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(false), 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"));
218 // Forward declarations.
219 class LoopVectorizeHints;
220 class LoopVectorizationLegality;
221 class LoopVectorizationCostModel;
222 class LoopVectorizationRequirements;
224 // A traits type that is intended to be used in graph algorithms. The graph it
225 // models starts at the loop header, and traverses the BasicBlocks that are in
226 // the loop body, but not the loop header. Since the loop header is skipped,
227 // the back edges are excluded.
228 struct LoopBodyTraits {
229 using NodeRef = std::pair<const Loop *, BasicBlock *>;
231 // This wraps a const Loop * into the iterator, so we know which edges to
233 class WrappedSuccIterator
234 : public iterator_adaptor_base<
235 WrappedSuccIterator, succ_iterator,
236 typename std::iterator_traits<succ_iterator>::iterator_category,
237 NodeRef, std::ptrdiff_t, NodeRef *, NodeRef> {
238 using BaseT = iterator_adaptor_base<
239 WrappedSuccIterator, succ_iterator,
240 typename std::iterator_traits<succ_iterator>::iterator_category,
241 NodeRef, std::ptrdiff_t, NodeRef *, NodeRef>;
246 WrappedSuccIterator(succ_iterator Begin, const Loop *L)
247 : BaseT(Begin), L(L) {}
249 NodeRef operator*() const { return {L, *I}; }
252 struct LoopBodyFilter {
253 bool operator()(NodeRef N) const {
254 const Loop *L = N.first;
255 return N.second != L->getHeader() && L->contains(N.second);
259 using ChildIteratorType =
260 filter_iterator<WrappedSuccIterator, LoopBodyFilter>;
262 static NodeRef getEntryNode(const Loop &G) { return {&G, G.getHeader()}; }
264 static ChildIteratorType child_begin(NodeRef Node) {
265 return make_filter_range(make_range<WrappedSuccIterator>(
266 {succ_begin(Node.second), Node.first},
267 {succ_end(Node.second), Node.first}),
272 static ChildIteratorType child_end(NodeRef Node) {
273 return make_filter_range(make_range<WrappedSuccIterator>(
274 {succ_begin(Node.second), Node.first},
275 {succ_end(Node.second), Node.first}),
281 /// Returns true if the given loop body has a cycle, excluding the loop
283 static bool hasCyclesInLoopBody(const Loop &L) {
287 for (const auto SCC :
288 make_range(scc_iterator<Loop, LoopBodyTraits>::begin(L),
289 scc_iterator<Loop, LoopBodyTraits>::end(L))) {
290 if (SCC.size() > 1) {
291 DEBUG(dbgs() << "LVL: Detected a cycle in the loop body:\n");
299 /// \brief This modifies LoopAccessReport to initialize message with
300 /// loop-vectorizer-specific part.
301 class VectorizationReport : public LoopAccessReport {
303 VectorizationReport(Instruction *I = nullptr)
304 : LoopAccessReport("loop not vectorized: ", I) {}
306 /// \brief This allows promotion of the loop-access analysis report into the
307 /// loop-vectorizer report. It modifies the message to add the
308 /// loop-vectorizer-specific part of the message.
309 explicit VectorizationReport(const LoopAccessReport &R)
310 : LoopAccessReport(Twine("loop not vectorized: ") + R.str(),
314 /// A helper function for converting Scalar types to vector types.
315 /// If the incoming type is void, we return void. If the VF is 1, we return
317 static Type *ToVectorTy(Type *Scalar, unsigned VF) {
318 if (Scalar->isVoidTy() || VF == 1)
320 return VectorType::get(Scalar, VF);
323 /// A helper function that returns GEP instruction and knows to skip a
324 /// 'bitcast'. The 'bitcast' may be skipped if the source and the destination
325 /// pointee types of the 'bitcast' have the same size.
327 /// bitcast double** %var to i64* - can be skipped
328 /// bitcast double** %var to i8* - can not
329 static GetElementPtrInst *getGEPInstruction(Value *Ptr) {
331 if (isa<GetElementPtrInst>(Ptr))
332 return cast<GetElementPtrInst>(Ptr);
334 if (isa<BitCastInst>(Ptr) &&
335 isa<GetElementPtrInst>(cast<BitCastInst>(Ptr)->getOperand(0))) {
336 Type *BitcastTy = Ptr->getType();
337 Type *GEPTy = cast<BitCastInst>(Ptr)->getSrcTy();
338 if (!isa<PointerType>(BitcastTy) || !isa<PointerType>(GEPTy))
340 Type *Pointee1Ty = cast<PointerType>(BitcastTy)->getPointerElementType();
341 Type *Pointee2Ty = cast<PointerType>(GEPTy)->getPointerElementType();
342 const DataLayout &DL = cast<BitCastInst>(Ptr)->getModule()->getDataLayout();
343 if (DL.getTypeSizeInBits(Pointee1Ty) == DL.getTypeSizeInBits(Pointee2Ty))
344 return cast<GetElementPtrInst>(cast<BitCastInst>(Ptr)->getOperand(0));
349 /// InnerLoopVectorizer vectorizes loops which contain only one basic
350 /// block to a specified vectorization factor (VF).
351 /// This class performs the widening of scalars into vectors, or multiple
352 /// scalars. This class also implements the following features:
353 /// * It inserts an epilogue loop for handling loops that don't have iteration
354 /// counts that are known to be a multiple of the vectorization factor.
355 /// * It handles the code generation for reduction variables.
356 /// * Scalarization (implementation using scalars) of un-vectorizable
358 /// InnerLoopVectorizer does not perform any vectorization-legality
359 /// checks, and relies on the caller to check for the different legality
360 /// aspects. The InnerLoopVectorizer relies on the
361 /// LoopVectorizationLegality class to provide information about the induction
362 /// and reduction variables that were found to a given vectorization factor.
363 class InnerLoopVectorizer {
365 InnerLoopVectorizer(Loop *OrigLoop, PredicatedScalarEvolution &PSE,
366 LoopInfo *LI, DominatorTree *DT,
367 const TargetLibraryInfo *TLI,
368 const TargetTransformInfo *TTI, AssumptionCache *AC,
369 unsigned VecWidth, unsigned UnrollFactor)
370 : OrigLoop(OrigLoop), PSE(PSE), LI(LI), DT(DT), TLI(TLI), TTI(TTI),
371 AC(AC), VF(VecWidth), UF(UnrollFactor),
372 Builder(PSE.getSE()->getContext()), Induction(nullptr),
373 OldInduction(nullptr), WidenMap(UnrollFactor), TripCount(nullptr),
374 VectorTripCount(nullptr), Legal(nullptr), AddedSafetyChecks(false) {}
376 // Perform the actual loop widening (vectorization).
377 // MinimumBitWidths maps scalar integer values to the smallest bitwidth they
378 // can be validly truncated to. The cost model has assumed this truncation
379 // will happen when vectorizing. VecValuesToIgnore contains scalar values
380 // that the cost model has chosen to ignore because they will not be
382 void vectorize(LoopVectorizationLegality *L,
383 const MapVector<Instruction *, uint64_t> &MinimumBitWidths,
384 SmallPtrSetImpl<const Value *> &VecValuesToIgnore) {
385 MinBWs = &MinimumBitWidths;
386 ValuesNotWidened = &VecValuesToIgnore;
388 // Create a new empty loop. Unlink the old loop and connect the new one.
390 // Widen each instruction in the old loop to a new one in the new loop.
391 // Use the Legality module to find the induction and reduction variables.
395 // Return true if any runtime check is added.
396 bool areSafetyChecksAdded() { return AddedSafetyChecks; }
398 virtual ~InnerLoopVectorizer() {}
401 /// A small list of PHINodes.
402 typedef SmallVector<PHINode *, 4> PhiVector;
403 /// When we unroll loops we have multiple vector values for each scalar.
404 /// This data structure holds the unrolled and vectorized values that
405 /// originated from one scalar instruction.
406 typedef SmallVector<Value *, 2> VectorParts;
408 // When we if-convert we need to create edge masks. We have to cache values
409 // so that we don't end up with exponential recursion/IR.
410 typedef DenseMap<std::pair<BasicBlock *, BasicBlock *>, VectorParts>
413 /// Create an empty loop, based on the loop ranges of the old loop.
414 void createEmptyLoop();
416 /// Set up the values of the IVs correctly when exiting the vector loop.
417 void fixupIVUsers(PHINode *OrigPhi, const InductionDescriptor &II,
418 Value *CountRoundDown, Value *EndValue,
419 BasicBlock *MiddleBlock);
421 /// Create a new induction variable inside L.
422 PHINode *createInductionVariable(Loop *L, Value *Start, Value *End,
423 Value *Step, Instruction *DL);
424 /// Copy and widen the instructions from the old loop.
425 virtual void vectorizeLoop();
427 /// Fix a first-order recurrence. This is the second phase of vectorizing
429 void fixFirstOrderRecurrence(PHINode *Phi);
431 /// \brief The Loop exit block may have single value PHI nodes where the
432 /// incoming value is 'Undef'. While vectorizing we only handled real values
433 /// that were defined inside the loop. Here we fix the 'undef case'.
437 /// Shrinks vector element sizes based on information in "MinBWs".
438 void truncateToMinimalBitwidths();
440 /// A helper function that computes the predicate of the block BB, assuming
441 /// that the header block of the loop is set to True. It returns the *entry*
442 /// mask for the block BB.
443 VectorParts createBlockInMask(BasicBlock *BB);
444 /// A helper function that computes the predicate of the edge between SRC
446 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
448 /// A helper function to vectorize a single BB within the innermost loop.
449 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
451 /// Vectorize a single PHINode in a block. This method handles the induction
452 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
453 /// arbitrary length vectors.
454 void widenPHIInstruction(Instruction *PN, VectorParts &Entry, unsigned UF,
455 unsigned VF, PhiVector *PV);
457 /// Insert the new loop to the loop hierarchy and pass manager
458 /// and update the analysis passes.
459 void updateAnalysis();
461 /// This instruction is un-vectorizable. Implement it as a sequence
462 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
463 /// scalarized instruction behind an if block predicated on the control
464 /// dependence of the instruction.
465 virtual void scalarizeInstruction(Instruction *Instr,
466 bool IfPredicateStore = false);
468 /// Vectorize Load and Store instructions,
469 virtual void vectorizeMemoryInstruction(Instruction *Instr);
471 /// Create a broadcast instruction. This method generates a broadcast
472 /// instruction (shuffle) for loop invariant values and for the induction
473 /// value. If this is the induction variable then we extend it to N, N+1, ...
474 /// this is needed because each iteration in the loop corresponds to a SIMD
476 virtual Value *getBroadcastInstrs(Value *V);
478 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
479 /// to each vector element of Val. The sequence starts at StartIndex.
480 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step);
482 /// Compute scalar induction steps. \p ScalarIV is the scalar induction
483 /// variable on which to base the steps, \p Step is the size of the step, and
484 /// \p EntryVal is the value from the original loop that maps to the steps.
485 /// Note that \p EntryVal doesn't have to be an induction variable (e.g., it
486 /// can be a truncate instruction).
