelement <2 x i32> %wide.load, i32 0
4432 // %11 = extractelement <2 x i32> %wide.load51, i32 0
4433 // %12 = sdiv i32 %10, %11
4434 // %13 = insertelement <2 x i32> undef, i32 %12, i32 0
4435 // br label %pred.sdiv.continue
4437 // pred.sdiv.continue:
4438 // %14 = phi <2 x i32> [ undef, %vector.body ], [ %13, %pred.sdiv.if ]
4439 // %15 = extractelement <2 x i1> %8, i32 1
4440 // br i1 %15, label %pred.sdiv.if54, label %pred.sdiv.continue55
4443 // %16 = extractelement <2 x i32> %wide.load, i32 1
4444 // %17 = extractelement <2 x i32> %wide.load51, i32 1
4445 // %18 = sdiv i32 %16, %17
4446 // %19 = insertelement <2 x i32> %14, i32 %18, i32 1
4447 // br label %pred.sdiv.continue55
4449 // pred.sdiv.continue55:
4450 // %20 = phi <2 x i32> [ %14, %pred.sdiv.continue ], [ %19, %pred.sdiv.if54 ]
4451 // %predphi = select <2 x i1> %8, <2 x i32> %20, <2 x i32> %5
4453 for (auto KV : PredicatedInstructions) {
4454 BasicBlock::iterator I(KV.first);
4455 BasicBlock *Head = I->getParent();
4456 auto *T = SplitBlockAndInsertIfThen(KV.second, &*I, /*Unreachable=*/false,
4457 /*BranchWeights=*/nullptr, DT, LI);
4459 sinkScalarOperands(&*I);
4461 BasicBlock *PredicatedBlock = I->getParent();
4462 Twine BBNamePrefix = Twine("pred.") + I->getOpcodeName();
4463 PredicatedBlock->setName(BBNamePrefix + ".if");
4464 PredicatedBlock->getSingleSuccessor()->setName(BBNamePrefix + ".continue");
4466 // If the instruction is non-void create a Phi node at reconvergence point.
4467 if (!I->getType()->isVoidTy()) {
4468 Value *IncomingTrue = nullptr;
4469 Value *IncomingFalse = nullptr;
4471 if (I->hasOneUse() && isa<InsertElementInst>(*I->user_begin())) {
4472 // If the predicated instruction is feeding an insert-element, move it
4473 // into the Then block; Phi node will be created for the vector.
4474 InsertElementInst *IEI = cast<InsertElementInst>(*I->user_begin());
4476 IncomingTrue = IEI; // the new vector with the inserted element.
4477 IncomingFalse = IEI->getOperand(0); // the unmodified vector
4479 // Phi node will be created for the scalar predicated instruction.
4481 IncomingFalse = UndefValue::get(I->getType());
4484 BasicBlock *PostDom = I->getParent()->getSingleSuccessor();
4485 assert(PostDom && "Then block has multiple successors");
4487 PHINode::Create(IncomingTrue->getType(), 2, "", &PostDom->front());
4488 IncomingTrue->replaceAllUsesWith(Phi);
4489 Phi->addIncoming(IncomingFalse, Head);
4490 Phi->addIncoming(IncomingTrue, I->getParent());
4494 DEBUG(DT->verifyDomTree());
4497 InnerLoopVectorizer::VectorParts
4498 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
4499 assert(is_contained(predecessors(Dst), Src) && "Invalid edge");
4501 // Look for cached value.
4502 std::pair<BasicBlock *, BasicBlock *> Edge(Src, Dst);
4503 EdgeMaskCacheTy::iterator ECEntryIt = EdgeMaskCache.find(Edge);
4504 if (ECEntryIt != EdgeMaskCache.end())
4505 return ECEntryIt->second;
4507 VectorParts SrcMask = createBlockInMask(Src);
4509 // The terminator has to be a branch inst!
4510 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
4511 assert(BI && "Unexpected terminator found");
4513 if (BI->isConditional()) {
4514 VectorParts EdgeMask = getVectorValue(BI->getCondition());
4516 if (BI->getSuccessor(0) != Dst)
4517 for (unsigned part = 0; part < UF; ++part)
4518 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
4520 for (unsigned part = 0; part < UF; ++part)
4521 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
4523 EdgeMaskCache[Edge] = EdgeMask;
4527 EdgeMaskCache[Edge] = SrcMask;
4531 InnerLoopVectorizer::VectorParts
4532 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
4533 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
4535 // Look for cached value.
4536 BlockMaskCacheTy::iterator BCEntryIt = BlockMaskCache.find(BB);
4537 if (BCEntryIt != BlockMaskCache.end())
4538 return BCEntryIt->second;
4540 // Loop incoming mask is all-one.
4541 if (OrigLoop->getHeader() == BB) {
4542 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
4543 const VectorParts &BlockMask = getVectorValue(C);
4544 BlockMaskCache[BB] = BlockMask;
4548 // This is the block mask. We OR all incoming edges, and with zero.
4549 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
4550 VectorParts BlockMask = getVectorValue(Zero);
4553 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
4554 VectorParts EM = createEdgeMask(*it, BB);
4555 for (unsigned part = 0; part < UF; ++part)
4556 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
4559 BlockMaskCache[BB] = BlockMask;
4563 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN, unsigned UF,
4565 PHINode *P = cast<PHINode>(PN);
4566 // In order to support recurrences we need to be able to vectorize Phi nodes.
4567 // Phi nodes have cycles, so we need to vectorize them in two stages. This is
4568 // stage #1: We create a new vector PHI node with no incoming edges. We'll use
4569 // this value when we vectorize all of the instructions that use the PHI.
4570 if (Legal->isReductionVariable(P) || Legal->isFirstOrderRecurrence(P)) {
4571 VectorParts Entry(UF);
4572 for (unsigned part = 0; part < UF; ++part) {
4573 // This is phase one of vectorizing PHIs.
4575 (VF == 1) ? PN->getType() : VectorType::get(PN->getType(), VF);
4576 Entry[part] = PHINode::Create(
4577 VecTy, 2, "vec.phi", &*LoopVectorBody->getFirstInsertionPt());
4579 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 case InductionDescriptor::IK_FpInduction:
4635 return widenIntOrFpInduction(P);
4636 case InductionDescriptor::IK_PtrInduction: {
4637 // Handle the pointer induction variable case.
4638 assert(P->getType()->isPointerTy() && "Unexpected type.");
4639 // This is the normalized GEP that starts counting at zero.
4640 Value *PtrInd = Induction;
4641 PtrInd = Builder.CreateSExtOrTrunc(PtrInd, II.getStep()->getType());
4642 // Determine the number of scalars we need to generate for each unroll
4643 // iteration. If the instruction is uniform, we only need to generate the
4644 // first lane. Otherwise, we generate all VF values.
4645 unsigned Lanes = Cost->isUniformAfterVectorization(P, VF) ? 1 : VF;
4646 // These are the scalar results. Notice that we don't generate vector GEPs
4647 // because scalar GEPs result in better code.
4648 ScalarParts Entry(UF);
4649 for (unsigned Part = 0; Part < UF; ++Part) {
4650 Entry[Part].resize(VF);
4651 for (unsigned Lane = 0; Lane < Lanes; ++Lane) {
4652 Constant *Idx = ConstantInt::get(PtrInd->getType(), Lane + Part * VF);
4653 Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx);
4654 Value *SclrGep = II.transform(Builder, GlobalIdx, PSE.getSE(), DL);
4655 SclrGep->setName("next.gep");
4656 Entry[Part][Lane] = SclrGep;
4659 VectorLoopValueMap.initScalar(P, Entry);
4665 /// A helper function for checking whether an integer division-related
4666 /// instruction may divide by zero (in which case it must be predicated if
4667 /// executed conditionally in the scalar code).
4668 /// TODO: It may be worthwhile to generalize and check isKnownNonZero().
4669 /// Non-zero divisors that are non compile-time constants will not be
4670 /// converted into multiplication, so we will still end up scalarizing
4671 /// the division, but can do so w/o predication.
4672 static bool mayDivideByZero(Instruction &I) {
4673 assert((I.getOpcode() == Instruction::UDiv ||
4674 I.getOpcode() == Instruction::SDiv ||
4675 I.getOpcode() == Instruction::URem ||
4676 I.getOpcode() == Instruction::SRem) &&
4677 "Unexpected instruction");
4678 Value *Divisor = I.getOperand(1);
4679 auto *CInt = dyn_cast<ConstantInt>(Divisor);
4680 return !CInt || CInt->isZero();
4683 void InnerLoopVectorizer::vectorizeInstruction(Instruction &I) {
4684 // Scalarize instructions that should remain scalar after vectorization.
4686 !(isa<BranchInst>(&I) || isa<PHINode>(&I) || isa<DbgInfoIntrinsic>(&I)) &&
4687 shouldScalarizeInstruction(&I)) {
4688 scalarizeInstruction(&I, Legal->isScalarWithPredication(&I));
4692 switch (I.getOpcode()) {
4693 case Instruction::Br:
4694 // Nothing to do for PHIs and BR, since we already took care of the
4695 // loop control flow instructions.
4697 case Instruction::PHI: {
4698 // Vectorize PHINodes.
4699 widenPHIInstruction(&I, UF, VF);
4702 case Instruction::GetElementPtr: {
4703 // Construct a vector GEP by widening the operands of the scalar GEP as
4704 // necessary. We mark the vector GEP 'inbounds' if appropriate. A GEP
4705 // results in a vector of pointers when at least one operand of the GEP
4706 // is vector-typed. Thus, to keep the representation compact, we only use
4707 // vector-typed operands for loop-varying values.
4708 auto *GEP = cast<GetElementPtrInst>(&I);
4709 VectorParts Entry(UF);
4711 if (VF > 1 && OrigLoop->hasLoopInvariantOperands(GEP)) {
4712 // If we are vectorizing, but the GEP has only loop-invariant operands,
4713 // the GEP we build (by only using vector-typed operands for
4714 // loop-varying values) would be a scalar pointer. Thus, to ensure we
4715 // produce a vector of pointers, we need to either arbitrarily pick an
4716 // operand to broadcast, or broadcast a clone of the original GEP.
4717 // Here, we broadcast a clone of the original.
4719 // TODO: If at some point we decide to scalarize instructions having
4720 // loop-invariant operands, this special case will no longer be
4721 // required. We would add the scalarization decision to
4722 // collectLoopScalars() and teach getVectorValue() to broadcast
4723 // the lane-zero scalar value.
4724 auto *Clone = Builder.Insert(GEP->clone());
4725 for (unsigned Part = 0; Part < UF; ++Part)
4726 Entry[Part] = Builder.CreateVectorSplat(VF, Clone);
4728 // If the GEP has at least one loop-varying operand, we are sure to
4729 // produce a vector of pointers. But if we are only unrolling, we want
4730 // to produce a scalar GEP for each unroll part. Thus, the GEP we
4731 // produce with the code below will be scalar (if VF == 1) or vector
4732 // (otherwise). Note that for the unroll-only case, we still maintain
4733 // values in the vector mapping with initVector, as we do for other
4735 for (unsigned Part = 0; Part < UF; ++Part) {
4737 // The pointer operand of the new GEP. If it's loop-invariant, we
4738 // won't broadcast it.
4739 auto *Ptr = OrigLoop->isLoopInvariant(GEP->getPointerOperand())
4740 ? GEP->getPointerOperand()
4741 : getVectorValue(GEP->getPointerOperand())[Part];
4743 // Collect all the indices for the new GEP. If any index is
4744 // loop-invariant, we won't broadcast it.
4745 SmallVector<Value *, 4> Indices;
4746 for (auto &U : make_range(GEP->idx_begin(), GEP->idx_end())) {
4747 if (OrigLoop->isLoopInvariant(U.get()))
4748 Indices.push_back(U.get());
4750 Indices.push_back(getVectorValue(U.get())[Part]);
4753 // Create the new GEP. Note that this GEP may be a scalar if VF == 1,
4754 // but it should be a vector, otherwise.
4755 auto *NewGEP = GEP->isInBounds()
4756 ? Builder.CreateInBoundsGEP(Ptr, Indices)
4757 : Builder.CreateGEP(Ptr, Indices);
4758 assert((VF == 1 || NewGEP->getType()->isVectorTy()) &&
4759 "NewGEP is not a pointer vector");
4760 Entry[Part] = NewGEP;
4764 VectorLoopValueMap.initVector(&I, Entry);
4765 addMetadata(Entry, GEP);
4768 case Instruction::UDiv:
4769 case Instruction::SDiv:
4770 case Instruction::SRem:
4771 case Instruction::URem:
4772 // Scalarize with predication if this instruction may divide by zero and
4773 // block execution is conditional, otherwise fallthrough.
4774 if (Legal->isScalarWithPredication(&I)) {
4775 scalarizeInstruction(&I, true);
4778 case Instruction::Add:
4779 case Instruction::FAdd:
4780 case Instruction::Sub:
4781 case Instruction::FSub:
4782 case Instruction::Mul:
4783 case Instruction::FMul:
4784 case Instruction::FDiv:
4785 case Instruction::FRem:
4786 case Instruction::Shl:
4787 case Instruction::LShr:
4788 case Instruction::AShr:
4789 case Instruction::And:
4790 case Instruction::Or:
4791 case Instruction::Xor: {
4792 // Just widen binops.
4793 auto *BinOp = cast<BinaryOperator>(&I);
4794 setDebugLocFromInst(Builder, BinOp);
4795 const VectorParts &A = getVectorValue(BinOp->getOperand(0));
4796 const VectorParts &B = getVectorValue(BinOp->getOperand(1));
4798 // Use this vector value for all users of the original instruction.
4799 VectorParts Entry(UF);
4800 for (unsigned Part = 0; Part < UF; ++Part) {
4801 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
4803 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
4804 VecOp->copyIRFlags(BinOp);
4809 VectorLoopValueMap.initVector(&I, Entry);
4810 addMetadata(Entry, BinOp);
4813 case Instruction::Select: {
4815 // If the selector is loop invariant we can create a select
4816 // instruction with a scalar condition. Otherwise, use vector-select.
4817 auto *SE = PSE.getSE();
4818 bool InvariantCond =
4819 SE->isLoopInvariant(PSE.getSCEV(I.getOperand(0)), OrigLoop);
4820 setDebugLocFromInst(Builder, &I);
4822 // The condition can be loop invariant but still defined inside the
4823 // loop. This means that we can't just use the original 'cond' value.
4824 // We have to take the 'vectorized' value and pick the first lane.
4825 // Instcombine will make this a no-op.
4826 const VectorParts &Cond = getVectorValue(I.getOperand(0));
4827 const VectorParts &Op0 = getVectorValue(I.getOperand(1));
4828 const VectorParts &Op1 = getVectorValue(I.getOperand(2));
4830 auto *ScalarCond = getScalarValue(I.getOperand(0), 0, 0);
4832 VectorParts Entry(UF);
4833 for (unsigned Part = 0; Part < UF; ++Part) {
4834 Entry[Part] = Builder.CreateSelect(
4835 InvariantCond ? ScalarCond : Cond[Part], Op0[Part], Op1[Part]);
4838 VectorLoopValueMap.initVector(&I, Entry);
4839 addMetadata(Entry, &I);
4843 case Instruction::ICmp:
4844 case Instruction::FCmp: {
4845 // Widen compares. Generate vector compares.
4846 bool FCmp = (I.getOpcode() == Instruction::FCmp);
4847 auto *Cmp = dyn_cast<CmpInst>(&I);
4848 setDebugLocFromInst(Builder, Cmp);
4849 const VectorParts &A = getVectorValue(Cmp->getOperand(0));
4850 const VectorParts &B = getVectorValue(Cmp->getOperand(1));
4851 VectorParts Entry(UF);
4852 for (unsigned Part = 0; Part < UF; ++Part) {
4855 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
4856 cast<FCmpInst>(C)->copyFastMathFlags(Cmp);
4858 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
4863 VectorLoopValueMap.initVector(&I, Entry);
4864 addMetadata(Entry, &I);
4868 case Instruction::Store:
4869 case Instruction::Load:
4870 vectorizeMemoryInstruction(&I);
4872 case Instruction::ZExt:
4873 case Instruction::SExt:
4874 case Instruction::FPToUI:
4875 case Instruction::FPToSI:
4876 case Instruction::FPExt:
4877 case Instruction::PtrToInt:
4878 case Instruction::IntToPtr:
4879 case Instruction::SIToFP:
4880 case Instruction::UIToFP:
4881 case Instruction::Trunc:
4882 case Instruction::FPTrunc:
4883 case Instruction::BitCast: {
4884 auto *CI = dyn_cast<CastInst>(&I);
4885 setDebugLocFromInst(Builder, CI);
4887 // Optimize the special case where the source is a constant integer
4888 // induction variable. Notice that we can only optimize the 'trunc' case
4889 // because (a) FP conversions lose precision, (b) sext/zext may wrap, and
4890 // (c) other casts depend on pointer size.
