1 //===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops
11 // and generates target-independent LLVM-IR.
12 // The vectorizer uses the TargetTransformInfo analysis to estimate the costs
13 // of instructions in order to estimate the profitability of vectorization.
15 // The loop vectorizer combines consecutive loop iterations into a single
16 // 'wide' iteration. After this transformation the index is incremented
17 // by the SIMD vector width, and not by one.
19 // This pass has three parts:
20 // 1. The main loop pass that drives the different parts.
21 // 2. LoopVectorizationLegality - A unit that checks for the legality
22 // of the vectorization.
23 // 3. InnerLoopVectorizer - A unit that performs the actual
24 // widening of instructions.
25 // 4. LoopVectorizationCostModel - A unit that checks for the profitability
26 // of vectorization. It decides on the optimal vector width, which
27 // can be one, if vectorization is not profitable.
29 //===----------------------------------------------------------------------===//
31 // The reduction-variable vectorization is based on the paper:
32 // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
34 // Variable uniformity checks are inspired by:
35 // Karrenberg, R. and Hack, S. Whole Function Vectorization.
37 // The interleaved access vectorization is based on the paper:
38 // Dorit Nuzman, Ira Rosen and Ayal Zaks. Auto-Vectorization of Interleaved
41 // Other ideas/concepts are from:
42 // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
44 // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of
45 // Vectorizing Compilers.
47 //===----------------------------------------------------------------------===//
49 #include "llvm/Transforms/Vectorize/LoopVectorize.h"
50 #include "llvm/ADT/DenseMap.h"
51 #include "llvm/ADT/Hashing.h"
52 #include "llvm/ADT/MapVector.h"
53 #include "llvm/ADT/SCCIterator.h"
54 #include "llvm/ADT/SetVector.h"
55 #include "llvm/ADT/SmallPtrSet.h"
56 #include "llvm/ADT/SmallSet.h"
57 #include "llvm/ADT/SmallVector.h"
58 #include "llvm/ADT/Statistic.h"
59 #include "llvm/ADT/StringExtras.h"
60 #include "llvm/Analysis/CodeMetrics.h"
61 #include "llvm/Analysis/GlobalsModRef.h"
62 #include "llvm/Analysis/LoopInfo.h"
63 #include "llvm/Analysis/LoopIterator.h"
64 #include "llvm/Analysis/LoopPass.h"
65 #include "llvm/Analysis/ScalarEvolutionExpander.h"
66 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
67 #include "llvm/Analysis/ValueTracking.h"
68 #include "llvm/Analysis/VectorUtils.h"
69 #include "llvm/IR/Constants.h"
70 #include "llvm/IR/DataLayout.h"
71 #include "llvm/IR/DebugInfo.h"
72 #include "llvm/IR/DerivedTypes.h"
73 #include "llvm/IR/DiagnosticInfo.h"
74 #include "llvm/IR/Dominators.h"
75 #include "llvm/IR/Function.h"
76 #include "llvm/IR/IRBuilder.h"
77 #include "llvm/IR/Instructions.h"
78 #include "llvm/IR/IntrinsicInst.h"
79 #include "llvm/IR/LLVMContext.h"
80 #include "llvm/IR/Module.h"
81 #include "llvm/IR/PatternMatch.h"
82 #include "llvm/IR/Type.h"
83 #include "llvm/IR/Value.h"
84 #include "llvm/IR/ValueHandle.h"
85 #include "llvm/IR/Verifier.h"
86 #include "llvm/Pass.h"
87 #include "llvm/Support/BranchProbability.h"
88 #include "llvm/Support/CommandLine.h"
89 #include "llvm/Support/Debug.h"
90 #include "llvm/Support/raw_ostream.h"
91 #include "llvm/Transforms/Scalar.h"
92 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
93 #include "llvm/Transforms/Utils/Local.h"
94 #include "llvm/Transforms/Utils/LoopUtils.h"
95 #include "llvm/Transforms/Utils/LoopVersioning.h"
96 #include "llvm/Transforms/Vectorize.h"
101 using namespace llvm;
102 using namespace llvm::PatternMatch;
104 #define LV_NAME "loop-vectorize"
105 #define DEBUG_TYPE LV_NAME
107 STATISTIC(LoopsVectorized, "Number of loops vectorized");
108 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
111 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
112 cl::desc("Enable if-conversion during vectorization."));
114 /// We don't vectorize loops with a known constant trip count below this number.
115 static cl::opt<unsigned> TinyTripCountVectorThreshold(
116 "vectorizer-min-trip-count", cl::init(16), cl::Hidden,
117 cl::desc("Don't vectorize loops with a constant "
118 "trip count that is smaller than this "
121 static cl::opt<bool> MaximizeBandwidth(
122 "vectorizer-maximize-bandwidth", cl::init(false), cl::Hidden,
123 cl::desc("Maximize bandwidth when selecting vectorization factor which "
124 "will be determined by the smallest type in loop."));
126 static cl::opt<bool> EnableInterleavedMemAccesses(
127 "enable-interleaved-mem-accesses", cl::init(false), cl::Hidden,
128 cl::desc("Enable vectorization on interleaved memory accesses in a loop"));
130 /// Maximum factor for an interleaved memory access.
131 static cl::opt<unsigned> MaxInterleaveGroupFactor(
132 "max-interleave-group-factor", cl::Hidden,
133 cl::desc("Maximum factor for an interleaved access group (default = 8)"),
136 /// We don't interleave loops with a known constant trip count below this
138 static const unsigned TinyTripCountInterleaveThreshold = 128;
140 static cl::opt<unsigned> ForceTargetNumScalarRegs(
141 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
142 cl::desc("A flag that overrides the target's number of scalar registers."));
144 static cl::opt<unsigned> ForceTargetNumVectorRegs(
145 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
146 cl::desc("A flag that overrides the target's number of vector registers."));
148 /// Maximum vectorization interleave count.
149 static const unsigned MaxInterleaveFactor = 16;
151 static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
152 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
153 cl::desc("A flag that overrides the target's max interleave factor for "
156 static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
157 "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
158 cl::desc("A flag that overrides the target's max interleave factor for "
159 "vectorized loops."));
161 static cl::opt<unsigned> ForceTargetInstructionCost(
162 "force-target-instruction-cost", cl::init(0), cl::Hidden,
163 cl::desc("A flag that overrides the target's expected cost for "
164 "an instruction to a single constant value. Mostly "
165 "useful for getting consistent testing."));
167 static cl::opt<unsigned> SmallLoopCost(
168 "small-loop-cost", cl::init(20), cl::Hidden,
170 "The cost of a loop that is considered 'small' by the interleaver."));
172 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
173 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
174 cl::desc("Enable the use of the block frequency analysis to access PGO "
175 "heuristics minimizing code growth in cold regions and being more "
176 "aggressive in hot regions."));
178 // Runtime interleave loops for load/store throughput.
179 static cl::opt<bool> EnableLoadStoreRuntimeInterleave(
180 "enable-loadstore-runtime-interleave", cl::init(true), cl::Hidden,
182 "Enable runtime interleaving until load/store ports are saturated"));
184 /// The number of stores in a loop that are allowed to need predication.
185 static cl::opt<unsigned> NumberOfStoresToPredicate(
186 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
187 cl::desc("Max number of stores to be predicated behind an if."));
189 static cl::opt<bool> EnableIndVarRegisterHeur(
190 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
191 cl::desc("Count the induction variable only once when interleaving"));
193 static cl::opt<bool> EnableCondStoresVectorization(
194 "enable-cond-stores-vec", cl::init(true), cl::Hidden,
195 cl::desc("Enable if predication of stores during vectorization."));
197 static cl::opt<unsigned> MaxNestedScalarReductionIC(
198 "max-nested-scalar-reduction-interleave", cl::init(2), cl::Hidden,
199 cl::desc("The maximum interleave count to use when interleaving a scalar "
200 "reduction in a nested loop."));
202 static cl::opt<unsigned> PragmaVectorizeMemoryCheckThreshold(
203 "pragma-vectorize-memory-check-threshold", cl::init(128), cl::Hidden,
204 cl::desc("The maximum allowed number of runtime memory checks with a "
205 "vectorize(enable) pragma."));
207 static cl::opt<unsigned> VectorizeSCEVCheckThreshold(
208 "vectorize-scev-check-threshold", cl::init(16), cl::Hidden,
209 cl::desc("The maximum number of SCEV checks allowed."));
211 static cl::opt<unsigned> PragmaVectorizeSCEVCheckThreshold(
212 "pragma-vectorize-scev-check-threshold", cl::init(128), cl::Hidden,
213 cl::desc("The maximum number of SCEV checks allowed with a "
214 "vectorize(enable) pragma"));
216 /// Create an analysis remark that explains why vectorization failed
218 /// \p PassName is the name of the pass (e.g. can be AlwaysPrint). \p
219 /// RemarkName is the identifier for the remark. If \p I is passed it is an
220 /// instruction that prevents vectorization. Otherwise \p TheLoop is used for
221 /// the location of the remark. \return the remark object that can be
223 static OptimizationRemarkAnalysis
224 createMissedAnalysis(const char *PassName, StringRef RemarkName, Loop *TheLoop,
225 Instruction *I = nullptr) {
226 Value *CodeRegion = TheLoop->getHeader();
227 DebugLoc DL = TheLoop->getStartLoc();
230 CodeRegion = I->getParent();
231 // If there is no debug location attached to the instruction, revert back to
233 if (I->getDebugLoc())
234 DL = I->getDebugLoc();
237 OptimizationRemarkAnalysis R(PassName, RemarkName, DL, CodeRegion);
238 R << "loop not vectorized: ";
244 // Forward declarations.
245 class LoopVectorizeHints;
246 class LoopVectorizationLegality;
247 class LoopVectorizationCostModel;
248 class LoopVectorizationRequirements;
250 /// Returns true if the given loop body has a cycle, excluding the loop
252 static bool hasCyclesInLoopBody(const Loop &L) {
256 for (const auto &SCC :
257 make_range(scc_iterator<Loop, LoopBodyTraits>::begin(L),
258 scc_iterator<Loop, LoopBodyTraits>::end(L))) {
259 if (SCC.size() > 1) {
260 DEBUG(dbgs() << "LVL: Detected a cycle in the loop body:\n");
268 /// \brief This modifies LoopAccessReport to initialize message with
269 /// loop-vectorizer-specific part.
270 class VectorizationReport : public LoopAccessReport {
272 VectorizationReport(Instruction *I = nullptr)
273 : LoopAccessReport("loop not vectorized: ", I) {}
275 /// \brief This allows promotion of the loop-access analysis report into the
276 /// loop-vectorizer report. It modifies the message to add the
277 /// loop-vectorizer-specific part of the message.
278 explicit VectorizationReport(const LoopAccessReport &R)
279 : LoopAccessReport(Twine("loop not vectorized: ") + R.str(),
283 /// A helper function for converting Scalar types to vector types.
284 /// If the incoming type is void, we return void. If the VF is 1, we return
286 static Type *ToVectorTy(Type *Scalar, unsigned VF) {
287 if (Scalar->isVoidTy() || VF == 1)
289 return VectorType::get(Scalar, VF);
292 /// A helper function that returns GEP instruction and knows to skip a
293 /// 'bitcast'. The 'bitcast' may be skipped if the source and the destination
294 /// pointee types of the 'bitcast' have the same size.
296 /// bitcast double** %var to i64* - can be skipped
297 /// bitcast double** %var to i8* - can not
298 static GetElementPtrInst *getGEPInstruction(Value *Ptr) {
300 if (isa<GetElementPtrInst>(Ptr))
301 return cast<GetElementPtrInst>(Ptr);
303 if (isa<BitCastInst>(Ptr) &&
304 isa<GetElementPtrInst>(cast<BitCastInst>(Ptr)->getOperand(0))) {
305 Type *BitcastTy = Ptr->getType();
306 Type *GEPTy = cast<BitCastInst>(Ptr)->getSrcTy();
307 if (!isa<PointerType>(BitcastTy) || !isa<PointerType>(GEPTy))
309 Type *Pointee1Ty = cast<PointerType>(BitcastTy)->getPointerElementType();
310 Type *Pointee2Ty = cast<PointerType>(GEPTy)->getPointerElementType();
311 const DataLayout &DL = cast<BitCastInst>(Ptr)->getModule()->getDataLayout();
312 if (DL.getTypeSizeInBits(Pointee1Ty) == DL.getTypeSizeInBits(Pointee2Ty))
313 return cast<GetElementPtrInst>(cast<BitCastInst>(Ptr)->getOperand(0));
318 /// A helper function that returns the pointer operand of a load or store
320 static Value *getPointerOperand(Value *I) {
321 if (auto *LI = dyn_cast<LoadInst>(I))
322 return LI->getPointerOperand();
323 if (auto *SI = dyn_cast<StoreInst>(I))
324 return SI->getPointerOperand();
328 /// A helper function that returns true if the given type is irregular. The
329 /// type is irregular if its allocated size doesn't equal the store size of an
330 /// element of the corresponding vector type at the given vectorization factor.
331 static bool hasIrregularType(Type *Ty, const DataLayout &DL, unsigned VF) {
333 // Determine if an array of VF elements of type Ty is "bitcast compatible"
334 // with a <VF x Ty> vector.
336 auto *VectorTy = VectorType::get(Ty, VF);
337 return VF * DL.getTypeAllocSize(Ty) != DL.getTypeStoreSize(VectorTy);
340 // If the vectorization factor is one, we just check if an array of type Ty
341 // requires padding between elements.
342 return DL.getTypeAllocSizeInBits(Ty) != DL.getTypeSizeInBits(Ty);
345 /// A helper function that returns the reciprocal of the block probability of
346 /// predicated blocks. If we return X, we are assuming the predicated block
347 /// will execute once for for every X iterations of the loop header.
349 /// TODO: We should use actual block probability here, if available. Currently,
350 /// we always assume predicated blocks have a 50% chance of executing.
351 static unsigned getReciprocalPredBlockProb() { return 2; }
353 /// InnerLoopVectorizer vectorizes loops which contain only one basic
354 /// block to a specified vectorization factor (VF).
355 /// This class performs the widening of scalars into vectors, or multiple
356 /// scalars. This class also implements the following features:
357 /// * It inserts an epilogue loop for handling loops that don't have iteration
358 /// counts that are known to be a multiple of the vectorization factor.
359 /// * It handles the code generation for reduction variables.
360 /// * Scalarization (implementation using scalars) of un-vectorizable
362 /// InnerLoopVectorizer does not perform any vectorization-legality
363 /// checks, and relies on the caller to check for the different legality
364 /// aspects. The InnerLoopVectorizer relies on the
365 /// LoopVectorizationLegality class to provide information about the induction
366 /// and reduction variables that were found to a given vectorization factor.
367 class InnerLoopVectorizer {
369 InnerLoopVectorizer(Loop *OrigLoop, PredicatedScalarEvolution &PSE,
370 LoopInfo *LI, DominatorTree *DT,
371 const TargetLibraryInfo *TLI,
372 const TargetTransformInfo *TTI, AssumptionCache *AC,
373 OptimizationRemarkEmitter *ORE, unsigned VecWidth,
374 unsigned UnrollFactor, LoopVectorizationLegality *LVL,
375 LoopVectorizationCostModel *CM)
376 : OrigLoop(OrigLoop), PSE(PSE), LI(LI), DT(DT), TLI(TLI), TTI(TTI),
377 AC(AC), ORE(ORE), VF(VecWidth), UF(UnrollFactor),
378 Builder(PSE.getSE()->getContext()), Induction(nullptr),
379 OldInduction(nullptr), VectorLoopValueMap(UnrollFactor, VecWidth),
380 TripCount(nullptr), VectorTripCount(nullptr), Legal(LVL), Cost(CM),
381 AddedSafetyChecks(false) {}
383 // Perform the actual loop widening (vectorization).
385 // Create a new empty loop. Unlink the old loop and connect the new one.
387 // Widen each instruction in the old loop to a new one in the new loop.
391 // Return true if any runtime check is added.
392 bool areSafetyChecksAdded() { return AddedSafetyChecks; }
394 virtual ~InnerLoopVectorizer() {}
397 /// A small list of PHINodes.
398 typedef SmallVector<PHINode *, 4> PhiVector;
400 /// A type for vectorized values in the new loop. Each value from the
401 /// original loop, when vectorized, is represented by UF vector values in the
402 /// new unrolled loop, where UF is the unroll factor.
403 typedef SmallVector<Value *, 2> VectorParts;
405 /// A type for scalarized values in the new loop. Each value from the
406 /// original loop, when scalarized, is represented by UF x VF scalar values
407 /// in the new unrolled loop, where UF is the unroll factor and VF is the
408 /// vectorization factor.
409 typedef SmallVector<SmallVector<Value *, 4>, 2> ScalarParts;
411 // When we if-convert we need to create edge masks. We have to cache values
412 // so that we don't end up with exponential recursion/IR.
413 typedef DenseMap<std::pair<BasicBlock *, BasicBlock *>, VectorParts>
416 /// Create an empty loop, based on the loop ranges of the old loop.
417 void createEmptyLoop();
419 /// Set up the values of the IVs correctly when exiting the vector loop.
420 void fixupIVUsers(PHINode *OrigPhi, const InductionDescriptor &II,
421 Value *CountRoundDown, Value *EndValue,
422 BasicBlock *MiddleBlock);
424 /// Create a new induction variable inside L.
425 PHINode *createInductionVariable(Loop *L, Value *Start, Value *End,
426 Value *Step, Instruction *DL);
427 /// Copy and widen the instructions from the old loop.
428 virtual void vectorizeLoop();
430 /// Fix a first-order recurrence. This is the second phase of vectorizing
432 void fixFirstOrderRecurrence(PHINode *Phi);
434 /// \brief The Loop exit block may have single value PHI nodes where the
435 /// incoming value is 'Undef'. While vectorizing we only handled real values
436 /// that were defined inside the loop. Here we fix the 'undef case'.
440 /// Iteratively sink the scalarized operands of a predicated instruction into
441 /// the block that was created for it.
442 void sinkScalarOperands(Instruction *PredInst);
444 /// Predicate conditional instructions that require predication on their
445 /// respective conditions.
446 void predicateInstructions();
448 /// Collect the instructions from the original loop that would be trivially
449 /// dead in the vectorized loop if generated.
450 void collectTriviallyDeadInstructions();
452 /// Shrinks vector element sizes to the smallest bitwidth they can be legally
454 void truncateToMinimalBitwidths();
456 /// A helper function that computes the predicate of the block BB, assuming
457 /// that the header block of the loop is set to True. It returns the *entry*
458 /// mask for the block BB.
459 VectorParts createBlockInMask(BasicBlock *BB);
460 /// A helper function that computes the predicate of the edge between SRC
462 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
464 /// A helper function to vectorize a single BB within the innermost loop.
465 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
467 /// Vectorize a single PHINode in a block. This method handles the induction
468 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
469 /// arbitrary length vectors.
470 void widenPHIInstruction(Instruction *PN, unsigned UF, unsigned VF,
473 /// Insert the new loop to the loop hierarchy and pass manager
474 /// and update the analysis passes.
475 void updateAnalysis();
477 /// This instruction is un-vectorizable. Implement it as a sequence
478 /// of scalars. If \p IfPredicateInstr is true we need to 'hide' each
479 /// scalarized instruction behind an if block predicated on the control
480 /// dependence of the instruction.
481 virtual void scalarizeInstruction(Instruction *Instr,
482 bool IfPredicateInstr = false);
484 /// Vectorize Load and Store instructions,
485 virtual void vectorizeMemoryInstruction(Instruction *Instr);
487 /// Create a broadcast instruction. This method generates a broadcast
488 /// instruction (shuffle) for loop invariant values and for the induction
489 /// value. If this is the induction variable then we extend it to N, N+1, ...
490 /// this is needed because each iteration in the loop corresponds to a SIMD
492 virtual Value *getBroadcastInstrs(Value *V);
494 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
495 /// to each vector element of Val. The sequence starts at StartIndex.
496 /// \p Opcode is relevant for FP induction variable.
497 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step,
498 Instruction::BinaryOps Opcode =
499 Instruction::BinaryOpsEnd);
501 /// Compute scalar induction steps. \p ScalarIV is the scalar induction
502 /// variable on which to base the steps, \p Step is the size of the step, and
503 /// \p EntryVal is the value from the original loop that maps to the steps.
504 /// Note that \p EntryVal doesn't have to be an induction variable (e.g., it
505 /// can be a truncate instruction).
506 void buildScalarSteps(Value *ScalarIV, Value *Step, Value *EntryVal);
508 /// Create a vector induction phi node based on an existing scalar one. This
509 /// currently only works for integer induction variables with a constant
510 /// step. \p EntryVal is the value from the original loop that maps to the
511 /// vector phi node. If \p EntryVal is a truncate instruction, instead of
512 /// widening the original IV, we widen a version of the IV truncated to \p
514 void createVectorIntInductionPHI(const InductionDescriptor &II,
515 Instruction *EntryVal);
517 /// Widen an integer induction variable \p IV. If \p Trunc is provided, the
518 /// induction variable will first be truncated to the corresponding type.
519 void widenIntInduction(PHINode *IV, TruncInst *Trunc = nullptr);
521 /// Returns true if an instruction \p I should be scalarized instead of
522 /// vectorized for the chosen vectorization factor.
523 bool shouldScalarizeInstruction(Instruction *I) const;
525 /// Returns true if we should generate a scalar version of \p IV.
526 bool needsScalarInduction(Instruction *IV) const;
528 /// Return a constant reference to the VectorParts corresponding to \p V from
529 /// the original loop. If the value has already been vectorized, the
530 /// corresponding vector entry in VectorLoopValueMap is returned. If,
531 /// however, the value has a scalar entry in VectorLoopValueMap, we construct
532 /// new vector values on-demand by inserting the scalar values into vectors
533 /// with an insertelement sequence. If the value has been neither vectorized
534 /// nor scalarized, it must be loop invariant, so we simply broadcast the
535 /// value into vectors.
536 const VectorParts &getVectorValue(Value *V);
538 /// Return a value in the new loop corresponding to \p V from the original
539 /// loop at unroll index \p Part and vector index \p Lane. If the value has
540 /// been vectorized but not scalarized, the necessary extractelement
541 /// instruction will be generated.
542 Value *getScalarValue(Value *V, unsigned Part, unsigned Lane);
544 /// Try to vectorize the interleaved access group that \p Instr belongs to.
545 void vectorizeInterleaveGroup(Instruction *Instr);
547 /// Generate a shuffle sequence that will reverse the vector Vec.
548 virtual Value *reverseVector(Value *Vec);
550 /// Returns (and creates if needed) the original loop trip count.
551 Value *getOrCreateTripCount(Loop *NewLoop);
553 /// Returns (and creates if needed) the trip count of the widened loop.
554 Value *getOrCreateVectorTripCount(Loop *NewLoop);
556 /// Emit a bypass check to see if the trip count would overflow, or we
557 /// wouldn't have enough iterations to execute one vector loop.
558 void emitMinimumIterationCountCheck(Loop *L, BasicBlock *Bypass);
559 /// Emit a bypass check to see if the vector trip count is nonzero.
560 void emitVectorLoopEnteredCheck(Loop *L, BasicBlock *Bypass);
561 /// Emit a bypass check to see if all of the SCEV assumptions we've
562 /// had to make are correct.
563 void emitSCEVChecks(Loop *L, BasicBlock *Bypass);
564 /// Emit bypass checks to check any memory assumptions we may have made.
565 void emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass);
567 /// Add additional metadata to \p To that was not present on \p Orig.
569 /// Currently this is used to add the noalias annotations based on the
570 /// inserted memchecks. Use this for instructions that are *cloned* into the
572 void addNewMetadata(Instruction *To, const Instruction *Orig);
574 /// Add metadata from one instruction to another.
576 /// This includes both the original MDs from \p From and additional ones (\see
577 /// addNewMetadata). Use this for *newly created* instructions in the vector
579 void addMetadata(Instruction *To, Instruction *From);
581 /// \brief Similar to the previous function but it adds the metadata to a
582 /// vector of instructions.
583 void addMetadata(ArrayRef<Value *> To, Instruction *From);
585 /// This is a helper class for maintaining vectorization state. It's used for
586 /// mapping values from the original loop to their corresponding values in
587 /// the new loop. Two mappings are maintained: one for vectorized values and
588 /// one for scalarized values. Vectorized values are represented with UF
589 /// vector values in the new loop, and scalarized values are represented with
590 /// UF x VF scalar values in the new loop. UF and VF are the unroll and
591 /// vectorization factors, respectively.
593 /// Entries can be added to either map with initVector and initScalar, which
594 /// initialize and return a constant reference to the new entry. If a
595 /// non-constant reference to a vector entry is required, getVector can be
596 /// used to retrieve a mutable entry. We currently directly modify the mapped
597 /// values during "fix-up" operations that occur once the first phase of
598 /// widening is complete. These operations include type truncation and the
599 /// second phase of recurrence widening.
601 /// Otherwise, entries from either map should be accessed using the
602 /// getVectorValue or getScalarValue functions from InnerLoopVectorizer.
603 /// getVectorValue and getScalarValue coordinate to generate a vector or
604 /// scalar value on-demand if one is not yet available. When vectorizing a
605 /// loop, we visit the definition of an instruction before its uses. When
606 /// visiting the definition, we either vectorize or scalarize the
607 /// instruction, creating an entry for it in the corresponding map. (In some
608 /// cases, such as induction variables, we will create both vector and scalar
609 /// entries.) Then, as we encounter uses of the definition, we derive values
610 /// for each scalar or vector use unless such a value is already available.
611 /// For example, if we scalarize a definition and one of its uses is vector,
612 /// we build the required vector on-demand with an insertelement sequence
613 /// when visiting the use. Otherwise, if the use is scalar, we can use the
614 /// existing scalar definition.
617 /// Construct an empty map with the given unroll and vectorization factors.
618 ValueMap(unsigned UnrollFactor, unsigned VecWidth)
619 : UF(UnrollFactor), VF(VecWidth) {
620 // The unroll and vectorization factors are only used in asserts builds
621 // to verify map entries are sized appropriately.
626 /// \return True if the map has a vector entry for \p Key.
627 bool hasVector(Value *Key) const { return VectorMapStorage.count(Key); }
629 /// \return True if the map has a scalar entry for \p Key.
630 bool hasScalar(Value *Key) const { return ScalarMapStorage.count(Key); }
632 /// \brief Map \p Key to the given VectorParts \p Entry, and return a
633 /// constant reference to the new vector map entry. The given key should
634 /// not already be in the map, and the given VectorParts should be
635 /// correctly sized for the current unroll factor.
636 const VectorParts &initVector(Value *Key, const VectorParts &Entry) {
637 assert(!hasVector(Key) && "Vector entry already initialized");
638 assert(Entry.size() == UF && "VectorParts has wrong dimensions");
639 VectorMapStorage[Key] = Entry;
640 return VectorMapStorage[Key];
643 /// \brief Map \p Key to the given ScalarParts \p Entry, and return a
644 /// constant reference to the new scalar map entry. The given key should
645 /// not already be in the map, and the given ScalarParts should be
646 /// correctly sized for the current unroll and vectorization factors.
647 const ScalarParts &initScalar(Value *Key, const ScalarParts &Entry) {
648 assert(!hasScalar(Key) && "Scalar entry already initialized");
649 assert(Entry.size() == UF &&
650 all_of(make_range(Entry.begin(), Entry.end()),
651 [&](const SmallVectorImpl<Value *> &Values) -> bool {
652 return Values.size() == VF;
654 "ScalarParts has wrong dimensions");
655 ScalarMapStorage[Key] = Entry;
656 return ScalarMapStorage[Key];
659 /// \return A reference to the vector map entry corresponding to \p Key.
660 /// The key should already be in the map. This function should only be used
661 /// when it's necessary to update values that have already been vectorized.
662 /// This is the case for "fix-up" operations including type truncation and
663 /// the second phase of recurrence vectorization. If a non-const reference
664 /// isn't required, getVectorValue should be used instead.
665 VectorParts &getVector(Value *Key) {
666 assert(hasVector(Key) && "Vector entry not initialized");
667 return VectorMapStorage.find(Key)->second;
670 /// Retrieve an entry from the vector or scalar maps. The preferred way to
671 /// access an existing mapped entry is with getVectorValue or
672 /// getScalarValue from InnerLoopVectorizer. Until those functions can be
673 /// moved inside ValueMap, we have to declare them as friends.
674 friend const VectorParts &InnerLoopVectorizer::getVectorValue(Value *V);
675 friend Value *InnerLoopVectorizer::getScalarValue(Value *V, unsigned Part,
679 /// The unroll factor. Each entry in the vector map contains UF vector
683 /// The vectorization factor. Each entry in the scalar map contains UF x VF
687 /// The vector and scalar map storage. We use std::map and not DenseMap
688 /// because insertions to DenseMap invalidate its iterators.
689 std::map<Value *, VectorParts> VectorMapStorage;
690 std::map<Value *, ScalarParts> ScalarMapStorage;
693 /// The original loop.
695 /// A wrapper around ScalarEvolution used to add runtime SCEV checks. Applies
696 /// dynamic knowledge to simplify SCEV expressions and converts them to a
697 /// more usable form.
698 PredicatedScalarEvolution &PSE;
705 /// Target Library Info.
706 const TargetLibraryInfo *TLI;
707 /// Target Transform Info.
708 const TargetTransformInfo *TTI;
709 /// Assumption Cache.
711 /// Interface to emit optimization remarks.
712 OptimizationRemarkEmitter *ORE;
714 /// \brief LoopVersioning. It's only set up (non-null) if memchecks were
717 /// This is currently only used to add no-alias metadata based on the
718 /// memchecks. The actually versioning is performed manually.
719 std::unique_ptr<LoopVersioning> LVer;
721 /// The vectorization SIMD factor to use. Each vector will have this many
726 /// The vectorization unroll factor to use. Each scalar is vectorized to this
727 /// many different vector instructions.
730 /// The builder that we use
733 // --- Vectorization state ---
735 /// The vector-loop preheader.
736 BasicBlock *LoopVectorPreHeader;
737 /// The scalar-loop preheader.
738 BasicBlock *LoopScalarPreHeader;
739 /// Middle Block between the vector and the scalar.
740 BasicBlock *LoopMiddleBlock;
741 /// The ExitBlock of the scalar loop.
742 BasicBlock *LoopExitBlock;
743 /// The vector loop body.
744 BasicBlock *LoopVectorBody;
745 /// The scalar loop body.
746 BasicBlock *LoopScalarBody;
747 /// A list of all bypass blocks. The first block is the entry of the loop.
748 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
750 /// The new Induction variable which was added to the new block.
752 /// The induction variable of the old basic block.
753 PHINode *OldInduction;
755 /// Maps values from the original loop to their corresponding values in the
756 /// vectorized loop. A key value can map to either vector values, scalar
757 /// values or both kinds of values, depending on whether the key was
758 /// vectorized and scalarized.
759 ValueMap VectorLoopValueMap;
761 /// Store instructions that should be predicated, as a pair
762 /// <StoreInst, Predicate>
763 SmallVector<std::pair<Instruction *, Value *>, 4> PredicatedInstructions;
764 EdgeMaskCache MaskCache;
765 /// Trip count of the original loop.
767 /// Trip count of the widened loop (TripCount - TripCount % (VF*UF))
768 Value *VectorTripCount;
770 /// The legality analysis.
771 LoopVectorizationLegality *Legal;
773 /// The profitablity analysis.
774 LoopVectorizationCostModel *Cost;
776 // Record whether runtime checks are added.
777 bool AddedSafetyChecks;
779 // Holds instructions from the original loop whose counterparts in the
780 // vectorized loop would be trivially dead if generated. For example,
781 // original induction update instructions can become dead because we
782 // separately emit induction "steps" when generating code for the new loop.
783 // Similarly, we create a new latch condition when setting up the structure
784 // of the new loop, so the old one can become dead.
785 SmallPtrSet<Instruction *, 4> DeadInstructions;
788 class InnerLoopUnroller : public InnerLoopVectorizer {
790 InnerLoopUnroller(Loop *OrigLoop, PredicatedScalarEvolution &PSE,
791 LoopInfo *LI, DominatorTree *DT,
792 const TargetLibraryInfo *TLI,
793 const TargetTransformInfo *TTI, AssumptionCache *AC,
794 OptimizationRemarkEmitter *ORE, unsigned UnrollFactor,
795 LoopVectorizationLegality *LVL,
796 LoopVectorizationCostModel *CM)
797 : InnerLoopVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, 1,
798 UnrollFactor, LVL, CM) {}
801 void scalarizeInstruction(Instruction *Instr,
802 bool IfPredicateInstr = false) override;
803 void vectorizeMemoryInstruction(Instruction *Instr) override;
804 Value *getBroadcastInstrs(Value *V) override;
805 Value *getStepVector(Value *Val, int StartIdx, Value *Step,
806 Instruction::BinaryOps Opcode =
807 Instruction::BinaryOpsEnd) override;
808 Value *reverseVector(Value *Vec) override;
811 /// \brief Look for a meaningful debug location on the instruction or it's
813 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
818 if (I->getDebugLoc() != Empty)
821 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
822 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
823 if (OpInst->getDebugLoc() != Empty)
830 /// \brief Set the debug location in the builder using the debug location in the
832 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
833 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
834 B.SetCurrentDebugLocation(Inst->getDebugLoc());
836 B.SetCurrentDebugLocation(DebugLoc());
840 /// \return string containing a file name and a line # for the given loop.
841 static std::string getDebugLocString(const Loop *L) {
844 raw_string_ostream OS(Result);
845 if (const DebugLoc LoopDbgLoc = L->getStartLoc())
846 LoopDbgLoc.print(OS);
848 // Just print the module name.
849 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
856 void InnerLoopVectorizer::addNewMetadata(Instruction *To,
857 const Instruction *Orig) {
858 // If the loop was versioned with memchecks, add the corresponding no-alias
860 if (LVer && (isa<LoadInst>(Orig) || isa<StoreInst>(Orig)))
861 LVer->annotateInstWithNoAlias(To, Orig);
864 void InnerLoopVectorizer::addMetadata(Instruction *To,
866 propagateMetadata(To, From);
867 addNewMetadata(To, From);
870 void InnerLoopVectorizer::addMetadata(ArrayRef<Value *> To,
872 for (Value *V : To) {
873 if (Instruction *I = dyn_cast<Instruction>(V))
874 addMetadata(I, From);
878 /// \brief The group of interleaved loads/stores sharing the same stride and
879 /// close to each other.
881 /// Each member in this group has an index starting from 0, and the largest
882 /// index should be less than interleaved factor, which is equal to the absolute
883 /// value of the access's stride.
885 /// E.g. An interleaved load group of factor 4:
886 /// for (unsigned i = 0; i < 1024; i+=4) {
887 /// a = A[i]; // Member of index 0
888 /// b = A[i+1]; // Member of index 1
889 /// d = A[i+3]; // Member of index 3
893 /// An interleaved store group of factor 4:
894 /// for (unsigned i = 0; i < 1024; i+=4) {
896 /// A[i] = a; // Member of index 0
897 /// A[i+1] = b; // Member of index 1
898 /// A[i+2] = c; // Member of index 2
899 /// A[i+3] = d; // Member of index 3
902 /// Note: the interleaved load group could have gaps (missing members), but
903 /// the interleaved store group doesn't allow gaps.
904 class InterleaveGroup {
906 InterleaveGroup(Instruction *Instr, int Stride, unsigned Align)
907 : Align(Align), SmallestKey(0), LargestKey(0), InsertPos(Instr) {
908 assert(Align && "The alignment should be non-zero");
910 Factor = std::abs(Stride);
911 assert(Factor > 1 && "Invalid interleave factor");
913 Reverse = Stride < 0;
917 bool isReverse() const { return Reverse; }
918 unsigned getFactor() const { return Factor; }
919 unsigned getAlignment() const { return Align; }
920 unsigned getNumMembers() const { return Members.size(); }
922 /// \brief Try to insert a new member \p Instr with index \p Index and
923 /// alignment \p NewAlign. The index is related to the leader and it could be
924 /// negative if it is the new leader.
926 /// \returns false if the instruction doesn't belong to the group.
927 bool insertMember(Instruction *Instr, int Index, unsigned NewAlign) {
928 assert(NewAlign && "The new member's alignment should be non-zero");
930 int Key = Index + SmallestKey;
932 // Skip if there is already a member with the same index.
933 if (Members.count(Key))
936 if (Key > LargestKey) {
937 // The largest index is always less than the interleave factor.