487 void buildScalarSteps(Value *ScalarIV, Value *Step, Value *EntryVal);
489 /// Create a vector induction phi node based on an existing scalar one. This
490 /// currently only works for integer induction variables with a constant
491 /// step. If \p TruncType is non-null, instead of widening the original IV,
492 /// we widen a version of the IV truncated to \p TruncType.
493 void createVectorIntInductionPHI(const InductionDescriptor &II,
494 VectorParts &Entry, IntegerType *TruncType);
496 /// Widen an integer induction variable \p IV. If \p Trunc is provided, the
497 /// induction variable will first be truncated to the corresponding type. The
498 /// widened values are placed in \p Entry.
499 void widenIntInduction(PHINode *IV, VectorParts &Entry,
500 TruncInst *Trunc = nullptr);
502 /// When we go over instructions in the basic block we rely on previous
503 /// values within the current basic block or on loop invariant values.
504 /// When we widen (vectorize) values we place them in the map. If the values
505 /// are not within the map, they have to be loop invariant, so we simply
506 /// broadcast them into a vector.
507 VectorParts &getVectorValue(Value *V);
509 /// Try to vectorize the interleaved access group that \p Instr belongs to.
510 void vectorizeInterleaveGroup(Instruction *Instr);
512 /// Generate a shuffle sequence that will reverse the vector Vec.
513 virtual Value *reverseVector(Value *Vec);
515 /// Returns (and creates if needed) the original loop trip count.
516 Value *getOrCreateTripCount(Loop *NewLoop);
518 /// Returns (and creates if needed) the trip count of the widened loop.
519 Value *getOrCreateVectorTripCount(Loop *NewLoop);
521 /// Emit a bypass check to see if the trip count would overflow, or we
522 /// wouldn't have enough iterations to execute one vector loop.
523 void emitMinimumIterationCountCheck(Loop *L, BasicBlock *Bypass);
524 /// Emit a bypass check to see if the vector trip count is nonzero.
525 void emitVectorLoopEnteredCheck(Loop *L, BasicBlock *Bypass);
526 /// Emit a bypass check to see if all of the SCEV assumptions we've
527 /// had to make are correct.
528 void emitSCEVChecks(Loop *L, BasicBlock *Bypass);
529 /// Emit bypass checks to check any memory assumptions we may have made.
530 void emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass);
532 /// Add additional metadata to \p To that was not present on \p Orig.
534 /// Currently this is used to add the noalias annotations based on the
535 /// inserted memchecks. Use this for instructions that are *cloned* into the
537 void addNewMetadata(Instruction *To, const Instruction *Orig);
539 /// Add metadata from one instruction to another.
541 /// This includes both the original MDs from \p From and additional ones (\see
542 /// addNewMetadata). Use this for *newly created* instructions in the vector
544 void addMetadata(Instruction *To, Instruction *From);
546 /// \brief Similar to the previous function but it adds the metadata to a
547 /// vector of instructions.
548 void addMetadata(ArrayRef<Value *> To, Instruction *From);
550 /// This is a helper class that holds the vectorizer state. It maps scalar
551 /// instructions to vector instructions. When the code is 'unrolled' then
552 /// then a single scalar value is mapped to multiple vector parts. The parts
553 /// are stored in the VectorPart type.
555 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
557 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
559 /// \return True if 'Key' is saved in the Value Map.
560 bool has(Value *Key) const { return MapStorage.count(Key); }
562 /// Initializes a new entry in the map. Sets all of the vector parts to the
563 /// save value in 'Val'.
564 /// \return A reference to a vector with splat values.
565 VectorParts &splat(Value *Key, Value *Val) {
566 VectorParts &Entry = MapStorage[Key];
567 Entry.assign(UF, Val);
571 ///\return A reference to the value that is stored at 'Key'.
572 VectorParts &get(Value *Key) {
573 VectorParts &Entry = MapStorage[Key];
576 assert(Entry.size() == UF);
581 /// The unroll factor. Each entry in the map stores this number of vector
585 /// Map storage. We use std::map and not DenseMap because insertions to a
586 /// dense map invalidates its iterators.
587 std::map<Value *, VectorParts> MapStorage;
590 /// The original loop.
592 /// A wrapper around ScalarEvolution used to add runtime SCEV checks. Applies
593 /// dynamic knowledge to simplify SCEV expressions and converts them to a
594 /// more usable form.
595 PredicatedScalarEvolution &PSE;
602 /// Target Library Info.
603 const TargetLibraryInfo *TLI;
604 /// Target Transform Info.
605 const TargetTransformInfo *TTI;
606 /// Assumption Cache.
609 /// \brief LoopVersioning. It's only set up (non-null) if memchecks were
612 /// This is currently only used to add no-alias metadata based on the
613 /// memchecks. The actually versioning is performed manually.
614 std::unique_ptr<LoopVersioning> LVer;
616 /// The vectorization SIMD factor to use. Each vector will have this many
621 /// The vectorization unroll factor to use. Each scalar is vectorized to this
622 /// many different vector instructions.
625 /// The builder that we use
628 // --- Vectorization state ---
630 /// The vector-loop preheader.
631 BasicBlock *LoopVectorPreHeader;
632 /// The scalar-loop preheader.
633 BasicBlock *LoopScalarPreHeader;
634 /// Middle Block between the vector and the scalar.
635 BasicBlock *LoopMiddleBlock;
636 /// The ExitBlock of the scalar loop.
637 BasicBlock *LoopExitBlock;
638 /// The vector loop body.
639 BasicBlock *LoopVectorBody;
640 /// The scalar loop body.
641 BasicBlock *LoopScalarBody;
642 /// A list of all bypass blocks. The first block is the entry of the loop.
643 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
645 /// The new Induction variable which was added to the new block.
647 /// The induction variable of the old basic block.
648 PHINode *OldInduction;
649 /// Maps scalars to widened vectors.
652 /// A map of induction variables from the original loop to their
653 /// corresponding VF * UF scalarized values in the vectorized loop. The
654 /// purpose of ScalarIVMap is similar to that of WidenMap. Whereas WidenMap
655 /// maps original loop values to their vector versions in the new loop,
656 /// ScalarIVMap maps induction variables from the original loop that are not
657 /// vectorized to their scalar equivalents in the vector loop. Maintaining a
658 /// separate map for scalarized induction variables allows us to avoid
659 /// unnecessary scalar-to-vector-to-scalar conversions.
660 DenseMap<Value *, SmallVector<Value *, 8>> ScalarIVMap;
662 /// Store instructions that should be predicated, as a pair
663 /// <StoreInst, Predicate>
664 SmallVector<std::pair<StoreInst *, Value *>, 4> PredicatedStores;
665 EdgeMaskCache MaskCache;
666 /// Trip count of the original loop.
668 /// Trip count of the widened loop (TripCount - TripCount % (VF*UF))
669 Value *VectorTripCount;
671 /// Map of scalar integer values to the smallest bitwidth they can be legally
672 /// represented as. The vector equivalents of these values should be truncated
674 const MapVector<Instruction *, uint64_t> *MinBWs;
676 /// A set of values that should not be widened. This is taken from
677 /// VecValuesToIgnore in the cost model.
678 SmallPtrSetImpl<const Value *> *ValuesNotWidened;
680 LoopVectorizationLegality *Legal;
682 // Record whether runtime checks are added.
683 bool AddedSafetyChecks;
686 class InnerLoopUnroller : public InnerLoopVectorizer {
688 InnerLoopUnroller(Loop *OrigLoop, PredicatedScalarEvolution &PSE,
689 LoopInfo *LI, DominatorTree *DT,
690 const TargetLibraryInfo *TLI,
691 const TargetTransformInfo *TTI, AssumptionCache *AC,
692 unsigned UnrollFactor)
693 : InnerLoopVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, 1,
697 void scalarizeInstruction(Instruction *Instr,
698 bool IfPredicateStore = false) override;
699 void vectorizeMemoryInstruction(Instruction *Instr) override;
700 Value *getBroadcastInstrs(Value *V) override;
701 Value *getStepVector(Value *Val, int StartIdx, Value *Step) override;
702 Value *reverseVector(Value *Vec) override;
705 /// \brief Look for a meaningful debug location on the instruction or it's
707 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
712 if (I->getDebugLoc() != Empty)
715 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
716 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
717 if (OpInst->getDebugLoc() != Empty)
724 /// \brief Set the debug location in the builder using the debug location in the
726 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
727 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
728 B.SetCurrentDebugLocation(Inst->getDebugLoc());
730 B.SetCurrentDebugLocation(DebugLoc());
734 /// \return string containing a file name and a line # for the given loop.
735 static std::string getDebugLocString(const Loop *L) {
738 raw_string_ostream OS(Result);
739 if (const DebugLoc LoopDbgLoc = L->getStartLoc())
740 LoopDbgLoc.print(OS);
742 // Just print the module name.
743 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
750 void InnerLoopVectorizer::addNewMetadata(Instruction *To,
751 const Instruction *Orig) {
752 // If the loop was versioned with memchecks, add the corresponding no-alias
754 if (LVer && (isa<LoadInst>(Orig) || isa<StoreInst>(Orig)))
755 LVer->annotateInstWithNoAlias(To, Orig);
758 void InnerLoopVectorizer::addMetadata(Instruction *To,
760 propagateMetadata(To, From);
761 addNewMetadata(To, From);
764 void InnerLoopVectorizer::addMetadata(ArrayRef<Value *> To,
766 for (Value *V : To) {
767 if (Instruction *I = dyn_cast<Instruction>(V))
768 addMetadata(I, From);
772 /// \brief The group of interleaved loads/stores sharing the same stride and
773 /// close to each other.
775 /// Each member in this group has an index starting from 0, and the largest
776 /// index should be less than interleaved factor, which is equal to the absolute
777 /// value of the access's stride.
779 /// E.g. An interleaved load group of factor 4:
780 /// for (unsigned i = 0; i < 1024; i+=4) {
781 /// a = A[i]; // Member of index 0
782 /// b = A[i+1]; // Member of index 1
783 /// d = A[i+3]; // Member of index 3
787 /// An interleaved store group of factor 4:
788 /// for (unsigned i = 0; i < 1024; i+=4) {
790 /// A[i] = a; // Member of index 0
791 /// A[i+1] = b; // Member of index 1
792 /// A[i+2] = c; // Member of index 2
793 /// A[i+3] = d; // Member of index 3
796 /// Note: the interleaved load group could have gaps (missing members), but
797 /// the interleaved store group doesn't allow gaps.
798 class InterleaveGroup {
800 InterleaveGroup(Instruction *Instr, int Stride, unsigned Align)
801 : Align(Align), SmallestKey(0), LargestKey(0), InsertPos(Instr) {
802 assert(Align && "The alignment should be non-zero");
804 Factor = std::abs(Stride);
805 assert(Factor > 1 && "Invalid interleave factor");
807 Reverse = Stride < 0;
811 bool isReverse() const { return Reverse; }
812 unsigned getFactor() const { return Factor; }
813 unsigned getAlignment() const { return Align; }
814 unsigned getNumMembers() const { return Members.size(); }
816 /// \brief Try to insert a new member \p Instr with index \p Index and
817 /// alignment \p NewAlign. The index is related to the leader and it could be
818 /// negative if it is the new leader.
820 /// \returns false if the instruction doesn't belong to the group.