4891 if (Cost->isOptimizableIVTruncate(CI, VF)) {
4892 widenIntOrFpInduction(cast<PHINode>(CI->getOperand(0)),
4893 cast<TruncInst>(CI));
4897 /// Vectorize casts.
4899 (VF == 1) ? CI->getType() : VectorType::get(CI->getType(), VF);
4901 const VectorParts &A = getVectorValue(CI->getOperand(0));
4902 VectorParts Entry(UF);
4903 for (unsigned Part = 0; Part < UF; ++Part)
4904 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
4905 VectorLoopValueMap.initVector(&I, Entry);
4906 addMetadata(Entry, &I);
4910 case Instruction::Call: {
4911 // Ignore dbg intrinsics.
4912 if (isa<DbgInfoIntrinsic>(I))
4914 setDebugLocFromInst(Builder, &I);
4916 Module *M = I.getParent()->getParent()->getParent();
4917 auto *CI = cast<CallInst>(&I);
4919 StringRef FnName = CI->getCalledFunction()->getName();
4920 Function *F = CI->getCalledFunction();
4921 Type *RetTy = ToVectorTy(CI->getType(), VF);
4922 SmallVector<Type *, 4> Tys;
4923 for (Value *ArgOperand : CI->arg_operands())
4924 Tys.push_back(ToVectorTy(ArgOperand->getType(), VF));
4926 Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
4927 if (ID && (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end ||
4928 ID == Intrinsic::lifetime_start)) {
4929 scalarizeInstruction(&I);
4932 // The flag shows whether we use Intrinsic or a usual Call for vectorized
4933 // version of the instruction.
4934 // Is it beneficial to perform intrinsic call compared to lib call?
4935 bool NeedToScalarize;
4936 unsigned CallCost = getVectorCallCost(CI, VF, *TTI, TLI, NeedToScalarize);
4937 bool UseVectorIntrinsic =
4938 ID && getVectorIntrinsicCost(CI, VF, *TTI, TLI) <= CallCost;
4939 if (!UseVectorIntrinsic && NeedToScalarize) {
4940 scalarizeInstruction(&I);
4944 VectorParts Entry(UF);
4945 for (unsigned Part = 0; Part < UF; ++Part) {
4946 SmallVector<Value *, 4> Args;
4947 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
4948 Value *Arg = CI->getArgOperand(i);
4949 // Some intrinsics have a scalar argument - don't replace it with a
4951 if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, i)) {
4952 const VectorParts &VectorArg = getVectorValue(CI->getArgOperand(i));
4953 Arg = VectorArg[Part];
4955 Args.push_back(Arg);
4959 if (UseVectorIntrinsic) {
4960 // Use vector version of the intrinsic.
4961 Type *TysForDecl[] = {CI->getType()};
4963 TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF);
4964 VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl);
4966 // Use vector version of the library call.
4967 StringRef VFnName = TLI->getVectorizedFunction(FnName, VF);
4968 assert(!VFnName.empty() && "Vector function name is empty.");
4969 VectorF = M->getFunction(VFnName);
4971 // Generate a declaration
4972 FunctionType *FTy = FunctionType::get(RetTy, Tys, false);
4974 Function::Create(FTy, Function::ExternalLinkage, VFnName, M);
4975 VectorF->copyAttributesFrom(F);
4978 assert(VectorF && "Can't create vector function.");
4980 SmallVector<OperandBundleDef, 1> OpBundles;
4981 CI->getOperandBundlesAsDefs(OpBundles);
4982 CallInst *V = Builder.CreateCall(VectorF, Args, OpBundles);
4984 if (isa<FPMathOperator>(V))
4985 V->copyFastMathFlags(CI);
4990 VectorLoopValueMap.initVector(&I, Entry);
4991 addMetadata(Entry, &I);
4996 // All other instructions are unsupported. Scalarize them.
4997 scalarizeInstruction(&I);
5002 void InnerLoopVectorizer::updateAnalysis() {
5003 // Forget the original basic block.
5004 PSE.getSE()->forgetLoop(OrigLoop);
5006 // Update the dominator tree information.
5007 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
5008 "Entry does not dominate exit.");
5010 DT->addNewBlock(LI->getLoopFor(LoopVectorBody)->getHeader(),
5011 LoopVectorPreHeader);
5012 DT->addNewBlock(LoopMiddleBlock,
5013 LI->getLoopFor(LoopVectorBody)->getLoopLatch());
5014 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
5015 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
5016 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
5018 DEBUG(DT->verifyDomTree());
5021 /// \brief Check whether it is safe to if-convert this phi node.
5023 /// Phi nodes with constant expressions that can trap are not safe to if
5025 static bool canIfConvertPHINodes(BasicBlock *BB) {
5026 for (Instruction &I : *BB) {
5027 auto *Phi = dyn_cast<PHINode>(&I);
5030 for (Value *V : Phi->incoming_values())
5031 if (auto *C = dyn_cast<Constant>(V))
5038 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
5039 if (!EnableIfConversion) {
5040 ORE->emit(createMissedAnalysis("IfConversionDisabled")
5041 << "if-conversion is disabled");
5045 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
5047 // A list of pointers that we can safely read and write to.
5048 SmallPtrSet<Value *, 8> SafePointes;
5050 // Collect safe addresses.
5051 for (BasicBlock *BB : TheLoop->blocks()) {
5052 if (blockNeedsPredication(BB))
5055 for (Instruction &I : *BB)
5056 if (auto *Ptr = getPointerOperand(&I))
5057 SafePointes.insert(Ptr);
5060 // Collect the blocks that need predication.
5061 BasicBlock *Header = TheLoop->getHeader();
5062 for (BasicBlock *BB : TheLoop->blocks()) {
5063 // We don't support switch statements inside loops.
5064 if (!isa<BranchInst>(BB->getTerminator())) {
5065 ORE->emit(createMissedAnalysis("LoopContainsSwitch", BB->getTerminator())
5066 << "loop contains a switch statement");
5070 // We must be able to predicate all blocks that need to be predicated.
5071 if (blockNeedsPredication(BB)) {
5072 if (!blockCanBePredicated(BB, SafePointes)) {
5073 ORE->emit(createMissedAnalysis("NoCFGForSelect", BB->getTerminator())
5074 << "control flow cannot be substituted for a select");
5077 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
5078 ORE->emit(createMissedAnalysis("NoCFGForSelect", BB->getTerminator())
5079 << "control flow cannot be substituted for a select");
5084 // We can if-convert this loop.
5088 bool LoopVectorizationLegality::canVectorize() {
5089 // We must have a loop in canonical form. Loops with indirectbr in them cannot
5090 // be canonicalized.
5091 if (!TheLoop->getLoopPreheader()) {
5092 ORE->emit(createMissedAnalysis("CFGNotUnderstood")
5093 << "loop control flow is not understood by vectorizer");
5097 // FIXME: The code is currently dead, since the loop gets sent to
5098 // LoopVectorizationLegality is already an innermost loop.
5100 // We can only vectorize innermost loops.
5101 if (!TheLoop->empty()) {
5102 ORE->emit(createMissedAnalysis("NotInnermostLoop")
5103 << "loop is not the innermost loop");
5107 // We must have a single backedge.
5108 if (TheLoop->getNumBackEdges() != 1) {
5109 ORE->emit(createMissedAnalysis("CFGNotUnderstood")
5110 << "loop control flow is not understood by vectorizer");
5114 // We must have a single exiting block.
5115 if (!TheLoop->getExitingBlock()) {
5116 ORE->emit(createMissedAnalysis("CFGNotUnderstood")
5117 << "loop control flow is not understood by vectorizer");
5121 // We only handle bottom-tested loops, i.e. loop in which the condition is
5122 // checked at the end of each iteration. With that we can assume that all
5123 // instructions in the loop are executed the same number of times.
5124 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
5125 ORE->emit(createMissedAnalysis("CFGNotUnderstood")
5126 << "loop control flow is not understood by vectorizer");
5130 // We need to have a loop header.
5131 DEBUG(dbgs() << "LV: Found a loop: " << TheLoop->getHeader()->getName()
5134 // Check if we can if-convert non-single-bb loops.
5135 unsigned NumBlocks = TheLoop->getNumBlocks();
5136 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
5137 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
5141 // ScalarEvolution needs to be able to find the exit count.
5142 const SCEV *ExitCount = PSE.getBackedgeTakenCount();
5143 if (ExitCount == PSE.getSE()->getCouldNotCompute()) {
5144 ORE->emit(createMissedAnalysis("CantComputeNumberOfIterations")
5145 << "could not determine number of loop iterations");
5146 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
5150 // Check if we can vectorize the instructions and CFG in this loop.
5151 if (!canVectorizeInstrs()) {
5152 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
5156 // Go over each instruction and look at memory deps.
5157 if (!canVectorizeMemory()) {
5158 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
5162 DEBUG(dbgs() << "LV: We can vectorize this loop"
5163 << (LAI->getRuntimePointerChecking()->Need
5164 ? " (with a runtime bound check)"
5168 bool UseInterleaved = TTI->enableInterleavedAccessVectorization();
5170 // If an override option has been passed in for interleaved accesses, use it.
5171 if (EnableInterleavedMemAccesses.getNumOccurrences() > 0)
5172 UseInterleaved = EnableInterleavedMemAccesses;
5174 // Analyze interleaved memory accesses.
5176 InterleaveInfo.analyzeInterleaving(*getSymbolicStrides());
5178 unsigned SCEVThreshold = VectorizeSCEVCheckThreshold;
5179 if (Hints->getForce() == LoopVectorizeHints::FK_Enabled)
5180 SCEVThreshold = PragmaVectorizeSCEVCheckThreshold;
5182 if (PSE.getUnionPredicate().getComplexity() > SCEVThreshold) {
5183 ORE->emit(createMissedAnalysis("TooManySCEVRunTimeChecks")
5184 << "Too many SCEV assumptions need to be made and checked "
5186 DEBUG(dbgs() << "LV: Too many SCEV checks needed.\n");
5190 // Okay! We can vectorize. At this point we don't have any other mem analysis
5191 // which may limit our maximum vectorization factor, so just return true with
5196 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
5197 if (Ty->isPointerTy())
5198 return DL.getIntPtrType(Ty);
5200 // It is possible that char's or short's overflow when we ask for the loop's
5201 // trip count, work around this by changing the type size.
5202 if (Ty->getScalarSizeInBits() < 32)
5203 return Type::getInt32Ty(Ty->getContext());
5208 static Type *getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
5209 Ty0 = convertPointerToIntegerType(DL, Ty0);
5210 Ty1 = convertPointerToIntegerType(DL, Ty1);
5211 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
5216 /// \brief Check that the instruction has outside loop users and is not an
5217 /// identified reduction variable.
5218 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
5219 SmallPtrSetImpl<Value *> &AllowedExit) {
5220 // Reduction and Induction instructions are allowed to have exit users. All
5221 // other instructions must not have external users.
5222 if (!AllowedExit.count(Inst))
5223 // Check that all of the users of the loop are inside the BB.
5224 for (User *U : Inst->users()) {
5225 Instruction *UI = cast<Instruction>(U);
5226 // This user may be a reduction exit value.
5227 if (!TheLoop->contains(UI)) {
5228 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
5235 void LoopVectorizationLegality::addInductionPhi(
5236 PHINode *Phi, const InductionDescriptor &ID,
5237 SmallPtrSetImpl<Value *> &AllowedExit) {
5238 Inductions[Phi] = ID;
5239 Type *PhiTy = Phi->getType();
5240 const DataLayout &DL = Phi->getModule()->getDataLayout();
5242 // Get the widest type.
5243 if (!PhiTy->isFloatingPointTy()) {
5245 WidestIndTy = convertPointerToIntegerType(DL, PhiTy);
5247 WidestIndTy = getWiderType(DL, PhiTy, WidestIndTy);
5250 // Int inductions are special because we only allow one IV.
5251 if (ID.getKind() == InductionDescriptor::IK_IntInduction &&
5252 ID.getConstIntStepValue() &&
5253 ID.getConstIntStepValue()->isOne() &&
5254 isa<Constant>(ID.getStartValue()) &&
5255 cast<Constant>(ID.getStartValue())->isNullValue()) {
5257 // Use the phi node with the widest type as induction. Use the last
5258 // one if there are multiple (no good reason for doing this other
5259 // than it is expedient). We've checked that it begins at zero and
5260 // steps by one, so this is a canonical induction variable.
5261 if (!PrimaryInduction || PhiTy == WidestIndTy)
5262 PrimaryInduction = Phi;
5265 // Both the PHI node itself, and the "post-increment" value feeding
5266 // back into the PHI node may have external users.
5267 AllowedExit.insert(Phi);
5268 AllowedExit.insert(Phi->getIncomingValueForBlock(TheLoop->getLoopLatch()));
5270 DEBUG(dbgs() << "LV: Found an induction variable.\n");
5274 bool LoopVectorizationLegality::canVectorizeInstrs() {
5275 BasicBlock *Header = TheLoop->getHeader();
5277 // Look for the attribute signaling the absence of NaNs.
5278 Function &F = *Header->getParent();
5280 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
5282 // For each block in the loop.
5283 for (BasicBlock *BB : TheLoop->blocks()) {
5284 // Scan the instructions in the block and look for hazards.
5285 for (Instruction &I : *BB) {
5286 if (auto *Phi = dyn_cast<PHINode>(&I)) {
5287 Type *PhiTy = Phi->getType();
5288 // Check that this PHI type is allowed.
5289 if (!PhiTy->isIntegerTy() && !PhiTy->isFloatingPointTy() &&
5290 !PhiTy->isPointerTy()) {
5291 ORE->emit(createMissedAnalysis("CFGNotUnderstood", Phi)
5292 << "loop control flow is not understood by vectorizer");
5293 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
5297 // If this PHINode is not in the header block, then we know that we
5298 // can convert it to select during if-conversion. No need to check if
5299 // the PHIs in this block are induction or reduction variables.
5301 // Check that this instruction has no outside users or is an
5302 // identified reduction value with an outside user.
5303 if (!hasOutsideLoopUser(TheLoop, Phi, AllowedExit))
5305 ORE->emit(createMissedAnalysis("NeitherInductionNorReduction", Phi)
5306 << "value could not be identified as "
5307 "an induction or reduction variable");
5311 // We only allow if-converted PHIs with exactly two incoming values.
5312 if (Phi->getNumIncomingValues() != 2) {
5313 ORE->emit(createMissedAnalysis("CFGNotUnderstood", Phi)
5314 << "control flow not understood by vectorizer");
5315 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
5319 RecurrenceDescriptor RedDes;
5320 if (RecurrenceDescriptor::isReductionPHI(Phi, TheLoop, RedDes)) {
5321 if (RedDes.hasUnsafeAlgebra())
5322 Requirements->addUnsafeAlgebraInst(RedDes.getUnsafeAlgebraInst());
5323 AllowedExit.insert(RedDes.getLoopExitInstr());
5324 Reductions[Phi] = RedDes;
5328 InductionDescriptor ID;
5329 if (InductionDescriptor::isInductionPHI(Phi, TheLoop, PSE, ID)) {
5330 addInductionPhi(Phi, ID, AllowedExit);
5331 if (ID.hasUnsafeAlgebra() && !HasFunNoNaNAttr)
5332 Requirements->addUnsafeAlgebraInst(ID.getUnsafeAlgebraInst());
5336 if (RecurrenceDescriptor::isFirstOrderRecurrence(Phi, TheLoop, DT)) {
5337 FirstOrderRecurrences.insert(Phi);
5341 // As a last resort, coerce the PHI to a AddRec expression
5342 // and re-try classifying it a an induction PHI.
5343 if (InductionDescriptor::isInductionPHI(Phi, TheLoop, PSE, ID, true)) {
5344 addInductionPhi(Phi, ID, AllowedExit);
5348 ORE->emit(createMissedAnalysis("NonReductionValueUsedOutsideLoop", Phi)
5349 << "value that could not be identified as "
5350 "reduction is used outside the loop");
5351 DEBUG(dbgs() << "LV: Found an unidentified PHI." << *Phi << "\n");
5353 } // end of PHI handling
5355 // We handle calls that:
5356 // * Are debug info intrinsics.
5357 // * Have a mapping to an IR intrinsic.
5358 // * Have a vector version available.