938 if (Index >= static_cast<int>(Factor))
942 } else if (Key < SmallestKey) {
943 // The largest index is always less than the interleave factor.
944 if (LargestKey - Key >= static_cast<int>(Factor))
950 // It's always safe to select the minimum alignment.
951 Align = std::min(Align, NewAlign);
952 Members[Key] = Instr;
956 /// \brief Get the member with the given index \p Index
958 /// \returns nullptr if contains no such member.
959 Instruction *getMember(unsigned Index) const {
960 int Key = SmallestKey + Index;
961 if (!Members.count(Key))
964 return Members.find(Key)->second;
967 /// \brief Get the index for the given member. Unlike the key in the member
968 /// map, the index starts from 0.
969 unsigned getIndex(Instruction *Instr) const {
970 for (auto I : Members)
971 if (I.second == Instr)
972 return I.first - SmallestKey;
974 llvm_unreachable("InterleaveGroup contains no such member");
977 Instruction *getInsertPos() const { return InsertPos; }
978 void setInsertPos(Instruction *Inst) { InsertPos = Inst; }
981 unsigned Factor; // Interleave Factor.
984 DenseMap<int, Instruction *> Members;
988 // To avoid breaking dependences, vectorized instructions of an interleave
989 // group should be inserted at either the first load or the last store in
992 // E.g. %even = load i32 // Insert Position
993 // %add = add i32 %even // Use of %even
997 // %odd = add i32 // Def of %odd
998 // store i32 %odd // Insert Position
999 Instruction *InsertPos;
1002 /// \brief Drive the analysis of interleaved memory accesses in the loop.
1004 /// Use this class to analyze interleaved accesses only when we can vectorize
1005 /// a loop. Otherwise it's meaningless to do analysis as the vectorization
1006 /// on interleaved accesses is unsafe.
1008 /// The analysis collects interleave groups and records the relationships
1009 /// between the member and the group in a map.
1010 class InterleavedAccessInfo {
1012 InterleavedAccessInfo(PredicatedScalarEvolution &PSE, Loop *L,
1013 DominatorTree *DT, LoopInfo *LI)
1014 : PSE(PSE), TheLoop(L), DT(DT), LI(LI), LAI(nullptr),
1015 RequiresScalarEpilogue(false) {}
1017 ~InterleavedAccessInfo() {
1018 SmallSet<InterleaveGroup *, 4> DelSet;
1019 // Avoid releasing a pointer twice.
1020 for (auto &I : InterleaveGroupMap)
1021 DelSet.insert(I.second);
1022 for (auto *Ptr : DelSet)
1026 /// \brief Analyze the interleaved accesses and collect them in interleave
1027 /// groups. Substitute symbolic strides using \p Strides.
1028 void analyzeInterleaving(const ValueToValueMap &Strides);
1030 /// \brief Check if \p Instr belongs to any interleave group.
1031 bool isInterleaved(Instruction *Instr) const {
1032 return InterleaveGroupMap.count(Instr);
1035 /// \brief Return the maximum interleave factor of all interleaved groups.
1036 unsigned getMaxInterleaveFactor() const {
1037 unsigned MaxFactor = 1;
1038 for (auto &Entry : InterleaveGroupMap)
1039 MaxFactor = std::max(MaxFactor, Entry.second->getFactor());
1043 /// \brief Get the interleave group that \p Instr belongs to.
1045 /// \returns nullptr if doesn't have such group.
1046 InterleaveGroup *getInterleaveGroup(Instruction *Instr) const {
1047 if (InterleaveGroupMap.count(Instr))
1048 return InterleaveGroupMap.find(Instr)->second;
1052 /// \brief Returns true if an interleaved group that may access memory
1053 /// out-of-bounds requires a scalar epilogue iteration for correctness.
1054 bool requiresScalarEpilogue() const { return RequiresScalarEpilogue; }
1056 /// \brief Initialize the LoopAccessInfo used for dependence checking.
1057 void setLAI(const LoopAccessInfo *Info) { LAI = Info; }
1060 /// A wrapper around ScalarEvolution, used to add runtime SCEV checks.
1061 /// Simplifies SCEV expressions in the context of existing SCEV assumptions.
1062 /// The interleaved access analysis can also add new predicates (for example
1063 /// by versioning strides of pointers).
1064 PredicatedScalarEvolution &PSE;
1068 const LoopAccessInfo *LAI;
1070 /// True if the loop may contain non-reversed interleaved groups with
1071 /// out-of-bounds accesses. We ensure we don't speculatively access memory
1072 /// out-of-bounds by executing at least one scalar epilogue iteration.
1073 bool RequiresScalarEpilogue;
1075 /// Holds the relationships between the members and the interleave group.
1076 DenseMap<Instruction *, InterleaveGroup *> InterleaveGroupMap;
1078 /// Holds dependences among the memory accesses in the loop. It maps a source
1079 /// access to a set of dependent sink accesses.
1080 DenseMap<Instruction *, SmallPtrSet<Instruction *, 2>> Dependences;
1082 /// \brief The descriptor for a strided memory access.
1083 struct StrideDescriptor {
1084 StrideDescriptor(int64_t Stride, const SCEV *Scev, uint64_t Size,
1086 : Stride(Stride), Scev(Scev), Size(Size), Align(Align) {}
1088 StrideDescriptor() = default;
1090 // The access's stride. It is negative for a reverse access.
1092 const SCEV *Scev = nullptr; // The scalar expression of this access
1093 uint64_t Size = 0; // The size of the memory object.
1094 unsigned Align = 0; // The alignment of this access.
1097 /// \brief A type for holding instructions and their stride descriptors.
1098 typedef std::pair<Instruction *, StrideDescriptor> StrideEntry;
1100 /// \brief Create a new interleave group with the given instruction \p Instr,
1101 /// stride \p Stride and alignment \p Align.
1103 /// \returns the newly created interleave group.
1104 InterleaveGroup *createInterleaveGroup(Instruction *Instr, int Stride,
1106 assert(!InterleaveGroupMap.count(Instr) &&
1107 "Already in an interleaved access group");
1108 InterleaveGroupMap[Instr] = new InterleaveGroup(Instr, Stride, Align);
1109 return InterleaveGroupMap[Instr];
1112 /// \brief Release the group and remove all the relationships.
1113 void releaseGroup(InterleaveGroup *Group) {
1114 for (unsigned i = 0; i < Group->getFactor(); i++)
1115 if (Instruction *Member = Group->getMember(i))
1116 InterleaveGroupMap.erase(Member);
1121 /// \brief Collect all the accesses with a constant stride in program order.
1122 void collectConstStrideAccesses(
1123 MapVector<Instruction *, StrideDescriptor> &AccessStrideInfo,
1124 const ValueToValueMap &Strides);
1126 /// \brief Returns true if \p Stride is allowed in an interleaved group.
1127 static bool isStrided(int Stride) {
1128 unsigned Factor = std::abs(Stride);
1129 return Factor >= 2 && Factor <= MaxInterleaveGroupFactor;
1132 /// \brief Returns true if \p BB is a predicated block.
1133 bool isPredicated(BasicBlock *BB) const {
1134 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
1137 /// \brief Returns true if LoopAccessInfo can be used for dependence queries.
1138 bool areDependencesValid() const {
1139 return LAI && LAI->getDepChecker().getDependences();
1142 /// \brief Returns true if memory accesses \p A and \p B can be reordered, if
1143 /// necessary, when constructing interleaved groups.
1145 /// \p A must precede \p B in program order. We return false if reordering is
1146 /// not necessary or is prevented because \p A and \p B may be dependent.
1147 bool canReorderMemAccessesForInterleavedGroups(StrideEntry *A,
1148 StrideEntry *B) const {
1150 // Code motion for interleaved accesses can potentially hoist strided loads
1151 // and sink strided stores. The code below checks the legality of the
1152 // following two conditions:
1154 // 1. Potentially moving a strided load (B) before any store (A) that
1157 // 2. Potentially moving a strided store (A) after any load or store (B)
1160 // It's legal to reorder A and B if we know there isn't a dependence from A
1161 // to B. Note that this determination is conservative since some
1162 // dependences could potentially be reordered safely.
1164 // A is potentially the source of a dependence.
1165 auto *Src = A->first;
1166 auto SrcDes = A->second;
1168 // B is potentially the sink of a dependence.
1169 auto *Sink = B->first;
1170 auto SinkDes = B->second;
1172 // Code motion for interleaved accesses can't violate WAR dependences.
1173 // Thus, reordering is legal if the source isn't a write.
1174 if (!Src->mayWriteToMemory())
1177 // At least one of the accesses must be strided.
1178 if (!isStrided(SrcDes.Stride) && !isStrided(SinkDes.Stride))
1181 // If dependence information is not available from LoopAccessInfo,
1182 // conservatively assume the instructions can't be reordered.
1183 if (!areDependencesValid())
1186 // If we know there is a dependence from source to sink, assume the
1187 // instructions can't be reordered. Otherwise, reordering is legal.
1188 return !Dependences.count(Src) || !Dependences.lookup(Src).count(Sink);
1191 /// \brief Collect the dependences from LoopAccessInfo.
1193 /// We process the dependences once during the interleaved access analysis to
1194 /// enable constant-time dependence queries.
1195 void collectDependences() {
1196 if (!areDependencesValid())
1198 auto *Deps = LAI->getDepChecker().getDependences();
1199 for (auto Dep : *Deps)
1200 Dependences[Dep.getSource(*LAI)].insert(Dep.getDestination(*LAI));
1204 /// Utility class for getting and setting loop vectorizer hints in the form
1205 /// of loop metadata.
1206 /// This class keeps a number of loop annotations locally (as member variables)
1207 /// and can, upon request, write them back as metadata on the loop. It will
1208 /// initially scan the loop for existing metadata, and will update the local
1209 /// values based on information in the loop.
1210 /// We cannot write all values to metadata, as the mere presence of some info,
1211 /// for example 'force', means a decision has been made. So, we need to be
1212 /// careful NOT to add them if the user hasn't specifically asked so.
1213 class LoopVectorizeHints {
1214 enum HintKind { HK_WIDTH, HK_UNROLL, HK_FORCE };
1216 /// Hint - associates name and validation with the hint value.
1219 unsigned Value; // This may have to change for non-numeric values.
1222 Hint(const char *Name, unsigned Value, HintKind Kind)
1223 : Name(Name), Value(Value), Kind(Kind) {}
1225 bool validate(unsigned Val) {
1228 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
1230 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1238 /// Vectorization width.
1240 /// Vectorization interleave factor.
1242 /// Vectorization forced
1245 /// Return the loop metadata prefix.
1246 static StringRef Prefix() { return "llvm.loop."; }
1248 /// True if there is any unsafe math in the loop.
1249 bool PotentiallyUnsafe;
1253 FK_Undefined = -1, ///< Not selected.
1254 FK_Disabled = 0, ///< Forcing disabled.
1255 FK_Enabled = 1, ///< Forcing enabled.
1258 LoopVectorizeHints(const Loop *L, bool DisableInterleaving,
1259 OptimizationRemarkEmitter &ORE)
1260 : Width("vectorize.width", VectorizerParams::VectorizationFactor,
1262 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1263 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1264 PotentiallyUnsafe(false), TheLoop(L), ORE(ORE) {
1265 // Populate values with existing loop metadata.
1266 getHintsFromMetadata();
1268 // force-vector-interleave overrides DisableInterleaving.
1269 if (VectorizerParams::isInterleaveForced())
1270 Interleave.Value = VectorizerParams::VectorizationInterleave;
1272 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
1273 << "LV: Interleaving disabled by the pass manager\n");
1276 /// Mark the loop L as already vectorized by setting the width to 1.
1277 void setAlreadyVectorized() {
1278 Width.Value = Interleave.Value = 1;
1279 Hint Hints[] = {Width, Interleave};
1280 writeHintsToMetadata(Hints);
1283 bool allowVectorization(Function *F, Loop *L, bool AlwaysVectorize) const {
1284 if (getForce() == LoopVectorizeHints::FK_Disabled) {
1285 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1286 emitRemarkWithHints();
1290 if (!AlwaysVectorize && getForce() != LoopVectorizeHints::FK_Enabled) {
1291 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1292 emitRemarkWithHints();
1296 if (getWidth() == 1 && getInterleave() == 1) {
1297 // FIXME: Add a separate metadata to indicate when the loop has already
1298 // been vectorized instead of setting width and count to 1.
1299 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1300 // FIXME: Add interleave.disable metadata. This will allow
1301 // vectorize.disable to be used without disabling the pass and errors
1302 // to differentiate between disabled vectorization and a width of 1.
1303 ORE.emit(OptimizationRemarkAnalysis(vectorizeAnalysisPassName(),
1304 "AllDisabled", L->getStartLoc(),
1306 << "loop not vectorized: vectorization and interleaving are "
1307 "explicitly disabled, or vectorize width and interleave "
1308 "count are both set to 1");
1315 /// Dumps all the hint information.
1316 void emitRemarkWithHints() const {
1317 using namespace ore;
1318 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1319 ORE.emit(OptimizationRemarkMissed(LV_NAME, "MissedExplicitlyDisabled",
1320 TheLoop->getStartLoc(),
1321 TheLoop->getHeader())
1322 << "loop not vectorized: vectorization is explicitly disabled");
1324 OptimizationRemarkMissed R(LV_NAME, "MissedDetails",
1325 TheLoop->getStartLoc(), TheLoop->getHeader());
1326 R << "loop not vectorized";
1327 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1328 R << " (Force=" << NV("Force", true);
1329 if (Width.Value != 0)
1330 R << ", Vector Width=" << NV("VectorWidth", Width.Value);
1331 if (Interleave.Value != 0)
1332 R << ", Interleave Count=" << NV("InterleaveCount", Interleave.Value);
1339 unsigned getWidth() const { return Width.Value; }
1340 unsigned getInterleave() const { return Interleave.Value; }
1341 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1343 /// \brief If hints are provided that force vectorization, use the AlwaysPrint
1344 /// pass name to force the frontend to print the diagnostic.
1345 const char *vectorizeAnalysisPassName() const {
1346 if (getWidth() == 1)
1348 if (getForce() == LoopVectorizeHints::FK_Disabled)
1350 if (getForce() == LoopVectorizeHints::FK_Undefined && getWidth() == 0)
1352 return OptimizationRemarkAnalysis::AlwaysPrint;
1355 bool allowReordering() const {
1356 // When enabling loop hints are provided we allow the vectorizer to change
1357 // the order of operations that is given by the scalar loop. This is not
1358 // enabled by default because can be unsafe or inefficient. For example,
1359 // reordering floating-point operations will change the way round-off
1360 // error accumulates in the loop.
1361 return getForce() == LoopVectorizeHints::FK_Enabled || getWidth() > 1;
1364 bool isPotentiallyUnsafe() const {
1365 // Avoid FP vectorization if the target is unsure about proper support.
1366 // This may be related to the SIMD unit in the target not handling
1367 // IEEE 754 FP ops properly, or bad single-to-double promotions.
1368 // Otherwise, a sequence of vectorized loops, even without reduction,
1369 // could lead to different end results on the destination vectors.
1370 return getForce() != LoopVectorizeHints::FK_Enabled && PotentiallyUnsafe;
1373 void setPotentiallyUnsafe() { PotentiallyUnsafe = true; }
1376 /// Find hints specified in the loop metadata and update local values.
1377 void getHintsFromMetadata() {
1378 MDNode *LoopID = TheLoop->getLoopID();
1382 // First operand should refer to the loop id itself.
1383 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1384 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1386 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1387 const MDString *S = nullptr;
1388 SmallVector<Metadata *, 4> Args;
1390 // The expected hint is either a MDString or a MDNode with the first
1391 // operand a MDString.
1392 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1393 if (!MD || MD->getNumOperands() == 0)
1395 S = dyn_cast<MDString>(MD->getOperand(0));
1396 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1397 Args.push_back(MD->getOperand(i));
1399 S = dyn_cast<MDString>(LoopID->getOperand(i));
1400 assert(Args.size() == 0 && "too many arguments for MDString");
1406 // Check if the hint starts with the loop metadata prefix.
1407 StringRef Name = S->getString();
1408 if (Args.size() == 1)
1409 setHint(Name, Args[0]);
1413 /// Checks string hint with one operand and set value if valid.
1414 void setHint(StringRef Name, Metadata *Arg) {
1415 if (!Name.startswith(Prefix()))
1417 Name = Name.substr(Prefix().size(), StringRef::npos);
1419 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1422 unsigned Val = C->getZExtValue();
1424 Hint *Hints[] = {&Width, &Interleave, &Force};
1425 for (auto H : Hints) {
1426 if (Name == H->Name) {
1427 if (H->validate(Val))
1430 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1436 /// Create a new hint from name / value pair.
1437 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1438 LLVMContext &Context = TheLoop->getHeader()->getContext();
1439 Metadata *MDs[] = {MDString::get(Context, Name),
1440 ConstantAsMetadata::get(
1441 ConstantInt::get(Type::getInt32Ty(Context), V))};
1442 return MDNode::get(Context, MDs);
1445 /// Matches metadata with hint name.
1446 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1447 MDString *Name = dyn_cast<MDString>(Node->getOperand(0));
1451 for (auto H : HintTypes)
1452 if (Name->getString().endswith(H.Name))
1457 /// Sets current hints into loop metadata, keeping other values intact.
1458 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1459 if (HintTypes.size() == 0)
1462 // Reserve the first element to LoopID (see below).
1463 SmallVector<Metadata *, 4> MDs(1);
1464 // If the loop already has metadata, then ignore the existing operands.
1465 MDNode *LoopID = TheLoop->getLoopID();
1467 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1468 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1469 // If node in update list, ignore old value.
1470 if (!matchesHintMetadataName(Node, HintTypes))
1471 MDs.push_back(Node);
1475 // Now, add the missing hints.
1476 for (auto H : HintTypes)
1477 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1479 // Replace current metadata node with new one.
1480 LLVMContext &Context = TheLoop->getHeader()->getContext();
1481 MDNode *NewLoopID = MDNode::get(Context, MDs);
1482 // Set operand 0 to refer to the loop id itself.
1483 NewLoopID->replaceOperandWith(0, NewLoopID);
1485 TheLoop->setLoopID(NewLoopID);
1488 /// The loop these hints belong to.
1489 const Loop *TheLoop;
1491 /// Interface to emit optimization remarks.
1492 OptimizationRemarkEmitter &ORE;
1495 static void emitAnalysisDiag(const Loop *TheLoop,
1496 const LoopVectorizeHints &Hints,
1497 OptimizationRemarkEmitter &ORE,
1498 const LoopAccessReport &Message) {
1499 const char *Name = Hints.vectorizeAnalysisPassName();
1500 LoopAccessReport::emitAnalysis(Message, TheLoop, Name, ORE);
1503 static void emitMissedWarning(Function *F, Loop *L,
1504 const LoopVectorizeHints &LH,
1505 OptimizationRemarkEmitter *ORE) {
1506 LH.emitRemarkWithHints();
1508 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1509 if (LH.getWidth() != 1)
1510 emitLoopVectorizeWarning(
1511 F->getContext(), *F, L->getStartLoc(),
1512 "failed explicitly specified loop vectorization");
1513 else if (LH.getInterleave() != 1)
1514 emitLoopInterleaveWarning(
1515 F->getContext(), *F, L->getStartLoc(),
1516 "failed explicitly specified loop interleaving");
1520 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
1521 /// to what vectorization factor.
1522 /// This class does not look at the profitability of vectorization, only the
1523 /// legality. This class has two main kinds of checks:
1524 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
1525 /// will change the order of memory accesses in a way that will change the
1526 /// correctness of the program.
1527 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
1528 /// checks for a number of different conditions, such as the availability of a
1529 /// single induction variable, that all types are supported and vectorize-able,
1530 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
1531 /// This class is also used by InnerLoopVectorizer for identifying
1532 /// induction variable and the different reduction variables.
1533 class LoopVectorizationLegality {
1535 LoopVectorizationLegality(
1536 Loop *L, PredicatedScalarEvolution &PSE, DominatorTree *DT,
1537 TargetLibraryInfo *TLI, AliasAnalysis *AA, Function *F,
1538 const TargetTransformInfo *TTI,
1539 std::function<const LoopAccessInfo &(Loop &)> *GetLAA, LoopInfo *LI,
1540 OptimizationRemarkEmitter *ORE, LoopVectorizationRequirements *R,
1541 LoopVectorizeHints *H)
1542 : NumPredStores(0), TheLoop(L), PSE(PSE), TLI(TLI), TTI(TTI), DT(DT),
1543 GetLAA(GetLAA), LAI(nullptr), ORE(ORE), InterleaveInfo(PSE, L, DT, LI),
1544 Induction(nullptr), WidestIndTy(nullptr), HasFunNoNaNAttr(false),
1545 Requirements(R), Hints(H) {}
1547 /// ReductionList contains the reduction descriptors for all
1548 /// of the reductions that were found in the loop.
1549 typedef DenseMap<PHINode *, RecurrenceDescriptor> ReductionList;
1551 /// InductionList saves induction variables and maps them to the
1552 /// induction descriptor.
1553 typedef MapVector<PHINode *, InductionDescriptor> InductionList;
1555 /// RecurrenceSet contains the phi nodes that are recurrences other than
1556 /// inductions and reductions.
1557 typedef SmallPtrSet<const PHINode *, 8> RecurrenceSet;
1559 /// Returns true if it is legal to vectorize this loop.
1560 /// This does not mean that it is profitable to vectorize this
1561 /// loop, only that it is legal to do so.
1562 bool canVectorize();
1564 /// Returns the Induction variable.
1565 PHINode *getInduction() { return Induction; }
1567 /// Returns the reduction variables found in the loop.
1568 ReductionList *getReductionVars() { return &Reductions; }
1570 /// Returns the induction variables found in the loop.
1571 InductionList *getInductionVars() { return &Inductions; }
1573 /// Return the first-order recurrences found in the loop.
1574 RecurrenceSet *getFirstOrderRecurrences() { return &FirstOrderRecurrences; }
1576 /// Returns the widest induction type.
1577 Type *getWidestInductionType() { return WidestIndTy; }
1579 /// Returns True if V is an induction variable in this loop.
1580 bool isInductionVariable(const Value *V);
1582 /// Returns True if PN is a reduction variable in this loop.
1583 bool isReductionVariable(PHINode *PN) { return Reductions.count(PN); }
1585 /// Returns True if Phi is a first-order recurrence in this loop.
1586 bool isFirstOrderRecurrence(const PHINode *Phi);
1588 /// Return true if the block BB needs to be predicated in order for the loop
1589 /// to be vectorized.
1590 bool blockNeedsPredication(BasicBlock *BB);
1592 /// Check if this pointer is consecutive when vectorizing. This happens
1593 /// when the last index of the GEP is the induction variable, or that the
1594 /// pointer itself is an induction variable.
1595 /// This check allows us to vectorize A[idx] into a wide load/store.
1597 /// 0 - Stride is unknown or non-consecutive.
1598 /// 1 - Address is consecutive.
1599 /// -1 - Address is consecutive, and decreasing.
1600 int isConsecutivePtr(Value *Ptr);
1602 /// Returns true if the value V is uniform within the loop.
1603 bool isUniform(Value *V);
1605 /// Returns true if \p I is known to be uniform after vectorization.
1606 bool isUniformAfterVectorization(Instruction *I) { return Uniforms.count(I); }
1608 /// Returns true if \p I is known to be scalar after vectorization.
1609 bool isScalarAfterVectorization(Instruction *I) { return Scalars.count(I); }
1611 /// Returns the information that we collected about runtime memory check.
1612 const RuntimePointerChecking *getRuntimePointerChecking() const {
1613 return LAI->getRuntimePointerChecking();
1616 const LoopAccessInfo *getLAI() const { return LAI; }
1618 /// \brief Check if \p Instr belongs to any interleaved access group.
1619 bool isAccessInterleaved(Instruction *Instr) {
1620 return InterleaveInfo.isInterleaved(Instr);
1623 /// \brief Return the maximum interleave factor of all interleaved groups.
1624 unsigned getMaxInterleaveFactor() const {
1625 return InterleaveInfo.getMaxInterleaveFactor();
1628 /// \brief Get the interleaved access group that \p Instr belongs to.
1629 const InterleaveGroup *getInterleavedAccessGroup(Instruction *Instr) {
1630 return InterleaveInfo.getInterleaveGroup(Instr);
1633 /// \brief Returns true if an interleaved group requires a scalar iteration
1634 /// to handle accesses with gaps.
1635 bool requiresScalarEpilogue() const {
1636 return InterleaveInfo.requiresScalarEpilogue();
1639 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
1641 bool hasStride(Value *V) { return LAI->hasStride(V); }
1643 /// Returns true if the target machine supports masked store operation
1644 /// for the given \p DataType and kind of access to \p Ptr.
1645 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
1646 return isConsecutivePtr(Ptr) && TTI->isLegalMaskedStore(DataType);
1648 /// Returns true if the target machine supports masked load operation
1649 /// for the given \p DataType and kind of access to \p Ptr.
1650 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
1651 return isConsecutivePtr(Ptr) && TTI->isLegalMaskedLoad(DataType);
1653 /// Returns true if the target machine supports masked scatter operation
1654 /// for the given \p DataType.
1655 bool isLegalMaskedScatter(Type *DataType) {
1656 return TTI->isLegalMaskedScatter(DataType);
1658 /// Returns true if the target machine supports masked gather operation
1659 /// for the given \p DataType.
1660 bool isLegalMaskedGather(Type *DataType) {
1661 return TTI->isLegalMaskedGather(DataType);
1663 /// Returns true if the target machine can represent \p V as a masked gather
1664 /// or scatter operation.
1665 bool isLegalGatherOrScatter(Value *V) {
1666 auto *LI = dyn_cast<LoadInst>(V);
1667 auto *SI = dyn_cast<StoreInst>(V);
1670 auto *Ptr = getPointerOperand(V);
1671 auto *Ty = cast<PointerType>(Ptr->getType())->getElementType();
1672 return (LI && isLegalMaskedGather(Ty)) || (SI && isLegalMaskedScatter(Ty));
1675 /// Returns true if vector representation of the instruction \p I
1677 bool isMaskRequired(const Instruction *I) { return (MaskedOp.count(I) != 0); }
1678 unsigned getNumStores() const { return LAI->getNumStores(); }
1679 unsigned getNumLoads() const { return LAI->getNumLoads(); }
1680 unsigned getNumPredStores() const { return NumPredStores; }
1682 /// Returns true if \p I is an instruction that will be scalarized with
1683 /// predication. Such instructions include conditional stores and
1684 /// instructions that may divide by zero.
1685 bool isScalarWithPredication(Instruction *I);
1687 /// Returns true if \p I is a memory instruction that has a consecutive or
1688 /// consecutive-like pointer operand. Consecutive-like pointers are pointers
1689 /// that are treated like consecutive pointers during vectorization. The
1690 /// pointer operands of interleaved accesses are an example.
1691 bool hasConsecutiveLikePtrOperand(Instruction *I);
1693 /// Returns true if \p I is a memory instruction that must be scalarized
1694 /// during vectorization.
1695 bool memoryInstructionMustBeScalarized(Instruction *I, unsigned VF = 1);
1698 /// Check if a single basic block loop is vectorizable.
1699 /// At this point we know that this is a loop with a constant trip count
1700 /// and we only need to check individual instructions.
1701 bool canVectorizeInstrs();
1703 /// When we vectorize loops we may change the order in which
1704 /// we read and write from memory. This method checks if it is
1705 /// legal to vectorize the code, considering only memory constrains.
1706 /// Returns true if the loop is vectorizable
1707 bool canVectorizeMemory();
1709 /// Return true if we can vectorize this loop using the IF-conversion
1711 bool canVectorizeWithIfConvert();
1713 /// Collect the instructions that are uniform after vectorization. An
1714 /// instruction is uniform if we represent it with a single scalar value in
1715 /// the vectorized loop corresponding to each vector iteration. Examples of
1716 /// uniform instructions include pointer operands of consecutive or
1717 /// interleaved memory accesses. Note that although uniformity implies an
1718 /// instruction will be scalar, the reverse is not true. In general, a
1719 /// scalarized instruction will be represented by VF scalar values in the
1720 /// vectorized loop, each corresponding to an iteration of the original
1722 void collectLoopUniforms();
1724 /// Collect the instructions that are scalar after vectorization. An
1725 /// instruction is scalar if it is known to be uniform or will be scalarized
1726 /// during vectorization. Non-uniform scalarized instructions will be
1727 /// represented by VF values in the vectorized loop, each corresponding to an
1728 /// iteration of the original scalar loop.
1729 void collectLoopScalars();
1731 /// Return true if all of the instructions in the block can be speculatively
1732 /// executed. \p SafePtrs is a list of addresses that are known to be legal
1733 /// and we know that we can read from them without segfault.
1734 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
1736 /// Updates the vectorization state by adding \p Phi to the inductions list.
1737 /// This can set \p Phi as the main induction of the loop if \p Phi is a
1738 /// better choice for the main induction than the existing one.
1739 void addInductionPhi(PHINode *Phi, const InductionDescriptor &ID,
1740 SmallPtrSetImpl<Value *> &AllowedExit);
1742 /// Report an analysis message to assist the user in diagnosing loops that are
1743 /// not vectorized. These are handled as LoopAccessReport rather than
1744 /// VectorizationReport because the << operator of VectorizationReport returns
1745 /// LoopAccessReport.
1746 void emitAnalysis(const LoopAccessReport &Message) const {
1747 emitAnalysisDiag(TheLoop, *Hints, *ORE, Message);
1750 /// Create an analysis remark that explains why vectorization failed
1752 /// \p RemarkName is the identifier for the remark. If \p I is passed it is
1753 /// an instruction that prevents vectorization. Otherwise the loop is used
1754 /// for the location of the remark. \return the remark object that can be
1756 OptimizationRemarkAnalysis
1757 createMissedAnalysis(StringRef RemarkName, Instruction *I = nullptr) const {
1758 return ::createMissedAnalysis(Hints->vectorizeAnalysisPassName(),
1759 RemarkName, TheLoop, I);
1762 /// \brief If an access has a symbolic strides, this maps the pointer value to
1763 /// the stride symbol.
1764 const ValueToValueMap *getSymbolicStrides() {
1765 // FIXME: Currently, the set of symbolic strides is sometimes queried before
1766 // it's collected. This happens from canVectorizeWithIfConvert, when the
1767 // pointer is checked to reference consecutive elements suitable for a
1769 return LAI ? &LAI->getSymbolicStrides() : nullptr;
1772 unsigned NumPredStores;
1774 /// The loop that we evaluate.
1776 /// A wrapper around ScalarEvolution used to add runtime SCEV checks.
1777 /// Applies dynamic knowledge to simplify SCEV expressions in the context
1778 /// of existing SCEV assumptions. The analysis will also add a minimal set
1779 /// of new predicates if this is required to enable vectorization and
1781 PredicatedScalarEvolution &PSE;
1782 /// Target Library Info.
1783 TargetLibraryInfo *TLI;
1784 /// Target Transform Info
1785 const TargetTransformInfo *TTI;
1788 // LoopAccess analysis.
1789 std::function<const LoopAccessInfo &(Loop &)> *GetLAA;
1790 // And the loop-accesses info corresponding to this loop. This pointer is
1791 // null until canVectorizeMemory sets it up.
1792 const LoopAccessInfo *LAI;
1793 /// Interface to emit optimization remarks.
1794 OptimizationRemarkEmitter *ORE;
1796 /// The interleave access information contains groups of interleaved accesses
1797 /// with the same stride and close to each other.
1798 InterleavedAccessInfo InterleaveInfo;
1800 // --- vectorization state --- //
1802 /// Holds the integer induction variable. This is the counter of the
1805 /// Holds the reduction variables.
1806 ReductionList Reductions;
1807 /// Holds all of the induction variables that we found in the loop.
1808 /// Notice that inductions don't need to start at zero and that induction
1809 /// variables can be pointers.
1810 InductionList Inductions;
1811 /// Holds the phi nodes that are first-order recurrences.
1812 RecurrenceSet FirstOrderRecurrences;
1813 /// Holds the widest induction type encountered.
1816 /// Allowed outside users. This holds the induction and reduction
1817 /// vars which can be accessed from outside the loop.
1818 SmallPtrSet<Value *, 4> AllowedExit;
1820 /// Holds the instructions known to be uniform after vectorization.
1821 SmallPtrSet<Instruction *, 4> Uniforms;
1823 /// Holds the instructions known to be scalar after vectorization.
1824 SmallPtrSet<Instruction *, 4> Scalars;
1826 /// Can we assume the absence of NaNs.
1827 bool HasFunNoNaNAttr;
1829 /// Vectorization requirements that will go through late-evaluation.
1830 LoopVectorizationRequirements *Requirements;
1832 /// Used to emit an analysis of any legality issues.
1833 LoopVectorizeHints *Hints;
1835 /// While vectorizing these instructions we have to generate a
1836 /// call to the appropriate masked intrinsic
1837 SmallPtrSet<const Instruction *, 8> MaskedOp;
1840 /// LoopVectorizationCostModel - estimates the expected speedups due to
1842 /// In many cases vectorization is not profitable. This can happen because of
1843 /// a number of reasons. In this class we mainly attempt to predict the
1844 /// expected speedup/slowdowns due to the supported instruction set. We use the
1845 /// TargetTransformInfo to query the different backends for the cost of
1846 /// different operations.
1847 class LoopVectorizationCostModel {
1849 LoopVectorizationCostModel(Loop *L, PredicatedScalarEvolution &PSE,
1850 LoopInfo *LI, LoopVectorizationLegality *Legal,
1851 const TargetTransformInfo &TTI,
1852 const TargetLibraryInfo *TLI, DemandedBits *DB,
1853 AssumptionCache *AC,
1854 OptimizationRemarkEmitter *ORE, const Function *F,
1855 const LoopVectorizeHints *Hints)
1856 : TheLoop(L), PSE(PSE), LI(LI), Legal(Legal), TTI(TTI), TLI(TLI), DB(DB),
1857 AC(AC), ORE(ORE), TheFunction(F), Hints(Hints) {}
1859 /// Information about vectorization costs
1860 struct VectorizationFactor {
1861 unsigned Width; // Vector width with best cost
1862 unsigned Cost; // Cost of the loop with that width
1864 /// \return The most profitable vectorization factor and the cost of that VF.
1865 /// This method checks every power of two up to VF. If UserVF is not ZERO
1866 /// then this vectorization factor will be selected if vectorization is
1868 VectorizationFactor selectVectorizationFactor(bool OptForSize);
1870 /// \return The size (in bits) of the smallest and widest types in the code
1871 /// that needs to be vectorized. We ignore values that remain scalar such as
1872 /// 64 bit loop indices.
1873 std::pair<unsigned, unsigned> getSmallestAndWidestTypes();
1875 /// \return The desired interleave count.
1876 /// If interleave count has been specified by metadata it will be returned.
1877 /// Otherwise, the interleave count is computed and returned. VF and LoopCost
1878 /// are the selected vectorization factor and the cost of the selected VF.
1879 unsigned selectInterleaveCount(bool OptForSize, unsigned VF,
1882 /// \return The most profitable unroll factor.
1883 /// This method finds the best unroll-factor based on register pressure and
1884 /// other parameters. VF and LoopCost are the selected vectorization factor
1885 /// and the cost of the selected VF.
1886 unsigned computeInterleaveCount(bool OptForSize, unsigned VF,
1889 /// \brief A struct that represents some properties of the register usage
1891 struct RegisterUsage {
1892 /// Holds the number of loop invariant values that are used in the loop.
1893 unsigned LoopInvariantRegs;
1894 /// Holds the maximum number of concurrent live intervals in the loop.
1895 unsigned MaxLocalUsers;
1896 /// Holds the number of instructions in the loop.
1897 unsigned NumInstructions;
1900 /// \return Returns information about the register usages of the loop for the
1901 /// given vectorization factors.
1902 SmallVector<RegisterUsage, 8> calculateRegisterUsage(ArrayRef<unsigned> VFs);
1904 /// Collect values we want to ignore in the cost model.
1905 void collectValuesToIgnore();
1907 /// \returns The smallest bitwidth each instruction can be represented with.
1908 /// The vector equivalents of these instructions should be truncated to this
1910 const MapVector<Instruction *, uint64_t> &getMinimalBitwidths() const {
1914 /// \returns True if it is more profitable to scalarize instruction \p I for
1915 /// vectorization factor \p VF.