821 bool insertMember(Instruction *Instr, int Index, unsigned NewAlign) {
822 assert(NewAlign && "The new member's alignment should be non-zero");
824 int Key = Index + SmallestKey;
826 // Skip if there is already a member with the same index.
827 if (Members.count(Key))
830 if (Key > LargestKey) {
831 // The largest index is always less than the interleave factor.
832 if (Index >= static_cast<int>(Factor))
836 } else if (Key < SmallestKey) {
837 // The largest index is always less than the interleave factor.
838 if (LargestKey - Key >= static_cast<int>(Factor))
844 // It's always safe to select the minimum alignment.
845 Align = std::min(Align, NewAlign);
846 Members[Key] = Instr;
850 /// \brief Get the member with the given index \p Index
852 /// \returns nullptr if contains no such member.
853 Instruction *getMember(unsigned Index) const {
854 int Key = SmallestKey + Index;
855 if (!Members.count(Key))
858 return Members.find(Key)->second;
861 /// \brief Get the index for the given member. Unlike the key in the member
862 /// map, the index starts from 0.
863 unsigned getIndex(Instruction *Instr) const {
864 for (auto I : Members)
865 if (I.second == Instr)
866 return I.first - SmallestKey;
868 llvm_unreachable("InterleaveGroup contains no such member");
871 Instruction *getInsertPos() const { return InsertPos; }
872 void setInsertPos(Instruction *Inst) { InsertPos = Inst; }
875 unsigned Factor; // Interleave Factor.
878 DenseMap<int, Instruction *> Members;
882 // To avoid breaking dependences, vectorized instructions of an interleave
883 // group should be inserted at either the first load or the last store in
886 // E.g. %even = load i32 // Insert Position
887 // %add = add i32 %even // Use of %even
891 // %odd = add i32 // Def of %odd
892 // store i32 %odd // Insert Position
893 Instruction *InsertPos;
896 /// \brief Drive the analysis of interleaved memory accesses in the loop.
898 /// Use this class to analyze interleaved accesses only when we can vectorize
899 /// a loop. Otherwise it's meaningless to do analysis as the vectorization
900 /// on interleaved accesses is unsafe.
902 /// The analysis collects interleave groups and records the relationships
903 /// between the member and the group in a map.
904 class InterleavedAccessInfo {
906 InterleavedAccessInfo(PredicatedScalarEvolution &PSE, Loop *L,
907 DominatorTree *DT, LoopInfo *LI)
908 : PSE(PSE), TheLoop(L), DT(DT), LI(LI), LAI(nullptr),
909 RequiresScalarEpilogue(false) {}
911 ~InterleavedAccessInfo() {
912 SmallSet<InterleaveGroup *, 4> DelSet;
913 // Avoid releasing a pointer twice.
914 for (auto &I : InterleaveGroupMap)
915 DelSet.insert(I.second);
916 for (auto *Ptr : DelSet)
920 /// \brief Analyze the interleaved accesses and collect them in interleave
921 /// groups. Substitute symbolic strides using \p Strides.
922 void analyzeInterleaving(const ValueToValueMap &Strides);
924 /// \brief Check if \p Instr belongs to any interleave group.
925 bool isInterleaved(Instruction *Instr) const {
926 return InterleaveGroupMap.count(Instr);
929 /// \brief Return the maximum interleave factor of all interleaved groups.
930 unsigned getMaxInterleaveFactor() const {
931 unsigned MaxFactor = 1;
932 for (auto &Entry : InterleaveGroupMap)
933 MaxFactor = std::max(MaxFactor, Entry.second->getFactor());
937 /// \brief Get the interleave group that \p Instr belongs to.
939 /// \returns nullptr if doesn't have such group.
940 InterleaveGroup *getInterleaveGroup(Instruction *Instr) const {
941 if (InterleaveGroupMap.count(Instr))
942 return InterleaveGroupMap.find(Instr)->second;
946 /// \brief Returns true if an interleaved group that may access memory
947 /// out-of-bounds requires a scalar epilogue iteration for correctness.
948 bool requiresScalarEpilogue() const { return RequiresScalarEpilogue; }
950 /// \brief Initialize the LoopAccessInfo used for dependence checking.
951 void setLAI(const LoopAccessInfo *Info) { LAI = Info; }
954 /// A wrapper around ScalarEvolution, used to add runtime SCEV checks.
955 /// Simplifies SCEV expressions in the context of existing SCEV assumptions.
956 /// The interleaved access analysis can also add new predicates (for example
957 /// by versioning strides of pointers).
958 PredicatedScalarEvolution &PSE;
962 const LoopAccessInfo *LAI;
964 /// True if the loop may contain non-reversed interleaved groups with
965 /// out-of-bounds accesses. We ensure we don't speculatively access memory
966 /// out-of-bounds by executing at least one scalar epilogue iteration.
967 bool RequiresScalarEpilogue;
969 /// Holds the relationships between the members and the interleave group.
970 DenseMap<Instruction *, InterleaveGroup *> InterleaveGroupMap;
972 /// Holds dependences among the memory accesses in the loop. It maps a source
973 /// access to a set of dependent sink accesses.
974 DenseMap<Instruction *, SmallPtrSet<Instruction *, 2>> Dependences;
976 /// \brief The descriptor for a strided memory access.
977 struct StrideDescriptor {
978 StrideDescriptor(int64_t Stride, const SCEV *Scev, uint64_t Size,
980 : Stride(Stride), Scev(Scev), Size(Size), Align(Align) {}
982 StrideDescriptor() = default;
984 // The access's stride. It is negative for a reverse access.
986 const SCEV *Scev = nullptr; // The scalar expression of this access
987 uint64_t Size = 0; // The size of the memory object.
988 unsigned Align = 0; // The alignment of this access.
991 /// \brief A type for holding instructions and their stride descriptors.
992 typedef std::pair<Instruction *, StrideDescriptor> StrideEntry;
994 /// \brief Create a new interleave group with the given instruction \p Instr,
995 /// stride \p Stride and alignment \p Align.
997 /// \returns the newly created interleave group.
998 InterleaveGroup *createInterleaveGroup(Instruction *Instr, int Stride,
1000 assert(!InterleaveGroupMap.count(Instr) &&
1001 "Already in an interleaved access group");
1002 InterleaveGroupMap[Instr] = new InterleaveGroup(Instr, Stride, Align);
1003 return InterleaveGroupMap[Instr];
1006 /// \brief Release the group and remove all the relationships.
1007 void releaseGroup(InterleaveGroup *Group) {
1008 for (unsigned i = 0; i < Group->getFactor(); i++)
1009 if (Instruction *Member = Group->getMember(i))
1010 InterleaveGroupMap.erase(Member);
1015 /// \brief Collect all the accesses with a constant stride in program order.
1016 void collectConstStrideAccesses(
1017 MapVector<Instruction *, StrideDescriptor> &AccessStrideInfo,
1018 const ValueToValueMap &Strides);
1020 /// \brief Returns true if \p Stride is allowed in an interleaved group.
1021 static bool isStrided(int Stride) {
1022 unsigned Factor = std::abs(Stride);
1023 return Factor >= 2 && Factor <= MaxInterleaveGroupFactor;
1026 /// \brief Returns true if \p BB is a predicated block.
1027 bool isPredicated(BasicBlock *BB) const {
1028 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
1031 /// \brief Returns true if LoopAccessInfo can be used for dependence queries.
1032 bool areDependencesValid() const {
1033 return LAI && LAI->getDepChecker().getDependences();
1036 /// \brief Returns true if memory accesses \p A and \p B can be reordered, if
1037 /// necessary, when constructing interleaved groups.
1039 /// \p A must precede \p B in program order. We return false if reordering is
1040 /// not necessary or is prevented because \p A and \p B may be dependent.
1041 bool canReorderMemAccessesForInterleavedGroups(StrideEntry *A,
1042 StrideEntry *B) const {
1044 // Code motion for interleaved accesses can potentially hoist strided loads
1045 // and sink strided stores. The code below checks the legality of the
1046 // following two conditions:
1048 // 1. Potentially moving a strided load (B) before any store (A) that
1051 // 2. Potentially moving a strided store (A) after any load or store (B)
1054 // It's legal to reorder A and B if we know there isn't a dependence from A
1055 // to B. Note that this determination is conservative since some
1056 // dependences could potentially be reordered safely.
1058 // A is potentially the source of a dependence.
1059 auto *Src = A->first;
1060 auto SrcDes = A->second;
1062 // B is potentially the sink of a dependence.
1063 auto *Sink = B->first;
1064 auto SinkDes = B->second;
1066 // Code motion for interleaved accesses can't violate WAR dependences.
1067 // Thus, reordering is legal if the source isn't a write.
1068 if (!Src->mayWriteToMemory())
1071 // At least one of the accesses must be strided.
1072 if (!isStrided(SrcDes.Stride) && !isStrided(SinkDes.Stride))
1075 // If dependence information is not available from LoopAccessInfo,
1076 // conservatively assume the instructions can't be reordered.
1077 if (!areDependencesValid())
1080 // If we know there is a dependence from source to sink, assume the
1081 // instructions can't be reordered. Otherwise, reordering is legal.
1082 return !Dependences.count(Src) || !Dependences.lookup(Src).count(Sink);
1085 /// \brief Collect the dependences from LoopAccessInfo.
1087 /// We process the dependences once during the interleaved access analysis to
1088 /// enable constant-time dependence queries.
1089 void collectDependences() {
1090 if (!areDependencesValid())
1092 auto *Deps = LAI->getDepChecker().getDependences();
1093 for (auto Dep : *Deps)
1094 Dependences[Dep.getSource(*LAI)].insert(Dep.getDestination(*LAI));
1098 /// Utility class for getting and setting loop vectorizer hints in the form
1099 /// of loop metadata.
1100 /// This class keeps a number of loop annotations locally (as member variables)
1101 /// and can, upon request, write them back as metadata on the loop. It will
1102 /// initially scan the loop for existing metadata, and will update the local
1103 /// values based on information in the loop.
1104 /// We cannot write all values to metadata, as the mere presence of some info,
1105 /// for example 'force', means a decision has been made. So, we need to be
1106 /// careful NOT to add them if the user hasn't specifically asked so.
1107 class LoopVectorizeHints {
1108 enum HintKind { HK_WIDTH, HK_UNROLL, HK_FORCE };
1110 /// Hint - associates name and validation with the hint value.
1113 unsigned Value; // This may have to change for non-numeric values.
1116 Hint(const char *Name, unsigned Value, HintKind Kind)
1117 : Name(Name), Value(Value), Kind(Kind) {}
1119 bool validate(unsigned Val) {
1122 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
1124 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1132 /// Vectorization width.
1134 /// Vectorization interleave factor.
1136 /// Vectorization forced
1139 /// Return the loop metadata prefix.
1140 static StringRef Prefix() { return "llvm.loop."; }
1142 /// True if there is any unsafe math in the loop.
1143 bool PotentiallyUnsafe;
1147 FK_Undefined = -1, ///< Not selected.
1148 FK_Disabled = 0, ///< Forcing disabled.
1149 FK_Enabled = 1, ///< Forcing enabled.
1152 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
1153 : Width("vectorize.width", VectorizerParams::VectorizationFactor,
1155 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1156 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1157 PotentiallyUnsafe(false), TheLoop(L) {
1158 // Populate values with existing loop metadata.