5359 auto *CI = dyn_cast<CallInst>(&I);
5360 if (CI && !getVectorIntrinsicIDForCall(CI, TLI) &&
5361 !isa<DbgInfoIntrinsic>(CI) &&
5362 !(CI->getCalledFunction() && TLI &&
5363 TLI->isFunctionVectorizable(CI->getCalledFunction()->getName()))) {
5364 ORE->emit(createMissedAnalysis("CantVectorizeCall", CI)
5365 << "call instruction cannot be vectorized");
5366 DEBUG(dbgs() << "LV: Found a non-intrinsic, non-libfunc callsite.\n");
5370 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
5371 // second argument is the same (i.e. loop invariant)
5372 if (CI && hasVectorInstrinsicScalarOpd(
5373 getVectorIntrinsicIDForCall(CI, TLI), 1)) {
5374 auto *SE = PSE.getSE();
5375 if (!SE->isLoopInvariant(PSE.getSCEV(CI->getOperand(1)), TheLoop)) {
5376 ORE->emit(createMissedAnalysis("CantVectorizeIntrinsic", CI)
5377 << "intrinsic instruction cannot be vectorized");
5378 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
5383 // Check that the instruction return type is vectorizable.
5384 // Also, we can't vectorize extractelement instructions.
5385 if ((!VectorType::isValidElementType(I.getType()) &&
5386 !I.getType()->isVoidTy()) ||
5387 isa<ExtractElementInst>(I)) {
5388 ORE->emit(createMissedAnalysis("CantVectorizeInstructionReturnType", &I)
5389 << "instruction return type cannot be vectorized");
5390 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
5394 // Check that the stored type is vectorizable.
5395 if (auto *ST = dyn_cast<StoreInst>(&I)) {
5396 Type *T = ST->getValueOperand()->getType();
5397 if (!VectorType::isValidElementType(T)) {
5398 ORE->emit(createMissedAnalysis("CantVectorizeStore", ST)
5399 << "store instruction cannot be vectorized");
5403 // FP instructions can allow unsafe algebra, thus vectorizable by
5404 // non-IEEE-754 compliant SIMD units.
5405 // This applies to floating-point math operations and calls, not memory
5406 // operations, shuffles, or casts, as they don't change precision or
5408 } else if (I.getType()->isFloatingPointTy() && (CI || I.isBinaryOp()) &&
5409 !I.hasUnsafeAlgebra()) {
5410 DEBUG(dbgs() << "LV: Found FP op with unsafe algebra.\n");
5411 Hints->setPotentiallyUnsafe();
5414 // Reduction instructions are allowed to have exit users.
5415 // All other instructions must not have external users.
5416 if (hasOutsideLoopUser(TheLoop, &I, AllowedExit)) {
5417 ORE->emit(createMissedAnalysis("ValueUsedOutsideLoop", &I)
5418 << "value cannot be used outside the loop");
5425 if (!PrimaryInduction) {
5426 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
5427 if (Inductions.empty()) {
5428 ORE->emit(createMissedAnalysis("NoInductionVariable")
5429 << "loop induction variable could not be identified");
5434 // Now we know the widest induction type, check if our found induction
5435 // is the same size. If it's not, unset it here and InnerLoopVectorizer
5436 // will create another.
5437 if (PrimaryInduction && WidestIndTy != PrimaryInduction->getType())
5438 PrimaryInduction = nullptr;
5443 void LoopVectorizationCostModel::collectLoopScalars(unsigned VF) {
5445 // We should not collect Scalars more than once per VF. Right now, this
5446 // function is called from collectUniformsAndScalars(), which already does
5447 // this check. Collecting Scalars for VF=1 does not make any sense.
5448 assert(VF >= 2 && !Scalars.count(VF) &&
5449 "This function should not be visited twice for the same VF");
5451 SmallSetVector<Instruction *, 8> Worklist;
5453 // These sets are used to seed the analysis with pointers used by memory
5454 // accesses that will remain scalar.
5455 SmallSetVector<Instruction *, 8> ScalarPtrs;
5456 SmallPtrSet<Instruction *, 8> PossibleNonScalarPtrs;
5458 // A helper that returns true if the use of Ptr by MemAccess will be scalar.
5459 // The pointer operands of loads and stores will be scalar as long as the
5460 // memory access is not a gather or scatter operation. The value operand of a
5461 // store will remain scalar if the store is scalarized.
5462 auto isScalarUse = [&](Instruction *MemAccess, Value *Ptr) {
5463 InstWidening WideningDecision = getWideningDecision(MemAccess, VF);
5464 assert(WideningDecision != CM_Unknown &&
5465 "Widening decision should be ready at this moment");
5466 if (auto *Store = dyn_cast<StoreInst>(MemAccess))
5467 if (Ptr == Store->getValueOperand())
5468 return WideningDecision == CM_Scalarize;
5469 assert(Ptr == getPointerOperand(MemAccess) &&
5470 "Ptr is neither a value or pointer operand");
5471 return WideningDecision != CM_GatherScatter;
5474 // A helper that returns true if the given value is a bitcast or
5475 // getelementptr instruction contained in the loop.
5476 auto isLoopVaryingBitCastOrGEP = [&](Value *V) {
5477 return ((isa<BitCastInst>(V) && V->getType()->isPointerTy()) ||
5478 isa<GetElementPtrInst>(V)) &&
5479 !TheLoop->isLoopInvariant(V);
5482 // A helper that evaluates a memory access's use of a pointer. If the use
5483 // will be a scalar use, and the pointer is only used by memory accesses, we
5484 // place the pointer in ScalarPtrs. Otherwise, the pointer is placed in
5485 // PossibleNonScalarPtrs.
5486 auto evaluatePtrUse = [&](Instruction *MemAccess, Value *Ptr) {
5488 // We only care about bitcast and getelementptr instructions contained in
5490 if (!isLoopVaryingBitCastOrGEP(Ptr))
5493 // If the pointer has already been identified as scalar (e.g., if it was
5494 // also identified as uniform), there's nothing to do.
5495 auto *I = cast<Instruction>(Ptr);
5496 if (Worklist.count(I))
5499 // If the use of the pointer will be a scalar use, and all users of the
5500 // pointer are memory accesses, place the pointer in ScalarPtrs. Otherwise,
5501 // place the pointer in PossibleNonScalarPtrs.
5502 if (isScalarUse(MemAccess, Ptr) && all_of(I->users(), [&](User *U) {
5503 return isa<LoadInst>(U) || isa<StoreInst>(U);
5505 ScalarPtrs.insert(I);
5507 PossibleNonScalarPtrs.insert(I);
5510 // We seed the scalars analysis with three classes of instructions: (1)
5511 // instructions marked uniform-after-vectorization, (2) bitcast and
5512 // getelementptr instructions used by memory accesses requiring a scalar use,
5513 // and (3) pointer induction variables and their update instructions (we
5514 // currently only scalarize these).
5516 // (1) Add to the worklist all instructions that have been identified as
5517 // uniform-after-vectorization.
5518 Worklist.insert(Uniforms[VF].begin(), Uniforms[VF].end());
5520 // (2) Add to the worklist all bitcast and getelementptr instructions used by
5521 // memory accesses requiring a scalar use. The pointer operands of loads and
5522 // stores will be scalar as long as the memory accesses is not a gather or
5523 // scatter operation. The value operand of a store will remain scalar if the
5524 // store is scalarized.
5525 for (auto *BB : TheLoop->blocks())
5526 for (auto &I : *BB) {
5527 if (auto *Load = dyn_cast<LoadInst>(&I)) {
5528 evaluatePtrUse(Load, Load->getPointerOperand());
5529 } else if (auto *Store = dyn_cast<StoreInst>(&I)) {
5530 evaluatePtrUse(Store, Store->getPointerOperand());
5531 evaluatePtrUse(Store, Store->getValueOperand());
5534 for (auto *I : ScalarPtrs)
5535 if (!PossibleNonScalarPtrs.count(I)) {
5536 DEBUG(dbgs() << "LV: Found scalar instruction: " << *I << "\n");
5540 // (3) Add to the worklist all pointer induction variables and their update
5543 // TODO: Once we are able to vectorize pointer induction variables we should
5544 // no longer insert them into the worklist here.
5545 auto *Latch = TheLoop->getLoopLatch();
5546 for (auto &Induction : *Legal->getInductionVars()) {
5547 auto *Ind = Induction.first;
5548 auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch));
5549 if (Induction.second.getKind() != InductionDescriptor::IK_PtrInduction)
5551 Worklist.insert(Ind);
5552 Worklist.insert(IndUpdate);
5553 DEBUG(dbgs() << "LV: Found scalar instruction: " << *Ind << "\n");
5554 DEBUG(dbgs() << "LV: Found scalar instruction: " << *IndUpdate << "\n");
5557 // Expand the worklist by looking through any bitcasts and getelementptr
5558 // instructions we've already identified as scalar. This is similar to the
5559 // expansion step in collectLoopUniforms(); however, here we're only
5560 // expanding to include additional bitcasts and getelementptr instructions.
5562 while (Idx != Worklist.size()) {
5563 Instruction *Dst = Worklist[Idx++];
5564 if (!isLoopVaryingBitCastOrGEP(Dst->getOperand(0)))
5566 auto *Src = cast<Instruction>(Dst->getOperand(0));
5567 if (all_of(Src->users(), [&](User *U) -> bool {
5568 auto *J = cast<Instruction>(U);
5569 return !TheLoop->contains(J) || Worklist.count(J) ||
5570 ((isa<LoadInst>(J) || isa<StoreInst>(J)) &&
5571 isScalarUse(J, Src));
5573 Worklist.insert(Src);
5574 DEBUG(dbgs() << "LV: Found scalar instruction: " << *Src << "\n");
5578 // An induction variable will remain scalar if all users of the induction
5579 // variable and induction variable update remain scalar.
5580 for (auto &Induction : *Legal->getInductionVars()) {
5581 auto *Ind = Induction.first;
5582 auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch));
5584 // We already considered pointer induction variables, so there's no reason
5585 // to look at their users again.
5587 // TODO: Once we are able to vectorize pointer induction variables we
5588 // should no longer skip over them here.
5589 if (Induction.second.getKind() == InductionDescriptor::IK_PtrInduction)
5592 // Determine if all users of the induction variable are scalar after
5594 auto ScalarInd = all_of(Ind->users(), [&](User *U) -> bool {
5595 auto *I = cast<Instruction>(U);
5596 return I == IndUpdate || !TheLoop->contains(I) || Worklist.count(I);
5601 // Determine if all users of the induction variable update instruction are
5602 // scalar after vectorization.
5603 auto ScalarIndUpdate = all_of(IndUpdate->users(), [&](User *U) -> bool {
5604 auto *I = cast<Instruction>(U);
5605 return I == Ind || !TheLoop->contains(I) || Worklist.count(I);
5607 if (!ScalarIndUpdate)
5610 // The induction variable and its update instruction will remain scalar.
5611 Worklist.insert(Ind);
5612 Worklist.insert(IndUpdate);
5613 DEBUG(dbgs() << "LV: Found scalar instruction: " << *Ind << "\n");
5614 DEBUG(dbgs() << "LV: Found scalar instruction: " << *IndUpdate << "\n");
5617 Scalars[VF].insert(Worklist.begin(), Worklist.end());
5620 bool LoopVectorizationLegality::isScalarWithPredication(Instruction *I) {
5621 if (!blockNeedsPredication(I->getParent()))
5623 switch(I->getOpcode()) {
5626 case Instruction::Store:
5627 return !isMaskRequired(I);
5628 case Instruction::UDiv:
5629 case Instruction::SDiv:
5630 case Instruction::SRem:
5631 case Instruction::URem:
5632 return mayDivideByZero(*I);
5637 bool LoopVectorizationLegality::memoryInstructionCanBeWidened(Instruction *I,
5639 // Get and ensure we have a valid memory instruction.
5640 LoadInst *LI = dyn_cast<LoadInst>(I);
5641 StoreInst *SI = dyn_cast<StoreInst>(I);
5642 assert((LI || SI) && "Invalid memory instruction");
5644 auto *Ptr = getPointerOperand(I);
5646 // In order to be widened, the pointer should be consecutive, first of all.
5647 if (!isConsecutivePtr(Ptr))
5650 // If the instruction is a store located in a predicated block, it will be
5652 if (isScalarWithPredication(I))
5655 // If the instruction's allocated size doesn't equal it's type size, it
5656 // requires padding and will be scalarized.
5657 auto &DL = I->getModule()->getDataLayout();
5658 auto *ScalarTy = LI ? LI->getType() : SI->getValueOperand()->getType();
5659 if (hasIrregularType(ScalarTy, DL, VF))
5665 void LoopVectorizationCostModel::collectLoopUniforms(unsigned VF) {
5667 // We should not collect Uniforms more than once per VF. Right now,
5668 // this function is called from collectUniformsAndScalars(), which
5669 // already does this check. Collecting Uniforms for VF=1 does not make any
5672 assert(VF >= 2 && !Uniforms.count(VF) &&
5673 "This function should not be visited twice for the same VF");
5675 // Visit the list of Uniforms. If we'll not find any uniform value, we'll
5676 // not analyze again. Uniforms.count(VF) will return 1.
5677 Uniforms[VF].clear();
5679 // We now know that the loop is vectorizable!
5680 // Collect instructions inside the loop that will remain uniform after
5683 // Global values, params and instructions outside of current loop are out of
5685 auto isOutOfScope = [&](Value *V) -> bool {
5686 Instruction *I = dyn_cast<Instruction>(V);
5687 return (!I || !TheLoop->contains(I));
5690 SetVector<Instruction *> Worklist;
5691 BasicBlock *Latch = TheLoop->getLoopLatch();
5693 // Start with the conditional branch. If the branch condition is an
5694 // instruction contained in the loop that is only used by the branch, it is
5696 auto *Cmp = dyn_cast<Instruction>(Latch->getTerminator()->getOperand(0));
5697 if (Cmp && TheLoop->contains(Cmp) && Cmp->hasOneUse()) {
5698 Worklist.insert(Cmp);
5699 DEBUG(dbgs() << "LV: Found uniform instruction: " << *Cmp << "\n");
5702 // Holds consecutive and consecutive-like pointers. Consecutive-like pointers
5703 // are pointers that are treated like consecutive pointers during
5704 // vectorization. The pointer operands of interleaved accesses are an
5706 SmallSetVector<Instruction *, 8> ConsecutiveLikePtrs;
5708 // Holds pointer operands of instructions that are possibly non-uniform.
5709 SmallPtrSet<Instruction *, 8> PossibleNonUniformPtrs;
5711 auto isUniformDecision = [&](Instruction *I, unsigned VF) {
5712 InstWidening WideningDecision = getWideningDecision(I, VF);
5713 assert(WideningDecision != CM_Unknown &&
5714 "Widening decision should be ready at this moment");
5716 return (WideningDecision == CM_Widen ||
5717 WideningDecision == CM_Interleave);
5719 // Iterate over the instructions in the loop, and collect all
5720 // consecutive-like pointer operands in ConsecutiveLikePtrs. If it's possible
5721 // that a consecutive-like pointer operand will be scalarized, we collect it
5722 // in PossibleNonUniformPtrs instead. We use two sets here because a single
5723 // getelementptr instruction can be used by both vectorized and scalarized
5724 // memory instructions. For example, if a loop loads and stores from the same
5725 // location, but the store is conditional, the store will be scalarized, and
5726 // the getelementptr won't remain uniform.
5727 for (auto *BB : TheLoop->blocks())
5728 for (auto &I : *BB) {
5730 // If there's no pointer operand, there's nothing to do.
5731 auto *Ptr = dyn_cast_or_null<Instruction>(getPointerOperand(&I));
5735 // True if all users of Ptr are memory accesses that have Ptr as their
5737 auto UsersAreMemAccesses = all_of(Ptr->users(), [&](User *U) -> bool {
5738 return getPointerOperand(U) == Ptr;
5741 // Ensure the memory instruction will not be scalarized or used by
5742 // gather/scatter, making its pointer operand non-uniform. If the pointer
5743 // operand is used by any instruction other than a memory access, we
5744 // conservatively assume the pointer operand may be non-uniform.
5745 if (!UsersAreMemAccesses || !isUniformDecision(&I, VF))
5746 PossibleNonUniformPtrs.insert(Ptr);
5748 // If the memory instruction will be vectorized and its pointer operand
5749 // is consecutive-like, or interleaving - the pointer operand should
5752 ConsecutiveLikePtrs.insert(Ptr);
5755 // Add to the Worklist all consecutive and consecutive-like pointers that
5756 // aren't also identified as possibly non-uniform.
5757 for (auto *V : ConsecutiveLikePtrs)
5758 if (!PossibleNonUniformPtrs.count(V)) {
5759 DEBUG(dbgs() << "LV: Found uniform instruction: " << *V << "\n");
5763 // Expand Worklist in topological order: whenever a new instruction
5764 // is added , its users should be either already inside Worklist, or
5765 // out of scope. It ensures a uniform instruction will only be used
5766 // by uniform instructions or out of scope instructions.
5768 while (idx != Worklist.size()) {
5769 Instruction *I = Worklist[idx++];
5771 for (auto OV : I->operand_values()) {
5772 if (isOutOfScope(OV))
5774 auto *OI = cast<Instruction>(OV);
5775 if (all_of(OI->users(), [&](User *U) -> bool {
5776 auto *J = cast<Instruction>(U);
5777 return !TheLoop->contains(J) || Worklist.count(J) ||
5778 (OI == getPointerOperand(J) && isUniformDecision(J, VF));
5780 Worklist.insert(OI);
5781 DEBUG(dbgs() << "LV: Found uniform instruction: " << *OI << "\n");
5786 // Returns true if Ptr is the pointer operand of a memory access instruction
5787 // I, and I is known to not require scalarization.