1916 bool isProfitableToScalarize(Instruction *I, unsigned VF) const {
1917 auto Scalars = InstsToScalarize.find(VF);
1918 assert(Scalars != InstsToScalarize.end() &&
1919 "VF not yet analyzed for scalarization profitability");
1920 return Scalars->second.count(I);
1923 /// \returns True if instruction \p I can be truncated to a smaller bitwidth
1924 /// for vectorization factor \p VF.
1925 bool canTruncateToMinimalBitwidth(Instruction *I, unsigned VF) const {
1926 return VF > 1 && MinBWs.count(I) && !isProfitableToScalarize(I, VF) &&
1927 !Legal->isScalarAfterVectorization(I);
1931 /// The vectorization cost is a combination of the cost itself and a boolean
1932 /// indicating whether any of the contributing operations will actually
1934 /// vector values after type legalization in the backend. If this latter value
1936 /// false, then all operations will be scalarized (i.e. no vectorization has
1937 /// actually taken place).
1938 typedef std::pair<unsigned, bool> VectorizationCostTy;
1940 /// Returns the expected execution cost. The unit of the cost does
1941 /// not matter because we use the 'cost' units to compare different
1942 /// vector widths. The cost that is returned is *not* normalized by
1943 /// the factor width.
1944 VectorizationCostTy expectedCost(unsigned VF);
1946 /// Returns the execution time cost of an instruction for a given vector
1947 /// width. Vector width of one means scalar.
1948 VectorizationCostTy getInstructionCost(Instruction *I, unsigned VF);
1950 /// The cost-computation logic from getInstructionCost which provides
1951 /// the vector type as an output parameter.
1952 unsigned getInstructionCost(Instruction *I, unsigned VF, Type *&VectorTy);
1954 /// Returns whether the instruction is a load or store and will be a emitted
1955 /// as a vector operation.
1956 bool isConsecutiveLoadOrStore(Instruction *I);
1958 /// Create an analysis remark that explains why vectorization failed
1960 /// \p RemarkName is the identifier for the remark. \return the remark object
1961 /// that can be streamed to.
1962 OptimizationRemarkAnalysis createMissedAnalysis(StringRef RemarkName) {
1963 return ::createMissedAnalysis(Hints->vectorizeAnalysisPassName(),
1964 RemarkName, TheLoop);
1967 /// Map of scalar integer values to the smallest bitwidth they can be legally
1968 /// represented as. The vector equivalents of these values should be truncated
1970 MapVector<Instruction *, uint64_t> MinBWs;
1972 /// A type representing the costs for instructions if they were to be
1973 /// scalarized rather than vectorized. The entries are Instruction-Cost
1975 typedef DenseMap<Instruction *, unsigned> ScalarCostsTy;
1977 /// A map holding scalar costs for different vectorization factors. The
1978 /// presence of a cost for an instruction in the mapping indicates that the
1979 /// instruction will be scalarized when vectorizing with the associated
1980 /// vectorization factor. The entries are VF-ScalarCostTy pairs.
1981 DenseMap<unsigned, ScalarCostsTy> InstsToScalarize;
1983 /// Returns the expected difference in cost from scalarizing the expression
1984 /// feeding a predicated instruction \p PredInst. The instructions to
1985 /// scalarize and their scalar costs are collected in \p ScalarCosts. A
1986 /// non-negative return value implies the expression will be scalarized.
1987 /// Currently, only single-use chains are considered for scalarization.
1988 int computePredInstDiscount(Instruction *PredInst, ScalarCostsTy &ScalarCosts,
1991 /// Collects the instructions to scalarize for each predicated instruction in
1993 void collectInstsToScalarize(unsigned VF);
1996 /// The loop that we evaluate.
1998 /// Predicated scalar evolution analysis.
1999 PredicatedScalarEvolution &PSE;
2000 /// Loop Info analysis.
2002 /// Vectorization legality.
2003 LoopVectorizationLegality *Legal;
2004 /// Vector target information.
2005 const TargetTransformInfo &TTI;
2006 /// Target Library Info.
2007 const TargetLibraryInfo *TLI;
2008 /// Demanded bits analysis.
2010 /// Assumption cache.
2011 AssumptionCache *AC;
2012 /// Interface to emit optimization remarks.
2013 OptimizationRemarkEmitter *ORE;
2015 const Function *TheFunction;
2016 /// Loop Vectorize Hint.
2017 const LoopVectorizeHints *Hints;
2018 /// Values to ignore in the cost model.
2019 SmallPtrSet<const Value *, 16> ValuesToIgnore;
2020 /// Values to ignore in the cost model when VF > 1.
2021 SmallPtrSet<const Value *, 16> VecValuesToIgnore;
2024 /// \brief This holds vectorization requirements that must be verified late in
2025 /// the process. The requirements are set by legalize and costmodel. Once
2026 /// vectorization has been determined to be possible and profitable the
2027 /// requirements can be verified by looking for metadata or compiler options.
2028 /// For example, some loops require FP commutativity which is only allowed if
2029 /// vectorization is explicitly specified or if the fast-math compiler option
2030 /// has been provided.
2031 /// Late evaluation of these requirements allows helpful diagnostics to be
2032 /// composed that tells the user what need to be done to vectorize the loop. For
2033 /// example, by specifying #pragma clang loop vectorize or -ffast-math. Late
2034 /// evaluation should be used only when diagnostics can generated that can be
2035 /// followed by a non-expert user.
2036 class LoopVectorizationRequirements {
2038 LoopVectorizationRequirements(OptimizationRemarkEmitter &ORE)
2039 : NumRuntimePointerChecks(0), UnsafeAlgebraInst(nullptr), ORE(ORE) {}
2041 void addUnsafeAlgebraInst(Instruction *I) {
2042 // First unsafe algebra instruction.
2043 if (!UnsafeAlgebraInst)
2044 UnsafeAlgebraInst = I;
2047 void addRuntimePointerChecks(unsigned Num) { NumRuntimePointerChecks = Num; }
2049 bool doesNotMeet(Function *F, Loop *L, const LoopVectorizeHints &Hints) {
2050 const char *PassName = Hints.vectorizeAnalysisPassName();
2051 bool Failed = false;
2052 if (UnsafeAlgebraInst && !Hints.allowReordering()) {
2054 OptimizationRemarkAnalysisFPCommute(PassName, "CantReorderFPOps",
2055 UnsafeAlgebraInst->getDebugLoc(),
2056 UnsafeAlgebraInst->getParent())
2057 << "loop not vectorized: cannot prove it is safe to reorder "
2058 "floating-point operations");
2062 // Test if runtime memcheck thresholds are exceeded.
2063 bool PragmaThresholdReached =
2064 NumRuntimePointerChecks > PragmaVectorizeMemoryCheckThreshold;
2065 bool ThresholdReached =
2066 NumRuntimePointerChecks > VectorizerParams::RuntimeMemoryCheckThreshold;
2067 if ((ThresholdReached && !Hints.allowReordering()) ||
2068 PragmaThresholdReached) {
2069 ORE.emit(OptimizationRemarkAnalysisAliasing(PassName, "CantReorderMemOps",
2072 << "loop not vectorized: cannot prove it is safe to reorder "
2073 "memory operations");
2074 DEBUG(dbgs() << "LV: Too many memory checks needed.\n");
2082 unsigned NumRuntimePointerChecks;
2083 Instruction *UnsafeAlgebraInst;
2085 /// Interface to emit optimization remarks.
2086 OptimizationRemarkEmitter &ORE;
2089 static void addAcyclicInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
2091 if (!hasCyclesInLoopBody(L))
2095 for (Loop *InnerL : L)
2096 addAcyclicInnerLoop(*InnerL, V);
2099 /// The LoopVectorize Pass.
2100 struct LoopVectorize : public FunctionPass {
2101 /// Pass identification, replacement for typeid
2104 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
2105 : FunctionPass(ID) {
2106 Impl.DisableUnrolling = NoUnrolling;
2107 Impl.AlwaysVectorize = AlwaysVectorize;
2108 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
2111 LoopVectorizePass Impl;
2113 bool runOnFunction(Function &F) override {
2114 if (skipFunction(F))
2117 auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
2118 auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
2119 auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2120 auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2121 auto *BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI();
2122 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
2123 auto *TLI = TLIP ? &TLIP->getTLI() : nullptr;
2124 auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
2125 auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
2126 auto *LAA = &getAnalysis<LoopAccessLegacyAnalysis>();
2127 auto *DB = &getAnalysis<DemandedBitsWrapperPass>().getDemandedBits();
2128 auto *ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
2130 std::function<const LoopAccessInfo &(Loop &)> GetLAA =
2131 [&](Loop &L) -> const LoopAccessInfo & { return LAA->getInfo(&L); };
2133 return Impl.runImpl(F, *SE, *LI, *TTI, *DT, *BFI, TLI, *DB, *AA, *AC,
2137 void getAnalysisUsage(AnalysisUsage &AU) const override {
2138 AU.addRequired<AssumptionCacheTracker>();
2139 AU.addRequiredID(LoopSimplifyID);
2140 AU.addRequiredID(LCSSAID);
2141 AU.addRequired<BlockFrequencyInfoWrapperPass>();
2142 AU.addRequired<DominatorTreeWrapperPass>();
2143 AU.addRequired<LoopInfoWrapperPass>();
2144 AU.addRequired<ScalarEvolutionWrapperPass>();
2145 AU.addRequired<TargetTransformInfoWrapperPass>();
2146 AU.addRequired<AAResultsWrapperPass>();
2147 AU.addRequired<LoopAccessLegacyAnalysis>();
2148 AU.addRequired<DemandedBitsWrapperPass>();
2149 AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
2150 AU.addPreserved<LoopInfoWrapperPass>();
2151 AU.addPreserved<DominatorTreeWrapperPass>();
2152 AU.addPreserved<BasicAAWrapperPass>();
2153 AU.addPreserved<GlobalsAAWrapperPass>();
2157 } // end anonymous namespace
2159 //===----------------------------------------------------------------------===//
2160 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
2161 // LoopVectorizationCostModel.
2162 //===----------------------------------------------------------------------===//
2164 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
2165 // We need to place the broadcast of invariant variables outside the loop.
2166 Instruction *Instr = dyn_cast<Instruction>(V);
2167 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
2168 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
2170 // Place the code for broadcasting invariant variables in the new preheader.
2171 IRBuilder<>::InsertPointGuard Guard(Builder);
2173 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
2175 // Broadcast the scalar into all locations in the vector.
2176 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
2181 void InnerLoopVectorizer::createVectorIntInductionPHI(
2182 const InductionDescriptor &II, Instruction *EntryVal) {
2183 Value *Start = II.getStartValue();
2184 ConstantInt *Step = II.getConstIntStepValue();
2185 assert(Step && "Can not widen an IV with a non-constant step");
2187 // Construct the initial value of the vector IV in the vector loop preheader
2188 auto CurrIP = Builder.saveIP();
2189 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
2190 if (isa<TruncInst>(EntryVal)) {
2191 auto *TruncType = cast<IntegerType>(EntryVal->getType());
2192 Step = ConstantInt::getSigned(TruncType, Step->getSExtValue());
2193 Start = Builder.CreateCast(Instruction::Trunc, Start, TruncType);
2195 Value *SplatStart = Builder.CreateVectorSplat(VF, Start);
2196 Value *SteppedStart = getStepVector(SplatStart, 0, Step);
2197 Builder.restoreIP(CurrIP);
2200 ConstantVector::getSplat(VF, ConstantInt::getSigned(Start->getType(),
2201 VF * Step->getSExtValue()));
2202 // We may need to add the step a number of times, depending on the unroll
2203 // factor. The last of those goes into the PHI.
2204 PHINode *VecInd = PHINode::Create(SteppedStart->getType(), 2, "vec.ind",
2205 &*LoopVectorBody->getFirstInsertionPt());
2206 Instruction *LastInduction = VecInd;
2207 VectorParts Entry(UF);
2208 for (unsigned Part = 0; Part < UF; ++Part) {
2209 Entry[Part] = LastInduction;
2210 LastInduction = cast<Instruction>(
2211 Builder.CreateAdd(LastInduction, SplatVF, "step.add"));
2213 VectorLoopValueMap.initVector(EntryVal, Entry);
2214 if (isa<TruncInst>(EntryVal))
2215 addMetadata(Entry, EntryVal);
2217 // Move the last step to the end of the latch block. This ensures consistent
2218 // placement of all induction updates.
2219 auto *LoopVectorLatch = LI->getLoopFor(LoopVectorBody)->getLoopLatch();
2220 auto *Br = cast<BranchInst>(LoopVectorLatch->getTerminator());
2221 auto *ICmp = cast<Instruction>(Br->getCondition());
2222 LastInduction->moveBefore(ICmp);
2223 LastInduction->setName("vec.ind.next");
2225 VecInd->addIncoming(SteppedStart, LoopVectorPreHeader);
2226 VecInd->addIncoming(LastInduction, LoopVectorLatch);
2229 bool InnerLoopVectorizer::shouldScalarizeInstruction(Instruction *I) const {
2230 return Legal->isScalarAfterVectorization(I) ||
2231 Cost->isProfitableToScalarize(I, VF);
2234 bool InnerLoopVectorizer::needsScalarInduction(Instruction *IV) const {
2235 if (shouldScalarizeInstruction(IV))
2237 auto isScalarInst = [&](User *U) -> bool {
2238 auto *I = cast<Instruction>(U);
2239 return (OrigLoop->contains(I) && shouldScalarizeInstruction(I));
2241 return any_of(IV->users(), isScalarInst);
2244 void InnerLoopVectorizer::widenIntInduction(PHINode *IV, TruncInst *Trunc) {
2246 auto II = Legal->getInductionVars()->find(IV);
2247 assert(II != Legal->getInductionVars()->end() && "IV is not an induction");
2249 auto ID = II->second;
2250 assert(IV->getType() == ID.getStartValue()->getType() && "Types must match");
2252 // The scalar value to broadcast. This will be derived from the canonical
2253 // induction variable.
2254 Value *ScalarIV = nullptr;
2256 // The step of the induction.
2257 Value *Step = nullptr;
2259 // The value from the original loop to which we are mapping the new induction
2261 Instruction *EntryVal = Trunc ? cast<Instruction>(Trunc) : IV;
2263 // True if we have vectorized the induction variable.
2264 auto VectorizedIV = false;
2266 // Determine if we want a scalar version of the induction variable. This is
2267 // true if the induction variable itself is not widened, or if it has at
2268 // least one user in the loop that is not widened.
2269 auto NeedsScalarIV = VF > 1 && needsScalarInduction(EntryVal);
2271 // If the induction variable has a constant integer step value, go ahead and
2273 if (ID.getConstIntStepValue())
2274 Step = ID.getConstIntStepValue();
2276 // Try to create a new independent vector induction variable. If we can't
2277 // create the phi node, we will splat the scalar induction variable in each
2279 if (VF > 1 && IV->getType() == Induction->getType() && Step &&
2280 !shouldScalarizeInstruction(EntryVal)) {
2281 createVectorIntInductionPHI(ID, EntryVal);
2282 VectorizedIV = true;
2285 // If we haven't yet vectorized the induction variable, or if we will create
2286 // a scalar one, we need to define the scalar induction variable and step
2287 // values. If we were given a truncation type, truncate the canonical
2288 // induction variable and constant step. Otherwise, derive these values from
2289 // the induction descriptor.
2290 if (!VectorizedIV || NeedsScalarIV) {
2292 auto *TruncType = cast<IntegerType>(Trunc->getType());
2293 assert(Step && "Truncation requires constant integer step");
2294 auto StepInt = cast<ConstantInt>(Step)->getSExtValue();
2295 ScalarIV = Builder.CreateCast(Instruction::Trunc, Induction, TruncType);
2296 Step = ConstantInt::getSigned(TruncType, StepInt);
2298 ScalarIV = Induction;
2299 auto &DL = OrigLoop->getHeader()->getModule()->getDataLayout();
2300 if (IV != OldInduction) {
2301 ScalarIV = Builder.CreateSExtOrTrunc(ScalarIV, IV->getType());
2302 ScalarIV = ID.transform(Builder, ScalarIV, PSE.getSE(), DL);
2303 ScalarIV->setName("offset.idx");
2306 SCEVExpander Exp(*PSE.getSE(), DL, "induction");
2307 Step = Exp.expandCodeFor(ID.getStep(), ID.getStep()->getType(),
2308 &*Builder.GetInsertPoint());
2313 // If we haven't yet vectorized the induction variable, splat the scalar
2314 // induction variable, and build the necessary step vectors.
2315 if (!VectorizedIV) {
2316 Value *Broadcasted = getBroadcastInstrs(ScalarIV);
2317 VectorParts Entry(UF);
2318 for (unsigned Part = 0; Part < UF; ++Part)
2319 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
2320 VectorLoopValueMap.initVector(EntryVal, Entry);
2322 addMetadata(Entry, Trunc);
2325 // If an induction variable is only used for counting loop iterations or
2326 // calculating addresses, it doesn't need to be widened. Create scalar steps
2327 // that can be used by instructions we will later scalarize. Note that the
2328 // addition of the scalar steps will not increase the number of instructions
2329 // in the loop in the common case prior to InstCombine. We will be trading
2330 // one vector extract for each scalar step.
2332 buildScalarSteps(ScalarIV, Step, EntryVal);
2335 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx, Value *Step,
2336 Instruction::BinaryOps BinOp) {
2337 // Create and check the types.
2338 assert(Val->getType()->isVectorTy() && "Must be a vector");
2339 int VLen = Val->getType()->getVectorNumElements();
2341 Type *STy = Val->getType()->getScalarType();
2342 assert((STy->isIntegerTy() || STy->isFloatingPointTy()) &&
2343 "Induction Step must be an integer or FP");
2344 assert(Step->getType() == STy && "Step has wrong type");
2346 SmallVector<Constant *, 8> Indices;
2348 if (STy->isIntegerTy()) {
2349 // Create a vector of consecutive numbers from zero to VF.
2350 for (int i = 0; i < VLen; ++i)
2351 Indices.push_back(ConstantInt::get(STy, StartIdx + i));
2353 // Add the consecutive indices to the vector value.
2354 Constant *Cv = ConstantVector::get(Indices);
2355 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
2356 Step = Builder.CreateVectorSplat(VLen, Step);
2357 assert(Step->getType() == Val->getType() && "Invalid step vec");
2358 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
2359 // which can be found from the original scalar operations.
2360 Step = Builder.CreateMul(Cv, Step);
2361 return Builder.CreateAdd(Val, Step, "induction");
2364 // Floating point induction.
2365 assert((BinOp == Instruction::FAdd || BinOp == Instruction::FSub) &&
2366 "Binary Opcode should be specified for FP induction");
2367 // Create a vector of consecutive numbers from zero to VF.
2368 for (int i = 0; i < VLen; ++i)
2369 Indices.push_back(ConstantFP::get(STy, (double)(StartIdx + i)));
2371 // Add the consecutive indices to the vector value.
2372 Constant *Cv = ConstantVector::get(Indices);
2374 Step = Builder.CreateVectorSplat(VLen, Step);
2376 // Floating point operations had to be 'fast' to enable the induction.
2377 FastMathFlags Flags;
2378 Flags.setUnsafeAlgebra();
2380 Value *MulOp = Builder.CreateFMul(Cv, Step);
2381 if (isa<Instruction>(MulOp))
2382 // Have to check, MulOp may be a constant
2383 cast<Instruction>(MulOp)->setFastMathFlags(Flags);
2385 Value *BOp = Builder.CreateBinOp(BinOp, Val, MulOp, "induction");
2386 if (isa<Instruction>(BOp))
2387 cast<Instruction>(BOp)->setFastMathFlags(Flags);
2391 void InnerLoopVectorizer::buildScalarSteps(Value *ScalarIV, Value *Step,
2394 // We shouldn't have to build scalar steps if we aren't vectorizing.
2395 assert(VF > 1 && "VF should be greater than one");
2397 // Get the value type and ensure it and the step have the same integer type.
2398 Type *ScalarIVTy = ScalarIV->getType()->getScalarType();
2399 assert(ScalarIVTy->isIntegerTy() && ScalarIVTy == Step->getType() &&
2400 "Val and Step should have the same integer type");
2402 // Determine the number of scalars we need to generate for each unroll
2403 // iteration. If EntryVal is uniform, we only need to generate the first
2404 // lane. Otherwise, we generate all VF values.
2406 Legal->isUniformAfterVectorization(cast<Instruction>(EntryVal)) ? 1 : VF;
2408 // Compute the scalar steps and save the results in VectorLoopValueMap.
2409 ScalarParts Entry(UF);
2410 for (unsigned Part = 0; Part < UF; ++Part) {
2411 Entry[Part].resize(VF);
2412 for (unsigned Lane = 0; Lane < Lanes; ++Lane) {
2413 auto *StartIdx = ConstantInt::get(ScalarIVTy, VF * Part + Lane);
2414 auto *Mul = Builder.CreateMul(StartIdx, Step);
2415 auto *Add = Builder.CreateAdd(ScalarIV, Mul);
2416 Entry[Part][Lane] = Add;
2419 VectorLoopValueMap.initScalar(EntryVal, Entry);
2422 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
2424 const ValueToValueMap &Strides = getSymbolicStrides() ? *getSymbolicStrides() :
2427 int Stride = getPtrStride(PSE, Ptr, TheLoop, Strides, true, false);
2428 if (Stride == 1 || Stride == -1)
2433 bool LoopVectorizationLegality::isUniform(Value *V) {
2434 return LAI->isUniform(V);
2437 const InnerLoopVectorizer::VectorParts &
2438 InnerLoopVectorizer::getVectorValue(Value *V) {
2439 assert(V != Induction && "The new induction variable should not be used.");
2440 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
2441 assert(!V->getType()->isVoidTy() && "Type does not produce a value");
2443 // If we have a stride that is replaced by one, do it here.
2444 if (Legal->hasStride(V))
2445 V = ConstantInt::get(V->getType(), 1);
2447 // If we have this scalar in the map, return it.
2448 if (VectorLoopValueMap.hasVector(V))
2449 return VectorLoopValueMap.VectorMapStorage[V];
2451 // If the value has not been vectorized, check if it has been scalarized
2452 // instead. If it has been scalarized, and we actually need the value in
2453 // vector form, we will construct the vector values on demand.
2454 if (VectorLoopValueMap.hasScalar(V)) {
2456 // Initialize a new vector map entry.
2457 VectorParts Entry(UF);
2459 // If we've scalarized a value, that value should be an instruction.
2460 auto *I = cast<Instruction>(V);
2462 // If we aren't vectorizing, we can just copy the scalar map values over to
2465 for (unsigned Part = 0; Part < UF; ++Part)
2466 Entry[Part] = getScalarValue(V, Part, 0);
2467 return VectorLoopValueMap.initVector(V, Entry);
2470 // Get the last scalar instruction we generated for V. If the value is
2471 // known to be uniform after vectorization, this corresponds to lane zero
2472 // of the last unroll iteration. Otherwise, the last instruction is the one
2473 // we created for the last vector lane of the last unroll iteration.
2474 unsigned LastLane = Legal->isUniformAfterVectorization(I) ? 0 : VF - 1;
2475 auto *LastInst = cast<Instruction>(getScalarValue(V, UF - 1, LastLane));
2477 // Set the insert point after the last scalarized instruction. This ensures
2478 // the insertelement sequence will directly follow the scalar definitions.
2479 auto OldIP = Builder.saveIP();
2480 auto NewIP = std::next(BasicBlock::iterator(LastInst));
2481 Builder.SetInsertPoint(&*NewIP);
2483 // However, if we are vectorizing, we need to construct the vector values.
2484 // If the value is known to be uniform after vectorization, we can just
2485 // broadcast the scalar value corresponding to lane zero for each unroll
2486 // iteration. Otherwise, we construct the vector values using insertelement
2487 // instructions. Since the resulting vectors are stored in
2488 // VectorLoopValueMap, we will only generate the insertelements once.
2489 for (unsigned Part = 0; Part < UF; ++Part) {
2490 Value *VectorValue = nullptr;
2491 if (Legal->isUniformAfterVectorization(I)) {
2492 VectorValue = getBroadcastInstrs(getScalarValue(V, Part, 0));
2494 VectorValue = UndefValue::get(VectorType::get(V->getType(), VF));
2495 for (unsigned Lane = 0; Lane < VF; ++Lane)
2496 VectorValue = Builder.CreateInsertElement(
2497 VectorValue, getScalarValue(V, Part, Lane),
2498 Builder.getInt32(Lane));
2500 Entry[Part] = VectorValue;
2502 Builder.restoreIP(OldIP);
2503 return VectorLoopValueMap.initVector(V, Entry);
2506 // If this scalar is unknown, assume that it is a constant or that it is
2507 // loop invariant. Broadcast V and save the value for future uses.
2508 Value *B = getBroadcastInstrs(V);
2509 return VectorLoopValueMap.initVector(V, VectorParts(UF, B));
2512 Value *InnerLoopVectorizer::getScalarValue(Value *V, unsigned Part,
2515 // If the value is not an instruction contained in the loop, it should
2516 // already be scalar.
2517 if (OrigLoop->isLoopInvariant(V))
2520 assert(Lane > 0 ? !Legal->isUniformAfterVectorization(cast<Instruction>(V))
2521 : true && "Uniform values only have lane zero");
2523 // If the value from the original loop has not been vectorized, it is
2524 // represented by UF x VF scalar values in the new loop. Return the requested
2526 if (VectorLoopValueMap.hasScalar(V))
2527 return VectorLoopValueMap.ScalarMapStorage[V][Part][Lane];
2529 // If the value has not been scalarized, get its entry in VectorLoopValueMap
2530 // for the given unroll part. If this entry is not a vector type (i.e., the
2531 // vectorization factor is one), there is no need to generate an
2532 // extractelement instruction.
2533 auto *U = getVectorValue(V)[Part];
2534 if (!U->getType()->isVectorTy()) {
2535 assert(VF == 1 && "Value not scalarized has non-vector type");
2539 // Otherwise, the value from the original loop has been vectorized and is
2540 // represented by UF vector values. Extract and return the requested scalar
2541 // value from the appropriate vector lane.
2542 return Builder.CreateExtractElement(U, Builder.getInt32(Lane));
2545 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
2546 assert(Vec->getType()->isVectorTy() && "Invalid type");
2547 SmallVector<Constant *, 8> ShuffleMask;
2548 for (unsigned i = 0; i < VF; ++i)
2549 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
2551 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
2552 ConstantVector::get(ShuffleMask),
2556 // Get a mask to interleave \p NumVec vectors into a wide vector.
2557 // I.e. <0, VF, VF*2, ..., VF*(NumVec-1), 1, VF+1, VF*2+1, ...>
2558 // E.g. For 2 interleaved vectors, if VF is 4, the mask is:
2559 // <0, 4, 1, 5, 2, 6, 3, 7>
2560 static Constant *getInterleavedMask(IRBuilder<> &Builder, unsigned VF,
2562 SmallVector<Constant *, 16> Mask;
2563 for (unsigned i = 0; i < VF; i++)
2564 for (unsigned j = 0; j < NumVec; j++)
2565 Mask.push_back(Builder.getInt32(j * VF + i));
2567 return ConstantVector::get(Mask);
2570 // Get the strided mask starting from index \p Start.
2571 // I.e. <Start, Start + Stride, ..., Start + Stride*(VF-1)>
2572 static Constant *getStridedMask(IRBuilder<> &Builder, unsigned Start,
2573 unsigned Stride, unsigned VF) {
2574 SmallVector<Constant *, 16> Mask;
2575 for (unsigned i = 0; i < VF; i++)
2576 Mask.push_back(Builder.getInt32(Start + i * Stride));
2578 return ConstantVector::get(Mask);
2581 // Get a mask of two parts: The first part consists of sequential integers
2582 // starting from 0, The second part consists of UNDEFs.
2583 // I.e. <0, 1, 2, ..., NumInt - 1, undef, ..., undef>
2584 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned NumInt,
2585 unsigned NumUndef) {
2586 SmallVector<Constant *, 16> Mask;
2587 for (unsigned i = 0; i < NumInt; i++)
2588 Mask.push_back(Builder.getInt32(i));
2590 Constant *Undef = UndefValue::get(Builder.getInt32Ty());
2591 for (unsigned i = 0; i < NumUndef; i++)
2592 Mask.push_back(Undef);
2594 return ConstantVector::get(Mask);
2597 // Concatenate two vectors with the same element type. The 2nd vector should
2598 // not have more elements than the 1st vector. If the 2nd vector has less
2599 // elements, extend it with UNDEFs.
2600 static Value *ConcatenateTwoVectors(IRBuilder<> &Builder, Value *V1,
2602 VectorType *VecTy1 = dyn_cast<VectorType>(V1->getType());
2603 VectorType *VecTy2 = dyn_cast<VectorType>(V2->getType());
2604 assert(VecTy1 && VecTy2 &&
2605 VecTy1->getScalarType() == VecTy2->getScalarType() &&
2606 "Expect two vectors with the same element type");
2608 unsigned NumElts1 = VecTy1->getNumElements();
2609 unsigned NumElts2 = VecTy2->getNumElements();
2610 assert(NumElts1 >= NumElts2 && "Unexpect the first vector has less elements");
2612 if (NumElts1 > NumElts2) {
2613 // Extend with UNDEFs.
2615 getSequentialMask(Builder, NumElts2, NumElts1 - NumElts2);
2616 V2 = Builder.CreateShuffleVector(V2, UndefValue::get(VecTy2), ExtMask);
2619 Constant *Mask = getSequentialMask(Builder, NumElts1 + NumElts2, 0);
2620 return Builder.CreateShuffleVector(V1, V2, Mask);
2623 // Concatenate vectors in the given list. All vectors have the same type.
2624 static Value *ConcatenateVectors(IRBuilder<> &Builder,
2625 ArrayRef<Value *> InputList) {
2626 unsigned NumVec = InputList.size();
2627 assert(NumVec > 1 && "Should be at least two vectors");
2629 SmallVector<Value *, 8> ResList;
2630 ResList.append(InputList.begin(), InputList.end());
2632 SmallVector<Value *, 8> TmpList;
2633 for (unsigned i = 0; i < NumVec - 1; i += 2) {
2634 Value *V0 = ResList[i], *V1 = ResList[i + 1];
2635 assert((V0->getType() == V1->getType() || i == NumVec - 2) &&
2636 "Only the last vector may have a different type");
2638 TmpList.push_back(ConcatenateTwoVectors(Builder, V0, V1));
2641 // Push the last vector if the total number of vectors is odd.
2642 if (NumVec % 2 != 0)
2643 TmpList.push_back(ResList[NumVec - 1]);
2646 NumVec = ResList.size();
2647 } while (NumVec > 1);
2652 // Try to vectorize the interleave group that \p Instr belongs to.
2654 // E.g. Translate following interleaved load group (factor = 3):
2655 // for (i = 0; i < N; i+=3) {
2656 // R = Pic[i]; // Member of index 0
2657 // G = Pic[i+1]; // Member of index 1
2658 // B = Pic[i+2]; // Member of index 2
2659 // ... // do something to R, G, B
2662 // %wide.vec = load <12 x i32> ; Read 4 tuples of R,G,B
2663 // %R.vec = shuffle %wide.vec, undef, <0, 3, 6, 9> ; R elements
2664 // %G.vec = shuffle %wide.vec, undef, <1, 4, 7, 10> ; G elements
2665 // %B.vec = shuffle %wide.vec, undef, <2, 5, 8, 11> ; B elements
2667 // Or translate following interleaved store group (factor = 3):
2668 // for (i = 0; i < N; i+=3) {
2669 // ... do something to R, G, B
2670 // Pic[i] = R; // Member of index 0
2671 // Pic[i+1] = G; // Member of index 1
2672 // Pic[i+2] = B; // Member of index 2
2675 // %R_G.vec = shuffle %R.vec, %G.vec, <0, 1, 2, ..., 7>
2676 // %B_U.vec = shuffle %B.vec, undef, <0, 1, 2, 3, u, u, u, u>
2677 // %interleaved.vec = shuffle %R_G.vec, %B_U.vec,
2678 // <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> ; Interleave R,G,B elements
2679 // store <12 x i32> %interleaved.vec ; Write 4 tuples of R,G,B
2680 void InnerLoopVectorizer::vectorizeInterleaveGroup(Instruction *Instr) {
2681 const InterleaveGroup *Group = Legal->getInterleavedAccessGroup(Instr);
2682 assert(Group && "Fail to get an interleaved access group.");
2684 // Skip if current instruction is not the insert position.
2685 if (Instr != Group->getInsertPos())
2688 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2689 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2690 Value *Ptr = getPointerOperand(Instr);
2692 // Prepare for the vector type of the interleaved load/store.
2693 Type *ScalarTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2694 unsigned InterleaveFactor = Group->getFactor();
2695 Type *VecTy = VectorType::get(ScalarTy, InterleaveFactor * VF);
2696 Type *PtrTy = VecTy->getPointerTo(Ptr->getType()->getPointerAddressSpace());
2698 // Prepare for the new pointers.
2699 setDebugLocFromInst(Builder, Ptr);
2700 SmallVector<Value *, 2> NewPtrs;
2701 unsigned Index = Group->getIndex(Instr);
2703 // If the group is reverse, adjust the index to refer to the last vector lane
2704 // instead of the first. We adjust the index from the first vector lane,
2705 // rather than directly getting the pointer for lane VF - 1, because the
2706 // pointer operand of the interleaved access is supposed to be uniform. For
2707 // uniform instructions, we're only required to generate a value for the
2708 // first vector lane in each unroll iteration.
2709 if (Group->isReverse())
2710 Index += (VF - 1) * Group->getFactor();
2712 for (unsigned Part = 0; Part < UF; Part++) {
2713 Value *NewPtr = getScalarValue(Ptr, Part, 0);
2715 // Notice current instruction could be any index. Need to adjust the address
2716 // to the member of index 0.
2718 // E.g. a = A[i+1]; // Member of index 1 (Current instruction)
2719 // b = A[i]; // Member of index 0
2720 // Current pointer is pointed to A[i+1], adjust it to A[i].
2722 // E.g. A[i+1] = a; // Member of index 1
2723 // A[i] = b; // Member of index 0
2724 // A[i+2] = c; // Member of index 2 (Current instruction)
2725 // Current pointer is pointed to A[i+2], adjust it to A[i].
2726 NewPtr = Builder.CreateGEP(NewPtr, Builder.getInt32(-Index));
2728 // Cast to the vector pointer type.
2729 NewPtrs.push_back(Builder.CreateBitCast(NewPtr, PtrTy));
2732 setDebugLocFromInst(Builder, Instr);
2733 Value *UndefVec = UndefValue::get(VecTy);
2735 // Vectorize the interleaved load group.
2738 // For each unroll part, create a wide load for the group.
2739 SmallVector<Value *, 2> NewLoads;
2740 for (unsigned Part = 0; Part < UF; Part++) {
2741 auto *NewLoad = Builder.CreateAlignedLoad(
2742 NewPtrs[Part], Group->getAlignment(), "wide.vec");
2743 addMetadata(NewLoad, Instr);
2744 NewLoads.push_back(NewLoad);
2747 // For each member in the group, shuffle out the appropriate data from the
2749 for (unsigned I = 0; I < InterleaveFactor; ++I) {
2750 Instruction *Member = Group->getMember(I);
2752 // Skip the gaps in the group.
2756 VectorParts Entry(UF);
2757 Constant *StrideMask = getStridedMask(Builder, I, InterleaveFactor, VF);
2758 for (unsigned Part = 0; Part < UF; Part++) {
2759 Value *StridedVec = Builder.CreateShuffleVector(
2760 NewLoads[Part], UndefVec, StrideMask, "strided.vec");
2762 // If this member has different type, cast the result type.
2763 if (Member->getType() != ScalarTy) {
2764 VectorType *OtherVTy = VectorType::get(Member->getType(), VF);
2765 StridedVec = Builder.CreateBitOrPointerCast(StridedVec, OtherVTy);
2769 Group->isReverse() ? reverseVector(StridedVec) : StridedVec;
2771 VectorLoopValueMap.initVector(Member, Entry);
2776 // The sub vector type for current instruction.
2777 VectorType *SubVT = VectorType::get(ScalarTy, VF);
2779 // Vectorize the interleaved store group.
2780 for (unsigned Part = 0; Part < UF; Part++) {
2781 // Collect the stored vector from each member.
2782 SmallVector<Value *, 4> StoredVecs;
2783 for (unsigned i = 0; i < InterleaveFactor; i++) {
2784 // Interleaved store group doesn't allow a gap, so each index has a member
2785 Instruction *Member = Group->getMember(i);
2786 assert(Member && "Fail to get a member from an interleaved store group");
2789 getVectorValue(cast<StoreInst>(Member)->getValueOperand())[Part];
2790 if (Group->isReverse())
2791 StoredVec = reverseVector(StoredVec);
2793 // If this member has different type, cast it to an unified type.