1159 getHintsFromMetadata();
1161 // force-vector-interleave overrides DisableInterleaving.
1162 if (VectorizerParams::isInterleaveForced())
1163 Interleave.Value = VectorizerParams::VectorizationInterleave;
1165 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
1166 << "LV: Interleaving disabled by the pass manager\n");
1169 /// Mark the loop L as already vectorized by setting the width to 1.
1170 void setAlreadyVectorized() {
1171 Width.Value = Interleave.Value = 1;
1172 Hint Hints[] = {Width, Interleave};
1173 writeHintsToMetadata(Hints);
1176 bool allowVectorization(Function *F, Loop *L, bool AlwaysVectorize) const {
1177 if (getForce() == LoopVectorizeHints::FK_Disabled) {
1178 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1179 emitOptimizationRemarkAnalysis(F->getContext(),
1180 vectorizeAnalysisPassName(), *F,
1181 L->getStartLoc(), emitRemark());
1185 if (!AlwaysVectorize && getForce() != LoopVectorizeHints::FK_Enabled) {
1186 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1187 emitOptimizationRemarkAnalysis(F->getContext(),
1188 vectorizeAnalysisPassName(), *F,
1189 L->getStartLoc(), emitRemark());
1193 if (getWidth() == 1 && getInterleave() == 1) {
1194 // FIXME: Add a separate metadata to indicate when the loop has already
1195 // been vectorized instead of setting width and count to 1.
1196 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1197 // FIXME: Add interleave.disable metadata. This will allow
1198 // vectorize.disable to be used without disabling the pass and errors
1199 // to differentiate between disabled vectorization and a width of 1.
1200 emitOptimizationRemarkAnalysis(
1201 F->getContext(), vectorizeAnalysisPassName(), *F, L->getStartLoc(),
1202 "loop not vectorized: vectorization and interleaving are explicitly "
1203 "disabled, or vectorize width and interleave count are both set to "
1211 /// Dumps all the hint information.
1212 std::string emitRemark() const {
1213 VectorizationReport R;
1214 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1215 R << "vectorization is explicitly disabled";
1217 R << "use -Rpass-analysis=loop-vectorize for more info";
1218 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1219 R << " (Force=true";
1220 if (Width.Value != 0)
1221 R << ", Vector Width=" << Width.Value;
1222 if (Interleave.Value != 0)
1223 R << ", Interleave Count=" << Interleave.Value;
1231 unsigned getWidth() const { return Width.Value; }
1232 unsigned getInterleave() const { return Interleave.Value; }
1233 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1235 /// \brief If hints are provided that force vectorization, use the AlwaysPrint
1236 /// pass name to force the frontend to print the diagnostic.
1237 const char *vectorizeAnalysisPassName() const {
1238 if (getWidth() == 1)
1240 if (getForce() == LoopVectorizeHints::FK_Disabled)
1242 if (getForce() == LoopVectorizeHints::FK_Undefined && getWidth() == 0)
1244 return DiagnosticInfoOptimizationRemarkAnalysis::AlwaysPrint;
1247 bool allowReordering() const {
1248 // When enabling loop hints are provided we allow the vectorizer to change
1249 // the order of operations that is given by the scalar loop. This is not
1250 // enabled by default because can be unsafe or inefficient. For example,
1251 // reordering floating-point operations will change the way round-off
1252 // error accumulates in the loop.
1253 return getForce() == LoopVectorizeHints::FK_Enabled || getWidth() > 1;
1256 bool isPotentiallyUnsafe() const {
1257 // Avoid FP vectorization if the target is unsure about proper support.
1258 // This may be related to the SIMD unit in the target not handling
1259 // IEEE 754 FP ops properly, or bad single-to-double promotions.
1260 // Otherwise, a sequence of vectorized loops, even without reduction,
1261 // could lead to different end results on the destination vectors.
1262 return getForce() != LoopVectorizeHints::FK_Enabled && PotentiallyUnsafe;
1265 void setPotentiallyUnsafe() { PotentiallyUnsafe = true; }
1268 /// Find hints specified in the loop metadata and update local values.
1269 void getHintsFromMetadata() {
1270 MDNode *LoopID = TheLoop->getLoopID();
1274 // First operand should refer to the loop id itself.
1275 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1276 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1278 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1279 const MDString *S = nullptr;
1280 SmallVector<Metadata *, 4> Args;
1282 // The expected hint is either a MDString or a MDNode with the first
1283 // operand a MDString.
1284 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1285 if (!MD || MD->getNumOperands() == 0)
1287 S = dyn_cast<MDString>(MD->getOperand(0));
1288 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1289 Args.push_back(MD->getOperand(i));
1291 S = dyn_cast<MDString>(LoopID->getOperand(i));
1292 assert(Args.size() == 0 && "too many arguments for MDString");
1298 // Check if the hint starts with the loop metadata prefix.
1299 StringRef Name = S->getString();
1300 if (Args.size() == 1)
1301 setHint(Name, Args[0]);
1305 /// Checks string hint with one operand and set value if valid.
1306 void setHint(StringRef Name, Metadata *Arg) {
1307 if (!Name.startswith(Prefix()))
1309 Name = Name.substr(Prefix().size(), StringRef::npos);
1311 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1314 unsigned Val = C->getZExtValue();
1316 Hint *Hints[] = {&Width, &Interleave, &Force};
1317 for (auto H : Hints) {
1318 if (Name == H->Name) {
1319 if (H->validate(Val))
1322 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1328 /// Create a new hint from name / value pair.
1329 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1330 LLVMContext &Context = TheLoop->getHeader()->getContext();
1331 Metadata *MDs[] = {MDString::get(Context, Name),
1332 ConstantAsMetadata::get(
1333 ConstantInt::get(Type::getInt32Ty(Context), V))};
1334 return MDNode::get(Context, MDs);
1337 /// Matches metadata with hint name.
1338 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1339 MDString *Name = dyn_cast<MDString>(Node->getOperand(0));
1343 for (auto H : HintTypes)
1344 if (Name->getString().endswith(H.Name))
1349 /// Sets current hints into loop metadata, keeping other values intact.
1350 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1351 if (HintTypes.size() == 0)
1354 // Reserve the first element to LoopID (see below).
1355 SmallVector<Metadata *, 4> MDs(1);
1356 // If the loop already has metadata, then ignore the existing operands.
1357 MDNode *LoopID = TheLoop->getLoopID();
1359 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1360 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1361 // If node in update list, ignore old value.
1362 if (!matchesHintMetadataName(Node, HintTypes))
1363 MDs.push_back(Node);
1367 // Now, add the missing hints.
1368 for (auto H : HintTypes)
1369 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1371 // Replace current metadata node with new one.
1372 LLVMContext &Context = TheLoop->getHeader()->getContext();
1373 MDNode *NewLoopID = MDNode::get(Context, MDs);
1374 // Set operand 0 to refer to the loop id itself.
1375 NewLoopID->replaceOperandWith(0, NewLoopID);
1377 TheLoop->setLoopID(NewLoopID);
1380 /// The loop these hints belong to.
1381 const Loop *TheLoop;
1384 static void emitAnalysisDiag(const Function *TheFunction, const Loop *TheLoop,
1385 const LoopVectorizeHints &Hints,
1386 const LoopAccessReport &Message) {
1387 const char *Name = Hints.vectorizeAnalysisPassName();
1388 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, Name);
1391 static void emitMissedWarning(Function *F, Loop *L,
1392 const LoopVectorizeHints &LH) {
1393 emitOptimizationRemarkMissed(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1396 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1397 if (LH.getWidth() != 1)
1398 emitLoopVectorizeWarning(
1399 F->getContext(), *F, L->getStartLoc(),
1400 "failed explicitly specified loop vectorization");
1401 else if (LH.getInterleave() != 1)
1402 emitLoopInterleaveWarning(
1403 F->getContext(), *F, L->getStartLoc(),
1404 "failed explicitly specified loop interleaving");
1408 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
1409 /// to what vectorization factor.
1410 /// This class does not look at the profitability of vectorization, only the
1411 /// legality. This class has two main kinds of checks:
1412 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
1413 /// will change the order of memory accesses in a way that will change the
1414 /// correctness of the program.
1415 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
1416 /// checks for a number of different conditions, such as the availability of a
1417 /// single induction variable, that all types are supported and vectorize-able,
1418 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
1419 /// This class is also used by InnerLoopVectorizer for identifying
1420 /// induction variable and the different reduction variables.
1421 class LoopVectorizationLegality {
1423 LoopVectorizationLegality(
1424 Loop *L, PredicatedScalarEvolution &PSE, DominatorTree *DT,
1425 TargetLibraryInfo *TLI, AliasAnalysis *AA, Function *F,
1426 const TargetTransformInfo *TTI,
1427 std::function<const LoopAccessInfo &(Loop &)> *GetLAA, LoopInfo *LI,
1428 LoopVectorizationRequirements *R, LoopVectorizeHints *H)
1429 : NumPredStores(0), TheLoop(L), PSE(PSE), TLI(TLI), TheFunction(F),
1430 TTI(TTI), DT(DT), GetLAA(GetLAA), LAI(nullptr),
1431 InterleaveInfo(PSE, L, DT, LI), Induction(nullptr),
1432 WidestIndTy(nullptr), HasFunNoNaNAttr(false), Requirements(R),
1435 /// ReductionList contains the reduction descriptors for all
1436 /// of the reductions that were found in the loop.
1437 typedef DenseMap<PHINode *, RecurrenceDescriptor> ReductionList;
1439 /// InductionList saves induction variables and maps them to the
1440 /// induction descriptor.
1441 typedef MapVector<PHINode *, InductionDescriptor> InductionList;
1443 /// RecurrenceSet contains the phi nodes that are recurrences other than
1444 /// inductions and reductions.
1445 typedef SmallPtrSet<const PHINode *, 8> RecurrenceSet;
1447 /// Returns true if it is legal to vectorize this loop.
1448 /// This does not mean that it is profitable to vectorize this
1449 /// loop, only that it is legal to do so.
1450 bool canVectorize();
1452 /// Returns the Induction variable.
1453 PHINode *getInduction() { return Induction; }
1455 /// Returns the reduction variables found in the loop.
1456 ReductionList *getReductionVars() { return &Reductions; }
1458 /// Returns the induction variables found in the loop.
1459 InductionList *getInductionVars() { return &Inductions; }
1461 /// Return the first-order recurrences found in the loop.
1462 RecurrenceSet *getFirstOrderRecurrences() { return &FirstOrderRecurrences; }
1464 /// Returns the widest induction type.
1465 Type *getWidestInductionType() { return WidestIndTy; }
1467 /// Returns True if V is an induction variable in this loop.
1468 bool isInductionVariable(const Value *V);
1470 /// Returns True if PN is a reduction variable in this loop.
1471 bool isReductionVariable(PHINode *PN) { return Reductions.count(PN); }
1473 /// Returns True if Phi is a first-order recurrence in this loop.
1474 bool isFirstOrderRecurrence(const PHINode *Phi);
1476 /// Return true if the block BB needs to be predicated in order for the loop
1477 /// to be vectorized.