5788 auto isVectorizedMemAccessUse = [&](Instruction *I, Value *Ptr) -> bool {
5789 return getPointerOperand(I) == Ptr && isUniformDecision(I, VF);
5792 // For an instruction to be added into Worklist above, all its users inside
5793 // the loop should also be in Worklist. However, this condition cannot be
5794 // true for phi nodes that form a cyclic dependence. We must process phi
5795 // nodes separately. An induction variable will remain uniform if all users
5796 // of the induction variable and induction variable update remain uniform.
5797 // The code below handles both pointer and non-pointer induction variables.
5798 for (auto &Induction : *Legal->getInductionVars()) {
5799 auto *Ind = Induction.first;
5800 auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch));
5802 // Determine if all users of the induction variable are uniform after
5804 auto UniformInd = all_of(Ind->users(), [&](User *U) -> bool {
5805 auto *I = cast<Instruction>(U);
5806 return I == IndUpdate || !TheLoop->contains(I) || Worklist.count(I) ||
5807 isVectorizedMemAccessUse(I, Ind);
5812 // Determine if all users of the induction variable update instruction are
5813 // uniform after vectorization.
5814 auto UniformIndUpdate = all_of(IndUpdate->users(), [&](User *U) -> bool {
5815 auto *I = cast<Instruction>(U);
5816 return I == Ind || !TheLoop->contains(I) || Worklist.count(I) ||
5817 isVectorizedMemAccessUse(I, IndUpdate);
5819 if (!UniformIndUpdate)
5822 // The induction variable and its update instruction will remain uniform.
5823 Worklist.insert(Ind);
5824 Worklist.insert(IndUpdate);
5825 DEBUG(dbgs() << "LV: Found uniform instruction: " << *Ind << "\n");
5826 DEBUG(dbgs() << "LV: Found uniform instruction: " << *IndUpdate << "\n");
5829 Uniforms[VF].insert(Worklist.begin(), Worklist.end());
5832 bool LoopVectorizationLegality::canVectorizeMemory() {
5833 LAI = &(*GetLAA)(*TheLoop);
5834 InterleaveInfo.setLAI(LAI);
5835 const OptimizationRemarkAnalysis *LAR = LAI->getReport();
5837 OptimizationRemarkAnalysis VR(Hints->vectorizeAnalysisPassName(),
5838 "loop not vectorized: ", *LAR);
5841 if (!LAI->canVectorizeMemory())
5844 if (LAI->hasStoreToLoopInvariantAddress()) {
5845 ORE->emit(createMissedAnalysis("CantVectorizeStoreToLoopInvariantAddress")
5846 << "write to a loop invariant address could not be vectorized");
5847 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
5851 Requirements->addRuntimePointerChecks(LAI->getNumRuntimePointerChecks());
5852 PSE.addPredicate(LAI->getPSE().getUnionPredicate());
5857 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
5858 Value *In0 = const_cast<Value *>(V);
5859 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
5863 return Inductions.count(PN);
5866 bool LoopVectorizationLegality::isFirstOrderRecurrence(const PHINode *Phi) {
5867 return FirstOrderRecurrences.count(Phi);
5870 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
5871 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
5874 bool LoopVectorizationLegality::blockCanBePredicated(
5875 BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs) {
5876 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
5878 for (Instruction &I : *BB) {
5879 // Check that we don't have a constant expression that can trap as operand.
5880 for (Value *Operand : I.operands()) {
5881 if (auto *C = dyn_cast<Constant>(Operand))
5885 // We might be able to hoist the load.
5886 if (I.mayReadFromMemory()) {
5887 auto *LI = dyn_cast<LoadInst>(&I);
5890 if (!SafePtrs.count(LI->getPointerOperand())) {
5891 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand()) ||
5892 isLegalMaskedGather(LI->getType())) {
5893 MaskedOp.insert(LI);
5896 // !llvm.mem.parallel_loop_access implies if-conversion safety.
5897 if (IsAnnotatedParallel)
5903 if (I.mayWriteToMemory()) {
5904 auto *SI = dyn_cast<StoreInst>(&I);
5905 // We only support predication of stores in basic blocks with one
5910 // Build a masked store if it is legal for the target.
5911 if (isLegalMaskedStore(SI->getValueOperand()->getType(),
5912 SI->getPointerOperand()) ||
5913 isLegalMaskedScatter(SI->getValueOperand()->getType())) {
5914 MaskedOp.insert(SI);
5918 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
5919 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
5921 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
5922 !isSinglePredecessor)
5932 void InterleavedAccessInfo::collectConstStrideAccesses(
5933 MapVector<Instruction *, StrideDescriptor> &AccessStrideInfo,
5934 const ValueToValueMap &Strides) {
5936 auto &DL = TheLoop->getHeader()->getModule()->getDataLayout();
5938 // Since it's desired that the load/store instructions be maintained in
5939 // "program order" for the interleaved access analysis, we have to visit the
5940 // blocks in the loop in reverse postorder (i.e., in a topological order).
5941 // Such an ordering will ensure that any load/store that may be executed
5942 // before a second load/store will precede the second load/store in
5943 // AccessStrideInfo.
5944 LoopBlocksDFS DFS(TheLoop);
5946 for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO()))
5947 for (auto &I : *BB) {
5948 auto *LI = dyn_cast<LoadInst>(&I);
5949 auto *SI = dyn_cast<StoreInst>(&I);
5953 Value *Ptr = getPointerOperand(&I);
5954 // We don't check wrapping here because we don't know yet if Ptr will be
5955 // part of a full group or a group with gaps. Checking wrapping for all
5956 // pointers (even those that end up in groups with no gaps) will be overly
5957 // conservative. For full groups, wrapping should be ok since if we would
5958 // wrap around the address space we would do a memory access at nullptr
5959 // even without the transformation. The wrapping checks are therefore
5960 // deferred until after we've formed the interleaved groups.
5961 int64_t Stride = getPtrStride(PSE, Ptr, TheLoop, Strides,
5962 /*Assume=*/true, /*ShouldCheckWrap=*/false);
5964 const SCEV *Scev = replaceSymbolicStrideSCEV(PSE, Strides, Ptr);
5965 PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
5966 uint64_t Size = DL.getTypeAllocSize(PtrTy->getElementType());
5968 // An alignment of 0 means target ABI alignment.
5969 unsigned Align = getMemInstAlignment(&I);
5971 Align = DL.getABITypeAlignment(PtrTy->getElementType());
5973 AccessStrideInfo[&I] = StrideDescriptor(Stride, Scev, Size, Align);
5977 // Analyze interleaved accesses and collect them into interleaved load and
5980 // When generating code for an interleaved load group, we effectively hoist all
5981 // loads in the group to the location of the first load in program order. When
5982 // generating code for an interleaved store group, we sink all stores to the
5983 // location of the last store. This code motion can change the order of load
5984 // and store instructions and may break dependences.
5986 // The code generation strategy mentioned above ensures that we won't violate
5987 // any write-after-read (WAR) dependences.
5989 // E.g., for the WAR dependence: a = A[i]; // (1)
5992 // The store group of (2) is always inserted at or below (2), and the load
5993 // group of (1) is always inserted at or above (1). Thus, the instructions will
5994 // never be reordered. All other dependences are checked to ensure the
5995 // correctness of the instruction reordering.
5997 // The algorithm visits all memory accesses in the loop in bottom-up program
5998 // order. Program order is established by traversing the blocks in the loop in
5999 // reverse postorder when collecting the accesses.
6001 // We visit the memory accesses in bottom-up order because it can simplify the
6002 // construction of store groups in the presence of write-after-write (WAW)
6005 // E.g., for the WAW dependence: A[i] = a; // (1)
6007 // A[i + 1] = c; // (3)
6009 // We will first create a store group with (3) and (2). (1) can't be added to
6010 // this group because it and (2) are dependent. However, (1) can be grouped
6011 // with other accesses that may precede it in program order. Note that a
6012 // bottom-up order does not imply that WAW dependences should not be checked.
6013 void InterleavedAccessInfo::analyzeInterleaving(
6014 const ValueToValueMap &Strides) {
6015 DEBUG(dbgs() << "LV: Analyzing interleaved accesses...\n");
6017 // Holds all accesses with a constant stride.
6018 MapVector<Instruction *, StrideDescriptor> AccessStrideInfo;
6019 collectConstStrideAccesses(AccessStrideInfo, Strides);
6021 if (AccessStrideInfo.empty())
6024 // Collect the dependences in the loop.
6025 collectDependences();
6027 // Holds all interleaved store groups temporarily.
6028 SmallSetVector<InterleaveGroup *, 4> StoreGroups;
6029 // Holds all interleaved load groups temporarily.
6030 SmallSetVector<InterleaveGroup *, 4> LoadGroups;
6032 // Search in bottom-up program order for pairs of accesses (A and B) that can
6033 // form interleaved load or store groups. In the algorithm below, access A
6034 // precedes access B in program order. We initialize a group for B in the
6035 // outer loop of the algorithm, and then in the inner loop, we attempt to
6036 // insert each A into B's group if:
6038 // 1. A and B have the same stride,
6039 // 2. A and B have the same memory object size, and
6040 // 3. A belongs in B's group according to its distance from B.
6042 // Special care is taken to ensure group formation will not break any
6044 for (auto BI = AccessStrideInfo.rbegin(), E = AccessStrideInfo.rend();
6046 Instruction *B = BI->first;
6047 StrideDescriptor DesB = BI->second;
6049 // Initialize a group for B if it has an allowable stride. Even if we don't
6050 // create a group for B, we continue with the bottom-up algorithm to ensure
6051 // we don't break any of B's dependences.
6052 InterleaveGroup *Group = nullptr;
6053 if (isStrided(DesB.Stride)) {
6054 Group = getInterleaveGroup(B);
6056 DEBUG(dbgs() << "LV: Creating an interleave group with:" << *B << '\n');
6057 Group = createInterleaveGroup(B, DesB.Stride, DesB.Align);
6059 if (B->mayWriteToMemory())
6060 StoreGroups.insert(Group);
6062 LoadGroups.insert(Group);
6065 for (auto AI = std::next(BI); AI != E; ++AI) {
6066 Instruction *A = AI->first;
6067 StrideDescriptor DesA = AI->second;
6069 // Our code motion strategy implies that we can't have dependences
6070 // between accesses in an interleaved group and other accesses located
6071 // between the first and last member of the group. Note that this also
6072 // means that a group can't have more than one member at a given offset.
6073 // The accesses in a group can have dependences with other accesses, but
6074 // we must ensure we don't extend the boundaries of the group such that
6075 // we encompass those dependent accesses.
6077 // For example, assume we have the sequence of accesses shown below in a
6080 // (1, 2) is a group | A[i] = a; // (1)
6081 // | A[i-1] = b; // (2) |
6082 // A[i-3] = c; // (3)
6083 // A[i] = d; // (4) | (2, 4) is not a group
6085 // Because accesses (2) and (3) are dependent, we can group (2) with (1)
6086 // but not with (4). If we did, the dependent access (3) would be within
6087 // the boundaries of the (2, 4) group.
6088 if (!canReorderMemAccessesForInterleavedGroups(&*AI, &*BI)) {
6090 // If a dependence exists and A is already in a group, we know that A
6091 // must be a store since A precedes B and WAR dependences are allowed.
6092 // Thus, A would be sunk below B. We release A's group to prevent this
6093 // illegal code motion. A will then be free to form another group with
6094 // instructions that precede it.
6095 if (isInterleaved(A)) {
6096 InterleaveGroup *StoreGroup = getInterleaveGroup(A);
6097 StoreGroups.remove(StoreGroup);
6098 releaseGroup(StoreGroup);
6101 // If a dependence exists and A is not already in a group (or it was
6102 // and we just released it), B might be hoisted above A (if B is a
6103 // load) or another store might be sunk below A (if B is a store). In
6104 // either case, we can't add additional instructions to B's group. B
6105 // will only form a group with instructions that it precedes.
6109 // At this point, we've checked for illegal code motion. If either A or B
6110 // isn't strided, there's nothing left to do.
6111 if (!isStrided(DesA.Stride) || !isStrided(DesB.Stride))
6114 // Ignore A if it's already in a group or isn't the same kind of memory
6116 if (isInterleaved(A) || A->mayReadFromMemory() != B->mayReadFromMemory())
6119 // Check rules 1 and 2. Ignore A if its stride or size is different from
6121 if (DesA.Stride != DesB.Stride || DesA.Size != DesB.Size)
6124 // Ignore A if the memory object of A and B don't belong to the same
6126 if (getMemInstAddressSpace(A) != getMemInstAddressSpace(B))
6129 // Calculate the distance from A to B.
6130 const SCEVConstant *DistToB = dyn_cast<SCEVConstant>(
6131 PSE.getSE()->getMinusSCEV(DesA.Scev, DesB.Scev));
6134 int64_t DistanceToB = DistToB->getAPInt().getSExtValue();
6136 // Check rule 3. Ignore A if its distance to B is not a multiple of the
6138 if (DistanceToB % static_cast<int64_t>(DesB.Size))
6141 // Ignore A if either A or B is in a predicated block. Although we
6142 // currently prevent group formation for predicated accesses, we may be
6143 // able to relax this limitation in the future once we handle more
6144 // complicated blocks.
6145 if (isPredicated(A->getParent()) || isPredicated(B->getParent()))
6148 // The index of A is the index of B plus A's distance to B in multiples
6151 Group->getIndex(B) + DistanceToB / static_cast<int64_t>(DesB.Size);
6153 // Try to insert A into B's group.
6154 if (Group->insertMember(A, IndexA, DesA.Align)) {
6155 DEBUG(dbgs() << "LV: Inserted:" << *A << '\n'
6156 << " into the interleave group with" << *B << '\n');
6157 InterleaveGroupMap[A] = Group;
6159 // Set the first load in program order as the insert position.
6160 if (A->mayReadFromMemory())
6161 Group->setInsertPos(A);
6163 } // Iteration over A accesses.
6164 } // Iteration over B accesses.
6166 // Remove interleaved store groups with gaps.
6167 for (InterleaveGroup *Group : StoreGroups)
6168 if (Group->getNumMembers() != Group->getFactor())
6169 releaseGroup(Group);
6171 // Remove interleaved groups with gaps (currently only loads) whose memory
6172 // accesses may wrap around. We have to revisit the getPtrStride analysis,
6173 // this time with ShouldCheckWrap=true, since collectConstStrideAccesses does
6174 // not check wrapping (see documentation there).
6175 // FORNOW we use Assume=false;
6176 // TODO: Change to Assume=true but making sure we don't exceed the threshold
6177 // of runtime SCEV assumptions checks (thereby potentially failing to
6178 // vectorize altogether).
6179 // Additional optional optimizations:
6180 // TODO: If we are peeling the loop and we know that the first pointer doesn't
6181 // wrap then we can deduce that all pointers in the group don't wrap.
6182 // This means that we can forcefully peel the loop in order to only have to
6183 // check the first pointer for no-wrap. When we'll change to use Assume=true
6184 // we'll only need at most one runtime check per interleaved group.
6186 for (InterleaveGroup *Group : LoadGroups) {
6188 // Case 1: A full group. Can Skip the checks; For full groups, if the wide
6189 // load would wrap around the address space we would do a memory access at
6190 // nullptr even without the transformation.
6191 if (Group->getNumMembers() == Group->getFactor())
6194 // Case 2: If first and last members of the group don't wrap this implies
6195 // that all the pointers in the group don't wrap.
6196 // So we check only group member 0 (which is always guaranteed to exist),
6197 // and group member Factor - 1; If the latter doesn't exist we rely on
6198 // peeling (if it is a non-reveresed accsess -- see Case 3).
6199 Value *FirstMemberPtr = getPointerOperand(Group->getMember(0));
6200 if (!getPtrStride(PSE, FirstMemberPtr, TheLoop, Strides, /*Assume=*/false,
6201 /*ShouldCheckWrap=*/true)) {
6202 DEBUG(dbgs() << "LV: Invalidate candidate interleaved group due to "
6203 "first group member potentially pointer-wrapping.\n");
6204 releaseGroup(Group);
6207 Instruction *LastMember = Group->getMember(Group->getFactor() - 1);
6209 Value *LastMemberPtr = getPointerOperand(LastMember);
6210 if (!getPtrStride(PSE, LastMemberPtr, TheLoop, Strides, /*Assume=*/false,
6211 /*ShouldCheckWrap=*/true)) {
6212 DEBUG(dbgs() << "LV: Invalidate candidate interleaved group due to "
6213 "last group member potentially pointer-wrapping.\n");
6214 releaseGroup(Group);
6217 // Case 3: A non-reversed interleaved load group with gaps: We need
6218 // to execute at least one scalar epilogue iteration. This will ensure
6219 // we don't speculatively access memory out-of-bounds. We only need
6220 // to look for a member at index factor - 1, since every group must have
6221 // a member at index zero.