2794 if (StoredVec->getType() != SubVT)
2795 StoredVec = Builder.CreateBitOrPointerCast(StoredVec, SubVT);
2797 StoredVecs.push_back(StoredVec);
2800 // Concatenate all vectors into a wide vector.
2801 Value *WideVec = ConcatenateVectors(Builder, StoredVecs);
2803 // Interleave the elements in the wide vector.
2804 Constant *IMask = getInterleavedMask(Builder, VF, InterleaveFactor);
2805 Value *IVec = Builder.CreateShuffleVector(WideVec, UndefVec, IMask,
2808 Instruction *NewStoreInstr =
2809 Builder.CreateAlignedStore(IVec, NewPtrs[Part], Group->getAlignment());
2810 addMetadata(NewStoreInstr, Instr);
2814 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
2815 // Attempt to issue a wide load.
2816 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2817 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2819 assert((LI || SI) && "Invalid Load/Store instruction");
2821 // Try to vectorize the interleave group if this access is interleaved.
2822 if (Legal->isAccessInterleaved(Instr))
2823 return vectorizeInterleaveGroup(Instr);
2825 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2826 Type *DataTy = VectorType::get(ScalarDataTy, VF);
2827 Value *Ptr = getPointerOperand(Instr);
2828 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
2829 // An alignment of 0 means target abi alignment. We need to use the scalar's
2830 // target abi alignment in such a case.
2831 const DataLayout &DL = Instr->getModule()->getDataLayout();
2833 Alignment = DL.getABITypeAlignment(ScalarDataTy);
2834 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
2836 // Scalarize the memory instruction if necessary.
2837 if (Legal->memoryInstructionMustBeScalarized(Instr, VF))
2838 return scalarizeInstruction(Instr, Legal->isScalarWithPredication(Instr));
2840 // Determine if the pointer operand of the access is either consecutive or
2841 // reverse consecutive.
2842 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
2843 bool Reverse = ConsecutiveStride < 0;
2845 // Determine if either a gather or scatter operation is legal.
2846 bool CreateGatherScatter =
2847 !ConsecutiveStride && Legal->isLegalGatherOrScatter(Instr);
2849 VectorParts VectorGep;
2851 // Handle consecutive loads/stores.
2852 GetElementPtrInst *Gep = getGEPInstruction(Ptr);
2853 if (ConsecutiveStride) {
2855 unsigned NumOperands = Gep->getNumOperands();
2857 // The original GEP that identified as a consecutive memory access
2858 // should have only one loop-variant operand.
2859 unsigned NumOfLoopVariantOps = 0;
2860 for (unsigned i = 0; i < NumOperands; ++i)
2861 if (!PSE.getSE()->isLoopInvariant(PSE.getSCEV(Gep->getOperand(i)),
2863 NumOfLoopVariantOps++;
2864 assert(NumOfLoopVariantOps == 1 &&
2865 "Consecutive GEP should have only one loop-variant operand");
2867 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2868 Gep2->setName("gep.indvar");
2870 // A new GEP is created for a 0-lane value of the first unroll iteration.
2871 // The GEPs for the rest of the unroll iterations are computed below as an
2872 // offset from this GEP.
2873 for (unsigned i = 0; i < NumOperands; ++i)
2874 // We can apply getScalarValue() for all GEP indices. It returns an
2875 // original value for loop-invariant operand and 0-lane for consecutive
2877 Gep2->setOperand(i, getScalarValue(Gep->getOperand(i),
2878 0, /* First unroll iteration */
2879 0 /* 0-lane of the vector */ ));
2880 setDebugLocFromInst(Builder, Gep);
2881 Ptr = Builder.Insert(Gep2);
2884 setDebugLocFromInst(Builder, Ptr);
2885 Ptr = getScalarValue(Ptr, 0, 0);
2888 // At this point we should vector version of GEP for Gather or Scatter
2889 assert(CreateGatherScatter && "The instruction should be scalarized");
2891 // Vectorizing GEP, across UF parts. We want to get a vector value for base
2892 // and each index that's defined inside the loop, even if it is
2893 // loop-invariant but wasn't hoisted out. Otherwise we want to keep them
2895 SmallVector<VectorParts, 4> OpsV;
2896 for (Value *Op : Gep->operands()) {
2897 Instruction *SrcInst = dyn_cast<Instruction>(Op);
2898 if (SrcInst && OrigLoop->contains(SrcInst))
2899 OpsV.push_back(getVectorValue(Op));
2901 OpsV.push_back(VectorParts(UF, Op));
2903 for (unsigned Part = 0; Part < UF; ++Part) {
2904 SmallVector<Value *, 4> Ops;
2905 Value *GEPBasePtr = OpsV[0][Part];
2906 for (unsigned i = 1; i < Gep->getNumOperands(); i++)
2907 Ops.push_back(OpsV[i][Part]);
2908 Value *NewGep = Builder.CreateGEP(GEPBasePtr, Ops, "VectorGep");
2909 cast<GetElementPtrInst>(NewGep)->setIsInBounds(Gep->isInBounds());
2910 assert(NewGep->getType()->isVectorTy() && "Expected vector GEP");
2913 Builder.CreateBitCast(NewGep, VectorType::get(Ptr->getType(), VF));
2914 VectorGep.push_back(NewGep);
2917 VectorGep = getVectorValue(Ptr);
2920 VectorParts Mask = createBlockInMask(Instr->getParent());
2923 assert(!Legal->isUniform(SI->getPointerOperand()) &&
2924 "We do not allow storing to uniform addresses");
2925 setDebugLocFromInst(Builder, SI);
2926 // We don't want to update the value in the map as it might be used in
2927 // another expression. So don't use a reference type for "StoredVal".
2928 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
2930 for (unsigned Part = 0; Part < UF; ++Part) {
2931 Instruction *NewSI = nullptr;
2932 if (CreateGatherScatter) {
2933 Value *MaskPart = Legal->isMaskRequired(SI) ? Mask[Part] : nullptr;
2934 NewSI = Builder.CreateMaskedScatter(StoredVal[Part], VectorGep[Part],
2935 Alignment, MaskPart);
2937 // Calculate the pointer for the specific unroll-part.
2939 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2942 // If we store to reverse consecutive memory locations, then we need
2943 // to reverse the order of elements in the stored value.
2944 StoredVal[Part] = reverseVector(StoredVal[Part]);
2945 // If the address is consecutive but reversed, then the
2946 // wide store needs to start at the last vector element.
2948 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2950 Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2951 Mask[Part] = reverseVector(Mask[Part]);
2955 Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace));
2957 if (Legal->isMaskRequired(SI))
2958 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
2962 Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
2964 addMetadata(NewSI, SI);
2970 assert(LI && "Must have a load instruction");
2971 setDebugLocFromInst(Builder, LI);
2972 VectorParts Entry(UF);
2973 for (unsigned Part = 0; Part < UF; ++Part) {
2975 if (CreateGatherScatter) {
2976 Value *MaskPart = Legal->isMaskRequired(LI) ? Mask[Part] : nullptr;
2977 NewLI = Builder.CreateMaskedGather(VectorGep[Part], Alignment, MaskPart,
2978 0, "wide.masked.gather");
2979 Entry[Part] = NewLI;
2981 // Calculate the pointer for the specific unroll-part.
2983 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2986 // If the address is consecutive but reversed, then the
2987 // wide load needs to start at the last vector element.
2988 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2989 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2990 Mask[Part] = reverseVector(Mask[Part]);
2994 Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace));
2995 if (Legal->isMaskRequired(LI))
2996 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
2997 UndefValue::get(DataTy),
2998 "wide.masked.load");
3000 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
3001 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
3003 addMetadata(NewLI, LI);
3005 VectorLoopValueMap.initVector(Instr, Entry);
3008 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr,
3009 bool IfPredicateInstr) {
3010 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
3011 DEBUG(dbgs() << "LV: Scalarizing"
3012 << (IfPredicateInstr ? " and predicating:" : ":") << *Instr
3014 // Holds vector parameters or scalars, in case of uniform vals.
3015 SmallVector<VectorParts, 4> Params;
3017 setDebugLocFromInst(Builder, Instr);
3019 // Does this instruction return a value ?
3020 bool IsVoidRetTy = Instr->getType()->isVoidTy();
3022 // Initialize a new scalar map entry.
3023 ScalarParts Entry(UF);
3026 if (IfPredicateInstr)
3027 Cond = createBlockInMask(Instr->getParent());
3029 // Determine the number of scalars we need to generate for each unroll
3030 // iteration. If the instruction is uniform, we only need to generate the
3031 // first lane. Otherwise, we generate all VF values.
3032 unsigned Lanes = Legal->isUniformAfterVectorization(Instr) ? 1 : VF;
3034 // For each vector unroll 'part':
3035 for (unsigned Part = 0; Part < UF; ++Part) {
3036 Entry[Part].resize(VF);
3037 // For each scalar that we create:
3038 for (unsigned Lane = 0; Lane < Lanes; ++Lane) {
3041 Value *Cmp = nullptr;
3042 if (IfPredicateInstr) {
3043 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Lane));
3044 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp,
3045 ConstantInt::get(Cmp->getType(), 1));
3048 Instruction *Cloned = Instr->clone();
3050 Cloned->setName(Instr->getName() + ".cloned");
3052 // Replace the operands of the cloned instructions with their scalar
3053 // equivalents in the new loop.
3054 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
3055 auto *NewOp = getScalarValue(Instr->getOperand(op), Part, Lane);
3056 Cloned->setOperand(op, NewOp);
3058 addNewMetadata(Cloned, Instr);
3060 // Place the cloned scalar in the new loop.
3061 Builder.Insert(Cloned);
3063 // Add the cloned scalar to the scalar map entry.
3064 Entry[Part][Lane] = Cloned;
3066 // If we just cloned a new assumption, add it the assumption cache.
3067 if (auto *II = dyn_cast<IntrinsicInst>(Cloned))
3068 if (II->getIntrinsicID() == Intrinsic::assume)
3069 AC->registerAssumption(II);
3072 if (IfPredicateInstr)
3073 PredicatedInstructions.push_back(std::make_pair(Cloned, Cmp));
3076 VectorLoopValueMap.initScalar(Instr, Entry);
3079 PHINode *InnerLoopVectorizer::createInductionVariable(Loop *L, Value *Start,
3080 Value *End, Value *Step,
3082 BasicBlock *Header = L->getHeader();
3083 BasicBlock *Latch = L->getLoopLatch();
3084 // As we're just creating this loop, it's possible no latch exists
3085 // yet. If so, use the header as this will be a single block loop.
3089 IRBuilder<> Builder(&*Header->getFirstInsertionPt());
3090 Instruction *OldInst = getDebugLocFromInstOrOperands(OldInduction);
3091 setDebugLocFromInst(Builder, OldInst);
3092 auto *Induction = Builder.CreatePHI(Start->getType(), 2, "index");
3094 Builder.SetInsertPoint(Latch->getTerminator());
3095 setDebugLocFromInst(Builder, OldInst);
3097 // Create i+1 and fill the PHINode.
3098 Value *Next = Builder.CreateAdd(Induction, Step, "index.next");
3099 Induction->addIncoming(Start, L->getLoopPreheader());
3100 Induction->addIncoming(Next, Latch);
3101 // Create the compare.
3102 Value *ICmp = Builder.CreateICmpEQ(Next, End);
3103 Builder.CreateCondBr(ICmp, L->getExitBlock(), Header);
3105 // Now we have two terminators. Remove the old one from the block.
3106 Latch->getTerminator()->eraseFromParent();
3111 Value *InnerLoopVectorizer::getOrCreateTripCount(Loop *L) {
3115 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
3116 // Find the loop boundaries.
3117 ScalarEvolution *SE = PSE.getSE();
3118 const SCEV *BackedgeTakenCount = PSE.getBackedgeTakenCount();
3119 assert(BackedgeTakenCount != SE->getCouldNotCompute() &&
3120 "Invalid loop count");
3122 Type *IdxTy = Legal->getWidestInductionType();
3124 // The exit count might have the type of i64 while the phi is i32. This can
3125 // happen if we have an induction variable that is sign extended before the
3126 // compare. The only way that we get a backedge taken count is that the
3127 // induction variable was signed and as such will not overflow. In such a case
3128 // truncation is legal.
3129 if (BackedgeTakenCount->getType()->getPrimitiveSizeInBits() >
3130 IdxTy->getPrimitiveSizeInBits())
3131 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount, IdxTy);
3132 BackedgeTakenCount = SE->getNoopOrZeroExtend(BackedgeTakenCount, IdxTy);
3134 // Get the total trip count from the count by adding 1.
3135 const SCEV *ExitCount = SE->getAddExpr(
3136 BackedgeTakenCount, SE->getOne(BackedgeTakenCount->getType()));
3138 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
3140 // Expand the trip count and place the new instructions in the preheader.
3141 // Notice that the pre-header does not change, only the loop body.
3142 SCEVExpander Exp(*SE, DL, "induction");
3144 // Count holds the overall loop count (N).
3145 TripCount = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
3146 L->getLoopPreheader()->getTerminator());
3148 if (TripCount->getType()->isPointerTy())
3150 CastInst::CreatePointerCast(TripCount, IdxTy, "exitcount.ptrcnt.to.int",
3151 L->getLoopPreheader()->getTerminator());
3156 Value *InnerLoopVectorizer::getOrCreateVectorTripCount(Loop *L) {
3157 if (VectorTripCount)
3158 return VectorTripCount;
3160 Value *TC = getOrCreateTripCount(L);
3161 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
3163 // Now we need to generate the expression for the part of the loop that the
3164 // vectorized body will execute. This is equal to N - (N % Step) if scalar
3165 // iterations are not required for correctness, or N - Step, otherwise. Step
3166 // is equal to the vectorization factor (number of SIMD elements) times the
3167 // unroll factor (number of SIMD instructions).
3168 Constant *Step = ConstantInt::get(TC->getType(), VF * UF);
3169 Value *R = Builder.CreateURem(TC, Step, "n.mod.vf");
3171 // If there is a non-reversed interleaved group that may speculatively access
3172 // memory out-of-bounds, we need to ensure that there will be at least one
3173 // iteration of the scalar epilogue loop. Thus, if the step evenly divides
3174 // the trip count, we set the remainder to be equal to the step. If the step
3175 // does not evenly divide the trip count, no adjustment is necessary since
3176 // there will already be scalar iterations. Note that the minimum iterations
3177 // check ensures that N >= Step.
3178 if (VF > 1 && Legal->requiresScalarEpilogue()) {
3179 auto *IsZero = Builder.CreateICmpEQ(R, ConstantInt::get(R->getType(), 0));
3180 R = Builder.CreateSelect(IsZero, Step, R);
3183 VectorTripCount = Builder.CreateSub(TC, R, "n.vec");
3185 return VectorTripCount;
3188 void InnerLoopVectorizer::emitMinimumIterationCountCheck(Loop *L,
3189 BasicBlock *Bypass) {
3190 Value *Count = getOrCreateTripCount(L);
3191 BasicBlock *BB = L->getLoopPreheader();
3192 IRBuilder<> Builder(BB->getTerminator());
3194 // Generate code to check that the loop's trip count that we computed by
3195 // adding one to the backedge-taken count will not overflow.
3196 Value *CheckMinIters = Builder.CreateICmpULT(
3197 Count, ConstantInt::get(Count->getType(), VF * UF), "min.iters.check");
3200 BB->splitBasicBlock(BB->getTerminator(), "min.iters.checked");
3201 // Update dominator tree immediately if the generated block is a
3202 // LoopBypassBlock because SCEV expansions to generate loop bypass
3203 // checks may query it before the current function is finished.
3204 DT->addNewBlock(NewBB, BB);
3205 if (L->getParentLoop())
3206 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
3207 ReplaceInstWithInst(BB->getTerminator(),
3208 BranchInst::Create(Bypass, NewBB, CheckMinIters));
3209 LoopBypassBlocks.push_back(BB);
3212 void InnerLoopVectorizer::emitVectorLoopEnteredCheck(Loop *L,
3213 BasicBlock *Bypass) {
3214 Value *TC = getOrCreateVectorTripCount(L);
3215 BasicBlock *BB = L->getLoopPreheader();
3216 IRBuilder<> Builder(BB->getTerminator());
3218 // Now, compare the new count to zero. If it is zero skip the vector loop and
3219 // jump to the scalar loop.
3220 Value *Cmp = Builder.CreateICmpEQ(TC, Constant::getNullValue(TC->getType()),
3223 // Generate code to check that the loop's trip count that we computed by
3224 // adding one to the backedge-taken count will not overflow.
3225 BasicBlock *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
3226 // Update dominator tree immediately if the generated block is a
3227 // LoopBypassBlock because SCEV expansions to generate loop bypass
3228 // checks may query it before the current function is finished.
3229 DT->addNewBlock(NewBB, BB);
3230 if (L->getParentLoop())
3231 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
3232 ReplaceInstWithInst(BB->getTerminator(),
3233 BranchInst::Create(Bypass, NewBB, Cmp));
3234 LoopBypassBlocks.push_back(BB);
3237 void InnerLoopVectorizer::emitSCEVChecks(Loop *L, BasicBlock *Bypass) {
3238 BasicBlock *BB = L->getLoopPreheader();
3240 // Generate the code to check that the SCEV assumptions that we made.
3241 // We want the new basic block to start at the first instruction in a
3242 // sequence of instructions that form a check.
3243 SCEVExpander Exp(*PSE.getSE(), Bypass->getModule()->getDataLayout(),
3246 Exp.expandCodeForPredicate(&PSE.getUnionPredicate(), BB->getTerminator());
3248 if (auto *C = dyn_cast<ConstantInt>(SCEVCheck))
3252 // Create a new block containing the stride check.
3253 BB->setName("vector.scevcheck");
3254 auto *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
3255 // Update dominator tree immediately if the generated block is a
3256 // LoopBypassBlock because SCEV expansions to generate loop bypass
3257 // checks may query it before the current function is finished.
3258 DT->addNewBlock(NewBB, BB);
3259 if (L->getParentLoop())
3260 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
3261 ReplaceInstWithInst(BB->getTerminator(),
3262 BranchInst::Create(Bypass, NewBB, SCEVCheck));
3263 LoopBypassBlocks.push_back(BB);
3264 AddedSafetyChecks = true;
3267 void InnerLoopVectorizer::emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass) {
3268 BasicBlock *BB = L->getLoopPreheader();
3270 // Generate the code that checks in runtime if arrays overlap. We put the
3271 // checks into a separate block to make the more common case of few elements
3273 Instruction *FirstCheckInst;
3274 Instruction *MemRuntimeCheck;
3275 std::tie(FirstCheckInst, MemRuntimeCheck) =
3276 Legal->getLAI()->addRuntimeChecks(BB->getTerminator());
3277 if (!MemRuntimeCheck)
3280 // Create a new block containing the memory check.
3281 BB->setName("vector.memcheck");
3282 auto *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
3283 // Update dominator tree immediately if the generated block is a
3284 // LoopBypassBlock because SCEV expansions to generate loop bypass
3285 // checks may query it before the current function is finished.
3286 DT->addNewBlock(NewBB, BB);
3287 if (L->getParentLoop())
3288 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
3289 ReplaceInstWithInst(BB->getTerminator(),
3290 BranchInst::Create(Bypass, NewBB, MemRuntimeCheck));
3291 LoopBypassBlocks.push_back(BB);
3292 AddedSafetyChecks = true;
3294 // We currently don't use LoopVersioning for the actual loop cloning but we
3295 // still use it to add the noalias metadata.
3296 LVer = llvm::make_unique<LoopVersioning>(*Legal->getLAI(), OrigLoop, LI, DT,
3298 LVer->prepareNoAliasMetadata();
3301 void InnerLoopVectorizer::createEmptyLoop() {
3303 In this function we generate a new loop. The new loop will contain
3304 the vectorized instructions while the old loop will continue to run the
3307 [ ] <-- loop iteration number check.
3310 | [ ] <-- vector loop bypass (may consist of multiple blocks).
3313 || [ ] <-- vector pre header.
3317 | [ ]_| <-- vector loop.
3320 | -[ ] <--- middle-block.
3323 -|- >[ ] <--- new preheader.
3327 | [ ]_| <-- old scalar loop to handle remainder.
3330 >[ ] <-- exit block.
3334 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
3335 BasicBlock *VectorPH = OrigLoop->getLoopPreheader();
3336 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
3337 assert(VectorPH && "Invalid loop structure");
3338 assert(ExitBlock && "Must have an exit block");
3340 // Some loops have a single integer induction variable, while other loops
3341 // don't. One example is c++ iterators that often have multiple pointer
3342 // induction variables. In the code below we also support a case where we
3343 // don't have a single induction variable.
3345 // We try to obtain an induction variable from the original loop as hard
3346 // as possible. However if we don't find one that:
3348 // - counts from zero, stepping by one
3349 // - is the size of the widest induction variable type
3350 // then we create a new one.
3351 OldInduction = Legal->getInduction();
3352 Type *IdxTy = Legal->getWidestInductionType();
3354 // Split the single block loop into the two loop structure described above.
3355 BasicBlock *VecBody =
3356 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
3357 BasicBlock *MiddleBlock =
3358 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
3359 BasicBlock *ScalarPH =
3360 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
3362 // Create and register the new vector loop.
3363 Loop *Lp = new Loop();
3364 Loop *ParentLoop = OrigLoop->getParentLoop();
3366 // Insert the new loop into the loop nest and register the new basic blocks
3367 // before calling any utilities such as SCEV that require valid LoopInfo.
3369 ParentLoop->addChildLoop(Lp);
3370 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
3371 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
3373 LI->addTopLevelLoop(Lp);
3375 Lp->addBasicBlockToLoop(VecBody, *LI);
3377 // Find the loop boundaries.
3378 Value *Count = getOrCreateTripCount(Lp);
3380 Value *StartIdx = ConstantInt::get(IdxTy, 0);
3382 // We need to test whether the backedge-taken count is uint##_max. Adding one
3383 // to it will cause overflow and an incorrect loop trip count in the vector
3384 // body. In case of overflow we want to directly jump to the scalar remainder
3386 emitMinimumIterationCountCheck(Lp, ScalarPH);
3387 // Now, compare the new count to zero. If it is zero skip the vector loop and
3388 // jump to the scalar loop.
3389 emitVectorLoopEnteredCheck(Lp, ScalarPH);
3390 // Generate the code to check any assumptions that we've made for SCEV
3392 emitSCEVChecks(Lp, ScalarPH);
3394 // Generate the code that checks in runtime if arrays overlap. We put the
3395 // checks into a separate block to make the more common case of few elements
3397 emitMemRuntimeChecks(Lp, ScalarPH);
3399 // Generate the induction variable.
3400 // The loop step is equal to the vectorization factor (num of SIMD elements)
3401 // times the unroll factor (num of SIMD instructions).
3402 Value *CountRoundDown = getOrCreateVectorTripCount(Lp);
3403 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
3405 createInductionVariable(Lp, StartIdx, CountRoundDown, Step,
3406 getDebugLocFromInstOrOperands(OldInduction));
3408 // We are going to resume the execution of the scalar loop.
3409 // Go over all of the induction variables that we found and fix the
3410 // PHIs that are left in the scalar version of the loop.
3411 // The starting values of PHI nodes depend on the counter of the last
3412 // iteration in the vectorized loop.
3413 // If we come from a bypass edge then we need to start from the original
3416 // This variable saves the new starting index for the scalar loop. It is used
3417 // to test if there are any tail iterations left once the vector loop has
3419 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
3420 for (auto &InductionEntry : *List) {
3421 PHINode *OrigPhi = InductionEntry.first;
3422 InductionDescriptor II = InductionEntry.second;
3424 // Create phi nodes to merge from the backedge-taken check block.
3425 PHINode *BCResumeVal = PHINode::Create(
3426 OrigPhi->getType(), 3, "bc.resume.val", ScalarPH->getTerminator());
3428 if (OrigPhi == OldInduction) {
3429 // We know what the end value is.
3430 EndValue = CountRoundDown;
3432 IRBuilder<> B(LoopBypassBlocks.back()->getTerminator());
3433 Type *StepType = II.getStep()->getType();
3434 Instruction::CastOps CastOp =
3435 CastInst::getCastOpcode(CountRoundDown, true, StepType, true);
3436 Value *CRD = B.CreateCast(CastOp, CountRoundDown, StepType, "cast.crd");
3437 const DataLayout &DL = OrigLoop->getHeader()->getModule()->getDataLayout();
3438 EndValue = II.transform(B, CRD, PSE.getSE(), DL);
3439 EndValue->setName("ind.end");
3442 // The new PHI merges the original incoming value, in case of a bypass,
3443 // or the value at the end of the vectorized loop.
3444 BCResumeVal->addIncoming(EndValue, MiddleBlock);
3446 // Fix up external users of the induction variable.
3447 fixupIVUsers(OrigPhi, II, CountRoundDown, EndValue, MiddleBlock);
3449 // Fix the scalar body counter (PHI node).
3450 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
3452 // The old induction's phi node in the scalar body needs the truncated
3454 for (BasicBlock *BB : LoopBypassBlocks)
3455 BCResumeVal->addIncoming(II.getStartValue(), BB);
3456 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
3459 // Add a check in the middle block to see if we have completed
3460 // all of the iterations in the first vector loop.
3461 // If (N - N%VF) == N, then we *don't* need to run the remainder.
3463 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, Count,
3464 CountRoundDown, "cmp.n", MiddleBlock->getTerminator());
3465 ReplaceInstWithInst(MiddleBlock->getTerminator(),
3466 BranchInst::Create(ExitBlock, ScalarPH, CmpN));
3468 // Get ready to start creating new instructions into the vectorized body.
3469 Builder.SetInsertPoint(&*VecBody->getFirstInsertionPt());
3472 LoopVectorPreHeader = Lp->getLoopPreheader();
3473 LoopScalarPreHeader = ScalarPH;
3474 LoopMiddleBlock = MiddleBlock;
3475 LoopExitBlock = ExitBlock;
3476 LoopVectorBody = VecBody;
3477 LoopScalarBody = OldBasicBlock;
3479 // Keep all loop hints from the original loop on the vector loop (we'll
3480 // replace the vectorizer-specific hints below).
3481 if (MDNode *LID = OrigLoop->getLoopID())
3484 LoopVectorizeHints Hints(Lp, true, *ORE);
3485 Hints.setAlreadyVectorized();
3488 // Fix up external users of the induction variable. At this point, we are
3489 // in LCSSA form, with all external PHIs that use the IV having one input value,
3490 // coming from the remainder loop. We need those PHIs to also have a correct
3491 // value for the IV when arriving directly from the middle block.
3492 void InnerLoopVectorizer::fixupIVUsers(PHINode *OrigPhi,
3493 const InductionDescriptor &II,
3494 Value *CountRoundDown, Value *EndValue,
3495 BasicBlock *MiddleBlock) {
3496 // There are two kinds of external IV usages - those that use the value
3497 // computed in the last iteration (the PHI) and those that use the penultimate
3498 // value (the value that feeds into the phi from the loop latch).
3499 // We allow both, but they, obviously, have different values.
3501 assert(OrigLoop->getExitBlock() && "Expected a single exit block");
3503 DenseMap<Value *, Value *> MissingVals;
3505 // An external user of the last iteration's value should see the value that
3506 // the remainder loop uses to initialize its own IV.
3507 Value *PostInc = OrigPhi->getIncomingValueForBlock(OrigLoop->getLoopLatch());
3508 for (User *U : PostInc->users()) {
3509 Instruction *UI = cast<Instruction>(U);
3510 if (!OrigLoop->contains(UI)) {
3511 assert(isa<PHINode>(UI) && "Expected LCSSA form");
3512 MissingVals[UI] = EndValue;
3516 // An external user of the penultimate value need to see EndValue - Step.
3517 // The simplest way to get this is to recompute it from the constituent SCEVs,
3518 // that is Start + (Step * (CRD - 1)).
3519 for (User *U : OrigPhi->users()) {
3520 auto *UI = cast<Instruction>(U);
3521 if (!OrigLoop->contains(UI)) {
3522 const DataLayout &DL =
3523 OrigLoop->getHeader()->getModule()->getDataLayout();
3524 assert(isa<PHINode>(UI) && "Expected LCSSA form");
3526 IRBuilder<> B(MiddleBlock->getTerminator());
3527 Value *CountMinusOne = B.CreateSub(
3528 CountRoundDown, ConstantInt::get(CountRoundDown->getType(), 1));
3529 Value *CMO = B.CreateSExtOrTrunc(CountMinusOne, II.getStep()->getType(),
3531 Value *Escape = II.transform(B, CMO, PSE.getSE(), DL);
3532 Escape->setName("ind.escape");
3533 MissingVals[UI] = Escape;
3537 for (auto &I : MissingVals) {
3538 PHINode *PHI = cast<PHINode>(I.first);
3539 // One corner case we have to handle is two IVs "chasing" each-other,
3540 // that is %IV2 = phi [...], [ %IV1, %latch ]
3541 // In this case, if IV1 has an external use, we need to avoid adding both
3542 // "last value of IV1" and "penultimate value of IV2". So, verify that we
3543 // don't already have an incoming value for the middle block.
3544 if (PHI->getBasicBlockIndex(MiddleBlock) == -1)
3545 PHI->addIncoming(I.second, MiddleBlock);
3550 struct CSEDenseMapInfo {
3551 static bool canHandle(Instruction *I) {
3552 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
3553 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
3555 static inline Instruction *getEmptyKey() {
3556 return DenseMapInfo<Instruction *>::getEmptyKey();
3558 static inline Instruction *getTombstoneKey() {
3559 return DenseMapInfo<Instruction *>::getTombstoneKey();
3561 static unsigned getHashValue(Instruction *I) {
3562 assert(canHandle(I) && "Unknown instruction!");
3563 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
3564 I->value_op_end()));
3566 static bool isEqual(Instruction *LHS, Instruction *RHS) {
3567 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
3568 LHS == getTombstoneKey() || RHS == getTombstoneKey())
3570 return LHS->isIdenticalTo(RHS);
3575 ///\brief Perform cse of induction variable instructions.
3576 static void cse(BasicBlock *BB) {
3577 // Perform simple cse.
3578 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
3579 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
3580 Instruction *In = &*I++;
3582 if (!CSEDenseMapInfo::canHandle(In))
3585 // Check if we can replace this instruction with any of the
3586 // visited instructions.
3587 if (Instruction *V = CSEMap.lookup(In)) {
3588 In->replaceAllUsesWith(V);
3589 In->eraseFromParent();
3597 /// \brief Adds a 'fast' flag to floating point operations.
3598 static Value *addFastMathFlag(Value *V) {
3599 if (isa<FPMathOperator>(V)) {
3600 FastMathFlags Flags;
3601 Flags.setUnsafeAlgebra();
3602 cast<Instruction>(V)->setFastMathFlags(Flags);
3607 /// \brief Estimate the overhead of scalarizing a value based on its type.
3608 /// Insert and Extract are set if the result needs to be inserted and/or
3609 /// extracted from vectors.
3610 static unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract,
3611 const TargetTransformInfo &TTI) {
3615 assert(Ty->isVectorTy() && "Can only scalarize vectors");
3618 for (unsigned I = 0, E = Ty->getVectorNumElements(); I < E; ++I) {
3620 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, Ty, I);
3622 Cost += TTI.getVectorInstrCost(Instruction::InsertElement, Ty, I);
3628 /// \brief Estimate the overhead of scalarizing an Instruction based on the
3629 /// types of its operands and return value.
3630 static unsigned getScalarizationOverhead(SmallVectorImpl<Type *> &OpTys,
3632 const TargetTransformInfo &TTI) {
3633 unsigned ScalarizationCost =
3634 getScalarizationOverhead(RetTy, true, false, TTI);
3636 for (Type *Ty : OpTys)
3637 ScalarizationCost += getScalarizationOverhead(Ty, false, true, TTI);
3639 return ScalarizationCost;
3642 /// \brief Estimate the overhead of scalarizing an instruction. This is a
3643 /// convenience wrapper for the type-based getScalarizationOverhead API.
3644 static unsigned getScalarizationOverhead(Instruction *I, unsigned VF,
3645 const TargetTransformInfo &TTI) {
3649 Type *RetTy = ToVectorTy(I->getType(), VF);
3651 SmallVector<Type *, 4> OpTys;
3652 unsigned OperandsNum = I->getNumOperands();
3653 for (unsigned OpInd = 0; OpInd < OperandsNum; ++OpInd)
3654 OpTys.push_back(ToVectorTy(I->getOperand(OpInd)->getType(), VF));
3656 return getScalarizationOverhead(OpTys, RetTy, TTI);
3659 // Estimate cost of a call instruction CI if it were vectorized with factor VF.
3660 // Return the cost of the instruction, including scalarization overhead if it's
3661 // needed. The flag NeedToScalarize shows if the call needs to be scalarized -
3662 // i.e. either vector version isn't available, or is too expensive.
3663 static unsigned getVectorCallCost(CallInst *CI, unsigned VF,
3664 const TargetTransformInfo &TTI,
3665 const TargetLibraryInfo *TLI,
3666 bool &NeedToScalarize) {
3667 Function *F = CI->getCalledFunction();
3668 StringRef FnName = CI->getCalledFunction()->getName();
3669 Type *ScalarRetTy = CI->getType();
3670 SmallVector<Type *, 4> Tys, ScalarTys;
3671 for (auto &ArgOp : CI->arg_operands())
3672 ScalarTys.push_back(ArgOp->getType());
3674 // Estimate cost of scalarized vector call. The source operands are assumed
3675 // to be vectors, so we need to extract individual elements from there,
3676 // execute VF scalar calls, and then gather the result into the vector return
3678 unsigned ScalarCallCost = TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys);
3680 return ScalarCallCost;
3682 // Compute corresponding vector type for return value and arguments.
3683 Type *RetTy = ToVectorTy(ScalarRetTy, VF);
3684 for (Type *ScalarTy : ScalarTys)
3685 Tys.push_back(ToVectorTy(ScalarTy, VF));
3687 // Compute costs of unpacking argument values for the scalar calls and
3688 // packing the return values to a vector.
3689 unsigned ScalarizationCost = getScalarizationOverhead(Tys, RetTy, TTI);
3691 unsigned Cost = ScalarCallCost * VF + ScalarizationCost;
3693 // If we can't emit a vector call for this function, then the currently found
3694 // cost is the cost we need to return.
3695 NeedToScalarize = true;
3696 if (!TLI || !TLI->isFunctionVectorizable(FnName, VF) || CI->isNoBuiltin())
3699 // If the corresponding vector cost is cheaper, return its cost.
3700 unsigned VectorCallCost = TTI.getCallInstrCost(nullptr, RetTy, Tys);
3701 if (VectorCallCost < Cost) {
3702 NeedToScalarize = false;
3703 return VectorCallCost;
3708 // Estimate cost of an intrinsic call instruction CI if it were vectorized with
3709 // factor VF. Return the cost of the instruction, including scalarization
3710 // overhead if it's needed.
3711 static unsigned getVectorIntrinsicCost(CallInst *CI, unsigned VF,
3712 const TargetTransformInfo &TTI,
3713 const TargetLibraryInfo *TLI) {
3714 Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
3715 assert(ID && "Expected intrinsic call!");
3717 Type *RetTy = ToVectorTy(CI->getType(), VF);
3718 SmallVector<Type *, 4> Tys;
3719 for (Value *ArgOperand : CI->arg_operands())
3720 Tys.push_back(ToVectorTy(ArgOperand->getType(), VF));
3723 if (auto *FPMO = dyn_cast<FPMathOperator>(CI))
3724 FMF = FPMO->getFastMathFlags();
3726 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys, FMF);
3729 static Type *smallestIntegerVectorType(Type *T1, Type *T2) {
3730 auto *I1 = cast<IntegerType>(T1->getVectorElementType());
3731 auto *I2 = cast<IntegerType>(T2->getVectorElementType());
3732 return I1->getBitWidth() < I2->getBitWidth() ? T1 : T2;
3734 static Type *largestIntegerVectorType(Type *T1, Type *T2) {
3735 auto *I1 = cast<IntegerType>(T1->getVectorElementType());
3736 auto *I2 = cast<IntegerType>(T2->getVectorElementType());
3737 return I1->getBitWidth() > I2->getBitWidth() ? T1 : T2;
3740 void InnerLoopVectorizer::truncateToMinimalBitwidths() {
3741 // For every instruction `I` in MinBWs, truncate the operands, create a
3742 // truncated version of `I` and reextend its result. InstCombine runs
3743 // later and will remove any ext/trunc pairs.