1478 bool blockNeedsPredication(BasicBlock *BB);
1480 /// Check if this pointer is consecutive when vectorizing. This happens
1481 /// when the last index of the GEP is the induction variable, or that the
1482 /// pointer itself is an induction variable.
1483 /// This check allows us to vectorize A[idx] into a wide load/store.
1485 /// 0 - Stride is unknown or non-consecutive.
1486 /// 1 - Address is consecutive.
1487 /// -1 - Address is consecutive, and decreasing.
1488 int isConsecutivePtr(Value *Ptr);
1490 /// Returns true if the value V is uniform within the loop.
1491 bool isUniform(Value *V);
1493 /// Returns true if this instruction will remain scalar after vectorization.
1494 bool isUniformAfterVectorization(Instruction *I) { return Uniforms.count(I); }
1496 /// Returns the information that we collected about runtime memory check.
1497 const RuntimePointerChecking *getRuntimePointerChecking() const {
1498 return LAI->getRuntimePointerChecking();
1501 const LoopAccessInfo *getLAI() const { return LAI; }
1503 /// \brief Check if \p Instr belongs to any interleaved access group.
1504 bool isAccessInterleaved(Instruction *Instr) {
1505 return InterleaveInfo.isInterleaved(Instr);
1508 /// \brief Return the maximum interleave factor of all interleaved groups.
1509 unsigned getMaxInterleaveFactor() const {
1510 return InterleaveInfo.getMaxInterleaveFactor();
1513 /// \brief Get the interleaved access group that \p Instr belongs to.
1514 const InterleaveGroup *getInterleavedAccessGroup(Instruction *Instr) {
1515 return InterleaveInfo.getInterleaveGroup(Instr);
1518 /// \brief Returns true if an interleaved group requires a scalar iteration
1519 /// to handle accesses with gaps.
1520 bool requiresScalarEpilogue() const {
1521 return InterleaveInfo.requiresScalarEpilogue();
1524 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
1526 bool hasStride(Value *V) { return LAI->hasStride(V); }
1528 /// Returns true if the target machine supports masked store operation
1529 /// for the given \p DataType and kind of access to \p Ptr.
1530 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
1531 return isConsecutivePtr(Ptr) && TTI->isLegalMaskedStore(DataType);
1533 /// Returns true if the target machine supports masked load operation
1534 /// for the given \p DataType and kind of access to \p Ptr.
1535 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
1536 return isConsecutivePtr(Ptr) && TTI->isLegalMaskedLoad(DataType);
1538 /// Returns true if the target machine supports masked scatter operation
1539 /// for the given \p DataType.
1540 bool isLegalMaskedScatter(Type *DataType) {
1541 return TTI->isLegalMaskedScatter(DataType);
1543 /// Returns true if the target machine supports masked gather operation
1544 /// for the given \p DataType.
1545 bool isLegalMaskedGather(Type *DataType) {
1546 return TTI->isLegalMaskedGather(DataType);
1549 /// Returns true if vector representation of the instruction \p I
1551 bool isMaskRequired(const Instruction *I) { return (MaskedOp.count(I) != 0); }
1552 unsigned getNumStores() const { return LAI->getNumStores(); }
1553 unsigned getNumLoads() const { return LAI->getNumLoads(); }
1554 unsigned getNumPredStores() const { return NumPredStores; }
1557 /// Check if a single basic block loop is vectorizable.
1558 /// At this point we know that this is a loop with a constant trip count
1559 /// and we only need to check individual instructions.
1560 bool canVectorizeInstrs();
1562 /// When we vectorize loops we may change the order in which
1563 /// we read and write from memory. This method checks if it is
1564 /// legal to vectorize the code, considering only memory constrains.
1565 /// Returns true if the loop is vectorizable
1566 bool canVectorizeMemory();
1568 /// Return true if we can vectorize this loop using the IF-conversion
1570 bool canVectorizeWithIfConvert();
1572 /// Collect the variables that need to stay uniform after vectorization.
1573 void collectLoopUniforms();
1575 /// Return true if all of the instructions in the block can be speculatively
1576 /// executed. \p SafePtrs is a list of addresses that are known to be legal
1577 /// and we know that we can read from them without segfault.
1578 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
1580 /// Updates the vectorization state by adding \p Phi to the inductions list.
1581 /// This can set \p Phi as the main induction of the loop if \p Phi is a
1582 /// better choice for the main induction than the existing one.
1583 void addInductionPhi(PHINode *Phi, const InductionDescriptor &ID,
1584 SmallPtrSetImpl<Value *> &AllowedExit);
1586 /// Report an analysis message to assist the user in diagnosing loops that are
1587 /// not vectorized. These are handled as LoopAccessReport rather than
1588 /// VectorizationReport because the << operator of VectorizationReport returns
1589 /// LoopAccessReport.
1590 void emitAnalysis(const LoopAccessReport &Message) const {
1591 emitAnalysisDiag(TheFunction, TheLoop, *Hints, Message);
1594 /// \brief If an access has a symbolic strides, this maps the pointer value to
1595 /// the stride symbol.
1596 const ValueToValueMap *getSymbolicStrides() {
1597 // FIXME: Currently, the set of symbolic strides is sometimes queried before
1598 // it's collected. This happens from canVectorizeWithIfConvert, when the
1599 // pointer is checked to reference consecutive elements suitable for a
1601 return LAI ? &LAI->getSymbolicStrides() : nullptr;
1604 unsigned NumPredStores;
1606 /// The loop that we evaluate.
1608 /// A wrapper around ScalarEvolution used to add runtime SCEV checks.
1609 /// Applies dynamic knowledge to simplify SCEV expressions in the context
1610 /// of existing SCEV assumptions. The analysis will also add a minimal set
1611 /// of new predicates if this is required to enable vectorization and
1613 PredicatedScalarEvolution &PSE;
1614 /// Target Library Info.
1615 TargetLibraryInfo *TLI;
1617 Function *TheFunction;
1618 /// Target Transform Info
1619 const TargetTransformInfo *TTI;
1622 // LoopAccess analysis.
1623 std::function<const LoopAccessInfo &(Loop &)> *GetLAA;
1624 // And the loop-accesses info corresponding to this loop. This pointer is
1625 // null until canVectorizeMemory sets it up.
1626 const LoopAccessInfo *LAI;
1628 /// The interleave access information contains groups of interleaved accesses
1629 /// with the same stride and close to each other.
1630 InterleavedAccessInfo InterleaveInfo;
1632 // --- vectorization state --- //
1634 /// Holds the integer induction variable. This is the counter of the
1637 /// Holds the reduction variables.
1638 ReductionList Reductions;
1639 /// Holds all of the induction variables that we found in the loop.
1640 /// Notice that inductions don't need to start at zero and that induction
1641 /// variables can be pointers.
1642 InductionList Inductions;
1643 /// Holds the phi nodes that are first-order recurrences.
1644 RecurrenceSet FirstOrderRecurrences;
1645 /// Holds the widest induction type encountered.
1648 /// Allowed outside users. This holds the induction and reduction
1649 /// vars which can be accessed from outside the loop.
1650 SmallPtrSet<Value *, 4> AllowedExit;
1651 /// This set holds the variables which are known to be uniform after
1653 SmallPtrSet<Instruction *, 4> Uniforms;
1655 /// Can we assume the absence of NaNs.
1656 bool HasFunNoNaNAttr;
1658 /// Vectorization requirements that will go through late-evaluation.
1659 LoopVectorizationRequirements *Requirements;
1661 /// Used to emit an analysis of any legality issues.
1662 LoopVectorizeHints *Hints;
1664 /// While vectorizing these instructions we have to generate a
1665 /// call to the appropriate masked intrinsic
1666 SmallPtrSet<const Instruction *, 8> MaskedOp;
1669 /// LoopVectorizationCostModel - estimates the expected speedups due to
1671 /// In many cases vectorization is not profitable. This can happen because of
1672 /// a number of reasons. In this class we mainly attempt to predict the
1673 /// expected speedup/slowdowns due to the supported instruction set. We use the
1674 /// TargetTransformInfo to query the different backends for the cost of
1675 /// different operations.
1676 class LoopVectorizationCostModel {
1678 LoopVectorizationCostModel(Loop *L, PredicatedScalarEvolution &PSE,
1679 LoopInfo *LI, LoopVectorizationLegality *Legal,
1680 const TargetTransformInfo &TTI,
1681 const TargetLibraryInfo *TLI, DemandedBits *DB,
1682 AssumptionCache *AC, const Function *F,
1683 const LoopVectorizeHints *Hints)
1684 : TheLoop(L), PSE(PSE), LI(LI), Legal(Legal), TTI(TTI), TLI(TLI), DB(DB),
1685 AC(AC), TheFunction(F), Hints(Hints) {}
1687 /// Information about vectorization costs
1688 struct VectorizationFactor {
1689 unsigned Width; // Vector width with best cost
1690 unsigned Cost; // Cost of the loop with that width
1692 /// \return The most profitable vectorization factor and the cost of that VF.
1693 /// This method checks every power of two up to VF. If UserVF is not ZERO
1694 /// then this vectorization factor will be selected if vectorization is
1696 VectorizationFactor selectVectorizationFactor(bool OptForSize);
1698 /// \return The size (in bits) of the smallest and widest types in the code
1699 /// that needs to be vectorized. We ignore values that remain scalar such as
1700 /// 64 bit loop indices.
1701 std::pair<unsigned, unsigned> getSmallestAndWidestTypes();
1703 /// \return The desired interleave count.
1704 /// If interleave count has been specified by metadata it will be returned.
1705 /// Otherwise, the interleave count is computed and returned. VF and LoopCost
1706 /// are the selected vectorization factor and the cost of the selected VF.
1707 unsigned selectInterleaveCount(bool OptForSize, unsigned VF,
1710 /// \return The most profitable unroll factor.
1711 /// This method finds the best unroll-factor based on register pressure and
1712 /// other parameters. VF and LoopCost are the selected vectorization factor
1713 /// and the cost of the selected VF.
1714 unsigned computeInterleaveCount(bool OptForSize, unsigned VF,
1717 /// \brief A struct that represents some properties of the register usage
1719 struct RegisterUsage {
1720 /// Holds the number of loop invariant values that are used in the loop.
1721 unsigned LoopInvariantRegs;
1722 /// Holds the maximum number of concurrent live intervals in the loop.
1723 unsigned MaxLocalUsers;
1724 /// Holds the number of instructions in the loop.
1725 unsigned NumInstructions;
1728 /// \return Returns information about the register usages of the loop for the
1729 /// given vectorization factors.
1730 SmallVector<RegisterUsage, 8> calculateRegisterUsage(ArrayRef<unsigned> VFs);
1732 /// Collect values we want to ignore in the cost model.
1733 void collectValuesToIgnore();
1736 /// The vectorization cost is a combination of the cost itself and a boolean
1737 /// indicating whether any of the contributing operations will actually
1739 /// vector values after type legalization in the backend. If this latter value
1741 /// false, then all operations will be scalarized (i.e. no vectorization has
1742 /// actually taken place).
1743 typedef std::pair<unsigned, bool> VectorizationCostTy;
1745 /// Returns the expected execution cost. The unit of the cost does
1746 /// not matter because we use the 'cost' units to compare different
1747 /// vector widths. The cost that is returned is *not* normalized by
1748 /// the factor width.