6222 if (Group->isReverse()) {
6223 releaseGroup(Group);
6226 DEBUG(dbgs() << "LV: Interleaved group requires epilogue iteration.\n");
6227 RequiresScalarEpilogue = true;
6232 Optional<unsigned> LoopVectorizationCostModel::computeMaxVF(bool OptForSize) {
6233 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
6234 ORE->emit(createMissedAnalysis("ConditionalStore")
6235 << "store that is conditionally executed prevents vectorization");
6236 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
6240 if (!OptForSize) // Remaining checks deal with scalar loop when OptForSize.
6241 return computeFeasibleMaxVF(OptForSize);
6243 if (Legal->getRuntimePointerChecking()->Need) {
6244 ORE->emit(createMissedAnalysis("CantVersionLoopWithOptForSize")
6245 << "runtime pointer checks needed. Enable vectorization of this "
6246 "loop with '#pragma clang loop vectorize(enable)' when "
6247 "compiling with -Os/-Oz");
6249 << "LV: Aborting. Runtime ptr check is required with -Os/-Oz.\n");
6253 // If we optimize the program for size, avoid creating the tail loop.
6254 unsigned TC = PSE.getSE()->getSmallConstantTripCount(TheLoop);
6255 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
6257 // If we don't know the precise trip count, don't try to vectorize.
6260 createMissedAnalysis("UnknownLoopCountComplexCFG")
6261 << "unable to calculate the loop count due to complex control flow");
6262 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n");
6266 unsigned MaxVF = computeFeasibleMaxVF(OptForSize);
6268 if (TC % MaxVF != 0) {
6269 // If the trip count that we found modulo the vectorization factor is not
6270 // zero then we require a tail.
6271 // FIXME: look for a smaller MaxVF that does divide TC rather than give up.
6272 // FIXME: return None if loop requiresScalarEpilog(<MaxVF>), or look for a
6273 // smaller MaxVF that does not require a scalar epilog.
6275 ORE->emit(createMissedAnalysis("NoTailLoopWithOptForSize")
6276 << "cannot optimize for size and vectorize at the "
6277 "same time. Enable vectorization of this loop "
6278 "with '#pragma clang loop vectorize(enable)' "
6279 "when compiling with -Os/-Oz");
6280 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n");
6287 unsigned LoopVectorizationCostModel::computeFeasibleMaxVF(bool OptForSize) {
6288 MinBWs = computeMinimumValueSizes(TheLoop->getBlocks(), *DB, &TTI);
6289 unsigned SmallestType, WidestType;
6290 std::tie(SmallestType, WidestType) = getSmallestAndWidestTypes();
6291 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
6292 unsigned MaxSafeDepDist = -1U;
6294 // Get the maximum safe dependence distance in bits computed by LAA. If the
6295 // loop contains any interleaved accesses, we divide the dependence distance
6296 // by the maximum interleave factor of all interleaved groups. Note that
6297 // although the division ensures correctness, this is a fairly conservative
6298 // computation because the maximum distance computed by LAA may not involve
6299 // any of the interleaved accesses.
6300 if (Legal->getMaxSafeDepDistBytes() != -1U)
6302 Legal->getMaxSafeDepDistBytes() * 8 / Legal->getMaxInterleaveFactor();
6305 ((WidestRegister < MaxSafeDepDist) ? WidestRegister : MaxSafeDepDist);
6306 unsigned MaxVectorSize = WidestRegister / WidestType;
6308 DEBUG(dbgs() << "LV: The Smallest and Widest types: " << SmallestType << " / "
6309 << WidestType << " bits.\n");
6310 DEBUG(dbgs() << "LV: The Widest register is: " << WidestRegister
6313 if (MaxVectorSize == 0) {
6314 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
6318 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
6319 " into one vector!");
6321 unsigned MaxVF = MaxVectorSize;
6322 if (MaximizeBandwidth && !OptForSize) {
6323 // Collect all viable vectorization factors.
6324 SmallVector<unsigned, 8> VFs;
6325 unsigned NewMaxVectorSize = WidestRegister / SmallestType;
6326 for (unsigned VS = MaxVectorSize; VS <= NewMaxVectorSize; VS *= 2)
6329 // For each VF calculate its register usage.
6330 auto RUs = calculateRegisterUsage(VFs);
6332 // Select the largest VF which doesn't require more registers than existing
6334 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(true);
6335 for (int i = RUs.size() - 1; i >= 0; --i) {
6336 if (RUs[i].MaxLocalUsers <= TargetNumRegisters) {
6345 LoopVectorizationCostModel::VectorizationFactor
6346 LoopVectorizationCostModel::selectVectorizationFactor(unsigned MaxVF) {
6347 float Cost = expectedCost(1).first;
6349 const float ScalarCost = Cost;
6352 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
6354 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
6355 // Ignore scalar width, because the user explicitly wants vectorization.
6356 if (ForceVectorization && MaxVF > 1) {
6358 Cost = expectedCost(Width).first / (float)Width;
6361 for (unsigned i = 2; i <= MaxVF; i *= 2) {
6362 // Notice that the vector loop needs to be executed less times, so
6363 // we need to divide the cost of the vector loops by the width of
6364 // the vector elements.
6365 VectorizationCostTy C = expectedCost(i);
6366 float VectorCost = C.first / (float)i;
6367 DEBUG(dbgs() << "LV: Vector loop of width " << i
6368 << " costs: " << (int)VectorCost << ".\n");
6369 if (!C.second && !ForceVectorization) {
6371 dbgs() << "LV: Not considering vector loop of width " << i
6372 << " because it will not generate any vector instructions.\n");
6375 if (VectorCost < Cost) {
6381 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
6382 << "LV: Vectorization seems to be not beneficial, "
6383 << "but was forced by a user.\n");
6384 DEBUG(dbgs() << "LV: Selecting VF: " << Width << ".\n");
6385 VectorizationFactor Factor = {Width, (unsigned)(Width * Cost)};
6389 std::pair<unsigned, unsigned>
6390 LoopVectorizationCostModel::getSmallestAndWidestTypes() {
6391 unsigned MinWidth = -1U;
6392 unsigned MaxWidth = 8;
6393 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
6396 for (BasicBlock *BB : TheLoop->blocks()) {
6397 // For each instruction in the loop.
6398 for (Instruction &I : *BB) {
6399 Type *T = I.getType();
6401 // Skip ignored values.
6402 if (ValuesToIgnore.count(&I))
6405 // Only examine Loads, Stores and PHINodes.
6406 if (!isa<LoadInst>(I) && !isa<StoreInst>(I) && !isa<PHINode>(I))
6409 // Examine PHI nodes that are reduction variables. Update the type to
6410 // account for the recurrence type.
6411 if (auto *PN = dyn_cast<PHINode>(&I)) {
6412 if (!Legal->isReductionVariable(PN))
6414 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[PN];
6415 T = RdxDesc.getRecurrenceType();
6418 // Examine the stored values.
6419 if (auto *ST = dyn_cast<StoreInst>(&I))
6420 T = ST->getValueOperand()->getType();
6422 // Ignore loaded pointer types and stored pointer types that are not
6425 // FIXME: The check here attempts to predict whether a load or store will
6426 // be vectorized. We only know this for certain after a VF has
6427 // been selected. Here, we assume that if an access can be
6428 // vectorized, it will be. We should also look at extending this
6429 // optimization to non-pointer types.
6431 if (T->isPointerTy() && !isConsecutiveLoadOrStore(&I) &&
6432 !Legal->isAccessInterleaved(&I) && !Legal->isLegalGatherOrScatter(&I))
6435 MinWidth = std::min(MinWidth,
6436 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
6437 MaxWidth = std::max(MaxWidth,
6438 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
6442 return {MinWidth, MaxWidth};
6445 unsigned LoopVectorizationCostModel::selectInterleaveCount(bool OptForSize,
6447 unsigned LoopCost) {
6449 // -- The interleave heuristics --
6450 // We interleave the loop in order to expose ILP and reduce the loop overhead.
6451 // There are many micro-architectural considerations that we can't predict
6452 // at this level. For example, frontend pressure (on decode or fetch) due to
6453 // code size, or the number and capabilities of the execution ports.
6455 // We use the following heuristics to select the interleave count:
6456 // 1. If the code has reductions, then we interleave to break the cross
6457 // iteration dependency.
6458 // 2. If the loop is really small, then we interleave to reduce the loop
6460 // 3. We don't interleave if we think that we will spill registers to memory
6461 // due to the increased register pressure.
6463 // When we optimize for size, we don't interleave.
6467 // We used the distance for the interleave count.
6468 if (Legal->getMaxSafeDepDistBytes() != -1U)
6471 // Do not interleave loops with a relatively small trip count.
6472 unsigned TC = PSE.getSE()->getSmallConstantTripCount(TheLoop);
6473 if (TC > 1 && TC < TinyTripCountInterleaveThreshold)
6476 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
6477 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters
6481 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
6482 TargetNumRegisters = ForceTargetNumScalarRegs;
6484 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
6485 TargetNumRegisters = ForceTargetNumVectorRegs;
6488 RegisterUsage R = calculateRegisterUsage({VF})[0];
6489 // We divide by these constants so assume that we have at least one
6490 // instruction that uses at least one register.
6491 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
6492 R.NumInstructions = std::max(R.NumInstructions, 1U);
6494 // We calculate the interleave count using the following formula.
6495 // Subtract the number of loop invariants from the number of available
6496 // registers. These registers are used by all of the interleaved instances.
6497 // Next, divide the remaining registers by the number of registers that is
6498 // required by the loop, in order to estimate how many parallel instances
6499 // fit without causing spills. All of this is rounded down if necessary to be
6500 // a power of two. We want power of two interleave count to simplify any
6501 // addressing operations or alignment considerations.
6502 unsigned IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
6505 // Don't count the induction variable as interleaved.
6506 if (EnableIndVarRegisterHeur)
6507 IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
6508 std::max(1U, (R.MaxLocalUsers - 1)));
6510 // Clamp the interleave ranges to reasonable counts.
6511 unsigned MaxInterleaveCount = TTI.getMaxInterleaveFactor(VF);
6513 // Check if the user has overridden the max.
6515 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
6516 MaxInterleaveCount = ForceTargetMaxScalarInterleaveFactor;
6518 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
6519 MaxInterleaveCount = ForceTargetMaxVectorInterleaveFactor;
6522 // If we did not calculate the cost for VF (because the user selected the VF)
6523 // then we calculate the cost of VF here.
6525 LoopCost = expectedCost(VF).first;
6527 // Clamp the calculated IC to be between the 1 and the max interleave count
6528 // that the target allows.
6529 if (IC > MaxInterleaveCount)
6530 IC = MaxInterleaveCount;
6534 // Interleave if we vectorized this loop and there is a reduction that could
6535 // benefit from interleaving.
6536 if (VF > 1 && Legal->getReductionVars()->size()) {
6537 DEBUG(dbgs() << "LV: Interleaving because of reductions.\n");
6541 // Note that if we've already vectorized the loop we will have done the
6542 // runtime check and so interleaving won't require further checks.
6543 bool InterleavingRequiresRuntimePointerCheck =
6544 (VF == 1 && Legal->getRuntimePointerChecking()->Need);
6546 // We want to interleave small loops in order to reduce the loop overhead and
6547 // potentially expose ILP opportunities.
6548 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
6549 if (!InterleavingRequiresRuntimePointerCheck && LoopCost < SmallLoopCost) {
6550 // We assume that the cost overhead is 1 and we use the cost model
6551 // to estimate the cost of the loop and interleave until the cost of the
6552 // loop overhead is about 5% of the cost of the loop.
6554 std::min(IC, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
6556 // Interleave until store/load ports (estimated by max interleave count) are
6558 unsigned NumStores = Legal->getNumStores();
6559 unsigned NumLoads = Legal->getNumLoads();
6560 unsigned StoresIC = IC / (NumStores ? NumStores : 1);
6561 unsigned LoadsIC = IC / (NumLoads ? NumLoads : 1);
6563 // If we have a scalar reduction (vector reductions are already dealt with
6564 // by this point), we can increase the critical path length if the loop
6565 // we're interleaving is inside another loop. Limit, by default to 2, so the
6566 // critical path only gets increased by one reduction operation.
6567 if (Legal->getReductionVars()->size() && TheLoop->getLoopDepth() > 1) {
6568 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionIC);
6569 SmallIC = std::min(SmallIC, F);
6570 StoresIC = std::min(StoresIC, F);
6571 LoadsIC = std::min(LoadsIC, F);
6574 if (EnableLoadStoreRuntimeInterleave &&
6575 std::max(StoresIC, LoadsIC) > SmallIC) {
6576 DEBUG(dbgs() << "LV: Interleaving to saturate store or load ports.\n");
6577 return std::max(StoresIC, LoadsIC);
6580 DEBUG(dbgs() << "LV: Interleaving to reduce branch cost.\n");
6584 // Interleave if this is a large loop (small loops are already dealt with by
6585 // this point) that could benefit from interleaving.
6586 bool HasReductions = (Legal->getReductionVars()->size() > 0);
6587 if (TTI.enableAggressiveInterleaving(HasReductions)) {
6588 DEBUG(dbgs() << "LV: Interleaving to expose ILP.\n");
6592 DEBUG(dbgs() << "LV: Not Interleaving.\n");
6596 SmallVector<LoopVectorizationCostModel::RegisterUsage, 8>
6597 LoopVectorizationCostModel::calculateRegisterUsage(ArrayRef<unsigned> VFs) {
6598 // This function calculates the register usage by measuring the highest number
6599 // of values that are alive at a single location. Obviously, this is a very
6600 // rough estimation. We scan the loop in a topological order in order and
6601 // assign a number to each instruction. We use RPO to ensure that defs are
6602 // met before their users. We assume that each instruction that has in-loop
6603 // users starts an interval. We record every time that an in-loop value is
6604 // used, so we have a list of the first and last occurrences of each
6605 // instruction. Next, we transpose this data structure into a multi map that
6606 // holds the list of intervals that *end* at a specific location. This multi
6607 // map allows us to perform a linear search. We scan the instructions linearly
6608 // and record each time that a new interval starts, by placing it in a set.
6609 // If we find this value in the multi-map then we remove it from the set.
6610 // The max register usage is the maximum size of the set.
6611 // We also search for instructions that are defined outside the loop, but are
6612 // used inside the loop. We need this number separately from the max-interval
6613 // usage number because when we unroll, loop-invariant values do not take
6615 LoopBlocksDFS DFS(TheLoop);
6619 RU.NumInstructions = 0;
6621 // Each 'key' in the map opens a new interval. The values
6622 // of the map are the index of the 'last seen' usage of the
6623 // instruction that is the key.
6624 typedef DenseMap<Instruction *, unsigned> IntervalMap;
6625 // Maps instruction to its index.
6626 DenseMap<unsigned, Instruction *> IdxToInstr;
6627 // Marks the end of each interval.
6628 IntervalMap EndPoint;
6629 // Saves the list of instruction indices that are used in the loop.
6630 SmallSet<Instruction *, 8> Ends;
6631 // Saves the list of values that are used in the loop but are
6632 // defined outside the loop, such as arguments and constants.
6633 SmallPtrSet<Value *, 8> LoopInvariants;
6636 for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO())) {
6637 RU.NumInstructions += BB->size();
6638 for (Instruction &I : *BB) {
6639 IdxToInstr[Index++] = &I;
6641 // Save the end location of each USE.
6642 for (Value *U : I.operands()) {
6643 auto *Instr = dyn_cast<Instruction>(U);
6645 // Ignore non-instruction values such as arguments, constants, etc.
6649 // If this instruction is outside the loop then record it and continue.
6650 if (!TheLoop->contains(Instr)) {
6651 LoopInvariants.insert(Instr);
6655 // Overwrite previous end points.
6656 EndPoint[Instr] = Index;
6662 // Saves the list of intervals that end with the index in 'key'.
6663 typedef SmallVector<Instruction *, 2> InstrList;
6664 DenseMap<unsigned, InstrList> TransposeEnds;
6666 // Transpose the EndPoints to a list of values that end at each index.
6667 for (auto &Interval : EndPoint)
6668 TransposeEnds[Interval.second].push_back(Interval.first);
6670 SmallSet<Instruction *, 8> OpenIntervals;
6672 // Get the size of the widest register.
6673 unsigned MaxSafeDepDist = -1U;
6674 if (Legal->getMaxSafeDepDistBytes() != -1U)
6675 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
6676 unsigned WidestRegister =
6677 std::min(TTI.getRegisterBitWidth(true), MaxSafeDepDist);
6678 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
6680 SmallVector<RegisterUsage, 8> RUs(VFs.size());
6681 SmallVector<unsigned, 8> MaxUsages(VFs.size(), 0);
6683 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
6685 // A lambda that gets the register usage for the given type and VF.