3745 SmallPtrSet<Value *, 4> Erased;
3746 for (const auto &KV : Cost->getMinimalBitwidths()) {
3747 // If the value wasn't vectorized, we must maintain the original scalar
3748 // type. The absence of the value from VectorLoopValueMap indicates that it
3749 // wasn't vectorized.
3750 if (!VectorLoopValueMap.hasVector(KV.first))
3752 VectorParts &Parts = VectorLoopValueMap.getVector(KV.first);
3753 for (Value *&I : Parts) {
3754 if (Erased.count(I) || I->use_empty() || !isa<Instruction>(I))
3756 Type *OriginalTy = I->getType();
3757 Type *ScalarTruncatedTy =
3758 IntegerType::get(OriginalTy->getContext(), KV.second);
3759 Type *TruncatedTy = VectorType::get(ScalarTruncatedTy,
3760 OriginalTy->getVectorNumElements());
3761 if (TruncatedTy == OriginalTy)
3764 IRBuilder<> B(cast<Instruction>(I));
3765 auto ShrinkOperand = [&](Value *V) -> Value * {
3766 if (auto *ZI = dyn_cast<ZExtInst>(V))
3767 if (ZI->getSrcTy() == TruncatedTy)
3768 return ZI->getOperand(0);
3769 return B.CreateZExtOrTrunc(V, TruncatedTy);
3772 // The actual instruction modification depends on the instruction type,
3774 Value *NewI = nullptr;
3775 if (auto *BO = dyn_cast<BinaryOperator>(I)) {
3776 NewI = B.CreateBinOp(BO->getOpcode(), ShrinkOperand(BO->getOperand(0)),
3777 ShrinkOperand(BO->getOperand(1)));
3778 cast<BinaryOperator>(NewI)->copyIRFlags(I);
3779 } else if (auto *CI = dyn_cast<ICmpInst>(I)) {
3781 B.CreateICmp(CI->getPredicate(), ShrinkOperand(CI->getOperand(0)),
3782 ShrinkOperand(CI->getOperand(1)));
3783 } else if (auto *SI = dyn_cast<SelectInst>(I)) {
3784 NewI = B.CreateSelect(SI->getCondition(),
3785 ShrinkOperand(SI->getTrueValue()),
3786 ShrinkOperand(SI->getFalseValue()));
3787 } else if (auto *CI = dyn_cast<CastInst>(I)) {
3788 switch (CI->getOpcode()) {
3790 llvm_unreachable("Unhandled cast!");
3791 case Instruction::Trunc:
3792 NewI = ShrinkOperand(CI->getOperand(0));
3794 case Instruction::SExt:
3795 NewI = B.CreateSExtOrTrunc(
3797 smallestIntegerVectorType(OriginalTy, TruncatedTy));
3799 case Instruction::ZExt:
3800 NewI = B.CreateZExtOrTrunc(
3802 smallestIntegerVectorType(OriginalTy, TruncatedTy));
3805 } else if (auto *SI = dyn_cast<ShuffleVectorInst>(I)) {
3806 auto Elements0 = SI->getOperand(0)->getType()->getVectorNumElements();
3807 auto *O0 = B.CreateZExtOrTrunc(
3808 SI->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements0));
3809 auto Elements1 = SI->getOperand(1)->getType()->getVectorNumElements();
3810 auto *O1 = B.CreateZExtOrTrunc(
3811 SI->getOperand(1), VectorType::get(ScalarTruncatedTy, Elements1));
3813 NewI = B.CreateShuffleVector(O0, O1, SI->getMask());
3814 } else if (isa<LoadInst>(I)) {
3815 // Don't do anything with the operands, just extend the result.
3817 } else if (auto *IE = dyn_cast<InsertElementInst>(I)) {
3818 auto Elements = IE->getOperand(0)->getType()->getVectorNumElements();
3819 auto *O0 = B.CreateZExtOrTrunc(
3820 IE->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements));
3821 auto *O1 = B.CreateZExtOrTrunc(IE->getOperand(1), ScalarTruncatedTy);
3822 NewI = B.CreateInsertElement(O0, O1, IE->getOperand(2));
3823 } else if (auto *EE = dyn_cast<ExtractElementInst>(I)) {
3824 auto Elements = EE->getOperand(0)->getType()->getVectorNumElements();
3825 auto *O0 = B.CreateZExtOrTrunc(
3826 EE->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements));
3827 NewI = B.CreateExtractElement(O0, EE->getOperand(2));
3829 llvm_unreachable("Unhandled instruction type!");
3832 // Lastly, extend the result.
3833 NewI->takeName(cast<Instruction>(I));
3834 Value *Res = B.CreateZExtOrTrunc(NewI, OriginalTy);
3835 I->replaceAllUsesWith(Res);
3836 cast<Instruction>(I)->eraseFromParent();
3842 // We'll have created a bunch of ZExts that are now parentless. Clean up.
3843 for (const auto &KV : Cost->getMinimalBitwidths()) {
3844 // If the value wasn't vectorized, we must maintain the original scalar
3845 // type. The absence of the value from VectorLoopValueMap indicates that it
3846 // wasn't vectorized.
3847 if (!VectorLoopValueMap.hasVector(KV.first))
3849 VectorParts &Parts = VectorLoopValueMap.getVector(KV.first);
3850 for (Value *&I : Parts) {
3851 ZExtInst *Inst = dyn_cast<ZExtInst>(I);
3852 if (Inst && Inst->use_empty()) {
3853 Value *NewI = Inst->getOperand(0);
3854 Inst->eraseFromParent();
3861 void InnerLoopVectorizer::vectorizeLoop() {
3862 //===------------------------------------------------===//
3864 // Notice: any optimization or new instruction that go
3865 // into the code below should be also be implemented in
3868 //===------------------------------------------------===//
3869 Constant *Zero = Builder.getInt32(0);
3871 // In order to support recurrences we need to be able to vectorize Phi nodes.
3872 // Phi nodes have cycles, so we need to vectorize them in two stages. First,
3873 // we create a new vector PHI node with no incoming edges. We use this value
3874 // when we vectorize all of the instructions that use the PHI. Next, after
3875 // all of the instructions in the block are complete we add the new incoming
3876 // edges to the PHI. At this point all of the instructions in the basic block
3877 // are vectorized, so we can use them to construct the PHI.
3878 PhiVector PHIsToFix;
3880 // Collect instructions from the original loop that will become trivially
3881 // dead in the vectorized loop. We don't need to vectorize these
3883 collectTriviallyDeadInstructions();
3885 // Scan the loop in a topological order to ensure that defs are vectorized
3887 LoopBlocksDFS DFS(OrigLoop);
3890 // Vectorize all of the blocks in the original loop.
3891 for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO()))
3892 vectorizeBlockInLoop(BB, &PHIsToFix);
3894 // Insert truncates and extends for any truncated instructions as hints to
3897 truncateToMinimalBitwidths();
3899 // At this point every instruction in the original loop is widened to a
3900 // vector form. Now we need to fix the recurrences in PHIsToFix. These PHI
3901 // nodes are currently empty because we did not want to introduce cycles.
3902 // This is the second stage of vectorizing recurrences.
3903 for (PHINode *Phi : PHIsToFix) {
3904 assert(Phi && "Unable to recover vectorized PHI");
3906 // Handle first-order recurrences that need to be fixed.
3907 if (Legal->isFirstOrderRecurrence(Phi)) {
3908 fixFirstOrderRecurrence(Phi);
3912 // If the phi node is not a first-order recurrence, it must be a reduction.
3913 // Get it's reduction variable descriptor.
3914 assert(Legal->isReductionVariable(Phi) &&
3915 "Unable to find the reduction variable");
3916 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[Phi];
3918 RecurrenceDescriptor::RecurrenceKind RK = RdxDesc.getRecurrenceKind();
3919 TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue();
3920 Instruction *LoopExitInst = RdxDesc.getLoopExitInstr();
3921 RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind =
3922 RdxDesc.getMinMaxRecurrenceKind();
3923 setDebugLocFromInst(Builder, ReductionStartValue);
3925 // We need to generate a reduction vector from the incoming scalar.
3926 // To do so, we need to generate the 'identity' vector and override
3927 // one of the elements with the incoming scalar reduction. We need
3928 // to do it in the vector-loop preheader.
3929 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
3931 // This is the vector-clone of the value that leaves the loop.
3932 const VectorParts &VectorExit = getVectorValue(LoopExitInst);
3933 Type *VecTy = VectorExit[0]->getType();
3935 // Find the reduction identity variable. Zero for addition, or, xor,
3936 // one for multiplication, -1 for And.
3939 if (RK == RecurrenceDescriptor::RK_IntegerMinMax ||
3940 RK == RecurrenceDescriptor::RK_FloatMinMax) {
3941 // MinMax reduction have the start value as their identify.
3943 VectorStart = Identity = ReductionStartValue;
3945 VectorStart = Identity =
3946 Builder.CreateVectorSplat(VF, ReductionStartValue, "minmax.ident");
3949 // Handle other reduction kinds:
3950 Constant *Iden = RecurrenceDescriptor::getRecurrenceIdentity(
3951 RK, VecTy->getScalarType());
3954 // This vector is the Identity vector where the first element is the
3955 // incoming scalar reduction.
3956 VectorStart = ReductionStartValue;
3958 Identity = ConstantVector::getSplat(VF, Iden);
3960 // This vector is the Identity vector where the first element is the
3961 // incoming scalar reduction.
3963 Builder.CreateInsertElement(Identity, ReductionStartValue, Zero);
3967 // Fix the vector-loop phi.
3969 // Reductions do not have to start at zero. They can start with
3970 // any loop invariant values.
3971 const VectorParts &VecRdxPhi = getVectorValue(Phi);
3972 BasicBlock *Latch = OrigLoop->getLoopLatch();
3973 Value *LoopVal = Phi->getIncomingValueForBlock(Latch);
3974 const VectorParts &Val = getVectorValue(LoopVal);
3975 for (unsigned part = 0; part < UF; ++part) {
3976 // Make sure to add the reduction stat value only to the
3977 // first unroll part.
3978 Value *StartVal = (part == 0) ? VectorStart : Identity;
3979 cast<PHINode>(VecRdxPhi[part])
3980 ->addIncoming(StartVal, LoopVectorPreHeader);
3981 cast<PHINode>(VecRdxPhi[part])
3982 ->addIncoming(Val[part], LoopVectorBody);
3985 // Before each round, move the insertion point right between
3986 // the PHIs and the values we are going to write.
3987 // This allows us to write both PHINodes and the extractelement
3989 Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt());
3991 VectorParts &RdxParts = VectorLoopValueMap.getVector(LoopExitInst);
3992 setDebugLocFromInst(Builder, LoopExitInst);
3994 // If the vector reduction can be performed in a smaller type, we truncate
3995 // then extend the loop exit value to enable InstCombine to evaluate the
3996 // entire expression in the smaller type.
3997 if (VF > 1 && Phi->getType() != RdxDesc.getRecurrenceType()) {
3998 Type *RdxVecTy = VectorType::get(RdxDesc.getRecurrenceType(), VF);
3999 Builder.SetInsertPoint(LoopVectorBody->getTerminator());
4000 for (unsigned part = 0; part < UF; ++part) {
4001 Value *Trunc = Builder.CreateTrunc(RdxParts[part], RdxVecTy);
4002 Value *Extnd = RdxDesc.isSigned() ? Builder.CreateSExt(Trunc, VecTy)
4003 : Builder.CreateZExt(Trunc, VecTy);
4004 for (Value::user_iterator UI = RdxParts[part]->user_begin();
4005 UI != RdxParts[part]->user_end();)
4007 (*UI++)->replaceUsesOfWith(RdxParts[part], Extnd);
4008 RdxParts[part] = Extnd;
4013 Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt());
4014 for (unsigned part = 0; part < UF; ++part)
4015 RdxParts[part] = Builder.CreateTrunc(RdxParts[part], RdxVecTy);
4018 // Reduce all of the unrolled parts into a single vector.
4019 Value *ReducedPartRdx = RdxParts[0];
4020 unsigned Op = RecurrenceDescriptor::getRecurrenceBinOp(RK);
4021 setDebugLocFromInst(Builder, ReducedPartRdx);
4022 for (unsigned part = 1; part < UF; ++part) {
4023 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
4024 // Floating point operations had to be 'fast' to enable the reduction.
4025 ReducedPartRdx = addFastMathFlag(
4026 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
4027 ReducedPartRdx, "bin.rdx"));
4029 ReducedPartRdx = RecurrenceDescriptor::createMinMaxOp(
4030 Builder, MinMaxKind, ReducedPartRdx, RdxParts[part]);
4034 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
4035 // and vector ops, reducing the set of values being computed by half each
4037 assert(isPowerOf2_32(VF) &&
4038 "Reduction emission only supported for pow2 vectors!");
4039 Value *TmpVec = ReducedPartRdx;
4040 SmallVector<Constant *, 32> ShuffleMask(VF, nullptr);
4041 for (unsigned i = VF; i != 1; i >>= 1) {
4042 // Move the upper half of the vector to the lower half.
4043 for (unsigned j = 0; j != i / 2; ++j)
4044 ShuffleMask[j] = Builder.getInt32(i / 2 + j);
4046 // Fill the rest of the mask with undef.
4047 std::fill(&ShuffleMask[i / 2], ShuffleMask.end(),
4048 UndefValue::get(Builder.getInt32Ty()));
4050 Value *Shuf = Builder.CreateShuffleVector(
4051 TmpVec, UndefValue::get(TmpVec->getType()),
4052 ConstantVector::get(ShuffleMask), "rdx.shuf");
4054 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
4055 // Floating point operations had to be 'fast' to enable the reduction.
4056 TmpVec = addFastMathFlag(Builder.CreateBinOp(
4057 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
4059 TmpVec = RecurrenceDescriptor::createMinMaxOp(Builder, MinMaxKind,
4063 // The result is in the first element of the vector.
4065 Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
4067 // If the reduction can be performed in a smaller type, we need to extend
4068 // the reduction to the wider type before we branch to the original loop.
4069 if (Phi->getType() != RdxDesc.getRecurrenceType())
4072 ? Builder.CreateSExt(ReducedPartRdx, Phi->getType())
4073 : Builder.CreateZExt(ReducedPartRdx, Phi->getType());
4076 // Create a phi node that merges control-flow from the backedge-taken check
4077 // block and the middle block.
4078 PHINode *BCBlockPhi = PHINode::Create(Phi->getType(), 2, "bc.merge.rdx",
4079 LoopScalarPreHeader->getTerminator());
4080 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
4081 BCBlockPhi->addIncoming(ReductionStartValue, LoopBypassBlocks[I]);
4082 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
4084 // Now, we need to fix the users of the reduction variable
4085 // inside and outside of the scalar remainder loop.
4086 // We know that the loop is in LCSSA form. We need to update the
4087 // PHI nodes in the exit blocks.
4088 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
4089 LEE = LoopExitBlock->end();
4090 LEI != LEE; ++LEI) {
4091 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
4095 // All PHINodes need to have a single entry edge, or two if
4096 // we already fixed them.
4097 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
4099 // We found our reduction value exit-PHI. Update it with the
4100 // incoming bypass edge.
4101 if (LCSSAPhi->getIncomingValue(0) == LoopExitInst) {
4102 // Add an edge coming from the bypass.
4103 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
4106 } // end of the LCSSA phi scan.
4108 // Fix the scalar loop reduction variable with the incoming reduction sum
4109 // from the vector body and from the backedge value.
4110 int IncomingEdgeBlockIdx =
4111 Phi->getBasicBlockIndex(OrigLoop->getLoopLatch());
4112 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
4113 // Pick the other block.
4114 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
4115 Phi->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
4116 Phi->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst);
4117 } // end of for each Phi in PHIsToFix.
4121 // Make sure DomTree is updated.
4124 predicateInstructions();
4126 // Remove redundant induction instructions.
4127 cse(LoopVectorBody);
4130 void InnerLoopVectorizer::fixFirstOrderRecurrence(PHINode *Phi) {
4132 // This is the second phase of vectorizing first-order recurrences. An
4133 // overview of the transformation is described below. Suppose we have the
4136 // for (int i = 0; i < n; ++i)
4137 // b[i] = a[i] - a[i - 1];
4139 // There is a first-order recurrence on "a". For this loop, the shorthand
4140 // scalar IR looks like:
4147 // i = phi [0, scalar.ph], [i+1, scalar.body]
4148 // s1 = phi [s_init, scalar.ph], [s2, scalar.body]
4151 // br cond, scalar.body, ...
4153 // In this example, s1 is a recurrence because it's value depends on the
4154 // previous iteration. In the first phase of vectorization, we created a
4155 // temporary value for s1. We now complete the vectorization and produce the
4156 // shorthand vector IR shown below (for VF = 4, UF = 1).
4159 // v_init = vector(..., ..., ..., a[-1])
4163 // i = phi [0, vector.ph], [i+4, vector.body]
4164 // v1 = phi [v_init, vector.ph], [v2, vector.body]
4165 // v2 = a[i, i+1, i+2, i+3];
4166 // v3 = vector(v1(3), v2(0, 1, 2))
4167 // b[i, i+1, i+2, i+3] = v2 - v3
4168 // br cond, vector.body, middle.block
4175 // s_init = phi [x, middle.block], [a[-1], otherwise]
4178 // After execution completes the vector loop, we extract the next value of
4179 // the recurrence (x) to use as the initial value in the scalar loop.
4181 // Get the original loop preheader and single loop latch.
4182 auto *Preheader = OrigLoop->getLoopPreheader();
4183 auto *Latch = OrigLoop->getLoopLatch();
4185 // Get the initial and previous values of the scalar recurrence.
4186 auto *ScalarInit = Phi->getIncomingValueForBlock(Preheader);
4187 auto *Previous = Phi->getIncomingValueForBlock(Latch);
4189 // Create a vector from the initial value.
4190 auto *VectorInit = ScalarInit;
4192 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
4193 VectorInit = Builder.CreateInsertElement(
4194 UndefValue::get(VectorType::get(VectorInit->getType(), VF)), VectorInit,
4195 Builder.getInt32(VF - 1), "vector.recur.init");
4198 // We constructed a temporary phi node in the first phase of vectorization.
4199 // This phi node will eventually be deleted.
4200 VectorParts &PhiParts = VectorLoopValueMap.getVector(Phi);
4201 Builder.SetInsertPoint(cast<Instruction>(PhiParts[0]));
4203 // Create a phi node for the new recurrence. The current value will either be
4204 // the initial value inserted into a vector or loop-varying vector value.
4205 auto *VecPhi = Builder.CreatePHI(VectorInit->getType(), 2, "vector.recur");
4206 VecPhi->addIncoming(VectorInit, LoopVectorPreHeader);
4208 // Get the vectorized previous value. We ensured the previous values was an
4209 // instruction when detecting the recurrence.
4210 auto &PreviousParts = getVectorValue(Previous);
4212 // Set the insertion point to be after this instruction. We ensured the
4213 // previous value dominated all uses of the phi when detecting the
4215 Builder.SetInsertPoint(
4216 &*++BasicBlock::iterator(cast<Instruction>(PreviousParts[UF - 1])));
4218 // We will construct a vector for the recurrence by combining the values for
4219 // the current and previous iterations. This is the required shuffle mask.
4220 SmallVector<Constant *, 8> ShuffleMask(VF);
4221 ShuffleMask[0] = Builder.getInt32(VF - 1);
4222 for (unsigned I = 1; I < VF; ++I)
4223 ShuffleMask[I] = Builder.getInt32(I + VF - 1);
4225 // The vector from which to take the initial value for the current iteration
4226 // (actual or unrolled). Initially, this is the vector phi node.
4227 Value *Incoming = VecPhi;
4229 // Shuffle the current and previous vector and update the vector parts.
4230 for (unsigned Part = 0; Part < UF; ++Part) {
4233 ? Builder.CreateShuffleVector(Incoming, PreviousParts[Part],
4234 ConstantVector::get(ShuffleMask))
4236 PhiParts[Part]->replaceAllUsesWith(Shuffle);
4237 cast<Instruction>(PhiParts[Part])->eraseFromParent();
4238 PhiParts[Part] = Shuffle;
4239 Incoming = PreviousParts[Part];
4242 // Fix the latch value of the new recurrence in the vector loop.
4243 VecPhi->addIncoming(Incoming, LI->getLoopFor(LoopVectorBody)->getLoopLatch());
4245 // Extract the last vector element in the middle block. This will be the
4246 // initial value for the recurrence when jumping to the scalar loop.
4247 auto *Extract = Incoming;
4249 Builder.SetInsertPoint(LoopMiddleBlock->getTerminator());
4250 Extract = Builder.CreateExtractElement(Extract, Builder.getInt32(VF - 1),
4251 "vector.recur.extract");
4254 // Fix the initial value of the original recurrence in the scalar loop.
4255 Builder.SetInsertPoint(&*LoopScalarPreHeader->begin());
4256 auto *Start = Builder.CreatePHI(Phi->getType(), 2, "scalar.recur.init");
4257 for (auto *BB : predecessors(LoopScalarPreHeader)) {
4258 auto *Incoming = BB == LoopMiddleBlock ? Extract : ScalarInit;
4259 Start->addIncoming(Incoming, BB);
4262 Phi->setIncomingValue(Phi->getBasicBlockIndex(LoopScalarPreHeader), Start);
4263 Phi->setName("scalar.recur");
4265 // Finally, fix users of the recurrence outside the loop. The users will need
4266 // either the last value of the scalar recurrence or the last value of the
4267 // vector recurrence we extracted in the middle block. Since the loop is in
4268 // LCSSA form, we just need to find the phi node for the original scalar
4269 // recurrence in the exit block, and then add an edge for the middle block.
4270 for (auto &I : *LoopExitBlock) {
4271 auto *LCSSAPhi = dyn_cast<PHINode>(&I);
4274 if (LCSSAPhi->getIncomingValue(0) == Phi) {
4275 LCSSAPhi->addIncoming(Extract, LoopMiddleBlock);
4281 void InnerLoopVectorizer::fixLCSSAPHIs() {
4282 for (Instruction &LEI : *LoopExitBlock) {
4283 auto *LCSSAPhi = dyn_cast<PHINode>(&LEI);
4286 if (LCSSAPhi->getNumIncomingValues() == 1)
4287 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
4292 void InnerLoopVectorizer::collectTriviallyDeadInstructions() {
4293 BasicBlock *Latch = OrigLoop->getLoopLatch();
4295 // We create new control-flow for the vectorized loop, so the original
4296 // condition will be dead after vectorization if it's only used by the
4298 auto *Cmp = dyn_cast<Instruction>(Latch->getTerminator()->getOperand(0));
4299 if (Cmp && Cmp->hasOneUse())
4300 DeadInstructions.insert(Cmp);
4302 // We create new "steps" for induction variable updates to which the original
4303 // induction variables map. An original update instruction will be dead if
4304 // all its users except the induction variable are dead.
4305 for (auto &Induction : *Legal->getInductionVars()) {
4306 PHINode *Ind = Induction.first;
4307 auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch));
4308 if (all_of(IndUpdate->users(), [&](User *U) -> bool {
4309 return U == Ind || DeadInstructions.count(cast<Instruction>(U));
4311 DeadInstructions.insert(IndUpdate);
4315 void InnerLoopVectorizer::sinkScalarOperands(Instruction *PredInst) {
4317 // The basic block and loop containing the predicated instruction.
4318 auto *PredBB = PredInst->getParent();
4319 auto *VectorLoop = LI->getLoopFor(PredBB);
4321 // Initialize a worklist with the operands of the predicated instruction.
4322 SetVector<Value *> Worklist(PredInst->op_begin(), PredInst->op_end());
4324 // Holds instructions that we need to analyze again. An instruction may be
4325 // reanalyzed if we don't yet know if we can sink it or not.
4326 SmallVector<Instruction *, 8> InstsToReanalyze;
4328 // Returns true if a given use occurs in the predicated block. Phi nodes use
4329 // their operands in their corresponding predecessor blocks.
4330 auto isBlockOfUsePredicated = [&](Use &U) -> bool {
4331 auto *I = cast<Instruction>(U.getUser());
4332 BasicBlock *BB = I->getParent();
4333 if (auto *Phi = dyn_cast<PHINode>(I))
4334 BB = Phi->getIncomingBlock(
4335 PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
4336 return BB == PredBB;
4339 // Iteratively sink the scalarized operands of the predicated instruction
4340 // into the block we created for it. When an instruction is sunk, it's
4341 // operands are then added to the worklist. The algorithm ends after one pass
4342 // through the worklist doesn't sink a single instruction.
4346 // Add the instructions that need to be reanalyzed to the worklist, and
4347 // reset the changed indicator.
4348 Worklist.insert(InstsToReanalyze.begin(), InstsToReanalyze.end());
4349 InstsToReanalyze.clear();
4352 while (!Worklist.empty()) {
4353 auto *I = dyn_cast<Instruction>(Worklist.pop_back_val());
4355 // We can't sink an instruction if it is a phi node, is already in the
4356 // predicated block, is not in the loop, or may have side effects.
4357 if (!I || isa<PHINode>(I) || I->getParent() == PredBB ||
4358 !VectorLoop->contains(I) || I->mayHaveSideEffects())
4361 // It's legal to sink the instruction if all its uses occur in the
4362 // predicated block. Otherwise, there's nothing to do yet, and we may
4363 // need to reanalyze the instruction.
4364 if (!all_of(I->uses(), isBlockOfUsePredicated)) {
4365 InstsToReanalyze.push_back(I);
4369 // Move the instruction to the beginning of the predicated block, and add
4370 // it's operands to the worklist.
4371 I->moveBefore(&*PredBB->getFirstInsertionPt());
4372 Worklist.insert(I->op_begin(), I->op_end());
4374 // The sinking may have enabled other instructions to be sunk, so we will
4381 void InnerLoopVectorizer::predicateInstructions() {
4383 // For each instruction I marked for predication on value C, split I into its
4384 // own basic block to form an if-then construct over C. Since I may be fed by
4385 // an extractelement instruction or other scalar operand, we try to
4386 // iteratively sink its scalar operands into the predicated block. If I feeds
4387 // an insertelement instruction, we try to move this instruction into the
4388 // predicated block as well. For non-void types, a phi node will be created
4389 // for the resulting value (either vector or scalar).
4391 // So for some predicated instruction, e.g. the conditional sdiv in:
4395 // %add = add nsw i32 %mul, %0
4396 // %cmp5 = icmp sgt i32 %2, 7
4397 // br i1 %cmp5, label %if.then, label %if.end
4400 // %div = sdiv i32 %0, %1
4404 // %x.0 = phi i32 [ %div, %if.then ], [ %add, %for.body ]
4406 // the sdiv at this point is scalarized and if-converted using a select.
4407 // The inactive elements in the vector are not used, but the predicated
4408 // instruction is still executed for all vector elements, essentially:
4412 // %17 = add nsw <2 x i32> %16, %wide.load
4413 // %29 = extractelement <2 x i32> %wide.load, i32 0
4414 // %30 = extractelement <2 x i32> %wide.load51, i32 0
4415 // %31 = sdiv i32 %29, %30
4416 // %32 = insertelement <2 x i32> undef, i32 %31, i32 0
4417 // %35 = extractelement <2 x i32> %wide.load, i32 1
4418 // %36 = extractelement <2 x i32> %wide.load51, i32 1
4419 // %37 = sdiv i32 %35, %36
4420 // %38 = insertelement <2 x i32> %32, i32 %37, i32 1
4421 // %predphi = select <2 x i1> %26, <2 x i32> %38, <2 x i32> %17
4423 // Predication will now re-introduce the original control flow to avoid false
4424 // side-effects by the sdiv instructions on the inactive elements, yielding
4429 // %5 = add nsw <2 x i32> %4, %wide.load
4430 // %8 = icmp sgt <2 x i32> %wide.load52, <i32 7, i32 7>
4431 // %9 = extractelement <2 x i1> %8, i32 0
4432 // br i1 %9, label %pred.sdiv.if, label %pred.sdiv.continue
4435 // %10 = extractelement <2 x i32> %wide.load, i32 0
4436 // %11 = extractelement <2 x i32> %wide.load51, i32 0
4437 // %12 = sdiv i32 %10, %11
4438 // %13 = insertelement <2 x i32> undef, i32 %12, i32 0
4439 // br label %pred.sdiv.continue
4441 // pred.sdiv.continue:
4442 // %14 = phi <2 x i32> [ undef, %vector.body ], [ %13, %pred.sdiv.if ]
4443 // %15 = extractelement <2 x i1> %8, i32 1
4444 // br i1 %15, label %pred.sdiv.if54, label %pred.sdiv.continue55
4447 // %16 = extractelement <2 x i32> %wide.load, i32 1
4448 // %17 = extractelement <2 x i32> %wide.load51, i32 1
4449 // %18 = sdiv i32 %16, %17
4450 // %19 = insertelement <2 x i32> %14, i32 %18, i32 1
4451 // br label %pred.sdiv.continue55
4453 // pred.sdiv.continue55:
4454 // %20 = phi <2 x i32> [ %14, %pred.sdiv.continue ], [ %19, %pred.sdiv.if54 ]
4455 // %predphi = select <2 x i1> %8, <2 x i32> %20, <2 x i32> %5
4457 for (auto KV : PredicatedInstructions) {
4458 BasicBlock::iterator I(KV.first);
4459 BasicBlock *Head = I->getParent();
4460 auto *BB = SplitBlock(Head, &*std::next(I), DT, LI);
4461 auto *T = SplitBlockAndInsertIfThen(KV.second, &*I, /*Unreachable=*/false,
4462 /*BranchWeights=*/nullptr, DT, LI);
4464 sinkScalarOperands(&*I);
4466 I->getParent()->setName(Twine("pred.") + I->getOpcodeName() + ".if");
4467 BB->setName(Twine("pred.") + I->getOpcodeName() + ".continue");
4469 // If the instruction is non-void create a Phi node at reconvergence point.
4470 if (!I->getType()->isVoidTy()) {
4471 Value *IncomingTrue = nullptr;
4472 Value *IncomingFalse = nullptr;
4474 if (I->hasOneUse() && isa<InsertElementInst>(*I->user_begin())) {
4475 // If the predicated instruction is feeding an insert-element, move it
4476 // into the Then block; Phi node will be created for the vector.
4477 InsertElementInst *IEI = cast<InsertElementInst>(*I->user_begin());
4479 IncomingTrue = IEI; // the new vector with the inserted element.
4480 IncomingFalse = IEI->getOperand(0); // the unmodified vector
4482 // Phi node will be created for the scalar predicated instruction.
4484 IncomingFalse = UndefValue::get(I->getType());
4487 BasicBlock *PostDom = I->getParent()->getSingleSuccessor();
4488 assert(PostDom && "Then block has multiple successors");
4490 PHINode::Create(IncomingTrue->getType(), 2, "", &PostDom->front());
4491 IncomingTrue->replaceAllUsesWith(Phi);
4492 Phi->addIncoming(IncomingFalse, Head);
4493 Phi->addIncoming(IncomingTrue, I->getParent());
4497 DEBUG(DT->verifyDomTree());
4500 InnerLoopVectorizer::VectorParts
4501 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
4502 assert(is_contained(predecessors(Dst), Src) && "Invalid edge");
4504 // Look for cached value.
4505 std::pair<BasicBlock *, BasicBlock *> Edge(Src, Dst);
4506 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
4507 if (ECEntryIt != MaskCache.end())
4508 return ECEntryIt->second;
4510 VectorParts SrcMask = createBlockInMask(Src);
4512 // The terminator has to be a branch inst!
4513 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
4514 assert(BI && "Unexpected terminator found");
4516 if (BI->isConditional()) {
4517 VectorParts EdgeMask = getVectorValue(BI->getCondition());
4519 if (BI->getSuccessor(0) != Dst)
4520 for (unsigned part = 0; part < UF; ++part)
4521 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
4523 for (unsigned part = 0; part < UF; ++part)
4524 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
4526 MaskCache[Edge] = EdgeMask;
4530 MaskCache[Edge] = SrcMask;
4534 InnerLoopVectorizer::VectorParts
4535 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
4536 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
4538 // Loop incoming mask is all-one.
4539 if (OrigLoop->getHeader() == BB) {
4540 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
4541 return getVectorValue(C);
4544 // This is the block mask. We OR all incoming edges, and with zero.
4545 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
4546 VectorParts BlockMask = getVectorValue(Zero);
4549 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
4550 VectorParts EM = createEdgeMask(*it, BB);
4551 for (unsigned part = 0; part < UF; ++part)
4552 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
4558 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN, unsigned UF,
4559 unsigned VF, PhiVector *PV) {
4560 PHINode *P = cast<PHINode>(PN);
4561 // Handle recurrences.
4562 if (Legal->isReductionVariable(P) || Legal->isFirstOrderRecurrence(P)) {
4563 VectorParts Entry(UF);
4564 for (unsigned part = 0; part < UF; ++part) {
4565 // This is phase one of vectorizing PHIs.
4567 (VF == 1) ? PN->getType() : VectorType::get(PN->getType(), VF);
4568 Entry[part] = PHINode::Create(
4569 VecTy, 2, "vec.phi", &*LoopVectorBody->getFirstInsertionPt());
4571 VectorLoopValueMap.initVector(P, Entry);
4576 setDebugLocFromInst(Builder, P);
4577 // Check for PHI nodes that are lowered to vector selects.
4578 if (P->getParent() != OrigLoop->getHeader()) {
4579 // We know that all PHIs in non-header blocks are converted into
4580 // selects, so we don't have to worry about the insertion order and we
4581 // can just use the builder.
4582 // At this point we generate the predication tree. There may be
4583 // duplications since this is a simple recursive scan, but future
4584 // optimizations will clean it up.
4586 unsigned NumIncoming = P->getNumIncomingValues();
4588 // Generate a sequence of selects of the form:
4589 // SELECT(Mask3, In3,
4590 // SELECT(Mask2, In2,
4592 VectorParts Entry(UF);
4593 for (unsigned In = 0; In < NumIncoming; In++) {
4595 createEdgeMask(P->getIncomingBlock(In), P->getParent());
4596 const VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
4598 for (unsigned part = 0; part < UF; ++part) {
4599 // We might have single edge PHIs (blocks) - use an identity
4600 // 'select' for the first PHI operand.
4602 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], In0[part]);
4604 // Select between the current value and the previous incoming edge
4605 // based on the incoming mask.
4606 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], Entry[part],
4610 VectorLoopValueMap.initVector(P, Entry);
4614 // This PHINode must be an induction variable.
4615 // Make sure that we know about it.
4616 assert(Legal->getInductionVars()->count(P) && "Not an induction variable");
4618 InductionDescriptor II = Legal->getInductionVars()->lookup(P);
4619 const DataLayout &DL = OrigLoop->getHeader()->getModule()->getDataLayout();
4621 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
4622 // which can be found from the original scalar operations.
4623 switch (II.getKind()) {
4624 case InductionDescriptor::IK_NoInduction:
4625 llvm_unreachable("Unknown induction");
4626 case InductionDescriptor::IK_IntInduction:
4627 return widenIntInduction(P);
4628 case InductionDescriptor::IK_PtrInduction: {
4629 // Handle the pointer induction variable case.
4630 assert(P->getType()->isPointerTy() && "Unexpected type.");
4631 // This is the normalized GEP that starts counting at zero.
4632 Value *PtrInd = Induction;
4633 PtrInd = Builder.CreateSExtOrTrunc(PtrInd, II.getStep()->getType());
4634 // Determine the number of scalars we need to generate for each unroll
4635 // iteration. If the instruction is uniform, we only need to generate the
4636 // first lane. Otherwise, we generate all VF values.
4637 unsigned Lanes = Legal->isUniformAfterVectorization(P) ? 1 : VF;
4638 // These are the scalar results. Notice that we don't generate vector GEPs
4639 // because scalar GEPs result in better code.