1749 VectorizationCostTy expectedCost(unsigned VF);
1751 /// Returns the execution time cost of an instruction for a given vector
1752 /// width. Vector width of one means scalar.
1753 VectorizationCostTy getInstructionCost(Instruction *I, unsigned VF);
1755 /// The cost-computation logic from getInstructionCost which provides
1756 /// the vector type as an output parameter.
1757 unsigned getInstructionCost(Instruction *I, unsigned VF, Type *&VectorTy);
1759 /// Returns whether the instruction is a load or store and will be a emitted
1760 /// as a vector operation.
1761 bool isConsecutiveLoadOrStore(Instruction *I);
1763 /// Report an analysis message to assist the user in diagnosing loops that are
1764 /// not vectorized. These are handled as LoopAccessReport rather than
1765 /// VectorizationReport because the << operator of VectorizationReport returns
1766 /// LoopAccessReport.
1767 void emitAnalysis(const LoopAccessReport &Message) const {
1768 emitAnalysisDiag(TheFunction, TheLoop, *Hints, Message);
1772 /// Map of scalar integer values to the smallest bitwidth they can be legally
1773 /// represented as. The vector equivalents of these values should be truncated
1775 MapVector<Instruction *, uint64_t> MinBWs;
1777 /// The loop that we evaluate.
1779 /// Predicated scalar evolution analysis.
1780 PredicatedScalarEvolution &PSE;
1781 /// Loop Info analysis.
1783 /// Vectorization legality.
1784 LoopVectorizationLegality *Legal;
1785 /// Vector target information.
1786 const TargetTransformInfo &TTI;
1787 /// Target Library Info.
1788 const TargetLibraryInfo *TLI;
1789 /// Demanded bits analysis.
1791 /// Assumption cache.
1792 AssumptionCache *AC;
1793 const Function *TheFunction;
1794 /// Loop Vectorize Hint.
1795 const LoopVectorizeHints *Hints;
1796 /// Values to ignore in the cost model.
1797 SmallPtrSet<const Value *, 16> ValuesToIgnore;
1798 /// Values to ignore in the cost model when VF > 1.
1799 SmallPtrSet<const Value *, 16> VecValuesToIgnore;
1802 /// \brief This holds vectorization requirements that must be verified late in
1803 /// the process. The requirements are set by legalize and costmodel. Once
1804 /// vectorization has been determined to be possible and profitable the
1805 /// requirements can be verified by looking for metadata or compiler options.
1806 /// For example, some loops require FP commutativity which is only allowed if
1807 /// vectorization is explicitly specified or if the fast-math compiler option
1808 /// has been provided.
1809 /// Late evaluation of these requirements allows helpful diagnostics to be
1810 /// composed that tells the user what need to be done to vectorize the loop. For
1811 /// example, by specifying #pragma clang loop vectorize or -ffast-math. Late
1812 /// evaluation should be used only when diagnostics can generated that can be
1813 /// followed by a non-expert user.
1814 class LoopVectorizationRequirements {
1816 LoopVectorizationRequirements()
1817 : NumRuntimePointerChecks(0), UnsafeAlgebraInst(nullptr) {}
1819 void addUnsafeAlgebraInst(Instruction *I) {
1820 // First unsafe algebra instruction.
1821 if (!UnsafeAlgebraInst)
1822 UnsafeAlgebraInst = I;
1825 void addRuntimePointerChecks(unsigned Num) { NumRuntimePointerChecks = Num; }
1827 bool doesNotMeet(Function *F, Loop *L, const LoopVectorizeHints &Hints) {
1828 const char *Name = Hints.vectorizeAnalysisPassName();
1829 bool Failed = false;
1830 if (UnsafeAlgebraInst && !Hints.allowReordering()) {
1831 emitOptimizationRemarkAnalysisFPCommute(
1832 F->getContext(), Name, *F, UnsafeAlgebraInst->getDebugLoc(),
1833 VectorizationReport() << "cannot prove it is safe to reorder "
1834 "floating-point operations");
1838 // Test if runtime memcheck thresholds are exceeded.
1839 bool PragmaThresholdReached =
1840 NumRuntimePointerChecks > PragmaVectorizeMemoryCheckThreshold;
1841 bool ThresholdReached =
1842 NumRuntimePointerChecks > VectorizerParams::RuntimeMemoryCheckThreshold;
1843 if ((ThresholdReached && !Hints.allowReordering()) ||
1844 PragmaThresholdReached) {
1845 emitOptimizationRemarkAnalysisAliasing(
1846 F->getContext(), Name, *F, L->getStartLoc(),
1847 VectorizationReport()
1848 << "cannot prove it is safe to reorder memory operations");
1849 DEBUG(dbgs() << "LV: Too many memory checks needed.\n");
1857 unsigned NumRuntimePointerChecks;
1858 Instruction *UnsafeAlgebraInst;
1861 static void addAcyclicInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1863 if (!hasCyclesInLoopBody(L))
1867 for (Loop *InnerL : L)
1868 addAcyclicInnerLoop(*InnerL, V);
1871 /// The LoopVectorize Pass.
1872 struct LoopVectorize : public FunctionPass {
1873 /// Pass identification, replacement for typeid
1876 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1877 : FunctionPass(ID) {
1878 Impl.DisableUnrolling = NoUnrolling;
1879 Impl.AlwaysVectorize = AlwaysVectorize;
1880 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1883 LoopVectorizePass Impl;
1885 bool runOnFunction(Function &F) override {
1886 if (skipFunction(F))
1889 auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1890 auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1891 auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1892 auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1893 auto *BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI();
1894 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1895 auto *TLI = TLIP ? &TLIP->getTLI() : nullptr;
1896 auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
1897 auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1898 auto *LAA = &getAnalysis<LoopAccessLegacyAnalysis>();
1899 auto *DB = &getAnalysis<DemandedBitsWrapperPass>().getDemandedBits();
1901 std::function<const LoopAccessInfo &(Loop &)> GetLAA =
1902 [&](Loop &L) -> const LoopAccessInfo & { return LAA->getInfo(&L); };
1904 return Impl.runImpl(F, *SE, *LI, *TTI, *DT, *BFI, TLI, *DB, *AA, *AC,
1908 void getAnalysisUsage(AnalysisUsage &AU) const override {
1909 AU.addRequired<AssumptionCacheTracker>();
1910 AU.addRequiredID(LoopSimplifyID);
1911 AU.addRequiredID(LCSSAID);
1912 AU.addRequired<BlockFrequencyInfoWrapperPass>();
1913 AU.addRequired<DominatorTreeWrapperPass>();
1914 AU.addRequired<LoopInfoWrapperPass>();
1915 AU.addRequired<ScalarEvolutionWrapperPass>();
1916 AU.addRequired<TargetTransformInfoWrapperPass>();
1917 AU.addRequired<AAResultsWrapperPass>();
1918 AU.addRequired<LoopAccessLegacyAnalysis>();
1919 AU.addRequired<DemandedBitsWrapperPass>();
1920 AU.addPreserved<LoopInfoWrapperPass>();
1921 AU.addPreserved<DominatorTreeWrapperPass>();
1922 AU.addPreserved<BasicAAWrapperPass>();
1923 AU.addPreserved<GlobalsAAWrapperPass>();
1927 } // end anonymous namespace
1929 //===----------------------------------------------------------------------===//
1930 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1931 // LoopVectorizationCostModel.
1932 //===----------------------------------------------------------------------===//
1934 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1935 // We need to place the broadcast of invariant variables outside the loop.
1936 Instruction *Instr = dyn_cast<Instruction>(V);
1937 bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
1938 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1940 // Place the code for broadcasting invariant variables in the new preheader.
1941 IRBuilder<>::InsertPointGuard Guard(Builder);
1943 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1945 // Broadcast the scalar into all locations in the vector.
1946 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1951 void InnerLoopVectorizer::createVectorIntInductionPHI(
1952 const InductionDescriptor &II, VectorParts &Entry, IntegerType *TruncType) {
1953 Value *Start = II.getStartValue();
1954 ConstantInt *Step = II.getConstIntStepValue();
1955 assert(Step && "Can not widen an IV with a non-constant step");
1957 // Construct the initial value of the vector IV in the vector loop preheader
1958 auto CurrIP = Builder.saveIP();
1959 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1961 Step = ConstantInt::getSigned(TruncType, Step->getSExtValue());
1962 Start = Builder.CreateCast(Instruction::Trunc, Start, TruncType);
1964 Value *SplatStart = Builder.CreateVectorSplat(VF, Start);
1965 Value *SteppedStart = getStepVector(SplatStart, 0, Step);
1966 Builder.restoreIP(CurrIP);
1969 ConstantVector::getSplat(VF, ConstantInt::getSigned(Start->getType(),
1970 VF * Step->getSExtValue()));
1971 // We may need to add the step a number of times, depending on the unroll
1972 // factor. The last of those goes into the PHI.
1973 PHINode *VecInd = PHINode::Create(SteppedStart->getType(), 2, "vec.ind",
1974 &*LoopVectorBody->getFirstInsertionPt());
1975 Value *LastInduction = VecInd;
1976 for (unsigned Part = 0; Part < UF; ++Part) {
1977 Entry[Part] = LastInduction;
1978 LastInduction = Builder.CreateAdd(LastInduction, SplatVF, "step.add");
1981 VecInd->addIncoming(SteppedStart, LoopVectorPreHeader);
1982 VecInd->addIncoming(LastInduction, LoopVectorBody);
1985 void InnerLoopVectorizer::widenIntInduction(PHINode *IV, VectorParts &Entry,
1988 auto II = Legal->getInductionVars()->find(IV);
1989 assert(II != Legal->getInductionVars()->end() && "IV is not an induction");
1991 auto ID = II->second;
1992 assert(IV->getType() == ID.getStartValue()->getType() && "Types must match");
1994 // If a truncate instruction was provided, get the smaller type.
1995 auto *TruncType = Trunc ? cast<IntegerType>(Trunc->getType()) : nullptr;
1997 // The step of the induction.
1998 Value *Step = nullptr;
2000 // If the induction variable has a constant integer step value, go ahead and
2002 if (ID.getConstIntStepValue())
2003 Step = ID.getConstIntStepValue();
2005 // Try to create a new independent vector induction variable. If we can't
2006 // create the phi node, we will splat the scalar induction variable in each
2008 if (VF > 1 && IV->getType() == Induction->getType() && Step &&
2009 !ValuesNotWidened->count(IV))
2010 return createVectorIntInductionPHI(ID, Entry, TruncType);
2012 // The scalar value to broadcast. This will be derived from the canonical
2013 // induction variable.
2014 Value *ScalarIV = nullptr;
2016 // Define the scalar induction variable and step values. If we were given a
2017 // truncation type, truncate the canonical induction variable and constant
2018 // step. Otherwise, derive these values from the induction descriptor.