6686 auto GetRegUsage = [&DL, WidestRegister](Type *Ty, unsigned VF) {
6687 if (Ty->isTokenTy())
6689 unsigned TypeSize = DL.getTypeSizeInBits(Ty->getScalarType());
6690 return std::max<unsigned>(1, VF * TypeSize / WidestRegister);
6693 for (unsigned int i = 0; i < Index; ++i) {
6694 Instruction *I = IdxToInstr[i];
6696 // Remove all of the instructions that end at this location.
6697 InstrList &List = TransposeEnds[i];
6698 for (Instruction *ToRemove : List)
6699 OpenIntervals.erase(ToRemove);
6701 // Ignore instructions that are never used within the loop.
6705 // Skip ignored values.
6706 if (ValuesToIgnore.count(I))
6709 // For each VF find the maximum usage of registers.
6710 for (unsigned j = 0, e = VFs.size(); j < e; ++j) {
6712 MaxUsages[j] = std::max(MaxUsages[j], OpenIntervals.size());
6715 collectUniformsAndScalars(VFs[j]);
6716 // Count the number of live intervals.
6717 unsigned RegUsage = 0;
6718 for (auto Inst : OpenIntervals) {
6719 // Skip ignored values for VF > 1.
6720 if (VecValuesToIgnore.count(Inst) ||
6721 isScalarAfterVectorization(Inst, VFs[j]))
6723 RegUsage += GetRegUsage(Inst->getType(), VFs[j]);
6725 MaxUsages[j] = std::max(MaxUsages[j], RegUsage);
6728 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # "
6729 << OpenIntervals.size() << '\n');
6731 // Add the current instruction to the list of open intervals.
6732 OpenIntervals.insert(I);
6735 for (unsigned i = 0, e = VFs.size(); i < e; ++i) {
6736 unsigned Invariant = 0;
6738 Invariant = LoopInvariants.size();
6740 for (auto Inst : LoopInvariants)
6741 Invariant += GetRegUsage(Inst->getType(), VFs[i]);
6744 DEBUG(dbgs() << "LV(REG): VF = " << VFs[i] << '\n');
6745 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsages[i] << '\n');
6746 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
6747 DEBUG(dbgs() << "LV(REG): LoopSize: " << RU.NumInstructions << '\n');
6749 RU.LoopInvariantRegs = Invariant;
6750 RU.MaxLocalUsers = MaxUsages[i];
6757 void LoopVectorizationCostModel::collectInstsToScalarize(unsigned VF) {
6759 // If we aren't vectorizing the loop, or if we've already collected the
6760 // instructions to scalarize, there's nothing to do. Collection may already
6761 // have occurred if we have a user-selected VF and are now computing the
6762 // expected cost for interleaving.
6763 if (VF < 2 || InstsToScalarize.count(VF))
6766 // Initialize a mapping for VF in InstsToScalalarize. If we find that it's
6767 // not profitable to scalarize any instructions, the presence of VF in the
6768 // map will indicate that we've analyzed it already.
6769 ScalarCostsTy &ScalarCostsVF = InstsToScalarize[VF];
6771 // Find all the instructions that are scalar with predication in the loop and
6772 // determine if it would be better to not if-convert the blocks they are in.
6773 // If so, we also record the instructions to scalarize.
6774 for (BasicBlock *BB : TheLoop->blocks()) {
6775 if (!Legal->blockNeedsPredication(BB))
6777 for (Instruction &I : *BB)
6778 if (Legal->isScalarWithPredication(&I)) {
6779 ScalarCostsTy ScalarCosts;
6780 if (computePredInstDiscount(&I, ScalarCosts, VF) >= 0)
6781 ScalarCostsVF.insert(ScalarCosts.begin(), ScalarCosts.end());
6783 // Remember that BB will remain after vectorization.
6784 PredicatedBBsAfterVectorization.insert(BB);
6789 int LoopVectorizationCostModel::computePredInstDiscount(
6790 Instruction *PredInst, DenseMap<Instruction *, unsigned> &ScalarCosts,
6793 assert(!isUniformAfterVectorization(PredInst, VF) &&
6794 "Instruction marked uniform-after-vectorization will be predicated");
6796 // Initialize the discount to zero, meaning that the scalar version and the
6797 // vector version cost the same.
6800 // Holds instructions to analyze. The instructions we visit are mapped in
6801 // ScalarCosts. Those instructions are the ones that would be scalarized if
6802 // we find that the scalar version costs less.
6803 SmallVector<Instruction *, 8> Worklist;
6805 // Returns true if the given instruction can be scalarized.
6806 auto canBeScalarized = [&](Instruction *I) -> bool {
6808 // We only attempt to scalarize instructions forming a single-use chain
6809 // from the original predicated block that would otherwise be vectorized.
6810 // Although not strictly necessary, we give up on instructions we know will
6811 // already be scalar to avoid traversing chains that are unlikely to be
6813 if (!I->hasOneUse() || PredInst->getParent() != I->getParent() ||
6814 isScalarAfterVectorization(I, VF))
6817 // If the instruction is scalar with predication, it will be analyzed
6818 // separately. We ignore it within the context of PredInst.
6819 if (Legal->isScalarWithPredication(I))
6822 // If any of the instruction's operands are uniform after vectorization,
6823 // the instruction cannot be scalarized. This prevents, for example, a
6824 // masked load from being scalarized.
6826 // We assume we will only emit a value for lane zero of an instruction
6827 // marked uniform after vectorization, rather than VF identical values.
6828 // Thus, if we scalarize an instruction that uses a uniform, we would
6829 // create uses of values corresponding to the lanes we aren't emitting code
6830 // for. This behavior can be changed by allowing getScalarValue to clone
6831 // the lane zero values for uniforms rather than asserting.
6832 for (Use &U : I->operands())
6833 if (auto *J = dyn_cast<Instruction>(U.get()))
6834 if (isUniformAfterVectorization(J, VF))
6837 // Otherwise, we can scalarize the instruction.
6841 // Returns true if an operand that cannot be scalarized must be extracted
6842 // from a vector. We will account for this scalarization overhead below. Note
6843 // that the non-void predicated instructions are placed in their own blocks,
6844 // and their return values are inserted into vectors. Thus, an extract would
6845 // still be required.
6846 auto needsExtract = [&](Instruction *I) -> bool {
6847 return TheLoop->contains(I) && !isScalarAfterVectorization(I, VF);
6850 // Compute the expected cost discount from scalarizing the entire expression
6851 // feeding the predicated instruction. We currently only consider expressions
6852 // that are single-use instruction chains.
6853 Worklist.push_back(PredInst);
6854 while (!Worklist.empty()) {
6855 Instruction *I = Worklist.pop_back_val();
6857 // If we've already analyzed the instruction, there's nothing to do.
6858 if (ScalarCosts.count(I))
6861 // Compute the cost of the vector instruction. Note that this cost already
6862 // includes the scalarization overhead of the predicated instruction.
6863 unsigned VectorCost = getInstructionCost(I, VF).first;
6865 // Compute the cost of the scalarized instruction. This cost is the cost of
6866 // the instruction as if it wasn't if-converted and instead remained in the
6867 // predicated block. We will scale this cost by block probability after
6868 // computing the scalarization overhead.
6869 unsigned ScalarCost = VF * getInstructionCost(I, 1).first;
6871 // Compute the scalarization overhead of needed insertelement instructions
6873 if (Legal->isScalarWithPredication(I) && !I->getType()->isVoidTy()) {
6874 ScalarCost += TTI.getScalarizationOverhead(ToVectorTy(I->getType(), VF),
6876 ScalarCost += VF * TTI.getCFInstrCost(Instruction::PHI);
6879 // Compute the scalarization overhead of needed extractelement
6880 // instructions. For each of the instruction's operands, if the operand can
6881 // be scalarized, add it to the worklist; otherwise, account for the
6883 for (Use &U : I->operands())
6884 if (auto *J = dyn_cast<Instruction>(U.get())) {
6885 assert(VectorType::isValidElementType(J->getType()) &&
6886 "Instruction has non-scalar type");
6887 if (canBeScalarized(J))
6888 Worklist.push_back(J);
6889 else if (needsExtract(J))
6890 ScalarCost += TTI.getScalarizationOverhead(
6891 ToVectorTy(J->getType(),VF), false, true);
6894 // Scale the total scalar cost by block probability.
6895 ScalarCost /= getReciprocalPredBlockProb();
6897 // Compute the discount. A non-negative discount means the vector version
6898 // of the instruction costs more, and scalarizing would be beneficial.
6899 Discount += VectorCost - ScalarCost;
6900 ScalarCosts[I] = ScalarCost;
6906 LoopVectorizationCostModel::VectorizationCostTy
6907 LoopVectorizationCostModel::expectedCost(unsigned VF) {
6908 VectorizationCostTy Cost;
6910 // Collect Uniform and Scalar instructions after vectorization with VF.
6911 collectUniformsAndScalars(VF);
6913 // Collect the instructions (and their associated costs) that will be more
6914 // profitable to scalarize.
6915 collectInstsToScalarize(VF);
6918 for (BasicBlock *BB : TheLoop->blocks()) {
6919 VectorizationCostTy BlockCost;
6921 // For each instruction in the old loop.
6922 for (Instruction &I : *BB) {
6923 // Skip dbg intrinsics.
6924 if (isa<DbgInfoIntrinsic>(I))
6927 // Skip ignored values.
6928 if (ValuesToIgnore.count(&I))
6931 VectorizationCostTy C = getInstructionCost(&I, VF);
6933 // Check if we should override the cost.
6934 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
6935 C.first = ForceTargetInstructionCost;
6937 BlockCost.first += C.first;
6938 BlockCost.second |= C.second;
6939 DEBUG(dbgs() << "LV: Found an estimated cost of " << C.first << " for VF "
6940 << VF << " For instruction: " << I << '\n');
6943 // If we are vectorizing a predicated block, it will have been
6944 // if-converted. This means that the block's instructions (aside from
6945 // stores and instructions that may divide by zero) will now be
6946 // unconditionally executed. For the scalar case, we may not always execute
6947 // the predicated block. Thus, scale the block's cost by the probability of
6949 if (VF == 1 && Legal->blockNeedsPredication(BB))
6950 BlockCost.first /= getReciprocalPredBlockProb();
6952 Cost.first += BlockCost.first;
6953 Cost.second |= BlockCost.second;
6959 /// \brief Gets Address Access SCEV after verifying that the access pattern
6960 /// is loop invariant except the induction variable dependence.
6962 /// This SCEV can be sent to the Target in order to estimate the address
6963 /// calculation cost.
6964 static const SCEV *getAddressAccessSCEV(
6966 LoopVectorizationLegality *Legal,
6967 ScalarEvolution *SE,
6968 const Loop *TheLoop) {
6969 auto *Gep = dyn_cast<GetElementPtrInst>(Ptr);
6973 // We are looking for a gep with all loop invariant indices except for one
6974 // which should be an induction variable.
6975 unsigned NumOperands = Gep->getNumOperands();
6976 for (unsigned i = 1; i < NumOperands; ++i) {
6977 Value *Opd = Gep->getOperand(i);
6978 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
6979 !Legal->isInductionVariable(Opd))
6983 // Now we know we have a GEP ptr, %inv, %ind, %inv. return the Ptr SCEV.
6984 return SE->getSCEV(Ptr);
6987 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
6988 return Legal->hasStride(I->getOperand(0)) ||
6989 Legal->hasStride(I->getOperand(1));
6992 unsigned LoopVectorizationCostModel::getMemInstScalarizationCost(Instruction *I,
6994 Type *ValTy = getMemInstValueType(I);
6995 auto SE = PSE.getSE();
6997 unsigned Alignment = getMemInstAlignment(I);
6998 unsigned AS = getMemInstAddressSpace(I);
6999 Value *Ptr = getPointerOperand(I);
7000 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
7002 // Figure out whether the access is strided and get the stride value
7003 // if it's known in compile time
7004 const SCEV *PtrSCEV = getAddressAccessSCEV(Ptr, Legal, SE, TheLoop);
7006 // Get the cost of the scalar memory instruction and address computation.
7007 unsigned Cost = VF * TTI.getAddressComputationCost(PtrTy, SE, PtrSCEV);
7010 TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(), Alignment,
7013 // Get the overhead of the extractelement and insertelement instructions
7014 // we might create due to scalarization.
7015 Cost += getScalarizationOverhead(I, VF, TTI);
7017 // If we have a predicated store, it may not be executed for each vector
7018 // lane. Scale the cost by the probability of executing the predicated
7020 if (Legal->isScalarWithPredication(I))
7021 Cost /= getReciprocalPredBlockProb();
7026 unsigned LoopVectorizationCostModel::getConsecutiveMemOpCost(Instruction *I,
7028 Type *ValTy = getMemInstValueType(I);
7029 Type *VectorTy = ToVectorTy(ValTy, VF);
7030 unsigned Alignment = getMemInstAlignment(I);
7031 Value *Ptr = getPointerOperand(I);
7032 unsigned AS = getMemInstAddressSpace(I);
7033 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
7035 assert((ConsecutiveStride == 1 || ConsecutiveStride == -1) &&
7036 "Stride should be 1 or -1 for consecutive memory access");
7038 if (Legal->isMaskRequired(I))
7039 Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
7041 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS, I);
7043 bool Reverse = ConsecutiveStride < 0;
7045 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
7049 unsigned LoopVectorizationCostModel::getUniformMemOpCost(Instruction *I,
7051 LoadInst *LI = cast<LoadInst>(I);
7052 Type *ValTy = LI->getType();
7053 Type *VectorTy = ToVectorTy(ValTy, VF);
7054 unsigned Alignment = LI->getAlignment();
7055 unsigned AS = LI->getPointerAddressSpace();
7057 return TTI.getAddressComputationCost(ValTy) +
7058 TTI.getMemoryOpCost(Instruction::Load, ValTy, Alignment, AS) +
7059 TTI.getShuffleCost(TargetTransformInfo::SK_Broadcast, VectorTy);
7062 unsigned LoopVectorizationCostModel::getGatherScatterCost(Instruction *I,
7064 Type *ValTy = getMemInstValueType(I);
7065 Type *VectorTy = ToVectorTy(ValTy, VF);
7066 unsigned Alignment = getMemInstAlignment(I);
7067 Value *Ptr = getPointerOperand(I);
7069 return TTI.getAddressComputationCost(VectorTy) +
7070 TTI.getGatherScatterOpCost(I->getOpcode(), VectorTy, Ptr,
7071 Legal->isMaskRequired(I), Alignment);
7074 unsigned LoopVectorizationCostModel::getInterleaveGroupCost(Instruction *I,
7076 Type *ValTy = getMemInstValueType(I);
7077 Type *VectorTy = ToVectorTy(ValTy, VF);
7078 unsigned AS = getMemInstAddressSpace(I);
7080 auto Group = Legal->getInterleavedAccessGroup(I);
7081 assert(Group && "Fail to get an interleaved access group.");
7083 unsigned InterleaveFactor = Group->getFactor();
7084 Type *WideVecTy = VectorType::get(ValTy, VF * InterleaveFactor);
7086 // Holds the indices of existing members in an interleaved load group.
7087 // An interleaved store group doesn't need this as it doesn't allow gaps.
7088 SmallVector<unsigned, 4> Indices;
7089 if (isa<LoadInst>(I)) {
7090 for (unsigned i = 0; i < InterleaveFactor; i++)
7091 if (Group->getMember(i))
7092 Indices.push_back(i);
7095 // Calculate the cost of the whole interleaved group.
7096 unsigned Cost = TTI.getInterleavedMemoryOpCost(I->getOpcode(), WideVecTy,
7097 Group->getFactor(), Indices,
7098 Group->getAlignment(), AS);
7100 if (Group->isReverse())
7101 Cost += Group->getNumMembers() *
7102 TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
7106 unsigned LoopVectorizationCostModel::getMemoryInstructionCost(Instruction *I,
7109 // Calculate scalar cost only. Vectorization cost should be ready at this
7112 Type *ValTy = getMemInstValueType(I);
7113 unsigned Alignment = getMemInstAlignment(I);
7114 unsigned AS = getMemInstAddressSpace(I);
7116 return TTI.getAddressComputationCost(ValTy) +
7117 TTI.getMemoryOpCost(I->getOpcode(), ValTy, Alignment, AS, I);
7119 return getWideningCost(I, VF);
7122 LoopVectorizationCostModel::VectorizationCostTy
7123 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
7124 // If we know that this instruction will remain uniform, check the cost of
7125 // the scalar version.