4640 ScalarParts Entry(UF);
4641 for (unsigned Part = 0; Part < UF; ++Part) {
4642 Entry[Part].resize(VF);
4643 for (unsigned Lane = 0; Lane < Lanes; ++Lane) {
4644 Constant *Idx = ConstantInt::get(PtrInd->getType(), Lane + Part * VF);
4645 Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx);
4646 Value *SclrGep = II.transform(Builder, GlobalIdx, PSE.getSE(), DL);
4647 SclrGep->setName("next.gep");
4648 Entry[Part][Lane] = SclrGep;
4651 VectorLoopValueMap.initScalar(P, Entry);
4654 case InductionDescriptor::IK_FpInduction: {
4655 assert(P->getType() == II.getStartValue()->getType() &&
4656 "Types must match");
4657 // Handle other induction variables that are now based on the
4659 assert(P != OldInduction && "Primary induction can be integer only");
4661 Value *V = Builder.CreateCast(Instruction::SIToFP, Induction, P->getType());
4662 V = II.transform(Builder, V, PSE.getSE(), DL);
4663 V->setName("fp.offset.idx");
4665 // Now we have scalar op: %fp.offset.idx = StartVal +/- Induction*StepVal
4667 Value *Broadcasted = getBroadcastInstrs(V);
4668 // After broadcasting the induction variable we need to make the vector
4669 // consecutive by adding StepVal*0, StepVal*1, StepVal*2, etc.
4670 Value *StepVal = cast<SCEVUnknown>(II.getStep())->getValue();
4671 VectorParts Entry(UF);
4672 for (unsigned part = 0; part < UF; ++part)
4673 Entry[part] = getStepVector(Broadcasted, VF * part, StepVal,
4674 II.getInductionOpcode());
4675 VectorLoopValueMap.initVector(P, Entry);
4681 /// A helper function for checking whether an integer division-related
4682 /// instruction may divide by zero (in which case it must be predicated if
4683 /// executed conditionally in the scalar code).
4684 /// TODO: It may be worthwhile to generalize and check isKnownNonZero().
4685 /// Non-zero divisors that are non compile-time constants will not be
4686 /// converted into multiplication, so we will still end up scalarizing
4687 /// the division, but can do so w/o predication.
4688 static bool mayDivideByZero(Instruction &I) {
4689 assert((I.getOpcode() == Instruction::UDiv ||
4690 I.getOpcode() == Instruction::SDiv ||
4691 I.getOpcode() == Instruction::URem ||
4692 I.getOpcode() == Instruction::SRem) &&
4693 "Unexpected instruction");
4694 Value *Divisor = I.getOperand(1);
4695 auto *CInt = dyn_cast<ConstantInt>(Divisor);
4696 return !CInt || CInt->isZero();
4699 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
4700 // For each instruction in the old loop.
4701 for (Instruction &I : *BB) {
4703 // If the instruction will become trivially dead when vectorized, we don't
4704 // need to generate it.
4705 if (DeadInstructions.count(&I))
4708 // Scalarize instructions that should remain scalar after vectorization.
4710 !(isa<BranchInst>(&I) || isa<PHINode>(&I) ||
4711 isa<DbgInfoIntrinsic>(&I)) &&
4712 shouldScalarizeInstruction(&I)) {
4713 scalarizeInstruction(&I, Legal->isScalarWithPredication(&I));
4717 switch (I.getOpcode()) {
4718 case Instruction::Br:
4719 // Nothing to do for PHIs and BR, since we already took care of the
4720 // loop control flow instructions.
4722 case Instruction::PHI: {
4723 // Vectorize PHINodes.
4724 widenPHIInstruction(&I, UF, VF, PV);
4728 case Instruction::UDiv:
4729 case Instruction::SDiv:
4730 case Instruction::SRem:
4731 case Instruction::URem:
4732 // Scalarize with predication if this instruction may divide by zero and
4733 // block execution is conditional, otherwise fallthrough.
4734 if (Legal->isScalarWithPredication(&I)) {
4735 scalarizeInstruction(&I, true);
4738 case Instruction::Add:
4739 case Instruction::FAdd:
4740 case Instruction::Sub:
4741 case Instruction::FSub:
4742 case Instruction::Mul:
4743 case Instruction::FMul:
4744 case Instruction::FDiv:
4745 case Instruction::FRem:
4746 case Instruction::Shl:
4747 case Instruction::LShr:
4748 case Instruction::AShr:
4749 case Instruction::And:
4750 case Instruction::Or:
4751 case Instruction::Xor: {
4752 // Just widen binops.
4753 auto *BinOp = cast<BinaryOperator>(&I);
4754 setDebugLocFromInst(Builder, BinOp);
4755 const VectorParts &A = getVectorValue(BinOp->getOperand(0));
4756 const VectorParts &B = getVectorValue(BinOp->getOperand(1));
4758 // Use this vector value for all users of the original instruction.
4759 VectorParts Entry(UF);
4760 for (unsigned Part = 0; Part < UF; ++Part) {
4761 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
4763 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
4764 VecOp->copyIRFlags(BinOp);
4769 VectorLoopValueMap.initVector(&I, Entry);
4770 addMetadata(Entry, BinOp);
4773 case Instruction::Select: {
4775 // If the selector is loop invariant we can create a select
4776 // instruction with a scalar condition. Otherwise, use vector-select.
4777 auto *SE = PSE.getSE();
4778 bool InvariantCond =
4779 SE->isLoopInvariant(PSE.getSCEV(I.getOperand(0)), OrigLoop);
4780 setDebugLocFromInst(Builder, &I);
4782 // The condition can be loop invariant but still defined inside the
4783 // loop. This means that we can't just use the original 'cond' value.
4784 // We have to take the 'vectorized' value and pick the first lane.
4785 // Instcombine will make this a no-op.
4786 const VectorParts &Cond = getVectorValue(I.getOperand(0));
4787 const VectorParts &Op0 = getVectorValue(I.getOperand(1));
4788 const VectorParts &Op1 = getVectorValue(I.getOperand(2));
4790 auto *ScalarCond = getScalarValue(I.getOperand(0), 0, 0);
4792 VectorParts Entry(UF);
4793 for (unsigned Part = 0; Part < UF; ++Part) {
4794 Entry[Part] = Builder.CreateSelect(
4795 InvariantCond ? ScalarCond : Cond[Part], Op0[Part], Op1[Part]);
4798 VectorLoopValueMap.initVector(&I, Entry);
4799 addMetadata(Entry, &I);
4803 case Instruction::ICmp:
4804 case Instruction::FCmp: {
4805 // Widen compares. Generate vector compares.
4806 bool FCmp = (I.getOpcode() == Instruction::FCmp);
4807 auto *Cmp = dyn_cast<CmpInst>(&I);
4808 setDebugLocFromInst(Builder, Cmp);
4809 const VectorParts &A = getVectorValue(Cmp->getOperand(0));
4810 const VectorParts &B = getVectorValue(Cmp->getOperand(1));
4811 VectorParts Entry(UF);
4812 for (unsigned Part = 0; Part < UF; ++Part) {
4815 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
4816 cast<FCmpInst>(C)->copyFastMathFlags(Cmp);
4818 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
4823 VectorLoopValueMap.initVector(&I, Entry);
4824 addMetadata(Entry, &I);
4828 case Instruction::Store:
4829 case Instruction::Load:
4830 vectorizeMemoryInstruction(&I);
4832 case Instruction::ZExt:
4833 case Instruction::SExt:
4834 case Instruction::FPToUI:
4835 case Instruction::FPToSI:
4836 case Instruction::FPExt:
4837 case Instruction::PtrToInt:
4838 case Instruction::IntToPtr:
4839 case Instruction::SIToFP:
4840 case Instruction::UIToFP:
4841 case Instruction::Trunc:
4842 case Instruction::FPTrunc:
4843 case Instruction::BitCast: {
4844 auto *CI = dyn_cast<CastInst>(&I);
4845 setDebugLocFromInst(Builder, CI);
4847 // Optimize the special case where the source is a constant integer
4848 // induction variable. Notice that we can only optimize the 'trunc' case
4849 // because (a) FP conversions lose precision, (b) sext/zext may wrap, and
4850 // (c) other casts depend on pointer size.
4851 auto ID = Legal->getInductionVars()->lookup(OldInduction);
4852 if (isa<TruncInst>(CI) && CI->getOperand(0) == OldInduction &&
4853 ID.getConstIntStepValue()) {
4854 widenIntInduction(OldInduction, cast<TruncInst>(CI));
4858 /// Vectorize casts.
4860 (VF == 1) ? CI->getType() : VectorType::get(CI->getType(), VF);
4862 const VectorParts &A = getVectorValue(CI->getOperand(0));
4863 VectorParts Entry(UF);
4864 for (unsigned Part = 0; Part < UF; ++Part)
4865 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
4866 VectorLoopValueMap.initVector(&I, Entry);
4867 addMetadata(Entry, &I);
4871 case Instruction::Call: {
4872 // Ignore dbg intrinsics.
4873 if (isa<DbgInfoIntrinsic>(I))
4875 setDebugLocFromInst(Builder, &I);
4877 Module *M = BB->getParent()->getParent();
4878 auto *CI = cast<CallInst>(&I);
4880 StringRef FnName = CI->getCalledFunction()->getName();
4881 Function *F = CI->getCalledFunction();
4882 Type *RetTy = ToVectorTy(CI->getType(), VF);
4883 SmallVector<Type *, 4> Tys;
4884 for (Value *ArgOperand : CI->arg_operands())
4885 Tys.push_back(ToVectorTy(ArgOperand->getType(), VF));
4887 Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
4888 if (ID && (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end ||
4889 ID == Intrinsic::lifetime_start)) {
4890 scalarizeInstruction(&I);
4893 // The flag shows whether we use Intrinsic or a usual Call for vectorized
4894 // version of the instruction.
4895 // Is it beneficial to perform intrinsic call compared to lib call?
4896 bool NeedToScalarize;
4897 unsigned CallCost = getVectorCallCost(CI, VF, *TTI, TLI, NeedToScalarize);
4898 bool UseVectorIntrinsic =
4899 ID && getVectorIntrinsicCost(CI, VF, *TTI, TLI) <= CallCost;
4900 if (!UseVectorIntrinsic && NeedToScalarize) {
4901 scalarizeInstruction(&I);
4905 VectorParts Entry(UF);
4906 for (unsigned Part = 0; Part < UF; ++Part) {
4907 SmallVector<Value *, 4> Args;
4908 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
4909 Value *Arg = CI->getArgOperand(i);
4910 // Some intrinsics have a scalar argument - don't replace it with a
4912 if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, i)) {
4913 const VectorParts &VectorArg = getVectorValue(CI->getArgOperand(i));
4914 Arg = VectorArg[Part];
4916 Args.push_back(Arg);
4920 if (UseVectorIntrinsic) {
4921 // Use vector version of the intrinsic.
4922 Type *TysForDecl[] = {CI->getType()};
4924 TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF);
4925 VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl);
4927 // Use vector version of the library call.
4928 StringRef VFnName = TLI->getVectorizedFunction(FnName, VF);
4929 assert(!VFnName.empty() && "Vector function name is empty.");
4930 VectorF = M->getFunction(VFnName);
4932 // Generate a declaration
4933 FunctionType *FTy = FunctionType::get(RetTy, Tys, false);
4935 Function::Create(FTy, Function::ExternalLinkage, VFnName, M);
4936 VectorF->copyAttributesFrom(F);
4939 assert(VectorF && "Can't create vector function.");
4941 SmallVector<OperandBundleDef, 1> OpBundles;
4942 CI->getOperandBundlesAsDefs(OpBundles);
4943 CallInst *V = Builder.CreateCall(VectorF, Args, OpBundles);
4945 if (isa<FPMathOperator>(V))
4946 V->copyFastMathFlags(CI);
4951 VectorLoopValueMap.initVector(&I, Entry);
4952 addMetadata(Entry, &I);
4957 // All other instructions are unsupported. Scalarize them.
4958 scalarizeInstruction(&I);
4961 } // end of for_each instr.
4964 void InnerLoopVectorizer::updateAnalysis() {
4965 // Forget the original basic block.
4966 PSE.getSE()->forgetLoop(OrigLoop);
4968 // Update the dominator tree information.
4969 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
4970 "Entry does not dominate exit.");
4972 // We don't predicate stores by this point, so the vector body should be a
4974 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
4976 DT->addNewBlock(LoopMiddleBlock, LoopVectorBody);
4977 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
4978 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
4979 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
4981 DEBUG(DT->verifyDomTree());
4984 /// \brief Check whether it is safe to if-convert this phi node.
4986 /// Phi nodes with constant expressions that can trap are not safe to if
4988 static bool canIfConvertPHINodes(BasicBlock *BB) {
4989 for (Instruction &I : *BB) {
4990 auto *Phi = dyn_cast<PHINode>(&I);
4993 for (Value *V : Phi->incoming_values())
4994 if (auto *C = dyn_cast<Constant>(V))
5001 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
5002 if (!EnableIfConversion) {
5003 ORE->emit(createMissedAnalysis("IfConversionDisabled")
5004 << "if-conversion is disabled");
5008 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
5010 // A list of pointers that we can safely read and write to.
5011 SmallPtrSet<Value *, 8> SafePointes;
5013 // Collect safe addresses.
5014 for (BasicBlock *BB : TheLoop->blocks()) {
5015 if (blockNeedsPredication(BB))
5018 for (Instruction &I : *BB)
5019 if (auto *Ptr = getPointerOperand(&I))
5020 SafePointes.insert(Ptr);
5023 // Collect the blocks that need predication.
5024 BasicBlock *Header = TheLoop->getHeader();
5025 for (BasicBlock *BB : TheLoop->blocks()) {
5026 // We don't support switch statements inside loops.
5027 if (!isa<BranchInst>(BB->getTerminator())) {
5028 ORE->emit(createMissedAnalysis("LoopContainsSwitch", BB->getTerminator())
5029 << "loop contains a switch statement");
5033 // We must be able to predicate all blocks that need to be predicated.
5034 if (blockNeedsPredication(BB)) {
5035 if (!blockCanBePredicated(BB, SafePointes)) {
5036 ORE->emit(createMissedAnalysis("NoCFGForSelect", BB->getTerminator())
5037 << "control flow cannot be substituted for a select");
5040 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
5041 ORE->emit(createMissedAnalysis("NoCFGForSelect", BB->getTerminator())
5042 << "control flow cannot be substituted for a select");
5047 // We can if-convert this loop.
5051 bool LoopVectorizationLegality::canVectorize() {
5052 // We must have a loop in canonical form. Loops with indirectbr in them cannot
5053 // be canonicalized.
5054 if (!TheLoop->getLoopPreheader()) {
5055 ORE->emit(createMissedAnalysis("CFGNotUnderstood")
5056 << "loop control flow is not understood by vectorizer");
5060 // FIXME: The code is currently dead, since the loop gets sent to
5061 // LoopVectorizationLegality is already an innermost loop.
5063 // We can only vectorize innermost loops.
5064 if (!TheLoop->empty()) {
5065 ORE->emit(createMissedAnalysis("NotInnermostLoop")
5066 << "loop is not the innermost loop");
5070 // We must have a single backedge.
5071 if (TheLoop->getNumBackEdges() != 1) {
5072 ORE->emit(createMissedAnalysis("CFGNotUnderstood")
5073 << "loop control flow is not understood by vectorizer");
5077 // We must have a single exiting block.
5078 if (!TheLoop->getExitingBlock()) {
5079 ORE->emit(createMissedAnalysis("CFGNotUnderstood")
5080 << "loop control flow is not understood by vectorizer");
5084 // We only handle bottom-tested loops, i.e. loop in which the condition is
5085 // checked at the end of each iteration. With that we can assume that all
5086 // instructions in the loop are executed the same number of times.
5087 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
5088 ORE->emit(createMissedAnalysis("CFGNotUnderstood")
5089 << "loop control flow is not understood by vectorizer");
5093 // We need to have a loop header.
5094 DEBUG(dbgs() << "LV: Found a loop: " << TheLoop->getHeader()->getName()
5097 // Check if we can if-convert non-single-bb loops.
5098 unsigned NumBlocks = TheLoop->getNumBlocks();
5099 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
5100 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
5104 // ScalarEvolution needs to be able to find the exit count.
5105 const SCEV *ExitCount = PSE.getBackedgeTakenCount();
5106 if (ExitCount == PSE.getSE()->getCouldNotCompute()) {
5107 ORE->emit(createMissedAnalysis("CantComputeNumberOfIterations")
5108 << "could not determine number of loop iterations");
5109 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
5113 // Check if we can vectorize the instructions and CFG in this loop.
5114 if (!canVectorizeInstrs()) {
5115 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
5119 // Go over each instruction and look at memory deps.
5120 if (!canVectorizeMemory()) {
5121 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
5125 DEBUG(dbgs() << "LV: We can vectorize this loop"
5126 << (LAI->getRuntimePointerChecking()->Need
5127 ? " (with a runtime bound check)"
5131 bool UseInterleaved = TTI->enableInterleavedAccessVectorization();
5133 // If an override option has been passed in for interleaved accesses, use it.
5134 if (EnableInterleavedMemAccesses.getNumOccurrences() > 0)
5135 UseInterleaved = EnableInterleavedMemAccesses;
5137 // Analyze interleaved memory accesses.
5139 InterleaveInfo.analyzeInterleaving(*getSymbolicStrides());
5141 // Collect all instructions that are known to be uniform after vectorization.
5142 collectLoopUniforms();
5144 // Collect all instructions that are known to be scalar after vectorization.
5145 collectLoopScalars();
5147 unsigned SCEVThreshold = VectorizeSCEVCheckThreshold;
5148 if (Hints->getForce() == LoopVectorizeHints::FK_Enabled)
5149 SCEVThreshold = PragmaVectorizeSCEVCheckThreshold;
5151 if (PSE.getUnionPredicate().getComplexity() > SCEVThreshold) {
5152 ORE->emit(createMissedAnalysis("TooManySCEVRunTimeChecks")
5153 << "Too many SCEV assumptions need to be made and checked "
5155 DEBUG(dbgs() << "LV: Too many SCEV checks needed.\n");
5159 // Okay! We can vectorize. At this point we don't have any other mem analysis
5160 // which may limit our maximum vectorization factor, so just return true with
5165 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
5166 if (Ty->isPointerTy())
5167 return DL.getIntPtrType(Ty);
5169 // It is possible that char's or short's overflow when we ask for the loop's
5170 // trip count, work around this by changing the type size.
5171 if (Ty->getScalarSizeInBits() < 32)
5172 return Type::getInt32Ty(Ty->getContext());
5177 static Type *getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
5178 Ty0 = convertPointerToIntegerType(DL, Ty0);
5179 Ty1 = convertPointerToIntegerType(DL, Ty1);
5180 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
5185 /// \brief Check that the instruction has outside loop users and is not an
5186 /// identified reduction variable.
5187 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
5188 SmallPtrSetImpl<Value *> &AllowedExit) {
5189 // Reduction and Induction instructions are allowed to have exit users. All
5190 // other instructions must not have external users.
5191 if (!AllowedExit.count(Inst))
5192 // Check that all of the users of the loop are inside the BB.
5193 for (User *U : Inst->users()) {
5194 Instruction *UI = cast<Instruction>(U);
5195 // This user may be a reduction exit value.
5196 if (!TheLoop->contains(UI)) {
5197 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
5204 void LoopVectorizationLegality::addInductionPhi(
5205 PHINode *Phi, const InductionDescriptor &ID,
5206 SmallPtrSetImpl<Value *> &AllowedExit) {
5207 Inductions[Phi] = ID;
5208 Type *PhiTy = Phi->getType();
5209 const DataLayout &DL = Phi->getModule()->getDataLayout();
5211 // Get the widest type.
5212 if (!PhiTy->isFloatingPointTy()) {
5214 WidestIndTy = convertPointerToIntegerType(DL, PhiTy);
5216 WidestIndTy = getWiderType(DL, PhiTy, WidestIndTy);
5219 // Int inductions are special because we only allow one IV.
5220 if (ID.getKind() == InductionDescriptor::IK_IntInduction &&
5221 ID.getConstIntStepValue() &&
5222 ID.getConstIntStepValue()->isOne() &&
5223 isa<Constant>(ID.getStartValue()) &&
5224 cast<Constant>(ID.getStartValue())->isNullValue()) {
5226 // Use the phi node with the widest type as induction. Use the last
5227 // one if there are multiple (no good reason for doing this other
5228 // than it is expedient). We've checked that it begins at zero and
5229 // steps by one, so this is a canonical induction variable.
5230 if (!Induction || PhiTy == WidestIndTy)
5234 // Both the PHI node itself, and the "post-increment" value feeding
5235 // back into the PHI node may have external users.
5236 AllowedExit.insert(Phi);
5237 AllowedExit.insert(Phi->getIncomingValueForBlock(TheLoop->getLoopLatch()));
5239 DEBUG(dbgs() << "LV: Found an induction variable.\n");
5243 bool LoopVectorizationLegality::canVectorizeInstrs() {
5244 BasicBlock *Header = TheLoop->getHeader();
5246 // Look for the attribute signaling the absence of NaNs.
5247 Function &F = *Header->getParent();
5249 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
5251 // For each block in the loop.
5252 for (BasicBlock *BB : TheLoop->blocks()) {
5253 // Scan the instructions in the block and look for hazards.
5254 for (Instruction &I : *BB) {
5255 if (auto *Phi = dyn_cast<PHINode>(&I)) {
5256 Type *PhiTy = Phi->getType();
5257 // Check that this PHI type is allowed.
5258 if (!PhiTy->isIntegerTy() && !PhiTy->isFloatingPointTy() &&
5259 !PhiTy->isPointerTy()) {
5260 ORE->emit(createMissedAnalysis("CFGNotUnderstood", Phi)
5261 << "loop control flow is not understood by vectorizer");
5262 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
5266 // If this PHINode is not in the header block, then we know that we
5267 // can convert it to select during if-conversion. No need to check if
5268 // the PHIs in this block are induction or reduction variables.
5270 // Check that this instruction has no outside users or is an
5271 // identified reduction value with an outside user.
5272 if (!hasOutsideLoopUser(TheLoop, Phi, AllowedExit))
5274 ORE->emit(createMissedAnalysis("NeitherInductionNorReduction", Phi)
5275 << "value could not be identified as "
5276 "an induction or reduction variable");
5280 // We only allow if-converted PHIs with exactly two incoming values.
5281 if (Phi->getNumIncomingValues() != 2) {
5282 ORE->emit(createMissedAnalysis("CFGNotUnderstood", Phi)
5283 << "control flow not understood by vectorizer");
5284 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
5288 RecurrenceDescriptor RedDes;
5289 if (RecurrenceDescriptor::isReductionPHI(Phi, TheLoop, RedDes)) {
5290 if (RedDes.hasUnsafeAlgebra())
5291 Requirements->addUnsafeAlgebraInst(RedDes.getUnsafeAlgebraInst());
5292 AllowedExit.insert(RedDes.getLoopExitInstr());
5293 Reductions[Phi] = RedDes;
5297 InductionDescriptor ID;
5298 if (InductionDescriptor::isInductionPHI(Phi, TheLoop, PSE, ID)) {
5299 addInductionPhi(Phi, ID, AllowedExit);
5300 if (ID.hasUnsafeAlgebra() && !HasFunNoNaNAttr)
5301 Requirements->addUnsafeAlgebraInst(ID.getUnsafeAlgebraInst());
5305 if (RecurrenceDescriptor::isFirstOrderRecurrence(Phi, TheLoop, DT)) {
5306 FirstOrderRecurrences.insert(Phi);
5310 // As a last resort, coerce the PHI to a AddRec expression
5311 // and re-try classifying it a an induction PHI.
5312 if (InductionDescriptor::isInductionPHI(Phi, TheLoop, PSE, ID, true)) {
5313 addInductionPhi(Phi, ID, AllowedExit);
5317 ORE->emit(createMissedAnalysis("NonReductionValueUsedOutsideLoop", Phi)
5318 << "value that could not be identified as "
5319 "reduction is used outside the loop");
5320 DEBUG(dbgs() << "LV: Found an unidentified PHI." << *Phi << "\n");
5322 } // end of PHI handling
5324 // We handle calls that:
5325 // * Are debug info intrinsics.
5326 // * Have a mapping to an IR intrinsic.
5327 // * Have a vector version available.
5328 auto *CI = dyn_cast<CallInst>(&I);
5329 if (CI && !getVectorIntrinsicIDForCall(CI, TLI) &&
5330 !isa<DbgInfoIntrinsic>(CI) &&
5331 !(CI->getCalledFunction() && TLI &&
5332 TLI->isFunctionVectorizable(CI->getCalledFunction()->getName()))) {
5333 ORE->emit(createMissedAnalysis("CantVectorizeCall", CI)
5334 << "call instruction cannot be vectorized");
5335 DEBUG(dbgs() << "LV: Found a non-intrinsic, non-libfunc callsite.\n");
5339 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
5340 // second argument is the same (i.e. loop invariant)
5341 if (CI && hasVectorInstrinsicScalarOpd(
5342 getVectorIntrinsicIDForCall(CI, TLI), 1)) {
5343 auto *SE = PSE.getSE();
5344 if (!SE->isLoopInvariant(PSE.getSCEV(CI->getOperand(1)), TheLoop)) {
5345 ORE->emit(createMissedAnalysis("CantVectorizeIntrinsic", CI)
5346 << "intrinsic instruction cannot be vectorized");
5347 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
5352 // Check that the instruction return type is vectorizable.
5353 // Also, we can't vectorize extractelement instructions.
5354 if ((!VectorType::isValidElementType(I.getType()) &&
5355 !I.getType()->isVoidTy()) ||
5356 isa<ExtractElementInst>(I)) {
5357 ORE->emit(createMissedAnalysis("CantVectorizeInstructionReturnType", &I)
5358 << "instruction return type cannot be vectorized");
5359 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
5363 // Check that the stored type is vectorizable.
5364 if (auto *ST = dyn_cast<StoreInst>(&I)) {
5365 Type *T = ST->getValueOperand()->getType();
5366 if (!VectorType::isValidElementType(T)) {
5367 ORE->emit(createMissedAnalysis("CantVectorizeStore", ST)
5368 << "store instruction cannot be vectorized");
5372 // FP instructions can allow unsafe algebra, thus vectorizable by
5373 // non-IEEE-754 compliant SIMD units.
5374 // This applies to floating-point math operations and calls, not memory
5375 // operations, shuffles, or casts, as they don't change precision or
5377 } else if (I.getType()->isFloatingPointTy() && (CI || I.isBinaryOp()) &&
5378 !I.hasUnsafeAlgebra()) {
5379 DEBUG(dbgs() << "LV: Found FP op with unsafe algebra.\n");
5380 Hints->setPotentiallyUnsafe();
5383 // Reduction instructions are allowed to have exit users.
5384 // All other instructions must not have external users.
5385 if (hasOutsideLoopUser(TheLoop, &I, AllowedExit)) {
5386 ORE->emit(createMissedAnalysis("ValueUsedOutsideLoop", &I)
5387 << "value cannot be used outside the loop");
5395 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
5396 if (Inductions.empty()) {
5397 ORE->emit(createMissedAnalysis("NoInductionVariable")
5398 << "loop induction variable could not be identified");
5403 // Now we know the widest induction type, check if our found induction
5404 // is the same size. If it's not, unset it here and InnerLoopVectorizer
5405 // will create another.
5406 if (Induction && WidestIndTy != Induction->getType())
5407 Induction = nullptr;
5412 void LoopVectorizationLegality::collectLoopScalars() {
5414 // If an instruction is uniform after vectorization, it will remain scalar.
5415 Scalars.insert(Uniforms.begin(), Uniforms.end());
5417 // Collect the getelementptr instructions that will not be vectorized. A
5418 // getelementptr instruction is only vectorized if it is used for a legal
5419 // gather or scatter operation.
5420 for (auto *BB : TheLoop->blocks())
5421 for (auto &I : *BB) {
5422 if (auto *GEP = dyn_cast<GetElementPtrInst>(&I)) {
5423 Scalars.insert(GEP);
5426 auto *Ptr = getPointerOperand(&I);
5429 auto *GEP = getGEPInstruction(Ptr);
5430 if (GEP && isLegalGatherOrScatter(&I))
5434 // An induction variable will remain scalar if all users of the induction
5435 // variable and induction variable update remain scalar.
5436 auto *Latch = TheLoop->getLoopLatch();
5437 for (auto &Induction : *getInductionVars()) {
5438 auto *Ind = Induction.first;
5439 auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch));
5441 // Determine if all users of the induction variable are scalar after
5443 auto ScalarInd = all_of(Ind->users(), [&](User *U) -> bool {
5444 auto *I = cast<Instruction>(U);
5445 return I == IndUpdate || !TheLoop->contains(I) || Scalars.count(I);
5450 // Determine if all users of the induction variable update instruction are
5451 // scalar after vectorization.
5452 auto ScalarIndUpdate = all_of(IndUpdate->users(), [&](User *U) -> bool {
5453 auto *I = cast<Instruction>(U);
5454 return I == Ind || !TheLoop->contains(I) || Scalars.count(I);
5456 if (!ScalarIndUpdate)
5459 // The induction variable and its update instruction will remain scalar.
5460 Scalars.insert(Ind);
5461 Scalars.insert(IndUpdate);
5465 bool LoopVectorizationLegality::hasConsecutiveLikePtrOperand(Instruction *I) {
5466 if (isAccessInterleaved(I))
5468 if (auto *Ptr = getPointerOperand(I))
5469 return isConsecutivePtr(Ptr);
5473 bool LoopVectorizationLegality::isScalarWithPredication(Instruction *I) {
5474 if (!blockNeedsPredication(I->getParent()))
5476 switch(I->getOpcode()) {
5479 case Instruction::Store:
5480 return !isMaskRequired(I);
5481 case Instruction::UDiv:
5482 case Instruction::SDiv:
5483 case Instruction::SRem:
5484 case Instruction::URem:
5485 return mayDivideByZero(*I);
5490 bool LoopVectorizationLegality::memoryInstructionMustBeScalarized(
5491 Instruction *I, unsigned VF) {
5493 // If the memory instruction is in an interleaved group, it will be
5494 // vectorized and its pointer will remain uniform.
5495 if (isAccessInterleaved(I))
5498 // Get and ensure we have a valid memory instruction.
5499 LoadInst *LI = dyn_cast<LoadInst>(I);
5500 StoreInst *SI = dyn_cast<StoreInst>(I);
5501 assert((LI || SI) && "Invalid memory instruction");
5503 // If the pointer operand is uniform (loop invariant), the memory instruction
5504 // will be scalarized.
5505 auto *Ptr = getPointerOperand(I);
5506 if (LI && isUniform(Ptr))
5509 // If the pointer operand is non-consecutive and neither a gather nor a
5510 // scatter operation is legal, the memory instruction will be scalarized.
5511 if (!isConsecutivePtr(Ptr) && !isLegalGatherOrScatter(I))
5514 // If the instruction is a store located in a predicated block, it will be
5516 if (isScalarWithPredication(I))
5519 // If the instruction's allocated size doesn't equal it's type size, it
5520 // requires padding and will be scalarized.
5521 auto &DL = I->getModule()->getDataLayout();
5522 auto *ScalarTy = LI ? LI->getType() : SI->getValueOperand()->getType();
5523 if (hasIrregularType(ScalarTy, DL, VF))
5526 // Otherwise, the memory instruction should be vectorized if the rest of the
5531 void LoopVectorizationLegality::collectLoopUniforms() {
5532 // We now know that the loop is vectorizable!
5533 // Collect instructions inside the loop that will remain uniform after
5536 // Global values, params and instructions outside of current loop are out of
5538 auto isOutOfScope = [&](Value *V) -> bool {
5539 Instruction *I = dyn_cast<Instruction>(V);
5540 return (!I || !TheLoop->contains(I));
5543 SetVector<Instruction *> Worklist;
5544 BasicBlock *Latch = TheLoop->getLoopLatch();
5546 // Start with the conditional branch. If the branch condition is an
5547 // instruction contained in the loop that is only used by the branch, it is
5549 auto *Cmp = dyn_cast<Instruction>(Latch->getTerminator()->getOperand(0));
5550 if (Cmp && TheLoop->contains(Cmp) && Cmp->hasOneUse()) {
5551 Worklist.insert(Cmp);
5552 DEBUG(dbgs() << "LV: Found uniform instruction: " << *Cmp << "\n");
5555 // Holds consecutive and consecutive-like pointers. Consecutive-like pointers
5556 // are pointers that are treated like consecutive pointers during
5557 // vectorization. The pointer operands of interleaved accesses are an
5559 SmallSetVector<Instruction *, 8> ConsecutiveLikePtrs;
5561 // Holds pointer operands of instructions that are possibly non-uniform.
5562 SmallPtrSet<Instruction *, 8> PossibleNonUniformPtrs;
5564 // Iterate over the instructions in the loop, and collect all
5565 // consecutive-like pointer operands in ConsecutiveLikePtrs. If it's possible
5566 // that a consecutive-like pointer operand will be scalarized, we collect it
5567 // in PossibleNonUniformPtrs instead. We use two sets here because a single
5568 // getelementptr instruction can be used by both vectorized and scalarized
5569 // memory instructions. For example, if a loop loads and stores from the same
5570 // location, but the store is conditional, the store will be scalarized, and
5571 // the getelementptr won't remain uniform.
5572 for (auto *BB : TheLoop->blocks())
5573 for (auto &I : *BB) {
5575 // If there's no pointer operand, there's nothing to do.
5576 auto *Ptr = dyn_cast_or_null<Instruction>(getPointerOperand(&I));
5580 // True if all users of Ptr are memory accesses that have Ptr as their
5582 auto UsersAreMemAccesses = all_of(Ptr->users(), [&](User *U) -> bool {
5583 return getPointerOperand(U) == Ptr;
5586 // Ensure the memory instruction will not be scalarized, making its
5587 // pointer operand non-uniform. If the pointer operand is used by some
5588 // instruction other than a memory access, we're not going to check if
5589 // that other instruction may be scalarized here. Thus, conservatively
5590 // assume the pointer operand may be non-uniform.
5591 if (!UsersAreMemAccesses || memoryInstructionMustBeScalarized(&I))
5592 PossibleNonUniformPtrs.insert(Ptr);
5594 // If the memory instruction will be vectorized and its pointer operand
5595 // is consecutive-like, the pointer operand should remain uniform.
5596 else if (hasConsecutiveLikePtrOperand(&I))
5597 ConsecutiveLikePtrs.insert(Ptr);
5600 // Add to the Worklist all consecutive and consecutive-like pointers that
5601 // aren't also identified as possibly non-uniform.
5602 for (auto *V : ConsecutiveLikePtrs)
5603 if (!PossibleNonUniformPtrs.count(V)) {
5604 DEBUG(dbgs() << "LV: Found uniform instruction: " << *V << "\n");
5608 // Expand Worklist in topological order: whenever a new instruction
5609 // is added , its users should be either already inside Worklist, or
5610 // out of scope. It ensures a uniform instruction will only be used
5611 // by uniform instructions or out of scope instructions.
5613 while (idx != Worklist.size()) {
5614 Instruction *I = Worklist[idx++];
5616 for (auto OV : I->operand_values()) {
5617 if (isOutOfScope(OV))
5619 auto *OI = cast<Instruction>(OV);
5620 if (all_of(OI->users(), [&](User *U) -> bool {
5621 return isOutOfScope(U) || Worklist.count(cast<Instruction>(U));
5623 Worklist.insert(OI);
5624 DEBUG(dbgs() << "LV: Found uniform instruction: " << *OI << "\n");
5629 // Returns true if Ptr is the pointer operand of a memory access instruction
5630 // I, and I is known to not require scalarization.
5631 auto isVectorizedMemAccessUse = [&](Instruction *I, Value *Ptr) -> bool {
5632 return getPointerOperand(I) == Ptr && !memoryInstructionMustBeScalarized(I);
5635 // For an instruction to be added into Worklist above, all its users inside
5636 // the loop should also be in Worklist. However, this condition cannot be
5637 // true for phi nodes that form a cyclic dependence. We must process phi
5638 // nodes separately. An induction variable will remain uniform if all users
5639 // of the induction variable and induction variable update remain uniform.
5640 // The code below handles both pointer and non-pointer induction variables.
5641 for (auto &Induction : Inductions) {
5642 auto *Ind = Induction.first;
5643 auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch));
5645 // Determine if all users of the induction variable are uniform after
5647 auto UniformInd = all_of(Ind->users(), [&](User *U) -> bool {
5648 auto *I = cast<Instruction>(U);
5649 return I == IndUpdate || !TheLoop->contains(I) || Worklist.count(I) ||
5650 isVectorizedMemAccessUse(I, Ind);
5655 // Determine if all users of the induction variable update instruction are
5656 // uniform after vectorization.
5657 auto UniformIndUpdate = all_of(IndUpdate->users(), [&](User *U) -> bool {
5658 auto *I = cast<Instruction>(U);
5659 return I == Ind || !TheLoop->contains(I) || Worklist.count(I) ||
5660 isVectorizedMemAccessUse(I, IndUpdate);
5662 if (!UniformIndUpdate)
5665 // The induction variable and its update instruction will remain uniform.