2020 assert(Step && "Truncation requires constant integer step");
2021 auto StepInt = cast<ConstantInt>(Step)->getSExtValue();
2022 ScalarIV = Builder.CreateCast(Instruction::Trunc, Induction, TruncType);
2023 Step = ConstantInt::getSigned(TruncType, StepInt);
2025 ScalarIV = Induction;
2026 auto &DL = OrigLoop->getHeader()->getModule()->getDataLayout();
2027 if (IV != OldInduction) {
2028 ScalarIV = Builder.CreateSExtOrTrunc(ScalarIV, IV->getType());
2029 ScalarIV = ID.transform(Builder, ScalarIV, PSE.getSE(), DL);
2030 ScalarIV->setName("offset.idx");
2033 SCEVExpander Exp(*PSE.getSE(), DL, "induction");
2034 Step = Exp.expandCodeFor(ID.getStep(), ID.getStep()->getType(),
2035 &*Builder.GetInsertPoint());
2039 // Splat the scalar induction variable, and build the necessary step vectors.
2040 Value *Broadcasted = getBroadcastInstrs(ScalarIV);
2041 for (unsigned Part = 0; Part < UF; ++Part)
2042 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
2044 // If an induction variable is only used for counting loop iterations or
2045 // calculating addresses, it doesn't need to be widened. Create scalar steps
2046 // that can be used by instructions we will later scalarize. Note that the
2047 // addition of the scalar steps will not increase the number of instructions
2048 // in the loop in the common case prior to InstCombine. We will be trading
2049 // one vector extract for each scalar step.
2050 if (VF > 1 && ValuesNotWidened->count(IV)) {
2051 auto *EntryVal = Trunc ? cast<Value>(Trunc) : IV;
2052 buildScalarSteps(ScalarIV, Step, EntryVal);
2056 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
2058 assert(Val->getType()->isVectorTy() && "Must be a vector");
2059 assert(Val->getType()->getScalarType()->isIntegerTy() &&
2060 "Elem must be an integer");
2061 assert(Step->getType() == Val->getType()->getScalarType() &&
2062 "Step has wrong type");
2063 // Create the types.
2064 Type *ITy = Val->getType()->getScalarType();
2065 VectorType *Ty = cast<VectorType>(Val->getType());
2066 int VLen = Ty->getNumElements();
2067 SmallVector<Constant *, 8> Indices;
2069 // Create a vector of consecutive numbers from zero to VF.
2070 for (int i = 0; i < VLen; ++i)
2071 Indices.push_back(ConstantInt::get(ITy, StartIdx + i));
2073 // Add the consecutive indices to the vector value.
2074 Constant *Cv = ConstantVector::get(Indices);
2075 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
2076 Step = Builder.CreateVectorSplat(VLen, Step);
2077 assert(Step->getType() == Val->getType() && "Invalid step vec");
2078 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
2079 // which can be found from the original scalar operations.
2080 Step = Builder.CreateMul(Cv, Step);
2081 return Builder.CreateAdd(Val, Step, "induction");
2084 void InnerLoopVectorizer::buildScalarSteps(Value *ScalarIV, Value *Step,
2087 // We shouldn't have to build scalar steps if we aren't vectorizing.
2088 assert(VF > 1 && "VF should be greater than one");
2090 // Get the value type and ensure it and the step have the same integer type.
2091 Type *ScalarIVTy = ScalarIV->getType()->getScalarType();
2092 assert(ScalarIVTy->isIntegerTy() && ScalarIVTy == Step->getType() &&
2093 "Val and Step should have the same integer type");
2095 // Compute the scalar steps and save the results in ScalarIVMap.
2096 for (unsigned Part = 0; Part < UF; ++Part)
2097 for (unsigned I = 0; I < VF; ++I) {
2098 auto *StartIdx = ConstantInt::get(ScalarIVTy, VF * Part + I);
2099 auto *Mul = Builder.CreateMul(StartIdx, Step);
2100 auto *Add = Builder.CreateAdd(ScalarIV, Mul);
2101 ScalarIVMap[EntryVal].push_back(Add);
2105 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
2106 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
2107 auto *SE = PSE.getSE();
2108 // Make sure that the pointer does not point to structs.
2109 if (Ptr->getType()->getPointerElementType()->isAggregateType())
2112 // If this value is a pointer induction variable, we know it is consecutive.
2113 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
2114 if (Phi && Inductions.count(Phi)) {
2115 InductionDescriptor II = Inductions[Phi];
2116 return II.getConsecutiveDirection();
2119 GetElementPtrInst *Gep = getGEPInstruction(Ptr);
2123 unsigned NumOperands = Gep->getNumOperands();
2124 Value *GpPtr = Gep->getPointerOperand();
2125 // If this GEP value is a consecutive pointer induction variable and all of
2126 // the indices are constant, then we know it is consecutive.
2127 Phi = dyn_cast<PHINode>(GpPtr);
2128 if (Phi && Inductions.count(Phi)) {
2130 // Make sure that the pointer does not point to structs.
2131 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
2132 if (GepPtrType->getElementType()->isAggregateType())
2135 // Make sure that all of the index operands are loop invariant.
2136 for (unsigned i = 1; i < NumOperands; ++i)
2137 if (!SE->isLoopInvariant(PSE.getSCEV(Gep->getOperand(i)), TheLoop))
2140 InductionDescriptor II = Inductions[Phi];
2141 return II.getConsecutiveDirection();
2144 unsigned InductionOperand = getGEPInductionOperand(Gep);
2146 // Check that all of the gep indices are uniform except for our induction
2148 for (unsigned i = 0; i != NumOperands; ++i)
2149 if (i != InductionOperand &&
2150 !SE->isLoopInvariant(PSE.getSCEV(Gep->getOperand(i)), TheLoop))
2153 // We can emit wide load/stores only if the last non-zero index is the
2154 // induction variable.
2155 const SCEV *Last = nullptr;
2156 if (!getSymbolicStrides() || !getSymbolicStrides()->count(Gep))
2157 Last = PSE.getSCEV(Gep->getOperand(InductionOperand));
2159 // Because of the multiplication by a stride we can have a s/zext cast.
2160 // We are going to replace this stride by 1 so the cast is safe to ignore.
2162 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
2163 // %0 = trunc i64 %indvars.iv to i32
2164 // %mul = mul i32 %0, %Stride1
2165 // %idxprom = zext i32 %mul to i64 << Safe cast.
2166 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
2168 Last = replaceSymbolicStrideSCEV(PSE, *getSymbolicStrides(),
2169 Gep->getOperand(InductionOperand), Gep);
2170 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
2172 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
2176 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
2177 const SCEV *Step = AR->getStepRecurrence(*SE);
2179 // The memory is consecutive because the last index is consecutive
2180 // and all other indices are loop invariant.
2183 if (Step->isAllOnesValue())
2190 bool LoopVectorizationLegality::isUniform(Value *V) {
2191 return LAI->isUniform(V);
2194 InnerLoopVectorizer::VectorParts &
2195 InnerLoopVectorizer::getVectorValue(Value *V) {
2196 assert(V != Induction && "The new induction variable should not be used.");
2197 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
2199 // If we have a stride that is replaced by one, do it here.
2200 if (Legal->hasStride(V))
2201 V = ConstantInt::get(V->getType(), 1);
2203 // If we have this scalar in the map, return it.
2204 if (WidenMap.has(V))
2205 return WidenMap.get(V);
2207 // If this scalar is unknown, assume that it is a constant or that it is
2208 // loop invariant. Broadcast V and save the value for future uses.
2209 Value *B = getBroadcastInstrs(V);
2210 return WidenMap.splat(V, B);
2213 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
2214 assert(Vec->getType()->isVectorTy() && "Invalid type");
2215 SmallVector<Constant *, 8> ShuffleMask;
2216 for (unsigned i = 0; i < VF; ++i)
2217 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
2219 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
2220 ConstantVector::get(ShuffleMask),
2224 // Get a mask to interleave \p NumVec vectors into a wide vector.
2225 // I.e. <0, VF, VF*2, ..., VF*(NumVec-1), 1, VF+1, VF*2+1, ...>
2226 // E.g. For 2 interleaved vectors, if VF is 4, the mask is:
2227 // <0, 4, 1, 5, 2, 6, 3, 7>
2228 static Constant *getInterleavedMask(IRBuilder<> &Builder, unsigned VF,
2230 SmallVector<Constant *, 16> Mask;
2231 for (unsigned i = 0; i < VF; i++)
2232 for (unsigned j = 0; j < NumVec; j++)
2233 Mask.push_back(Builder.getInt32(j * VF + i));
2235 return ConstantVector::get(Mask);
2238 // Get the strided mask starting from index \p Start.
2239 // I.e. <Start, Start + Stride, ..., Start + Stride*(VF-1)>
2240 static Constant *getStridedMask(IRBuilder<> &Builder, unsigned Start,
2241 unsigned Stride, unsigned VF) {
2242 SmallVector<Constant *, 16> Mask;
2243 for (unsigned i = 0; i < VF; i++)
2244 Mask.push_back(Builder.getInt32(Start + i * Stride));
2246 return ConstantVector::get(Mask);
2249 // Get a mask of two parts: The first part consists of sequential integers
2250 // starting from 0, The second part consists of UNDEFs.
2251 // I.e. <0, 1, 2, ..., NumInt - 1, undef, ..., undef>
2252 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned NumInt,
2253 unsigned NumUndef) {
2254 SmallVector<Constant *, 16> Mask;
2255 for (unsigned i = 0; i < NumInt; i++)
2256 Mask.push_back(Builder.getInt32(i));
2258 Constant *Undef = UndefValue::get(Builder.getInt32Ty());
2259 for (unsigned i = 0; i < NumUndef; i++)
2260 Mask.push_back(Undef);
2262 return ConstantVector::get(Mask);
2265 // Concatenate two vectors with the same element type. The 2nd vector should
2266 // not have more elements than the 1st vector. If the 2nd vector has less
2267 // elements, extend it with UNDEFs.
2268 static Value *ConcatenateTwoVectors(IRBuilder<> &Builder, Value *V1,
2270 VectorType *VecTy1 = dyn_cast<VectorType>(V1->getType());
2271 VectorType *VecTy2 = dyn_cast<VectorType>(V2->getType());
2272 assert(VecTy1 && VecTy2 &&
2273 VecTy1->getScalarType() == VecTy2->getScalarType() &&
2274 "Expect two vectors with the same element type");
2276 unsigned NumElts1 = VecTy1->getNumElements();
2277 unsigned NumElts2 = VecTy2->getNumElements();
2278 assert(NumElts1 >= NumElts2 && "Unexpect the first vector has less elements");
2280 if (NumElts1 > NumElts2) {
2281 // Extend with UNDEFs.
2283 getSequentialMask(Builder, NumElts2, NumElts1 - NumElts2);
2284 V2 = Builder.CreateShuffleVector(V2, UndefValue::get(VecTy2), ExtMask);
2287 Constant *Mask = getSequentialMask(Builder, NumElts1 + NumElts2, 0);
2288 return Builder.CreateShuffleVector(V1, V2, Mask);
2291 // Concatenate vectors in the given list. All vectors have the same type.