7126 if (isUniformAfterVectorization(I, VF))
7129 if (VF > 1 && isProfitableToScalarize(I, VF))
7130 return VectorizationCostTy(InstsToScalarize[VF][I], false);
7133 unsigned C = getInstructionCost(I, VF, VectorTy);
7135 bool TypeNotScalarized =
7136 VF > 1 && !VectorTy->isVoidTy() && TTI.getNumberOfParts(VectorTy) < VF;
7137 return VectorizationCostTy(C, TypeNotScalarized);
7140 void LoopVectorizationCostModel::setCostBasedWideningDecision(unsigned VF) {
7143 for (BasicBlock *BB : TheLoop->blocks()) {
7144 // For each instruction in the old loop.
7145 for (Instruction &I : *BB) {
7146 Value *Ptr = getPointerOperand(&I);
7150 if (isa<LoadInst>(&I) && Legal->isUniform(Ptr)) {
7151 // Scalar load + broadcast
7152 unsigned Cost = getUniformMemOpCost(&I, VF);
7153 setWideningDecision(&I, VF, CM_Scalarize, Cost);
7157 // We assume that widening is the best solution when possible.
7158 if (Legal->memoryInstructionCanBeWidened(&I, VF)) {
7159 unsigned Cost = getConsecutiveMemOpCost(&I, VF);
7160 setWideningDecision(&I, VF, CM_Widen, Cost);
7164 // Choose between Interleaving, Gather/Scatter or Scalarization.
7165 unsigned InterleaveCost = UINT_MAX;
7166 unsigned NumAccesses = 1;
7167 if (Legal->isAccessInterleaved(&I)) {
7168 auto Group = Legal->getInterleavedAccessGroup(&I);
7169 assert(Group && "Fail to get an interleaved access group.");
7171 // Make one decision for the whole group.
7172 if (getWideningDecision(&I, VF) != CM_Unknown)
7175 NumAccesses = Group->getNumMembers();
7176 InterleaveCost = getInterleaveGroupCost(&I, VF);
7179 unsigned GatherScatterCost =
7180 Legal->isLegalGatherOrScatter(&I)
7181 ? getGatherScatterCost(&I, VF) * NumAccesses
7184 unsigned ScalarizationCost =
7185 getMemInstScalarizationCost(&I, VF) * NumAccesses;
7187 // Choose better solution for the current VF,
7188 // write down this decision and use it during vectorization.
7190 InstWidening Decision;
7191 if (InterleaveCost <= GatherScatterCost &&
7192 InterleaveCost < ScalarizationCost) {
7193 Decision = CM_Interleave;
7194 Cost = InterleaveCost;
7195 } else if (GatherScatterCost < ScalarizationCost) {
7196 Decision = CM_GatherScatter;
7197 Cost = GatherScatterCost;
7199 Decision = CM_Scalarize;
7200 Cost = ScalarizationCost;
7202 // If the instructions belongs to an interleave group, the whole group
7203 // receives the same decision. The whole group receives the cost, but
7204 // the cost will actually be assigned to one instruction.
7205 if (auto Group = Legal->getInterleavedAccessGroup(&I))
7206 setWideningDecision(Group, VF, Decision, Cost);
7208 setWideningDecision(&I, VF, Decision, Cost);
7213 unsigned LoopVectorizationCostModel::getInstructionCost(Instruction *I,
7216 Type *RetTy = I->getType();
7217 if (canTruncateToMinimalBitwidth(I, VF))
7218 RetTy = IntegerType::get(RetTy->getContext(), MinBWs[I]);
7219 VectorTy = ToVectorTy(RetTy, VF);
7220 auto SE = PSE.getSE();
7222 // TODO: We need to estimate the cost of intrinsic calls.
7223 switch (I->getOpcode()) {
7224 case Instruction::GetElementPtr:
7225 // We mark this instruction as zero-cost because the cost of GEPs in
7226 // vectorized code depends on whether the corresponding memory instruction
7227 // is scalarized or not. Therefore, we handle GEPs with the memory
7228 // instruction cost.
7230 case Instruction::Br: {
7231 // In cases of scalarized and predicated instructions, there will be VF
7232 // predicated blocks in the vectorized loop. Each branch around these
7233 // blocks requires also an extract of its vector compare i1 element.
7234 bool ScalarPredicatedBB = false;
7235 BranchInst *BI = cast<BranchInst>(I);
7236 if (VF > 1 && BI->isConditional() &&
7237 (PredicatedBBsAfterVectorization.count(BI->getSuccessor(0)) ||
7238 PredicatedBBsAfterVectorization.count(BI->getSuccessor(1))))
7239 ScalarPredicatedBB = true;
7241 if (ScalarPredicatedBB) {
7242 // Return cost for branches around scalarized and predicated blocks.
7244 VectorType::get(IntegerType::getInt1Ty(RetTy->getContext()), VF);
7245 return (TTI.getScalarizationOverhead(Vec_i1Ty, false, true) +
7246 (TTI.getCFInstrCost(Instruction::Br) * VF));
7247 } else if (I->getParent() == TheLoop->getLoopLatch() || VF == 1)
7248 // The back-edge branch will remain, as will all scalar branches.
7249 return TTI.getCFInstrCost(Instruction::Br);
7251 // This branch will be eliminated by if-conversion.
7253 // Note: We currently assume zero cost for an unconditional branch inside
7254 // a predicated block since it will become a fall-through, although we
7255 // may decide in the future to call TTI for all branches.
7257 case Instruction::PHI: {
7258 auto *Phi = cast<PHINode>(I);
7260 // First-order recurrences are replaced by vector shuffles inside the loop.
7261 if (VF > 1 && Legal->isFirstOrderRecurrence(Phi))
7262 return TTI.getShuffleCost(TargetTransformInfo::SK_ExtractSubvector,
7263 VectorTy, VF - 1, VectorTy);
7265 // Phi nodes in non-header blocks (not inductions, reductions, etc.) are
7266 // converted into select instructions. We require N - 1 selects per phi
7267 // node, where N is the number of incoming values.
7268 if (VF > 1 && Phi->getParent() != TheLoop->getHeader())
7269 return (Phi->getNumIncomingValues() - 1) *
7270 TTI.getCmpSelInstrCost(
7271 Instruction::Select, ToVectorTy(Phi->getType(), VF),
7272 ToVectorTy(Type::getInt1Ty(Phi->getContext()), VF));
7274 return TTI.getCFInstrCost(Instruction::PHI);
7276 case Instruction::UDiv:
7277 case Instruction::SDiv:
7278 case Instruction::URem:
7279 case Instruction::SRem:
7280 // If we have a predicated instruction, it may not be executed for each
7281 // vector lane. Get the scalarization cost and scale this amount by the
7282 // probability of executing the predicated block. If the instruction is not
7283 // predicated, we fall through to the next case.
7284 if (VF > 1 && Legal->isScalarWithPredication(I)) {
7287 // These instructions have a non-void type, so account for the phi nodes
7288 // that we will create. This cost is likely to be zero. The phi node
7289 // cost, if any, should be scaled by the block probability because it
7290 // models a copy at the end of each predicated block.
7291 Cost += VF * TTI.getCFInstrCost(Instruction::PHI);
7293 // The cost of the non-predicated instruction.
7294 Cost += VF * TTI.getArithmeticInstrCost(I->getOpcode(), RetTy);
7296 // The cost of insertelement and extractelement instructions needed for
7298 Cost += getScalarizationOverhead(I, VF, TTI);
7300 // Scale the cost by the probability of executing the predicated blocks.
7301 // This assumes the predicated block for each vector lane is equally
7303 return Cost / getReciprocalPredBlockProb();
7305 case Instruction::Add:
7306 case Instruction::FAdd:
7307 case Instruction::Sub:
7308 case Instruction::FSub:
7309 case Instruction::Mul:
7310 case Instruction::FMul:
7311 case Instruction::FDiv:
7312 case Instruction::FRem:
7313 case Instruction::Shl:
7314 case Instruction::LShr:
7315 case Instruction::AShr:
7316 case Instruction::And:
7317 case Instruction::Or:
7318 case Instruction::Xor: {
7319 // Since we will replace the stride by 1 the multiplication should go away.
7320 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
7322 // Certain instructions can be cheaper to vectorize if they have a constant
7323 // second vector operand. One example of this are shifts on x86.
7324 TargetTransformInfo::OperandValueKind Op1VK =
7325 TargetTransformInfo::OK_AnyValue;
7326 TargetTransformInfo::OperandValueKind Op2VK =
7327 TargetTransformInfo::OK_AnyValue;
7328 TargetTransformInfo::OperandValueProperties Op1VP =
7329 TargetTransformInfo::OP_None;
7330 TargetTransformInfo::OperandValueProperties Op2VP =
7331 TargetTransformInfo::OP_None;
7332 Value *Op2 = I->getOperand(1);
7334 // Check for a splat or for a non uniform vector of constants.
7335 if (isa<ConstantInt>(Op2)) {
7336 ConstantInt *CInt = cast<ConstantInt>(Op2);
7337 if (CInt && CInt->getValue().isPowerOf2())
7338 Op2VP = TargetTransformInfo::OP_PowerOf2;
7339 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
7340 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
7341 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
7342 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
7344 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
7345 if (CInt && CInt->getValue().isPowerOf2())
7346 Op2VP = TargetTransformInfo::OP_PowerOf2;
7347 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
7349 } else if (Legal->isUniform(Op2)) {
7350 Op2VK = TargetTransformInfo::OK_UniformValue;
7352 SmallVector<const Value *, 4> Operands(I->operand_values());
7353 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK,
7354 Op2VK, Op1VP, Op2VP, Operands);
7356 case Instruction::Select: {
7357 SelectInst *SI = cast<SelectInst>(I);
7358 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
7359 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
7360 Type *CondTy = SI->getCondition()->getType();
7362 CondTy = VectorType::get(CondTy, VF);
7364 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy, I);
7366 case Instruction::ICmp:
7367 case Instruction::FCmp: {
7368 Type *ValTy = I->getOperand(0)->getType();
7369 Instruction *Op0AsInstruction = dyn_cast<Instruction>(I->getOperand(0));
7370 if (canTruncateToMinimalBitwidth(Op0AsInstruction, VF))
7371 ValTy = IntegerType::get(ValTy->getContext(), MinBWs[Op0AsInstruction]);
7372 VectorTy = ToVectorTy(ValTy, VF);
7373 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, nullptr, I);
7375 case Instruction::Store:
7376 case Instruction::Load: {
7377 VectorTy = ToVectorTy(getMemInstValueType(I), VF);
7378 return getMemoryInstructionCost(I, VF);
7380 case Instruction::ZExt:
7381 case Instruction::SExt:
7382 case Instruction::FPToUI:
7383 case Instruction::FPToSI:
7384 case Instruction::FPExt:
7385 case Instruction::PtrToInt:
7386 case Instruction::IntToPtr:
7387 case Instruction::SIToFP:
7388 case Instruction::UIToFP:
7389 case Instruction::Trunc:
7390 case Instruction::FPTrunc:
7391 case Instruction::BitCast: {
7392 // We optimize the truncation of induction variables having constant
7393 // integer steps. The cost of these truncations is the same as the scalar
7395 if (isOptimizableIVTruncate(I, VF)) {
7396 auto *Trunc = cast<TruncInst>(I);
7397 return TTI.getCastInstrCost(Instruction::Trunc, Trunc->getDestTy(),
7398 Trunc->getSrcTy(), Trunc);
7401 Type *SrcScalarTy = I->getOperand(0)->getType();
7402 Type *SrcVecTy = ToVectorTy(SrcScalarTy, VF);
7403 if (canTruncateToMinimalBitwidth(I, VF)) {
7404 // This cast is going to be shrunk. This may remove the cast or it might
7405 // turn it into slightly different cast. For example, if MinBW == 16,
7406 // "zext i8 %1 to i32" becomes "zext i8 %1 to i16".
7408 // Calculate the modified src and dest types.
7409 Type *MinVecTy = VectorTy;
7410 if (I->getOpcode() == Instruction::Trunc) {
7411 SrcVecTy = smallestIntegerVectorType(SrcVecTy, MinVecTy);
7413 largestIntegerVectorType(ToVectorTy(I->getType(), VF), MinVecTy);
7414 } else if (I->getOpcode() == Instruction::ZExt ||
7415 I->getOpcode() == Instruction::SExt) {
7416 SrcVecTy = largestIntegerVectorType(SrcVecTy, MinVecTy);
7418 smallestIntegerVectorType(ToVectorTy(I->getType(), VF), MinVecTy);
7422 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy, I);
7424 case Instruction::Call: {
7425 bool NeedToScalarize;
7426 CallInst *CI = cast<CallInst>(I);
7427 unsigned CallCost = getVectorCallCost(CI, VF, TTI, TLI, NeedToScalarize);
7428 if (getVectorIntrinsicIDForCall(CI, TLI))
7429 return std::min(CallCost, getVectorIntrinsicCost(CI, VF, TTI, TLI));
7433 // The cost of executing VF copies of the scalar instruction. This opcode
7434 // is unknown. Assume that it is the same as 'mul'.
7435 return VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy) +
7436 getScalarizationOverhead(I, VF, TTI);
7440 char LoopVectorize::ID = 0;
7441 static const char lv_name[] = "Loop Vectorization";
7442 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
7443 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
7444 INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass)
7445 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
7446 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
7447 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
7448 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
7449 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
7450 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
7451 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
7452 INITIALIZE_PASS_DEPENDENCY(LoopAccessLegacyAnalysis)
7453 INITIALIZE_PASS_DEPENDENCY(DemandedBitsWrapperPass)
7454 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
7455 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
7458 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
7459 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
7463 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
7465 // Check if the pointer operand of a load or store instruction is
7467 if (auto *Ptr = getPointerOperand(Inst))
7468 return Legal->isConsecutivePtr(Ptr);
7472 void LoopVectorizationCostModel::collectValuesToIgnore() {
7473 // Ignore ephemeral values.
7474 CodeMetrics::collectEphemeralValues(TheLoop, AC, ValuesToIgnore);
7476 // Ignore type-promoting instructions we identified during reduction
7478 for (auto &Reduction : *Legal->getReductionVars()) {
7479 RecurrenceDescriptor &RedDes = Reduction.second;
7480 SmallPtrSetImpl<Instruction *> &Casts = RedDes.getCastInsts();
7481 VecValuesToIgnore.insert(Casts.begin(), Casts.end());
7485 LoopVectorizationCostModel::VectorizationFactor
7486 LoopVectorizationPlanner::plan(bool OptForSize, unsigned UserVF) {
7488 // Width 1 means no vectorize, cost 0 means uncomputed cost.
7489 const LoopVectorizationCostModel::VectorizationFactor NoVectorization = {1U,
7491 Optional<unsigned> MaybeMaxVF = CM.computeMaxVF(OptForSize);
7492 if (!MaybeMaxVF.hasValue()) // Cases considered too costly to vectorize.
7493 return NoVectorization;
7496 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
7497 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
7498 // Collect the instructions (and their associated costs) that will be more
7499 // profitable to scalarize.
7500 CM.selectUserVectorizationFactor(UserVF);
7504 unsigned MaxVF = MaybeMaxVF.getValue();
7505 assert(MaxVF != 0 && "MaxVF is zero.");
7507 return NoVectorization;
7509 // Select the optimal vectorization factor.
7510 return CM.selectVectorizationFactor(MaxVF);
7513 void LoopVectorizationPlanner::executePlan(InnerLoopVectorizer &ILV) {
7514 // Perform the actual loop transformation.
7516 // 1. Create a new empty loop. Unlink the old loop and connect the new one.
7517 ILV.createVectorizedLoopSkeleton();
7519 //===------------------------------------------------===//
7521 // Notice: any optimization or new instruction that go
7522 // into the code below should also be implemented in
7525 //===------------------------------------------------===//
7527 // 2. Copy and widen instructions from the old loop into the new loop.
7529 // Collect instructions from the original loop that will become trivially dead
7530 // in the vectorized loop. We don't need to vectorize these instructions. For
7531 // example, original induction update instructions can become dead because we
7532 // separately emit induction "steps" when generating code for the new loop.
7533 // Similarly, we create a new latch condition when setting up the structure
7534 // of the new loop, so the old one can become dead.
7535 SmallPtrSet<Instruction *, 4> DeadInstructions;
7536 collectTriviallyDeadInstructions(DeadInstructions);
7538 // Scan the loop in a topological order to ensure that defs are vectorized
7540 LoopBlocksDFS DFS(OrigLoop);
7543 // Vectorize all instructions in the original loop that will not become
7544 // trivially dead when vectorized.
7545 for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO()))
7546 for (Instruction &I : *BB)
7547 if (!DeadInstructions.count(&I))
7548 ILV.vectorizeInstruction(I);
7550 // 3. Fix the vectorized code: take care of header phi's, live-outs,
7551 // predication, updating analyses.