5666 Worklist.insert(Ind);
5667 Worklist.insert(IndUpdate);
5668 DEBUG(dbgs() << "LV: Found uniform instruction: " << *Ind << "\n");
5669 DEBUG(dbgs() << "LV: Found uniform instruction: " << *IndUpdate << "\n");
5672 Uniforms.insert(Worklist.begin(), Worklist.end());
5675 bool LoopVectorizationLegality::canVectorizeMemory() {
5676 LAI = &(*GetLAA)(*TheLoop);
5677 InterleaveInfo.setLAI(LAI);
5678 const OptimizationRemarkAnalysis *LAR = LAI->getReport();
5680 OptimizationRemarkAnalysis VR(Hints->vectorizeAnalysisPassName(),
5681 "loop not vectorized: ", *LAR);
5684 if (!LAI->canVectorizeMemory())
5687 if (LAI->hasStoreToLoopInvariantAddress()) {
5688 ORE->emit(createMissedAnalysis("CantVectorizeStoreToLoopInvariantAddress")
5689 << "write to a loop invariant address could not be vectorized");
5690 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
5694 Requirements->addRuntimePointerChecks(LAI->getNumRuntimePointerChecks());
5695 PSE.addPredicate(LAI->getPSE().getUnionPredicate());
5700 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
5701 Value *In0 = const_cast<Value *>(V);
5702 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
5706 return Inductions.count(PN);
5709 bool LoopVectorizationLegality::isFirstOrderRecurrence(const PHINode *Phi) {
5710 return FirstOrderRecurrences.count(Phi);
5713 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
5714 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
5717 bool LoopVectorizationLegality::blockCanBePredicated(
5718 BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs) {
5719 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
5721 for (Instruction &I : *BB) {
5722 // Check that we don't have a constant expression that can trap as operand.
5723 for (Value *Operand : I.operands()) {
5724 if (auto *C = dyn_cast<Constant>(Operand))
5728 // We might be able to hoist the load.
5729 if (I.mayReadFromMemory()) {
5730 auto *LI = dyn_cast<LoadInst>(&I);
5733 if (!SafePtrs.count(LI->getPointerOperand())) {
5734 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand()) ||
5735 isLegalMaskedGather(LI->getType())) {
5736 MaskedOp.insert(LI);
5739 // !llvm.mem.parallel_loop_access implies if-conversion safety.
5740 if (IsAnnotatedParallel)
5746 if (I.mayWriteToMemory()) {
5747 auto *SI = dyn_cast<StoreInst>(&I);
5748 // We only support predication of stores in basic blocks with one
5753 // Build a masked store if it is legal for the target.
5754 if (isLegalMaskedStore(SI->getValueOperand()->getType(),
5755 SI->getPointerOperand()) ||
5756 isLegalMaskedScatter(SI->getValueOperand()->getType())) {
5757 MaskedOp.insert(SI);
5761 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
5762 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
5764 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
5765 !isSinglePredecessor)
5775 void InterleavedAccessInfo::collectConstStrideAccesses(
5776 MapVector<Instruction *, StrideDescriptor> &AccessStrideInfo,
5777 const ValueToValueMap &Strides) {
5779 auto &DL = TheLoop->getHeader()->getModule()->getDataLayout();
5781 // Since it's desired that the load/store instructions be maintained in
5782 // "program order" for the interleaved access analysis, we have to visit the
5783 // blocks in the loop in reverse postorder (i.e., in a topological order).
5784 // Such an ordering will ensure that any load/store that may be executed
5785 // before a second load/store will precede the second load/store in
5786 // AccessStrideInfo.
5787 LoopBlocksDFS DFS(TheLoop);
5789 for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO()))
5790 for (auto &I : *BB) {
5791 auto *LI = dyn_cast<LoadInst>(&I);
5792 auto *SI = dyn_cast<StoreInst>(&I);
5796 Value *Ptr = getPointerOperand(&I);
5797 // We don't check wrapping here because we don't know yet if Ptr will be
5798 // part of a full group or a group with gaps. Checking wrapping for all
5799 // pointers (even those that end up in groups with no gaps) will be overly
5800 // conservative. For full groups, wrapping should be ok since if we would
5801 // wrap around the address space we would do a memory access at nullptr
5802 // even without the transformation. The wrapping checks are therefore
5803 // deferred until after we've formed the interleaved groups.
5804 int64_t Stride = getPtrStride(PSE, Ptr, TheLoop, Strides,
5805 /*Assume=*/true, /*ShouldCheckWrap=*/false);
5807 const SCEV *Scev = replaceSymbolicStrideSCEV(PSE, Strides, Ptr);
5808 PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
5809 uint64_t Size = DL.getTypeAllocSize(PtrTy->getElementType());
5811 // An alignment of 0 means target ABI alignment.
5812 unsigned Align = LI ? LI->getAlignment() : SI->getAlignment();
5814 Align = DL.getABITypeAlignment(PtrTy->getElementType());
5816 AccessStrideInfo[&I] = StrideDescriptor(Stride, Scev, Size, Align);
5820 // Analyze interleaved accesses and collect them into interleaved load and
5823 // When generating code for an interleaved load group, we effectively hoist all
5824 // loads in the group to the location of the first load in program order. When
5825 // generating code for an interleaved store group, we sink all stores to the
5826 // location of the last store. This code motion can change the order of load
5827 // and store instructions and may break dependences.
5829 // The code generation strategy mentioned above ensures that we won't violate
5830 // any write-after-read (WAR) dependences.
5832 // E.g., for the WAR dependence: a = A[i]; // (1)
5835 // The store group of (2) is always inserted at or below (2), and the load
5836 // group of (1) is always inserted at or above (1). Thus, the instructions will
5837 // never be reordered. All other dependences are checked to ensure the
5838 // correctness of the instruction reordering.
5840 // The algorithm visits all memory accesses in the loop in bottom-up program
5841 // order. Program order is established by traversing the blocks in the loop in
5842 // reverse postorder when collecting the accesses.
5844 // We visit the memory accesses in bottom-up order because it can simplify the
5845 // construction of store groups in the presence of write-after-write (WAW)
5848 // E.g., for the WAW dependence: A[i] = a; // (1)
5850 // A[i + 1] = c; // (3)
5852 // We will first create a store group with (3) and (2). (1) can't be added to
5853 // this group because it and (2) are dependent. However, (1) can be grouped
5854 // with other accesses that may precede it in program order. Note that a
5855 // bottom-up order does not imply that WAW dependences should not be checked.
5856 void InterleavedAccessInfo::analyzeInterleaving(
5857 const ValueToValueMap &Strides) {
5858 DEBUG(dbgs() << "LV: Analyzing interleaved accesses...\n");
5860 // Holds all accesses with a constant stride.
5861 MapVector<Instruction *, StrideDescriptor> AccessStrideInfo;
5862 collectConstStrideAccesses(AccessStrideInfo, Strides);
5864 if (AccessStrideInfo.empty())
5867 // Collect the dependences in the loop.
5868 collectDependences();
5870 // Holds all interleaved store groups temporarily.
5871 SmallSetVector<InterleaveGroup *, 4> StoreGroups;
5872 // Holds all interleaved load groups temporarily.
5873 SmallSetVector<InterleaveGroup *, 4> LoadGroups;
5875 // Search in bottom-up program order for pairs of accesses (A and B) that can
5876 // form interleaved load or store groups. In the algorithm below, access A
5877 // precedes access B in program order. We initialize a group for B in the
5878 // outer loop of the algorithm, and then in the inner loop, we attempt to
5879 // insert each A into B's group if:
5881 // 1. A and B have the same stride,
5882 // 2. A and B have the same memory object size, and
5883 // 3. A belongs in B's group according to its distance from B.
5885 // Special care is taken to ensure group formation will not break any
5887 for (auto BI = AccessStrideInfo.rbegin(), E = AccessStrideInfo.rend();
5889 Instruction *B = BI->first;
5890 StrideDescriptor DesB = BI->second;
5892 // Initialize a group for B if it has an allowable stride. Even if we don't
5893 // create a group for B, we continue with the bottom-up algorithm to ensure
5894 // we don't break any of B's dependences.
5895 InterleaveGroup *Group = nullptr;
5896 if (isStrided(DesB.Stride)) {
5897 Group = getInterleaveGroup(B);
5899 DEBUG(dbgs() << "LV: Creating an interleave group with:" << *B << '\n');
5900 Group = createInterleaveGroup(B, DesB.Stride, DesB.Align);
5902 if (B->mayWriteToMemory())
5903 StoreGroups.insert(Group);
5905 LoadGroups.insert(Group);
5908 for (auto AI = std::next(BI); AI != E; ++AI) {
5909 Instruction *A = AI->first;
5910 StrideDescriptor DesA = AI->second;
5912 // Our code motion strategy implies that we can't have dependences
5913 // between accesses in an interleaved group and other accesses located
5914 // between the first and last member of the group. Note that this also
5915 // means that a group can't have more than one member at a given offset.
5916 // The accesses in a group can have dependences with other accesses, but
5917 // we must ensure we don't extend the boundaries of the group such that
5918 // we encompass those dependent accesses.
5920 // For example, assume we have the sequence of accesses shown below in a
5923 // (1, 2) is a group | A[i] = a; // (1)
5924 // | A[i-1] = b; // (2) |
5925 // A[i-3] = c; // (3)
5926 // A[i] = d; // (4) | (2, 4) is not a group
5928 // Because accesses (2) and (3) are dependent, we can group (2) with (1)
5929 // but not with (4). If we did, the dependent access (3) would be within
5930 // the boundaries of the (2, 4) group.
5931 if (!canReorderMemAccessesForInterleavedGroups(&*AI, &*BI)) {
5933 // If a dependence exists and A is already in a group, we know that A
5934 // must be a store since A precedes B and WAR dependences are allowed.
5935 // Thus, A would be sunk below B. We release A's group to prevent this
5936 // illegal code motion. A will then be free to form another group with
5937 // instructions that precede it.
5938 if (isInterleaved(A)) {
5939 InterleaveGroup *StoreGroup = getInterleaveGroup(A);
5940 StoreGroups.remove(StoreGroup);
5941 releaseGroup(StoreGroup);
5944 // If a dependence exists and A is not already in a group (or it was
5945 // and we just released it), B might be hoisted above A (if B is a
5946 // load) or another store might be sunk below A (if B is a store). In
5947 // either case, we can't add additional instructions to B's group. B
5948 // will only form a group with instructions that it precedes.
5952 // At this point, we've checked for illegal code motion. If either A or B
5953 // isn't strided, there's nothing left to do.
5954 if (!isStrided(DesA.Stride) || !isStrided(DesB.Stride))
5957 // Ignore A if it's already in a group or isn't the same kind of memory
5959 if (isInterleaved(A) || A->mayReadFromMemory() != B->mayReadFromMemory())
5962 // Check rules 1 and 2. Ignore A if its stride or size is different from
5964 if (DesA.Stride != DesB.Stride || DesA.Size != DesB.Size)
5967 // Calculate the distance from A to B.
5968 const SCEVConstant *DistToB = dyn_cast<SCEVConstant>(
5969 PSE.getSE()->getMinusSCEV(DesA.Scev, DesB.Scev));
5972 int64_t DistanceToB = DistToB->getAPInt().getSExtValue();
5974 // Check rule 3. Ignore A if its distance to B is not a multiple of the
5976 if (DistanceToB % static_cast<int64_t>(DesB.Size))
5979 // Ignore A if either A or B is in a predicated block. Although we
5980 // currently prevent group formation for predicated accesses, we may be
5981 // able to relax this limitation in the future once we handle more
5982 // complicated blocks.
5983 if (isPredicated(A->getParent()) || isPredicated(B->getParent()))
5986 // The index of A is the index of B plus A's distance to B in multiples
5989 Group->getIndex(B) + DistanceToB / static_cast<int64_t>(DesB.Size);
5991 // Try to insert A into B's group.
5992 if (Group->insertMember(A, IndexA, DesA.Align)) {
5993 DEBUG(dbgs() << "LV: Inserted:" << *A << '\n'
5994 << " into the interleave group with" << *B << '\n');
5995 InterleaveGroupMap[A] = Group;
5997 // Set the first load in program order as the insert position.
5998 if (A->mayReadFromMemory())
5999 Group->setInsertPos(A);
6001 } // Iteration over A accesses.
6002 } // Iteration over B accesses.
6004 // Remove interleaved store groups with gaps.
6005 for (InterleaveGroup *Group : StoreGroups)
6006 if (Group->getNumMembers() != Group->getFactor())
6007 releaseGroup(Group);
6009 // Remove interleaved groups with gaps (currently only loads) whose memory
6010 // accesses may wrap around. We have to revisit the getPtrStride analysis,
6011 // this time with ShouldCheckWrap=true, since collectConstStrideAccesses does
6012 // not check wrapping (see documentation there).
6013 // FORNOW we use Assume=false;
6014 // TODO: Change to Assume=true but making sure we don't exceed the threshold
6015 // of runtime SCEV assumptions checks (thereby potentially failing to
6016 // vectorize altogether).
6017 // Additional optional optimizations:
6018 // TODO: If we are peeling the loop and we know that the first pointer doesn't
6019 // wrap then we can deduce that all pointers in the group don't wrap.
6020 // This means that we can forcefully peel the loop in order to only have to
6021 // check the first pointer for no-wrap. When we'll change to use Assume=true
6022 // we'll only need at most one runtime check per interleaved group.
6024 for (InterleaveGroup *Group : LoadGroups) {
6026 // Case 1: A full group. Can Skip the checks; For full groups, if the wide
6027 // load would wrap around the address space we would do a memory access at
6028 // nullptr even without the transformation.
6029 if (Group->getNumMembers() == Group->getFactor())
6032 // Case 2: If first and last members of the group don't wrap this implies
6033 // that all the pointers in the group don't wrap.
6034 // So we check only group member 0 (which is always guaranteed to exist),
6035 // and group member Factor - 1; If the latter doesn't exist we rely on
6036 // peeling (if it is a non-reveresed accsess -- see Case 3).
6037 Value *FirstMemberPtr = getPointerOperand(Group->getMember(0));
6038 if (!getPtrStride(PSE, FirstMemberPtr, TheLoop, Strides, /*Assume=*/false,
6039 /*ShouldCheckWrap=*/true)) {
6040 DEBUG(dbgs() << "LV: Invalidate candidate interleaved group due to "
6041 "first group member potentially pointer-wrapping.\n");
6042 releaseGroup(Group);
6045 Instruction *LastMember = Group->getMember(Group->getFactor() - 1);
6047 Value *LastMemberPtr = getPointerOperand(LastMember);
6048 if (!getPtrStride(PSE, LastMemberPtr, TheLoop, Strides, /*Assume=*/false,
6049 /*ShouldCheckWrap=*/true)) {
6050 DEBUG(dbgs() << "LV: Invalidate candidate interleaved group due to "
6051 "last group member potentially pointer-wrapping.\n");
6052 releaseGroup(Group);
6056 // Case 3: A non-reversed interleaved load group with gaps: We need
6057 // to execute at least one scalar epilogue iteration. This will ensure
6058 // we don't speculatively access memory out-of-bounds. We only need
6059 // to look for a member at index factor - 1, since every group must have
6060 // a member at index zero.
6061 if (Group->isReverse()) {
6062 releaseGroup(Group);
6065 DEBUG(dbgs() << "LV: Interleaved group requires epilogue iteration.\n");
6066 RequiresScalarEpilogue = true;
6071 LoopVectorizationCostModel::VectorizationFactor
6072 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
6073 // Width 1 means no vectorize
6074 VectorizationFactor Factor = {1U, 0U};
6075 if (OptForSize && Legal->getRuntimePointerChecking()->Need) {
6076 ORE->emit(createMissedAnalysis("CantVersionLoopWithOptForSize")
6077 << "runtime pointer checks needed. Enable vectorization of this "
6078 "loop with '#pragma clang loop vectorize(enable)' when "
6079 "compiling with -Os/-Oz");
6081 << "LV: Aborting. Runtime ptr check is required with -Os/-Oz.\n");
6085 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
6086 ORE->emit(createMissedAnalysis("ConditionalStore")
6087 << "store that is conditionally executed prevents vectorization");
6088 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
6092 MinBWs = computeMinimumValueSizes(TheLoop->getBlocks(), *DB, &TTI);
6093 unsigned SmallestType, WidestType;
6094 std::tie(SmallestType, WidestType) = getSmallestAndWidestTypes();
6095 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
6096 unsigned MaxSafeDepDist = -1U;
6098 // Get the maximum safe dependence distance in bits computed by LAA. If the
6099 // loop contains any interleaved accesses, we divide the dependence distance
6100 // by the maximum interleave factor of all interleaved groups. Note that
6101 // although the division ensures correctness, this is a fairly conservative
6102 // computation because the maximum distance computed by LAA may not involve
6103 // any of the interleaved accesses.
6104 if (Legal->getMaxSafeDepDistBytes() != -1U)
6106 Legal->getMaxSafeDepDistBytes() * 8 / Legal->getMaxInterleaveFactor();
6109 ((WidestRegister < MaxSafeDepDist) ? WidestRegister : MaxSafeDepDist);
6110 unsigned MaxVectorSize = WidestRegister / WidestType;
6112 DEBUG(dbgs() << "LV: The Smallest and Widest types: " << SmallestType << " / "
6113 << WidestType << " bits.\n");
6114 DEBUG(dbgs() << "LV: The Widest register is: " << WidestRegister
6117 if (MaxVectorSize == 0) {
6118 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
6122 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
6123 " into one vector!");
6125 unsigned VF = MaxVectorSize;
6126 if (MaximizeBandwidth && !OptForSize) {
6127 // Collect all viable vectorization factors.
6128 SmallVector<unsigned, 8> VFs;
6129 unsigned NewMaxVectorSize = WidestRegister / SmallestType;
6130 for (unsigned VS = MaxVectorSize; VS <= NewMaxVectorSize; VS *= 2)
6133 // For each VF calculate its register usage.
6134 auto RUs = calculateRegisterUsage(VFs);
6136 // Select the largest VF which doesn't require more registers than existing
6138 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(true);
6139 for (int i = RUs.size() - 1; i >= 0; --i) {
6140 if (RUs[i].MaxLocalUsers <= TargetNumRegisters) {
6147 // If we optimize the program for size, avoid creating the tail loop.
6149 unsigned TC = PSE.getSE()->getSmallConstantTripCount(TheLoop);
6150 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
6152 // If we don't know the precise trip count, don't try to vectorize.
6155 createMissedAnalysis("UnknownLoopCountComplexCFG")
6156 << "unable to calculate the loop count due to complex control flow");
6157 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n");
6161 // Find the maximum SIMD width that can fit within the trip count.
6162 VF = TC % MaxVectorSize;
6167 // If the trip count that we found modulo the vectorization factor is not
6168 // zero then we require a tail.
6169 ORE->emit(createMissedAnalysis("NoTailLoopWithOptForSize")
6170 << "cannot optimize for size and vectorize at the "
6171 "same time. Enable vectorization of this loop "
6172 "with '#pragma clang loop vectorize(enable)' "
6173 "when compiling with -Os/-Oz");
6174 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n");
6179 int UserVF = Hints->getWidth();
6181 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
6182 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
6184 Factor.Width = UserVF;
6185 collectInstsToScalarize(UserVF);
6189 float Cost = expectedCost(1).first;
6191 const float ScalarCost = Cost;
6194 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
6196 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
6197 // Ignore scalar width, because the user explicitly wants vectorization.
6198 if (ForceVectorization && VF > 1) {
6200 Cost = expectedCost(Width).first / (float)Width;
6203 for (unsigned i = 2; i <= VF; i *= 2) {
6204 // Notice that the vector loop needs to be executed less times, so
6205 // we need to divide the cost of the vector loops by the width of
6206 // the vector elements.
6207 VectorizationCostTy C = expectedCost(i);
6208 float VectorCost = C.first / (float)i;
6209 DEBUG(dbgs() << "LV: Vector loop of width " << i
6210 << " costs: " << (int)VectorCost << ".\n");
6211 if (!C.second && !ForceVectorization) {
6213 dbgs() << "LV: Not considering vector loop of width " << i
6214 << " because it will not generate any vector instructions.\n");
6217 if (VectorCost < Cost) {
6223 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
6224 << "LV: Vectorization seems to be not beneficial, "
6225 << "but was forced by a user.\n");
6226 DEBUG(dbgs() << "LV: Selecting VF: " << Width << ".\n");
6227 Factor.Width = Width;
6228 Factor.Cost = Width * Cost;
6232 std::pair<unsigned, unsigned>
6233 LoopVectorizationCostModel::getSmallestAndWidestTypes() {
6234 unsigned MinWidth = -1U;
6235 unsigned MaxWidth = 8;
6236 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
6239 for (BasicBlock *BB : TheLoop->blocks()) {
6240 // For each instruction in the loop.
6241 for (Instruction &I : *BB) {
6242 Type *T = I.getType();
6244 // Skip ignored values.
6245 if (ValuesToIgnore.count(&I))
6248 // Only examine Loads, Stores and PHINodes.
6249 if (!isa<LoadInst>(I) && !isa<StoreInst>(I) && !isa<PHINode>(I))
6252 // Examine PHI nodes that are reduction variables. Update the type to
6253 // account for the recurrence type.
6254 if (auto *PN = dyn_cast<PHINode>(&I)) {
6255 if (!Legal->isReductionVariable(PN))
6257 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[PN];
6258 T = RdxDesc.getRecurrenceType();
6261 // Examine the stored values.
6262 if (auto *ST = dyn_cast<StoreInst>(&I))
6263 T = ST->getValueOperand()->getType();
6265 // Ignore loaded pointer types and stored pointer types that are not
6266 // consecutive. However, we do want to take consecutive stores/loads of
6267 // pointer vectors into account.
6268 if (T->isPointerTy() && !isConsecutiveLoadOrStore(&I))
6271 MinWidth = std::min(MinWidth,
6272 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
6273 MaxWidth = std::max(MaxWidth,
6274 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
6278 return {MinWidth, MaxWidth};
6281 unsigned LoopVectorizationCostModel::selectInterleaveCount(bool OptForSize,
6283 unsigned LoopCost) {
6285 // -- The interleave heuristics --
6286 // We interleave the loop in order to expose ILP and reduce the loop overhead.
6287 // There are many micro-architectural considerations that we can't predict
6288 // at this level. For example, frontend pressure (on decode or fetch) due to
6289 // code size, or the number and capabilities of the execution ports.
6291 // We use the following heuristics to select the interleave count:
6292 // 1. If the code has reductions, then we interleave to break the cross
6293 // iteration dependency.
6294 // 2. If the loop is really small, then we interleave to reduce the loop
6296 // 3. We don't interleave if we think that we will spill registers to memory
6297 // due to the increased register pressure.
6299 // When we optimize for size, we don't interleave.
6303 // We used the distance for the interleave count.
6304 if (Legal->getMaxSafeDepDistBytes() != -1U)
6307 // Do not interleave loops with a relatively small trip count.
6308 unsigned TC = PSE.getSE()->getSmallConstantTripCount(TheLoop);
6309 if (TC > 1 && TC < TinyTripCountInterleaveThreshold)
6312 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
6313 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters
6317 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
6318 TargetNumRegisters = ForceTargetNumScalarRegs;
6320 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
6321 TargetNumRegisters = ForceTargetNumVectorRegs;
6324 RegisterUsage R = calculateRegisterUsage({VF})[0];
6325 // We divide by these constants so assume that we have at least one
6326 // instruction that uses at least one register.
6327 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
6328 R.NumInstructions = std::max(R.NumInstructions, 1U);
6330 // We calculate the interleave count using the following formula.
6331 // Subtract the number of loop invariants from the number of available
6332 // registers. These registers are used by all of the interleaved instances.
6333 // Next, divide the remaining registers by the number of registers that is
6334 // required by the loop, in order to estimate how many parallel instances
6335 // fit without causing spills. All of this is rounded down if necessary to be
6336 // a power of two. We want power of two interleave count to simplify any
6337 // addressing operations or alignment considerations.
6338 unsigned IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
6341 // Don't count the induction variable as interleaved.
6342 if (EnableIndVarRegisterHeur)
6343 IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
6344 std::max(1U, (R.MaxLocalUsers - 1)));
6346 // Clamp the interleave ranges to reasonable counts.
6347 unsigned MaxInterleaveCount = TTI.getMaxInterleaveFactor(VF);
6349 // Check if the user has overridden the max.
6351 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
6352 MaxInterleaveCount = ForceTargetMaxScalarInterleaveFactor;
6354 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
6355 MaxInterleaveCount = ForceTargetMaxVectorInterleaveFactor;
6358 // If we did not calculate the cost for VF (because the user selected the VF)
6359 // then we calculate the cost of VF here.
6361 LoopCost = expectedCost(VF).first;
6363 // Clamp the calculated IC to be between the 1 and the max interleave count
6364 // that the target allows.
6365 if (IC > MaxInterleaveCount)
6366 IC = MaxInterleaveCount;
6370 // Interleave if we vectorized this loop and there is a reduction that could
6371 // benefit from interleaving.
6372 if (VF > 1 && Legal->getReductionVars()->size()) {
6373 DEBUG(dbgs() << "LV: Interleaving because of reductions.\n");
6377 // Note that if we've already vectorized the loop we will have done the
6378 // runtime check and so interleaving won't require further checks.
6379 bool InterleavingRequiresRuntimePointerCheck =
6380 (VF == 1 && Legal->getRuntimePointerChecking()->Need);
6382 // We want to interleave small loops in order to reduce the loop overhead and
6383 // potentially expose ILP opportunities.
6384 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
6385 if (!InterleavingRequiresRuntimePointerCheck && LoopCost < SmallLoopCost) {
6386 // We assume that the cost overhead is 1 and we use the cost model
6387 // to estimate the cost of the loop and interleave until the cost of the
6388 // loop overhead is about 5% of the cost of the loop.
6390 std::min(IC, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
6392 // Interleave until store/load ports (estimated by max interleave count) are
6394 unsigned NumStores = Legal->getNumStores();
6395 unsigned NumLoads = Legal->getNumLoads();
6396 unsigned StoresIC = IC / (NumStores ? NumStores : 1);
6397 unsigned LoadsIC = IC / (NumLoads ? NumLoads : 1);
6399 // If we have a scalar reduction (vector reductions are already dealt with
6400 // by this point), we can increase the critical path length if the loop
6401 // we're interleaving is inside another loop. Limit, by default to 2, so the
6402 // critical path only gets increased by one reduction operation.
6403 if (Legal->getReductionVars()->size() && TheLoop->getLoopDepth() > 1) {
6404 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionIC);
6405 SmallIC = std::min(SmallIC, F);
6406 StoresIC = std::min(StoresIC, F);
6407 LoadsIC = std::min(LoadsIC, F);
6410 if (EnableLoadStoreRuntimeInterleave &&
6411 std::max(StoresIC, LoadsIC) > SmallIC) {
6412 DEBUG(dbgs() << "LV: Interleaving to saturate store or load ports.\n");
6413 return std::max(StoresIC, LoadsIC);
6416 DEBUG(dbgs() << "LV: Interleaving to reduce branch cost.\n");
6420 // Interleave if this is a large loop (small loops are already dealt with by
6421 // this point) that could benefit from interleaving.
6422 bool HasReductions = (Legal->getReductionVars()->size() > 0);
6423 if (TTI.enableAggressiveInterleaving(HasReductions)) {
6424 DEBUG(dbgs() << "LV: Interleaving to expose ILP.\n");
6428 DEBUG(dbgs() << "LV: Not Interleaving.\n");
6432 SmallVector<LoopVectorizationCostModel::RegisterUsage, 8>
6433 LoopVectorizationCostModel::calculateRegisterUsage(ArrayRef<unsigned> VFs) {
6434 // This function calculates the register usage by measuring the highest number
6435 // of values that are alive at a single location. Obviously, this is a very
6436 // rough estimation. We scan the loop in a topological order in order and
6437 // assign a number to each instruction. We use RPO to ensure that defs are
6438 // met before their users. We assume that each instruction that has in-loop
6439 // users starts an interval. We record every time that an in-loop value is
6440 // used, so we have a list of the first and last occurrences of each
6441 // instruction. Next, we transpose this data structure into a multi map that
6442 // holds the list of intervals that *end* at a specific location. This multi
6443 // map allows us to perform a linear search. We scan the instructions linearly
6444 // and record each time that a new interval starts, by placing it in a set.
6445 // If we find this value in the multi-map then we remove it from the set.
6446 // The max register usage is the maximum size of the set.
6447 // We also search for instructions that are defined outside the loop, but are
6448 // used inside the loop. We need this number separately from the max-interval
6449 // usage number because when we unroll, loop-invariant values do not take
6451 LoopBlocksDFS DFS(TheLoop);
6455 RU.NumInstructions = 0;
6457 // Each 'key' in the map opens a new interval. The values
6458 // of the map are the index of the 'last seen' usage of the
6459 // instruction that is the key.
6460 typedef DenseMap<Instruction *, unsigned> IntervalMap;
6461 // Maps instruction to its index.
6462 DenseMap<unsigned, Instruction *> IdxToInstr;
6463 // Marks the end of each interval.
6464 IntervalMap EndPoint;
6465 // Saves the list of instruction indices that are used in the loop.
6466 SmallSet<Instruction *, 8> Ends;
6467 // Saves the list of values that are used in the loop but are
6468 // defined outside the loop, such as arguments and constants.
6469 SmallPtrSet<Value *, 8> LoopInvariants;
6472 for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO())) {
6473 RU.NumInstructions += BB->size();
6474 for (Instruction &I : *BB) {
6475 IdxToInstr[Index++] = &I;
6477 // Save the end location of each USE.
6478 for (Value *U : I.operands()) {
6479 auto *Instr = dyn_cast<Instruction>(U);
6481 // Ignore non-instruction values such as arguments, constants, etc.
6485 // If this instruction is outside the loop then record it and continue.
6486 if (!TheLoop->contains(Instr)) {
6487 LoopInvariants.insert(Instr);
6491 // Overwrite previous end points.
6492 EndPoint[Instr] = Index;
6498 // Saves the list of intervals that end with the index in 'key'.
6499 typedef SmallVector<Instruction *, 2> InstrList;
6500 DenseMap<unsigned, InstrList> TransposeEnds;
6502 // Transpose the EndPoints to a list of values that end at each index.
6503 for (auto &Interval : EndPoint)
6504 TransposeEnds[Interval.second].push_back(Interval.first);
6506 SmallSet<Instruction *, 8> OpenIntervals;
6508 // Get the size of the widest register.
6509 unsigned MaxSafeDepDist = -1U;
6510 if (Legal->getMaxSafeDepDistBytes() != -1U)
6511 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
6512 unsigned WidestRegister =
6513 std::min(TTI.getRegisterBitWidth(true), MaxSafeDepDist);
6514 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
6516 SmallVector<RegisterUsage, 8> RUs(VFs.size());
6517 SmallVector<unsigned, 8> MaxUsages(VFs.size(), 0);
6519 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
6521 // A lambda that gets the register usage for the given type and VF.
6522 auto GetRegUsage = [&DL, WidestRegister](Type *Ty, unsigned VF) {
6523 if (Ty->isTokenTy())
6525 unsigned TypeSize = DL.getTypeSizeInBits(Ty->getScalarType());
6526 return std::max<unsigned>(1, VF * TypeSize / WidestRegister);
6529 for (unsigned int i = 0; i < Index; ++i) {
6530 Instruction *I = IdxToInstr[i];
6532 // Remove all of the instructions that end at this location.
6533 InstrList &List = TransposeEnds[i];
6534 for (Instruction *ToRemove : List)
6535 OpenIntervals.erase(ToRemove);
6537 // Ignore instructions that are never used within the loop.
6541 // Skip ignored values.
6542 if (ValuesToIgnore.count(I))
6545 // For each VF find the maximum usage of registers.
6546 for (unsigned j = 0, e = VFs.size(); j < e; ++j) {
6548 MaxUsages[j] = std::max(MaxUsages[j], OpenIntervals.size());
6552 // Count the number of live intervals.
6553 unsigned RegUsage = 0;
6554 for (auto Inst : OpenIntervals) {
6555 // Skip ignored values for VF > 1.
6556 if (VecValuesToIgnore.count(Inst))
6558 RegUsage += GetRegUsage(Inst->getType(), VFs[j]);
6560 MaxUsages[j] = std::max(MaxUsages[j], RegUsage);
6563 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # "
6564 << OpenIntervals.size() << '\n');
6566 // Add the current instruction to the list of open intervals.
6567 OpenIntervals.insert(I);
6570 for (unsigned i = 0, e = VFs.size(); i < e; ++i) {
6571 unsigned Invariant = 0;
6573 Invariant = LoopInvariants.size();
6575 for (auto Inst : LoopInvariants)
6576 Invariant += GetRegUsage(Inst->getType(), VFs[i]);
6579 DEBUG(dbgs() << "LV(REG): VF = " << VFs[i] << '\n');
6580 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsages[i] << '\n');
6581 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
6582 DEBUG(dbgs() << "LV(REG): LoopSize: " << RU.NumInstructions << '\n');
6584 RU.LoopInvariantRegs = Invariant;
6585 RU.MaxLocalUsers = MaxUsages[i];
6592 void LoopVectorizationCostModel::collectInstsToScalarize(unsigned VF) {
6594 // If we aren't vectorizing the loop, or if we've already collected the
6595 // instructions to scalarize, there's nothing to do. Collection may already
6596 // have occurred if we have a user-selected VF and are now computing the
6597 // expected cost for interleaving.
6598 if (VF < 2 || InstsToScalarize.count(VF))
6601 // Initialize a mapping for VF in InstsToScalalarize. If we find that it's
6602 // not profitable to scalarize any instructions, the presence of VF in the
6603 // map will indicate that we've analyzed it already.
6604 ScalarCostsTy &ScalarCostsVF = InstsToScalarize[VF];
6606 // Find all the instructions that are scalar with predication in the loop and
6607 // determine if it would be better to not if-convert the blocks they are in.
6608 // If so, we also record the instructions to scalarize.
6609 for (BasicBlock *BB : TheLoop->blocks()) {
6610 if (!Legal->blockNeedsPredication(BB))
6612 for (Instruction &I : *BB)
6613 if (Legal->isScalarWithPredication(&I)) {
6614 ScalarCostsTy ScalarCosts;
6615 if (computePredInstDiscount(&I, ScalarCosts, VF) >= 0)
6616 ScalarCostsVF.insert(ScalarCosts.begin(), ScalarCosts.end());
6621 int LoopVectorizationCostModel::computePredInstDiscount(
6622 Instruction *PredInst, DenseMap<Instruction *, unsigned> &ScalarCosts,
6625 assert(!Legal->isUniformAfterVectorization(PredInst) &&
6626 "Instruction marked uniform-after-vectorization will be predicated");
6628 // Initialize the discount to zero, meaning that the scalar version and the
6629 // vector version cost the same.
6632 // Holds instructions to analyze. The instructions we visit are mapped in
6633 // ScalarCosts. Those instructions are the ones that would be scalarized if
6634 // we find that the scalar version costs less.
6635 SmallVector<Instruction *, 8> Worklist;
6637 // Returns true if the given instruction can be scalarized.
6638 auto canBeScalarized = [&](Instruction *I) -> bool {
6640 // We only attempt to scalarize instructions forming a single-use chain
6641 // from the original predicated block that would otherwise be vectorized.
6642 // Although not strictly necessary, we give up on instructions we know will
6643 // already be scalar to avoid traversing chains that are unlikely to be
6645 if (!I->hasOneUse() || PredInst->getParent() != I->getParent() ||
6646 Legal->isScalarAfterVectorization(I))
6649 // If the instruction is scalar with predication, it will be analyzed
6650 // separately. We ignore it within the context of PredInst.
6651 if (Legal->isScalarWithPredication(I))
6654 // If any of the instruction's operands are uniform after vectorization,
6655 // the instruction cannot be scalarized. This prevents, for example, a
6656 // masked load from being scalarized.
6658 // We assume we will only emit a value for lane zero of an instruction
6659 // marked uniform after vectorization, rather than VF identical values.
6660 // Thus, if we scalarize an instruction that uses a uniform, we would
6661 // create uses of values corresponding to the lanes we aren't emitting code
6662 // for. This behavior can be changed by allowing getScalarValue to clone
6663 // the lane zero values for uniforms rather than asserting.
6664 for (Use &U : I->operands())
6665 if (auto *J = dyn_cast<Instruction>(U.get()))
6666 if (Legal->isUniformAfterVectorization(J))
6669 // Otherwise, we can scalarize the instruction.