7552 ILV.fixVectorizedLoop();
7555 void LoopVectorizationPlanner::collectTriviallyDeadInstructions(
7556 SmallPtrSetImpl<Instruction *> &DeadInstructions) {
7557 BasicBlock *Latch = OrigLoop->getLoopLatch();
7559 // We create new control-flow for the vectorized loop, so the original
7560 // condition will be dead after vectorization if it's only used by the
7562 auto *Cmp = dyn_cast<Instruction>(Latch->getTerminator()->getOperand(0));
7563 if (Cmp && Cmp->hasOneUse())
7564 DeadInstructions.insert(Cmp);
7566 // We create new "steps" for induction variable updates to which the original
7567 // induction variables map. An original update instruction will be dead if
7568 // all its users except the induction variable are dead.
7569 for (auto &Induction : *Legal->getInductionVars()) {
7570 PHINode *Ind = Induction.first;
7571 auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch));
7572 if (all_of(IndUpdate->users(), [&](User *U) -> bool {
7573 return U == Ind || DeadInstructions.count(cast<Instruction>(U));
7575 DeadInstructions.insert(IndUpdate);
7579 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
7580 auto *SI = dyn_cast<StoreInst>(Instr);
7581 bool IfPredicateInstr = (SI && Legal->blockNeedsPredication(SI->getParent()));
7583 return scalarizeInstruction(Instr, IfPredicateInstr);
7586 Value *InnerLoopUnroller::reverseVector(Value *Vec) { return Vec; }
7588 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) { return V; }
7590 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step,
7591 Instruction::BinaryOps BinOp) {
7592 // When unrolling and the VF is 1, we only need to add a simple scalar.
7593 Type *Ty = Val->getType();
7594 assert(!Ty->isVectorTy() && "Val must be a scalar");
7596 if (Ty->isFloatingPointTy()) {
7597 Constant *C = ConstantFP::get(Ty, (double)StartIdx);
7599 // Floating point operations had to be 'fast' to enable the unrolling.
7600 Value *MulOp = addFastMathFlag(Builder.CreateFMul(C, Step));
7601 return addFastMathFlag(Builder.CreateBinOp(BinOp, Val, MulOp));
7603 Constant *C = ConstantInt::get(Ty, StartIdx);
7604 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");
7607 static void AddRuntimeUnrollDisableMetaData(Loop *L) {
7608 SmallVector<Metadata *, 4> MDs;
7609 // Reserve first location for self reference to the LoopID metadata node.
7610 MDs.push_back(nullptr);
7611 bool IsUnrollMetadata = false;
7612 MDNode *LoopID = L->getLoopID();
7614 // First find existing loop unrolling disable metadata.
7615 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
7616 auto *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
7618 const auto *S = dyn_cast<MDString>(MD->getOperand(0));
7620 S && S->getString().startswith("llvm.loop.unroll.disable");
7622 MDs.push_back(LoopID->getOperand(i));
7626 if (!IsUnrollMetadata) {
7627 // Add runtime unroll disable metadata.
7628 LLVMContext &Context = L->getHeader()->getContext();
7629 SmallVector<Metadata *, 1> DisableOperands;
7630 DisableOperands.push_back(
7631 MDString::get(Context, "llvm.loop.unroll.runtime.disable"));
7632 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
7633 MDs.push_back(DisableNode);
7634 MDNode *NewLoopID = MDNode::get(Context, MDs);
7635 // Set operand 0 to refer to the loop id itself.
7636 NewLoopID->replaceOperandWith(0, NewLoopID);
7637 L->setLoopID(NewLoopID);
7641 bool LoopVectorizePass::processLoop(Loop *L) {
7642 assert(L->empty() && "Only process inner loops.");
7645 const std::string DebugLocStr = getDebugLocString(L);
7648 DEBUG(dbgs() << "\nLV: Checking a loop in \""
7649 << L->getHeader()->getParent()->getName() << "\" from "
7650 << DebugLocStr << "\n");
7652 LoopVectorizeHints Hints(L, DisableUnrolling, *ORE);
7654 DEBUG(dbgs() << "LV: Loop hints:"
7656 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
7658 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
7661 << " width=" << Hints.getWidth()
7662 << " unroll=" << Hints.getInterleave() << "\n");
7664 // Function containing loop
7665 Function *F = L->getHeader()->getParent();
7667 // Looking at the diagnostic output is the only way to determine if a loop
7668 // was vectorized (other than looking at the IR or machine code), so it
7669 // is important to generate an optimization remark for each loop. Most of
7670 // these messages are generated as OptimizationRemarkAnalysis. Remarks
7671 // generated as OptimizationRemark and OptimizationRemarkMissed are
7672 // less verbose reporting vectorized loops and unvectorized loops that may
7673 // benefit from vectorization, respectively.
7675 if (!Hints.allowVectorization(F, L, AlwaysVectorize)) {
7676 DEBUG(dbgs() << "LV: Loop hints prevent vectorization.\n");
7680 // Check the loop for a trip count threshold:
7681 // do not vectorize loops with a tiny trip count.
7682 const unsigned MaxTC = SE->getSmallConstantMaxTripCount(L);
7683 if (MaxTC > 0u && MaxTC < TinyTripCountVectorThreshold) {
7684 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
7685 << "This loop is not worth vectorizing.");
7686 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
7687 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
7689 DEBUG(dbgs() << "\n");
7690 ORE->emit(createMissedAnalysis(Hints.vectorizeAnalysisPassName(),
7692 << "vectorization is not beneficial "
7693 "and is not explicitly forced");
7698 PredicatedScalarEvolution PSE(*SE, *L);
7700 // Check if it is legal to vectorize the loop.
7701 LoopVectorizationRequirements Requirements(*ORE);
7702 LoopVectorizationLegality LVL(L, PSE, DT, TLI, AA, F, TTI, GetLAA, LI, ORE,
7703 &Requirements, &Hints);
7704 if (!LVL.canVectorize()) {
7705 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
7706 emitMissedWarning(F, L, Hints, ORE);
7710 // Check the function attributes to find out if this function should be
7711 // optimized for size.
7713 Hints.getForce() != LoopVectorizeHints::FK_Enabled && F->optForSize();
7715 // Compute the weighted frequency of this loop being executed and see if it
7716 // is less than 20% of the function entry baseline frequency. Note that we
7717 // always have a canonical loop here because we think we *can* vectorize.
7718 // FIXME: This is hidden behind a flag due to pervasive problems with
7719 // exactly what block frequency models.
7720 if (LoopVectorizeWithBlockFrequency) {
7721 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
7722 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
7723 LoopEntryFreq < ColdEntryFreq)
7727 // Check the function attributes to see if implicit floats are allowed.
7728 // FIXME: This check doesn't seem possibly correct -- what if the loop is
7729 // an integer loop and the vector instructions selected are purely integer
7730 // vector instructions?
7731 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
7732 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
7733 "attribute is used.\n");
7734 ORE->emit(createMissedAnalysis(Hints.vectorizeAnalysisPassName(),
7735 "NoImplicitFloat", L)
7736 << "loop not vectorized due to NoImplicitFloat attribute");
7737 emitMissedWarning(F, L, Hints, ORE);
7741 // Check if the target supports potentially unsafe FP vectorization.
7742 // FIXME: Add a check for the type of safety issue (denormal, signaling)
7743 // for the target we're vectorizing for, to make sure none of the
7744 // additional fp-math flags can help.
7745 if (Hints.isPotentiallyUnsafe() &&
7746 TTI->isFPVectorizationPotentiallyUnsafe()) {
7747 DEBUG(dbgs() << "LV: Potentially unsafe FP op prevents vectorization.\n");
7749 createMissedAnalysis(Hints.vectorizeAnalysisPassName(), "UnsafeFP", L)
7750 << "loop not vectorized due to unsafe FP support.");
7751 emitMissedWarning(F, L, Hints, ORE);
7755 // Use the cost model.
7756 LoopVectorizationCostModel CM(L, PSE, LI, &LVL, *TTI, TLI, DB, AC, ORE, F,
7758 CM.collectValuesToIgnore();
7760 // Use the planner for vectorization.
7761 LoopVectorizationPlanner LVP(L, LI, &LVL, CM);
7763 // Get user vectorization factor.
7764 unsigned UserVF = Hints.getWidth();
7766 // Plan how to best vectorize, return the best VF and its cost.
7767 LoopVectorizationCostModel::VectorizationFactor VF =
7768 LVP.plan(OptForSize, UserVF);
7770 // Select the interleave count.
7771 unsigned IC = CM.selectInterleaveCount(OptForSize, VF.Width, VF.Cost);
7773 // Get user interleave count.
7774 unsigned UserIC = Hints.getInterleave();
7776 // Identify the diagnostic messages that should be produced.
7777 std::pair<StringRef, std::string> VecDiagMsg, IntDiagMsg;
7778 bool VectorizeLoop = true, InterleaveLoop = true;
7779 if (Requirements.doesNotMeet(F, L, Hints)) {
7780 DEBUG(dbgs() << "LV: Not vectorizing: loop did not meet vectorization "
7782 emitMissedWarning(F, L, Hints, ORE);
7786 if (VF.Width == 1) {
7787 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
7788 VecDiagMsg = std::make_pair(
7789 "VectorizationNotBeneficial",
7790 "the cost-model indicates that vectorization is not beneficial");
7791 VectorizeLoop = false;
7794 if (IC == 1 && UserIC <= 1) {
7795 // Tell the user interleaving is not beneficial.
7796 DEBUG(dbgs() << "LV: Interleaving is not beneficial.\n");
7797 IntDiagMsg = std::make_pair(
7798 "InterleavingNotBeneficial",
7799 "the cost-model indicates that interleaving is not beneficial");
7800 InterleaveLoop = false;
7802 IntDiagMsg.first = "InterleavingNotBeneficialAndDisabled";
7803 IntDiagMsg.second +=
7804 " and is explicitly disabled or interleave count is set to 1";
7806 } else if (IC > 1 && UserIC == 1) {
7807 // Tell the user interleaving is beneficial, but it explicitly disabled.
7809 << "LV: Interleaving is beneficial but is explicitly disabled.");
7810 IntDiagMsg = std::make_pair(
7811 "InterleavingBeneficialButDisabled",
7812 "the cost-model indicates that interleaving is beneficial "
7813 "but is explicitly disabled or interleave count is set to 1");
7814 InterleaveLoop = false;
7817 // Override IC if user provided an interleave count.
7818 IC = UserIC > 0 ? UserIC : IC;
7820 // Emit diagnostic messages, if any.
7821 const char *VAPassName = Hints.vectorizeAnalysisPassName();
7822 if (!VectorizeLoop && !InterleaveLoop) {
7823 // Do not vectorize or interleaving the loop.
7824 ORE->emit(OptimizationRemarkMissed(VAPassName, VecDiagMsg.first,
7825 L->getStartLoc(), L->getHeader())
7826 << VecDiagMsg.second);
7827 ORE->emit(OptimizationRemarkMissed(LV_NAME, IntDiagMsg.first,
7828 L->getStartLoc(), L->getHeader())
7829 << IntDiagMsg.second);
7831 } else if (!VectorizeLoop && InterleaveLoop) {
7832 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
7833 ORE->emit(OptimizationRemarkAnalysis(VAPassName, VecDiagMsg.first,
7834 L->getStartLoc(), L->getHeader())
7835 << VecDiagMsg.second);
7836 } else if (VectorizeLoop && !InterleaveLoop) {
7837 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
7838 << DebugLocStr << '\n');
7839 ORE->emit(OptimizationRemarkAnalysis(LV_NAME, IntDiagMsg.first,
7840 L->getStartLoc(), L->getHeader())
7841 << IntDiagMsg.second);
7842 } else if (VectorizeLoop && InterleaveLoop) {
7843 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
7844 << DebugLocStr << '\n');
7845 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
7848 using namespace ore;
7849 if (!VectorizeLoop) {
7850 assert(IC > 1 && "interleave count should not be 1 or 0");
7851 // If we decided that it is not legal to vectorize the loop, then
7853 InnerLoopUnroller Unroller(L, PSE, LI, DT, TLI, TTI, AC, ORE, IC, &LVL,
7855 LVP.executePlan(Unroller);
7857 ORE->emit(OptimizationRemark(LV_NAME, "Interleaved", L->getStartLoc(),
7859 << "interleaved loop (interleaved count: "
7860 << NV("InterleaveCount", IC) << ")");
7862 // If we decided that it is *legal* to vectorize the loop, then do it.
7863 InnerLoopVectorizer LB(L, PSE, LI, DT, TLI, TTI, AC, ORE, VF.Width, IC,
7865 LVP.executePlan(LB);
7868 // Add metadata to disable runtime unrolling a scalar loop when there are
7869 // no runtime checks about strides and memory. A scalar loop that is
7870 // rarely used is not worth unrolling.
7871 if (!LB.areSafetyChecksAdded())
7872 AddRuntimeUnrollDisableMetaData(L);
7874 // Report the vectorization decision.
7875 ORE->emit(OptimizationRemark(LV_NAME, "Vectorized", L->getStartLoc(),
7877 << "vectorized loop (vectorization width: "
7878 << NV("VectorizationFactor", VF.Width)
7879 << ", interleaved count: " << NV("InterleaveCount", IC) << ")");
7882 // Mark the loop as already vectorized to avoid vectorizing again.
7883 Hints.setAlreadyVectorized();
7885 DEBUG(verifyFunction(*L->getHeader()->getParent()));
7889 bool LoopVectorizePass::runImpl(
7890 Function &F, ScalarEvolution &SE_, LoopInfo &LI_, TargetTransformInfo &TTI_,
7891 DominatorTree &DT_, BlockFrequencyInfo &BFI_, TargetLibraryInfo *TLI_,
7892 DemandedBits &DB_, AliasAnalysis &AA_, AssumptionCache &AC_,
7893 std::function<const LoopAccessInfo &(Loop &)> &GetLAA_,
7894 OptimizationRemarkEmitter &ORE_) {
7908 // Compute some weights outside of the loop over the loops. Compute this
7909 // using a BranchProbability to re-use its scaling math.
7910 const BranchProbability ColdProb(1, 5); // 20%
7911 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
7914 // 1. the target claims to have no vector registers, and
7915 // 2. interleaving won't help ILP.
7917 // The second condition is necessary because, even if the target has no
7918 // vector registers, loop vectorization may still enable scalar
7920 if (!TTI->getNumberOfRegisters(true) && TTI->getMaxInterleaveFactor(1) < 2)
7923 bool Changed = false;
7925 // The vectorizer requires loops to be in simplified form.
7926 // Since simplification may add new inner loops, it has to run before the
7927 // legality and profitability checks. This means running the loop vectorizer
7928 // will simplify all loops, regardless of whether anything end up being
7931 Changed |= simplifyLoop(L, DT, LI, SE, AC, false /* PreserveLCSSA */);
7933 // Build up a worklist of inner-loops to vectorize. This is necessary as
7934 // the act of vectorizing or partially unrolling a loop creates new loops
7935 // and can invalidate iterators across the loops.
7936 SmallVector<Loop *, 8> Worklist;
7939 addAcyclicInnerLoop(*L, Worklist);
7941 LoopsAnalyzed += Worklist.size();
7943 // Now walk the identified inner loops.
7944 while (!Worklist.empty()) {
7945 Loop *L = Worklist.pop_back_val();
7947 // For the inner loops we actually process, form LCSSA to simplify the
7949 Changed |= formLCSSARecursively(*L, *DT, LI, SE);
7951 Changed |= processLoop(L);
7954 // Process each loop nest in the function.
7960 PreservedAnalyses LoopVectorizePass::run(Function &F,
7961 FunctionAnalysisManager &AM) {
7962 auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F);
7963 auto &LI = AM.getResult<LoopAnalysis>(F);
7964 auto &TTI = AM.getResult<TargetIRAnalysis>(F);
7965 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
7966 auto &BFI = AM.getResult<BlockFrequencyAnalysis>(F);
7967 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
7968 auto &AA = AM.getResult<AAManager>(F);
7969 auto &AC = AM.getResult<AssumptionAnalysis>(F);
7970 auto &DB = AM.getResult<DemandedBitsAnalysis>(F);
7971 auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
7973 auto &LAM = AM.getResult<LoopAnalysisManagerFunctionProxy>(F).getManager();
7974 std::function<const LoopAccessInfo &(Loop &)> GetLAA =
7975 [&](Loop &L) -> const LoopAccessInfo & {
7976 LoopStandardAnalysisResults AR = {AA, AC, DT, LI, SE, TLI, TTI};
7977 return LAM.getResult<LoopAccessAnalysis>(L, AR);
7980 runImpl(F, SE, LI, TTI, DT, BFI, &TLI, DB, AA, AC, GetLAA, ORE);
7982 return PreservedAnalyses::all();
7983 PreservedAnalyses PA;
7984 PA.preserve<LoopAnalysis>();
7985 PA.preserve<DominatorTreeAnalysis>();
7986 PA.preserve<BasicAA>();
7987 PA.preserve<GlobalsAA>();