6673 // Returns true if an operand that cannot be scalarized must be extracted
6674 // from a vector. We will account for this scalarization overhead below. Note
6675 // that the non-void predicated instructions are placed in their own blocks,
6676 // and their return values are inserted into vectors. Thus, an extract would
6677 // still be required.
6678 auto needsExtract = [&](Instruction *I) -> bool {
6679 return TheLoop->contains(I) && !Legal->isScalarAfterVectorization(I);
6682 // Compute the expected cost discount from scalarizing the entire expression
6683 // feeding the predicated instruction. We currently only consider expressions
6684 // that are single-use instruction chains.
6685 Worklist.push_back(PredInst);
6686 while (!Worklist.empty()) {
6687 Instruction *I = Worklist.pop_back_val();
6689 // If we've already analyzed the instruction, there's nothing to do.
6690 if (ScalarCosts.count(I))
6693 // Compute the cost of the vector instruction. Note that this cost already
6694 // includes the scalarization overhead of the predicated instruction.
6695 unsigned VectorCost = getInstructionCost(I, VF).first;
6697 // Compute the cost of the scalarized instruction. This cost is the cost of
6698 // the instruction as if it wasn't if-converted and instead remained in the
6699 // predicated block. We will scale this cost by block probability after
6700 // computing the scalarization overhead.
6701 unsigned ScalarCost = VF * getInstructionCost(I, 1).first;
6703 // Compute the scalarization overhead of needed insertelement instructions
6705 if (Legal->isScalarWithPredication(I) && !I->getType()->isVoidTy()) {
6706 ScalarCost += getScalarizationOverhead(ToVectorTy(I->getType(), VF), true,
6708 ScalarCost += VF * TTI.getCFInstrCost(Instruction::PHI);
6711 // Compute the scalarization overhead of needed extractelement
6712 // instructions. For each of the instruction's operands, if the operand can
6713 // be scalarized, add it to the worklist; otherwise, account for the
6715 for (Use &U : I->operands())
6716 if (auto *J = dyn_cast<Instruction>(U.get())) {
6717 assert(VectorType::isValidElementType(J->getType()) &&
6718 "Instruction has non-scalar type");
6719 if (canBeScalarized(J))
6720 Worklist.push_back(J);
6721 else if (needsExtract(J))
6722 ScalarCost += getScalarizationOverhead(ToVectorTy(J->getType(), VF),
6726 // Scale the total scalar cost by block probability.
6727 ScalarCost /= getReciprocalPredBlockProb();
6729 // Compute the discount. A non-negative discount means the vector version
6730 // of the instruction costs more, and scalarizing would be beneficial.
6731 Discount += VectorCost - ScalarCost;
6732 ScalarCosts[I] = ScalarCost;
6738 LoopVectorizationCostModel::VectorizationCostTy
6739 LoopVectorizationCostModel::expectedCost(unsigned VF) {
6740 VectorizationCostTy Cost;
6742 // Collect the instructions (and their associated costs) that will be more
6743 // profitable to scalarize.
6744 collectInstsToScalarize(VF);
6747 for (BasicBlock *BB : TheLoop->blocks()) {
6748 VectorizationCostTy BlockCost;
6750 // For each instruction in the old loop.
6751 for (Instruction &I : *BB) {
6752 // Skip dbg intrinsics.
6753 if (isa<DbgInfoIntrinsic>(I))
6756 // Skip ignored values.
6757 if (ValuesToIgnore.count(&I))
6760 VectorizationCostTy C = getInstructionCost(&I, VF);
6762 // Check if we should override the cost.
6763 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
6764 C.first = ForceTargetInstructionCost;
6766 BlockCost.first += C.first;
6767 BlockCost.second |= C.second;
6768 DEBUG(dbgs() << "LV: Found an estimated cost of " << C.first << " for VF "
6769 << VF << " For instruction: " << I << '\n');
6772 // If we are vectorizing a predicated block, it will have been
6773 // if-converted. This means that the block's instructions (aside from
6774 // stores and instructions that may divide by zero) will now be
6775 // unconditionally executed. For the scalar case, we may not always execute
6776 // the predicated block. Thus, scale the block's cost by the probability of
6778 if (VF == 1 && Legal->blockNeedsPredication(BB))
6779 BlockCost.first /= getReciprocalPredBlockProb();
6781 Cost.first += BlockCost.first;
6782 Cost.second |= BlockCost.second;
6788 /// \brief Gets Address Access SCEV after verifying that the access pattern
6789 /// is loop invariant except the induction variable dependence.
6791 /// This SCEV can be sent to the Target in order to estimate the address
6792 /// calculation cost.
6793 static const SCEV *getAddressAccessSCEV(
6795 LoopVectorizationLegality *Legal,
6796 ScalarEvolution *SE,
6797 const Loop *TheLoop) {
6798 auto *Gep = dyn_cast<GetElementPtrInst>(Ptr);
6802 // We are looking for a gep with all loop invariant indices except for one
6803 // which should be an induction variable.
6804 unsigned NumOperands = Gep->getNumOperands();
6805 for (unsigned i = 1; i < NumOperands; ++i) {
6806 Value *Opd = Gep->getOperand(i);
6807 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
6808 !Legal->isInductionVariable(Opd))
6812 // Now we know we have a GEP ptr, %inv, %ind, %inv. return the Ptr SCEV.
6813 return SE->getSCEV(Ptr);
6816 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
6817 return Legal->hasStride(I->getOperand(0)) ||
6818 Legal->hasStride(I->getOperand(1));
6821 LoopVectorizationCostModel::VectorizationCostTy
6822 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
6823 // If we know that this instruction will remain uniform, check the cost of
6824 // the scalar version.
6825 if (Legal->isUniformAfterVectorization(I))
6828 if (VF > 1 && isProfitableToScalarize(I, VF))
6829 return VectorizationCostTy(InstsToScalarize[VF][I], false);
6832 unsigned C = getInstructionCost(I, VF, VectorTy);
6834 bool TypeNotScalarized =
6835 VF > 1 && !VectorTy->isVoidTy() && TTI.getNumberOfParts(VectorTy) < VF;
6836 return VectorizationCostTy(C, TypeNotScalarized);
6839 unsigned LoopVectorizationCostModel::getInstructionCost(Instruction *I,
6842 Type *RetTy = I->getType();
6843 if (canTruncateToMinimalBitwidth(I, VF))
6844 RetTy = IntegerType::get(RetTy->getContext(), MinBWs[I]);
6845 VectorTy = ToVectorTy(RetTy, VF);
6846 auto SE = PSE.getSE();
6848 // TODO: We need to estimate the cost of intrinsic calls.
6849 switch (I->getOpcode()) {
6850 case Instruction::GetElementPtr:
6851 // We mark this instruction as zero-cost because the cost of GEPs in
6852 // vectorized code depends on whether the corresponding memory instruction
6853 // is scalarized or not. Therefore, we handle GEPs with the memory
6854 // instruction cost.
6856 case Instruction::Br: {
6857 return TTI.getCFInstrCost(I->getOpcode());
6859 case Instruction::PHI: {
6860 auto *Phi = cast<PHINode>(I);
6862 // First-order recurrences are replaced by vector shuffles inside the loop.
6863 if (VF > 1 && Legal->isFirstOrderRecurrence(Phi))
6864 return TTI.getShuffleCost(TargetTransformInfo::SK_ExtractSubvector,
6865 VectorTy, VF - 1, VectorTy);
6867 // TODO: IF-converted IFs become selects.
6870 case Instruction::UDiv:
6871 case Instruction::SDiv:
6872 case Instruction::URem:
6873 case Instruction::SRem:
6874 // If we have a predicated instruction, it may not be executed for each
6875 // vector lane. Get the scalarization cost and scale this amount by the
6876 // probability of executing the predicated block. If the instruction is not
6877 // predicated, we fall through to the next case.
6878 if (VF > 1 && Legal->isScalarWithPredication(I)) {
6881 // These instructions have a non-void type, so account for the phi nodes
6882 // that we will create. This cost is likely to be zero. The phi node
6883 // cost, if any, should be scaled by the block probability because it
6884 // models a copy at the end of each predicated block.
6885 Cost += VF * TTI.getCFInstrCost(Instruction::PHI);
6887 // The cost of the non-predicated instruction.
6888 Cost += VF * TTI.getArithmeticInstrCost(I->getOpcode(), RetTy);
6890 // The cost of insertelement and extractelement instructions needed for
6892 Cost += getScalarizationOverhead(I, VF, TTI);
6894 // Scale the cost by the probability of executing the predicated blocks.
6895 // This assumes the predicated block for each vector lane is equally
6897 return Cost / getReciprocalPredBlockProb();
6899 case Instruction::Add:
6900 case Instruction::FAdd:
6901 case Instruction::Sub:
6902 case Instruction::FSub:
6903 case Instruction::Mul:
6904 case Instruction::FMul:
6905 case Instruction::FDiv:
6906 case Instruction::FRem:
6907 case Instruction::Shl:
6908 case Instruction::LShr:
6909 case Instruction::AShr:
6910 case Instruction::And:
6911 case Instruction::Or:
6912 case Instruction::Xor: {
6913 // Since we will replace the stride by 1 the multiplication should go away.
6914 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
6916 // Certain instructions can be cheaper to vectorize if they have a constant
6917 // second vector operand. One example of this are shifts on x86.
6918 TargetTransformInfo::OperandValueKind Op1VK =
6919 TargetTransformInfo::OK_AnyValue;
6920 TargetTransformInfo::OperandValueKind Op2VK =
6921 TargetTransformInfo::OK_AnyValue;
6922 TargetTransformInfo::OperandValueProperties Op1VP =
6923 TargetTransformInfo::OP_None;
6924 TargetTransformInfo::OperandValueProperties Op2VP =
6925 TargetTransformInfo::OP_None;
6926 Value *Op2 = I->getOperand(1);
6928 // Check for a splat or for a non uniform vector of constants.
6929 if (isa<ConstantInt>(Op2)) {
6930 ConstantInt *CInt = cast<ConstantInt>(Op2);
6931 if (CInt && CInt->getValue().isPowerOf2())
6932 Op2VP = TargetTransformInfo::OP_PowerOf2;
6933 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
6934 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
6935 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
6936 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
6938 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
6939 if (CInt && CInt->getValue().isPowerOf2())
6940 Op2VP = TargetTransformInfo::OP_PowerOf2;
6941 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
6943 } else if (Legal->isUniform(Op2)) {
6944 Op2VK = TargetTransformInfo::OK_UniformValue;
6947 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
6950 case Instruction::Select: {
6951 SelectInst *SI = cast<SelectInst>(I);
6952 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
6953 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
6954 Type *CondTy = SI->getCondition()->getType();
6956 CondTy = VectorType::get(CondTy, VF);
6958 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
6960 case Instruction::ICmp:
6961 case Instruction::FCmp: {
6962 Type *ValTy = I->getOperand(0)->getType();
6963 Instruction *Op0AsInstruction = dyn_cast<Instruction>(I->getOperand(0));
6964 if (canTruncateToMinimalBitwidth(Op0AsInstruction, VF))
6965 ValTy = IntegerType::get(ValTy->getContext(), MinBWs[Op0AsInstruction]);
6966 VectorTy = ToVectorTy(ValTy, VF);
6967 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
6969 case Instruction::Store:
6970 case Instruction::Load: {
6971 StoreInst *SI = dyn_cast<StoreInst>(I);
6972 LoadInst *LI = dyn_cast<LoadInst>(I);
6973 Type *ValTy = (SI ? SI->getValueOperand()->getType() : LI->getType());
6974 VectorTy = ToVectorTy(ValTy, VF);
6976 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
6978 SI ? SI->getPointerAddressSpace() : LI->getPointerAddressSpace();
6979 Value *Ptr = getPointerOperand(I);
6980 // We add the cost of address computation here instead of with the gep
6981 // instruction because only here we know whether the operation is
6984 return TTI.getAddressComputationCost(VectorTy) +
6985 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
6987 if (LI && Legal->isUniform(Ptr)) {
6988 // Scalar load + broadcast
6989 unsigned Cost = TTI.getAddressComputationCost(ValTy->getScalarType());
6990 Cost += TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
6993 TTI.getShuffleCost(TargetTransformInfo::SK_Broadcast, ValTy);
6996 // For an interleaved access, calculate the total cost of the whole
6997 // interleave group.
6998 if (Legal->isAccessInterleaved(I)) {
6999 auto Group = Legal->getInterleavedAccessGroup(I);
7000 assert(Group && "Fail to get an interleaved access group.");
7002 // Only calculate the cost once at the insert position.
7003 if (Group->getInsertPos() != I)
7006 unsigned InterleaveFactor = Group->getFactor();
7008 VectorType::get(VectorTy->getVectorElementType(),
7009 VectorTy->getVectorNumElements() * InterleaveFactor);
7011 // Holds the indices of existing members in an interleaved load group.
7012 // An interleaved store group doesn't need this as it doesn't allow gaps.
7013 SmallVector<unsigned, 4> Indices;
7015 for (unsigned i = 0; i < InterleaveFactor; i++)
7016 if (Group->getMember(i))
7017 Indices.push_back(i);
7020 // Calculate the cost of the whole interleaved group.
7021 unsigned Cost = TTI.getInterleavedMemoryOpCost(
7022 I->getOpcode(), WideVecTy, Group->getFactor(), Indices,
7023 Group->getAlignment(), AS);
7025 if (Group->isReverse())
7027 Group->getNumMembers() *
7028 TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
7030 // FIXME: The interleaved load group with a huge gap could be even more
7031 // expensive than scalar operations. Then we could ignore such group and
7032 // use scalar operations instead.
7036 // Check if the memory instruction will be scalarized.
7037 if (Legal->memoryInstructionMustBeScalarized(I, VF)) {
7039 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
7041 // Figure out whether the access is strided and get the stride value
7042 // if it's known in compile time
7043 const SCEV *PtrSCEV = getAddressAccessSCEV(Ptr, Legal, SE, TheLoop);
7045 // Get the cost of the scalar memory instruction and address computation.
7046 Cost += VF * TTI.getAddressComputationCost(PtrTy, SE, PtrSCEV);
7048 TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
7051 // Get the overhead of the extractelement and insertelement instructions
7052 // we might create due to scalarization.
7053 Cost += getScalarizationOverhead(I, VF, TTI);
7055 // If we have a predicated store, it may not be executed for each vector
7056 // lane. Scale the cost by the probability of executing the predicated
7058 if (Legal->isScalarWithPredication(I))
7059 Cost /= getReciprocalPredBlockProb();
7064 // Determine if the pointer operand of the access is either consecutive or
7065 // reverse consecutive.
7066 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
7067 bool Reverse = ConsecutiveStride < 0;
7069 // Determine if either a gather or scatter operation is legal.
7070 bool UseGatherOrScatter =
7071 !ConsecutiveStride && Legal->isLegalGatherOrScatter(I);
7073 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
7074 if (UseGatherOrScatter) {
7075 assert(ConsecutiveStride == 0 &&
7076 "Gather/Scatter are not used for consecutive stride");
7078 TTI.getGatherScatterOpCost(I->getOpcode(), VectorTy, Ptr,
7079 Legal->isMaskRequired(I), Alignment);
7081 // Wide load/stores.
7082 if (Legal->isMaskRequired(I))
7084 TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
7086 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
7089 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
7092 case Instruction::ZExt:
7093 case Instruction::SExt:
7094 case Instruction::FPToUI:
7095 case Instruction::FPToSI:
7096 case Instruction::FPExt:
7097 case Instruction::PtrToInt:
7098 case Instruction::IntToPtr:
7099 case Instruction::SIToFP:
7100 case Instruction::UIToFP:
7101 case Instruction::Trunc:
7102 case Instruction::FPTrunc:
7103 case Instruction::BitCast: {
7104 // We optimize the truncation of induction variable.
7105 // The cost of these is the same as the scalar operation.
7106 if (I->getOpcode() == Instruction::Trunc &&
7107 Legal->isInductionVariable(I->getOperand(0)))
7108 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
7109 I->getOperand(0)->getType());
7111 Type *SrcScalarTy = I->getOperand(0)->getType();
7112 Type *SrcVecTy = ToVectorTy(SrcScalarTy, VF);
7113 if (canTruncateToMinimalBitwidth(I, VF)) {
7114 // This cast is going to be shrunk. This may remove the cast or it might
7115 // turn it into slightly different cast. For example, if MinBW == 16,
7116 // "zext i8 %1 to i32" becomes "zext i8 %1 to i16".
7118 // Calculate the modified src and dest types.
7119 Type *MinVecTy = VectorTy;
7120 if (I->getOpcode() == Instruction::Trunc) {
7121 SrcVecTy = smallestIntegerVectorType(SrcVecTy, MinVecTy);
7123 largestIntegerVectorType(ToVectorTy(I->getType(), VF), MinVecTy);
7124 } else if (I->getOpcode() == Instruction::ZExt ||
7125 I->getOpcode() == Instruction::SExt) {
7126 SrcVecTy = largestIntegerVectorType(SrcVecTy, MinVecTy);
7128 smallestIntegerVectorType(ToVectorTy(I->getType(), VF), MinVecTy);
7132 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
7134 case Instruction::Call: {
7135 bool NeedToScalarize;
7136 CallInst *CI = cast<CallInst>(I);
7137 unsigned CallCost = getVectorCallCost(CI, VF, TTI, TLI, NeedToScalarize);
7138 if (getVectorIntrinsicIDForCall(CI, TLI))
7139 return std::min(CallCost, getVectorIntrinsicCost(CI, VF, TTI, TLI));
7143 // The cost of executing VF copies of the scalar instruction. This opcode
7144 // is unknown. Assume that it is the same as 'mul'.
7145 return VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy) +
7146 getScalarizationOverhead(I, VF, TTI);
7150 char LoopVectorize::ID = 0;
7151 static const char lv_name[] = "Loop Vectorization";
7152 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
7153 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
7154 INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass)
7155 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
7156 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
7157 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
7158 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
7159 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
7160 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
7161 INITIALIZE_PASS_DEPENDENCY(LCSSAWrapperPass)
7162 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
7163 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
7164 INITIALIZE_PASS_DEPENDENCY(LoopAccessLegacyAnalysis)
7165 INITIALIZE_PASS_DEPENDENCY(DemandedBitsWrapperPass)
7166 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
7167 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
7170 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
7171 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
7175 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
7177 // Check if the pointer operand of a load or store instruction is
7179 if (auto *Ptr = getPointerOperand(Inst))
7180 return Legal->isConsecutivePtr(Ptr);
7184 void LoopVectorizationCostModel::collectValuesToIgnore() {
7185 // Ignore ephemeral values.
7186 CodeMetrics::collectEphemeralValues(TheLoop, AC, ValuesToIgnore);
7188 // Ignore type-promoting instructions we identified during reduction
7190 for (auto &Reduction : *Legal->getReductionVars()) {
7191 RecurrenceDescriptor &RedDes = Reduction.second;
7192 SmallPtrSetImpl<Instruction *> &Casts = RedDes.getCastInsts();
7193 VecValuesToIgnore.insert(Casts.begin(), Casts.end());
7196 // Insert values known to be scalar into VecValuesToIgnore. This is a
7197 // conservative estimation of the values that will later be scalarized.
7199 // FIXME: Even though an instruction is not scalar-after-vectoriztion, it may
7200 // still be scalarized. For example, we may find an instruction to be
7201 // more profitable for a given vectorization factor if it were to be
7202 // scalarized. But at this point, we haven't yet computed the
7203 // vectorization factor.
7204 for (auto *BB : TheLoop->getBlocks())
7206 if (Legal->isScalarAfterVectorization(&I))
7207 VecValuesToIgnore.insert(&I);
7210 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
7211 bool IfPredicateInstr) {
7212 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
7213 // Holds vector parameters or scalars, in case of uniform vals.
7214 SmallVector<VectorParts, 4> Params;
7216 setDebugLocFromInst(Builder, Instr);
7218 // Does this instruction return a value ?
7219 bool IsVoidRetTy = Instr->getType()->isVoidTy();
7221 // Initialize a new scalar map entry.
7222 ScalarParts Entry(UF);
7225 if (IfPredicateInstr)
7226 Cond = createBlockInMask(Instr->getParent());
7228 // For each vector unroll 'part':
7229 for (unsigned Part = 0; Part < UF; ++Part) {
7230 Entry[Part].resize(1);
7231 // For each scalar that we create:
7233 // Start an "if (pred) a[i] = ..." block.
7234 Value *Cmp = nullptr;
7235 if (IfPredicateInstr) {
7236 if (Cond[Part]->getType()->isVectorTy())
7238 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
7239 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
7240 ConstantInt::get(Cond[Part]->getType(), 1));
7243 Instruction *Cloned = Instr->clone();
7245 Cloned->setName(Instr->getName() + ".cloned");
7247 // Replace the operands of the cloned instructions with their scalar
7248 // equivalents in the new loop.
7249 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
7250 auto *NewOp = getScalarValue(Instr->getOperand(op), Part, 0);
7251 Cloned->setOperand(op, NewOp);
7254 // Place the cloned scalar in the new loop.
7255 Builder.Insert(Cloned);
7257 // Add the cloned scalar to the scalar map entry.
7258 Entry[Part][0] = Cloned;
7260 // If we just cloned a new assumption, add it the assumption cache.
7261 if (auto *II = dyn_cast<IntrinsicInst>(Cloned))
7262 if (II->getIntrinsicID() == Intrinsic::assume)
7263 AC->registerAssumption(II);
7266 if (IfPredicateInstr)
7267 PredicatedInstructions.push_back(std::make_pair(Cloned, Cmp));
7269 VectorLoopValueMap.initScalar(Instr, Entry);
7272 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
7273 auto *SI = dyn_cast<StoreInst>(Instr);
7274 bool IfPredicateInstr = (SI && Legal->blockNeedsPredication(SI->getParent()));
7276 return scalarizeInstruction(Instr, IfPredicateInstr);
7279 Value *InnerLoopUnroller::reverseVector(Value *Vec) { return Vec; }
7281 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) { return V; }
7283 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step,
7284 Instruction::BinaryOps BinOp) {
7285 // When unrolling and the VF is 1, we only need to add a simple scalar.
7286 Type *Ty = Val->getType();
7287 assert(!Ty->isVectorTy() && "Val must be a scalar");
7289 if (Ty->isFloatingPointTy()) {
7290 Constant *C = ConstantFP::get(Ty, (double)StartIdx);
7292 // Floating point operations had to be 'fast' to enable the unrolling.
7293 Value *MulOp = addFastMathFlag(Builder.CreateFMul(C, Step));
7294 return addFastMathFlag(Builder.CreateBinOp(BinOp, Val, MulOp));
7296 Constant *C = ConstantInt::get(Ty, StartIdx);
7297 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");
7300 static void AddRuntimeUnrollDisableMetaData(Loop *L) {
7301 SmallVector<Metadata *, 4> MDs;
7302 // Reserve first location for self reference to the LoopID metadata node.
7303 MDs.push_back(nullptr);
7304 bool IsUnrollMetadata = false;
7305 MDNode *LoopID = L->getLoopID();
7307 // First find existing loop unrolling disable metadata.
7308 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
7309 auto *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
7311 const auto *S = dyn_cast<MDString>(MD->getOperand(0));
7313 S && S->getString().startswith("llvm.loop.unroll.disable");
7315 MDs.push_back(LoopID->getOperand(i));
7319 if (!IsUnrollMetadata) {
7320 // Add runtime unroll disable metadata.
7321 LLVMContext &Context = L->getHeader()->getContext();
7322 SmallVector<Metadata *, 1> DisableOperands;
7323 DisableOperands.push_back(
7324 MDString::get(Context, "llvm.loop.unroll.runtime.disable"));
7325 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
7326 MDs.push_back(DisableNode);
7327 MDNode *NewLoopID = MDNode::get(Context, MDs);
7328 // Set operand 0 to refer to the loop id itself.
7329 NewLoopID->replaceOperandWith(0, NewLoopID);
7330 L->setLoopID(NewLoopID);
7334 bool LoopVectorizePass::processLoop(Loop *L) {
7335 assert(L->empty() && "Only process inner loops.");
7338 const std::string DebugLocStr = getDebugLocString(L);
7341 DEBUG(dbgs() << "\nLV: Checking a loop in \""
7342 << L->getHeader()->getParent()->getName() << "\" from "
7343 << DebugLocStr << "\n");
7345 LoopVectorizeHints Hints(L, DisableUnrolling, *ORE);
7347 DEBUG(dbgs() << "LV: Loop hints:"
7349 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
7351 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
7354 << " width=" << Hints.getWidth()
7355 << " unroll=" << Hints.getInterleave() << "\n");
7357 // Function containing loop
7358 Function *F = L->getHeader()->getParent();
7360 // Looking at the diagnostic output is the only way to determine if a loop
7361 // was vectorized (other than looking at the IR or machine code), so it
7362 // is important to generate an optimization remark for each loop. Most of
7363 // these messages are generated as OptimizationRemarkAnalysis. Remarks
7364 // generated as OptimizationRemark and OptimizationRemarkMissed are
7365 // less verbose reporting vectorized loops and unvectorized loops that may
7366 // benefit from vectorization, respectively.
7368 if (!Hints.allowVectorization(F, L, AlwaysVectorize)) {
7369 DEBUG(dbgs() << "LV: Loop hints prevent vectorization.\n");
7373 // Check the loop for a trip count threshold:
7374 // do not vectorize loops with a tiny trip count.
7375 const unsigned MaxTC = SE->getSmallConstantMaxTripCount(L);
7376 if (MaxTC > 0u && MaxTC < TinyTripCountVectorThreshold) {
7377 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
7378 << "This loop is not worth vectorizing.");
7379 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
7380 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
7382 DEBUG(dbgs() << "\n");
7383 ORE->emit(createMissedAnalysis(Hints.vectorizeAnalysisPassName(),
7385 << "vectorization is not beneficial "
7386 "and is not explicitly forced");
7391 PredicatedScalarEvolution PSE(*SE, *L);
7393 // Check if it is legal to vectorize the loop.
7394 LoopVectorizationRequirements Requirements(*ORE);
7395 LoopVectorizationLegality LVL(L, PSE, DT, TLI, AA, F, TTI, GetLAA, LI, ORE,
7396 &Requirements, &Hints);
7397 if (!LVL.canVectorize()) {
7398 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
7399 emitMissedWarning(F, L, Hints, ORE);
7403 // Use the cost model.
7404 LoopVectorizationCostModel CM(L, PSE, LI, &LVL, *TTI, TLI, DB, AC, ORE, F,
7406 CM.collectValuesToIgnore();
7408 // Check the function attributes to find out if this function should be
7409 // optimized for size.
7411 Hints.getForce() != LoopVectorizeHints::FK_Enabled && F->optForSize();
7413 // Compute the weighted frequency of this loop being executed and see if it
7414 // is less than 20% of the function entry baseline frequency. Note that we
7415 // always have a canonical loop here because we think we *can* vectorize.
7416 // FIXME: This is hidden behind a flag due to pervasive problems with
7417 // exactly what block frequency models.
7418 if (LoopVectorizeWithBlockFrequency) {
7419 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
7420 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
7421 LoopEntryFreq < ColdEntryFreq)
7425 // Check the function attributes to see if implicit floats are allowed.
7426 // FIXME: This check doesn't seem possibly correct -- what if the loop is
7427 // an integer loop and the vector instructions selected are purely integer
7428 // vector instructions?
7429 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
7430 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
7431 "attribute is used.\n");
7432 ORE->emit(createMissedAnalysis(Hints.vectorizeAnalysisPassName(),
7433 "NoImplicitFloat", L)
7434 << "loop not vectorized due to NoImplicitFloat attribute");
7435 emitMissedWarning(F, L, Hints, ORE);
7439 // Check if the target supports potentially unsafe FP vectorization.
7440 // FIXME: Add a check for the type of safety issue (denormal, signaling)
7441 // for the target we're vectorizing for, to make sure none of the
7442 // additional fp-math flags can help.
7443 if (Hints.isPotentiallyUnsafe() &&
7444 TTI->isFPVectorizationPotentiallyUnsafe()) {
7445 DEBUG(dbgs() << "LV: Potentially unsafe FP op prevents vectorization.\n");
7447 createMissedAnalysis(Hints.vectorizeAnalysisPassName(), "UnsafeFP", L)
7448 << "loop not vectorized due to unsafe FP support.");
7449 emitMissedWarning(F, L, Hints, ORE);
7453 // Select the optimal vectorization factor.
7454 const LoopVectorizationCostModel::VectorizationFactor VF =
7455 CM.selectVectorizationFactor(OptForSize);
7457 // Select the interleave count.
7458 unsigned IC = CM.selectInterleaveCount(OptForSize, VF.Width, VF.Cost);
7460 // Get user interleave count.
7461 unsigned UserIC = Hints.getInterleave();
7463 // Identify the diagnostic messages that should be produced.
7464 std::pair<StringRef, std::string> VecDiagMsg, IntDiagMsg;
7465 bool VectorizeLoop = true, InterleaveLoop = true;
7466 if (Requirements.doesNotMeet(F, L, Hints)) {
7467 DEBUG(dbgs() << "LV: Not vectorizing: loop did not meet vectorization "
7469 emitMissedWarning(F, L, Hints, ORE);
7473 if (VF.Width == 1) {
7474 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
7475 VecDiagMsg = std::make_pair(
7476 "VectorizationNotBeneficial",
7477 "the cost-model indicates that vectorization is not beneficial");
7478 VectorizeLoop = false;
7481 if (IC == 1 && UserIC <= 1) {
7482 // Tell the user interleaving is not beneficial.
7483 DEBUG(dbgs() << "LV: Interleaving is not beneficial.\n");
7484 IntDiagMsg = std::make_pair(
7485 "InterleavingNotBeneficial",
7486 "the cost-model indicates that interleaving is not beneficial");
7487 InterleaveLoop = false;
7489 IntDiagMsg.first = "InterleavingNotBeneficialAndDisabled";
7490 IntDiagMsg.second +=
7491 " and is explicitly disabled or interleave count is set to 1";
7493 } else if (IC > 1 && UserIC == 1) {
7494 // Tell the user interleaving is beneficial, but it explicitly disabled.
7496 << "LV: Interleaving is beneficial but is explicitly disabled.");
7497 IntDiagMsg = std::make_pair(
7498 "InterleavingBeneficialButDisabled",
7499 "the cost-model indicates that interleaving is beneficial "
7500 "but is explicitly disabled or interleave count is set to 1");
7501 InterleaveLoop = false;
7504 // Override IC if user provided an interleave count.
7505 IC = UserIC > 0 ? UserIC : IC;
7507 // Emit diagnostic messages, if any.
7508 const char *VAPassName = Hints.vectorizeAnalysisPassName();
7509 if (!VectorizeLoop && !InterleaveLoop) {
7510 // Do not vectorize or interleaving the loop.
7511 ORE->emit(OptimizationRemarkAnalysis(VAPassName, VecDiagMsg.first,
7512 L->getStartLoc(), L->getHeader())
7513 << VecDiagMsg.second);
7514 ORE->emit(OptimizationRemarkAnalysis(LV_NAME, IntDiagMsg.first,
7515 L->getStartLoc(), L->getHeader())
7516 << IntDiagMsg.second);
7518 } else if (!VectorizeLoop && InterleaveLoop) {
7519 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
7520 ORE->emit(OptimizationRemarkAnalysis(VAPassName, VecDiagMsg.first,
7521 L->getStartLoc(), L->getHeader())
7522 << VecDiagMsg.second);
7523 } else if (VectorizeLoop && !InterleaveLoop) {
7524 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
7525 << DebugLocStr << '\n');
7526 ORE->emit(OptimizationRemarkAnalysis(LV_NAME, IntDiagMsg.first,
7527 L->getStartLoc(), L->getHeader())
7528 << IntDiagMsg.second);
7529 } else if (VectorizeLoop && InterleaveLoop) {
7530 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
7531 << DebugLocStr << '\n');
7532 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
7535 using namespace ore;
7536 if (!VectorizeLoop) {
7537 assert(IC > 1 && "interleave count should not be 1 or 0");
7538 // If we decided that it is not legal to vectorize the loop, then
7540 InnerLoopUnroller Unroller(L, PSE, LI, DT, TLI, TTI, AC, ORE, IC, &LVL,
7542 Unroller.vectorize();
7544 ORE->emit(OptimizationRemark(LV_NAME, "Interleaved", L->getStartLoc(),
7546 << "interleaved loop (interleaved count: "
7547 << NV("InterleaveCount", IC) << ")");
7549 // If we decided that it is *legal* to vectorize the loop, then do it.
7550 InnerLoopVectorizer LB(L, PSE, LI, DT, TLI, TTI, AC, ORE, VF.Width, IC,
7555 // Add metadata to disable runtime unrolling a scalar loop when there are
7556 // no runtime checks about strides and memory. A scalar loop that is
7557 // rarely used is not worth unrolling.
7558 if (!LB.areSafetyChecksAdded())
7559 AddRuntimeUnrollDisableMetaData(L);
7561 // Report the vectorization decision.
7562 ORE->emit(OptimizationRemark(LV_NAME, "Vectorized", L->getStartLoc(),
7564 << "vectorized loop (vectorization width: "
7565 << NV("VectorizationFactor", VF.Width)
7566 << ", interleaved count: " << NV("InterleaveCount", IC) << ")");
7569 // Mark the loop as already vectorized to avoid vectorizing again.
7570 Hints.setAlreadyVectorized();
7572 DEBUG(verifyFunction(*L->getHeader()->getParent()));
7576 bool LoopVectorizePass::runImpl(
7577 Function &F, ScalarEvolution &SE_, LoopInfo &LI_, TargetTransformInfo &TTI_,
7578 DominatorTree &DT_, BlockFrequencyInfo &BFI_, TargetLibraryInfo *TLI_,
7579 DemandedBits &DB_, AliasAnalysis &AA_, AssumptionCache &AC_,
7580 std::function<const LoopAccessInfo &(Loop &)> &GetLAA_,
7581 OptimizationRemarkEmitter &ORE_) {
7595 // Compute some weights outside of the loop over the loops. Compute this
7596 // using a BranchProbability to re-use its scaling math.
7597 const BranchProbability ColdProb(1, 5); // 20%
7598 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
7601 // 1. the target claims to have no vector registers, and
7602 // 2. interleaving won't help ILP.
7604 // The second condition is necessary because, even if the target has no
7605 // vector registers, loop vectorization may still enable scalar
7607 if (!TTI->getNumberOfRegisters(true) && TTI->getMaxInterleaveFactor(1) < 2)
7610 // Build up a worklist of inner-loops to vectorize. This is necessary as
7611 // the act of vectorizing or partially unrolling a loop creates new loops
7612 // and can invalidate iterators across the loops.
7613 SmallVector<Loop *, 8> Worklist;
7616 addAcyclicInnerLoop(*L, Worklist);
7618 LoopsAnalyzed += Worklist.size();
7620 // Now walk the identified inner loops.
7621 bool Changed = false;
7622 while (!Worklist.empty())
7623 Changed |= processLoop(Worklist.pop_back_val());
7625 // Process each loop nest in the function.
7631 PreservedAnalyses LoopVectorizePass::run(Function &F,
7632 FunctionAnalysisManager &AM) {
7633 auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F);
7634 auto &LI = AM.getResult<LoopAnalysis>(F);
7635 auto &TTI = AM.getResult<TargetIRAnalysis>(F);
7636 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
7637 auto &BFI = AM.getResult<BlockFrequencyAnalysis>(F);
7638 auto *TLI = AM.getCachedResult<TargetLibraryAnalysis>(F);
7639 auto &AA = AM.getResult<AAManager>(F);
7640 auto &AC = AM.getResult<AssumptionAnalysis>(F);
7641 auto &DB = AM.getResult<DemandedBitsAnalysis>(F);
7642 auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
7644 auto &LAM = AM.getResult<LoopAnalysisManagerFunctionProxy>(F).getManager();
7645 std::function<const LoopAccessInfo &(Loop &)> GetLAA =
7646 [&](Loop &L) -> const LoopAccessInfo & {
7647 return LAM.getResult<LoopAccessAnalysis>(L);
7650 runImpl(F, SE, LI, TTI, DT, BFI, TLI, DB, AA, AC, GetLAA, ORE);
7652 return PreservedAnalyses::all();
7653 PreservedAnalyses PA;
7654 PA.preserve<LoopAnalysis>();
7655 PA.preserve<DominatorTreeAnalysis>();
7656 PA.preserve<BasicAA>();
7657 PA.preserve<GlobalsAA>();