2292 static Value *ConcatenateVectors(IRBuilder<> &Builder,
2293 ArrayRef<Value *> InputList) {
2294 unsigned NumVec = InputList.size();
2295 assert(NumVec > 1 && "Should be at least two vectors");
2297 SmallVector<Value *, 8> ResList;
2298 ResList.append(InputList.begin(), InputList.end());
2300 SmallVector<Value *, 8> TmpList;
2301 for (unsigned i = 0; i < NumVec - 1; i += 2) {
2302 Value *V0 = ResList[i], *V1 = ResList[i + 1];
2303 assert((V0->getType() == V1->getType() || i == NumVec - 2) &&
2304 "Only the last vector may have a different type");
2306 TmpList.push_back(ConcatenateTwoVectors(Builder, V0, V1));
2309 // Push the last vector if the total number of vectors is odd.
2310 if (NumVec % 2 != 0)
2311 TmpList.push_back(ResList[NumVec - 1]);
2314 NumVec = ResList.size();
2315 } while (NumVec > 1);
2320 // Try to vectorize the interleave group that \p Instr belongs to.
2322 // E.g. Translate following interleaved load group (factor = 3):
2323 // for (i = 0; i < N; i+=3) {
2324 // R = Pic[i]; // Member of index 0
2325 // G = Pic[i+1]; // Member of index 1
2326 // B = Pic[i+2]; // Member of index 2
2327 // ... // do something to R, G, B
2330 // %wide.vec = load <12 x i32> ; Read 4 tuples of R,G,B
2331 // %R.vec = shuffle %wide.vec, undef, <0, 3, 6, 9> ; R elements
2332 // %G.vec = shuffle %wide.vec, undef, <1, 4, 7, 10> ; G elements
2333 // %B.vec = shuffle %wide.vec, undef, <2, 5, 8, 11> ; B elements
2335 // Or translate following interleaved store group (factor = 3):
2336 // for (i = 0; i < N; i+=3) {
2337 // ... do something to R, G, B
2338 // Pic[i] = R; // Member of index 0
2339 // Pic[i+1] = G; // Member of index 1
2340 // Pic[i+2] = B; // Member of index 2
2343 // %R_G.vec = shuffle %R.vec, %G.vec, <0, 1, 2, ..., 7>
2344 // %B_U.vec = shuffle %B.vec, undef, <0, 1, 2, 3, u, u, u, u>
2345 // %interleaved.vec = shuffle %R_G.vec, %B_U.vec,
2346 // <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> ; Interleave R,G,B elements
2347 // store <12 x i32> %interleaved.vec ; Write 4 tuples of R,G,B
2348 void InnerLoopVectorizer::vectorizeInterleaveGroup(Instruction *Instr) {
2349 const InterleaveGroup *Group = Legal->getInterleavedAccessGroup(Instr);
2350 assert(Group && "Fail to get an interleaved access group.");
2352 // Skip if current instruction is not the insert position.
2353 if (Instr != Group->getInsertPos())
2356 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2357 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2358 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
2360 // Prepare for the vector type of the interleaved load/store.
2361 Type *ScalarTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2362 unsigned InterleaveFactor = Group->getFactor();
2363 Type *VecTy = VectorType::get(ScalarTy, InterleaveFactor * VF);
2364 Type *PtrTy = VecTy->getPointerTo(Ptr->getType()->getPointerAddressSpace());
2366 // Prepare for the new pointers.
2367 setDebugLocFromInst(Builder, Ptr);
2368 VectorParts &PtrParts = getVectorValue(Ptr);
2369 SmallVector<Value *, 2> NewPtrs;
2370 unsigned Index = Group->getIndex(Instr);
2371 for (unsigned Part = 0; Part < UF; Part++) {
2372 // Extract the pointer for current instruction from the pointer vector. A
2373 // reverse access uses the pointer in the last lane.
2374 Value *NewPtr = Builder.CreateExtractElement(
2376 Group->isReverse() ? Builder.getInt32(VF - 1) : Builder.getInt32(0));
2378 // Notice current instruction could be any index. Need to adjust the address
2379 // to the member of index 0.
2381 // E.g. a = A[i+1]; // Member of index 1 (Current instruction)
2382 // b = A[i]; // Member of index 0
2383 // Current pointer is pointed to A[i+1], adjust it to A[i].
2385 // E.g. A[i+1] = a; // Member of index 1
2386 // A[i] = b; // Member of index 0
2387 // A[i+2] = c; // Member of index 2 (Current instruction)
2388 // Current pointer is pointed to A[i+2], adjust it to A[i].
2389 NewPtr = Builder.CreateGEP(NewPtr, Builder.getInt32(-Index));
2391 // Cast to the vector pointer type.
2392 NewPtrs.push_back(Builder.CreateBitCast(NewPtr, PtrTy));
2395 setDebugLocFromInst(Builder, Instr);
2396 Value *UndefVec = UndefValue::get(VecTy);
2398 // Vectorize the interleaved load group.
2400 for (unsigned Part = 0; Part < UF; Part++) {
2401 Instruction *NewLoadInstr = Builder.CreateAlignedLoad(
2402 NewPtrs[Part], Group->getAlignment(), "wide.vec");
2404 for (unsigned i = 0; i < InterleaveFactor; i++) {
2405 Instruction *Member = Group->getMember(i);
2407 // Skip the gaps in the group.
2411 Constant *StrideMask = getStridedMask(Builder, i, InterleaveFactor, VF);
2412 Value *StridedVec = Builder.CreateShuffleVector(
2413 NewLoadInstr, UndefVec, StrideMask, "strided.vec");
2415 // If this member has different type, cast the result type.
2416 if (Member->getType() != ScalarTy) {
2417 VectorType *OtherVTy = VectorType::get(Member->getType(), VF);
2418 StridedVec = Builder.CreateBitOrPointerCast(StridedVec, OtherVTy);
2421 VectorParts &Entry = WidenMap.get(Member);
2423 Group->isReverse() ? reverseVector(StridedVec) : StridedVec;
2426 addMetadata(NewLoadInstr, Instr);
2431 // The sub vector type for current instruction.
2432 VectorType *SubVT = VectorType::get(ScalarTy, VF);
2434 // Vectorize the interleaved store group.
2435 for (unsigned Part = 0; Part < UF; Part++) {
2436 // Collect the stored vector from each member.
2437 SmallVector<Value *, 4> StoredVecs;
2438 for (unsigned i = 0; i < InterleaveFactor; i++) {
2439 // Interleaved store group doesn't allow a gap, so each index has a member
2440 Instruction *Member = Group->getMember(i);
2441 assert(Member && "Fail to get a member from an interleaved store group");
2444 getVectorValue(cast<StoreInst>(Member)->getValueOperand())[Part];
2445 if (Group->isReverse())
2446 StoredVec = reverseVector(StoredVec);
2448 // If this member has different type, cast it to an unified type.
2449 if (StoredVec->getType() != SubVT)
2450 StoredVec = Builder.CreateBitOrPointerCast(StoredVec, SubVT);
2452 StoredVecs.push_back(StoredVec);
2455 // Concatenate all vectors into a wide vector.
2456 Value *WideVec = ConcatenateVectors(Builder, StoredVecs);
2458 // Interleave the elements in the wide vector.
2459 Constant *IMask = getInterleavedMask(Builder, VF, InterleaveFactor);
2460 Value *IVec = Builder.CreateShuffleVector(WideVec, UndefVec, IMask,
2463 Instruction *NewStoreInstr =
2464 Builder.CreateAlignedStore(IVec, NewPtrs[Part], Group->getAlignment());
2465 addMetadata(NewStoreInstr, Instr);
2469 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
2470 // Attempt to issue a wide load.
2471 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2472 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2474 assert((LI || SI) && "Invalid Load/Store instruction");
2476 // Try to vectorize the interleave group if this access is interleaved.
2477 if (Legal->isAccessInterleaved(Instr))
2478 return vectorizeInterleaveGroup(Instr);
2480 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2481 Type *DataTy = VectorType::get(ScalarDataTy, VF);
2482 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
2483 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
2484 // An alignment of 0 means target abi alignment. We need to use the scalar's
2485 // target abi alignment in such a case.
2486 const DataLayout &DL = Instr->getModule()->getDataLayout();
2488 Alignment = DL.getABITypeAlignment(ScalarDataTy);
2489 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
2490 uint64_t ScalarAllocatedSize = DL.getTypeAllocSize(ScalarDataTy);
2491 uint64_t VectorElementSize = DL.getTypeStoreSize(DataTy) / VF;
2493 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
2494 !Legal->isMaskRequired(SI))
2495 return scalarizeInstruction(Instr, true);
2497 if (ScalarAllocatedSize != VectorElementSize)
2498 return scalarizeInstruction(Instr);
2500 // If the pointer is loop invariant scalarize the load.
2501 if (LI && Legal->isUniform(Ptr))
2502 return scalarizeInstruction(Instr);
2504 // If the pointer is non-consecutive and gather/scatter is not supported
2505 // scalarize the instruction.
2506 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
2507 bool Reverse = ConsecutiveStride < 0;
2508 bool CreateGatherScatter =
2509 !ConsecutiveStride && ((LI && Legal->isLegalMaskedGather(ScalarDataTy)) ||
2510 (SI && Legal->isLegalMaskedScatter(ScalarDataTy)));
2512 if (!ConsecutiveStride && !CreateGatherScatter)
2513 return scalarizeInstruction(Instr);
2515 Constant *Zero = Builder.getInt32(0);
2516 VectorParts &Entry = WidenMap.get(Instr);
2517 VectorParts VectorGep;
2519 // Handle consecutive loads/stores.
2520 GetElementPtrInst *Gep = getGEPInstruction(Ptr);
2521 if (ConsecutiveStride) {
2522 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
2523 setDebugLocFromInst(Builder, Gep);
2524 Value *PtrOperand = Gep->getPointerOperand();
2525 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
2526 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
2528 // Create the new GEP with the new induction variable.
2529 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2530 Gep2->setOperand(0, FirstBasePtr);
2531 Gep2->setName("gep.indvar.base");
2532 Ptr = Builder.Insert(Gep2);
2534 setDebugLocFromInst(Builder, Gep);
2535 assert(PSE.getSE()->isLoopInvariant(PSE.getSCEV(Gep->getPointerOperand()),
2537 "Base ptr must be invariant");
2538 // The last index does not have to be the induction. It can be
2539 // consecutive and be a function of the index. For example A[I+1];
2540 unsigned NumOperands = Gep->getNumOperands();
2541 unsigned InductionOperand = getGEPInductionOperand(Gep);
2542 // Create the new GEP with the new induction variable.
2543 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2545 for (unsigned i = 0; i < NumOperands; ++i) {
2546 Value *GepOperand = Gep->getOperand(i);
2547 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
2549 // Update last index or loop invariant instruction anchored in loop.
2550 if (i == InductionOperand ||
2551 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
2552 assert((i == InductionOperand ||
2553 PSE.getSE()->isLoopInvariant(PSE.getSCEV(GepOperandInst),
2555 "Must be last index or loop invariant");
2557 VectorParts &GEPParts = getVectorValue(GepOperand);
2559 // If GepOperand is an induction variable, and there's a scalarized
2560 // version of it available, use it. Otherwise, we will need to create
2561 // an extractelement instruction.
2562 Value *Index = ScalarIVMap.count(GepOperand)
2563 ? ScalarIVMap[GepOperand][0]
2564 : Builder.CreateExtractElement(GEPParts[0], Zero);
2566 Gep2->setOperand(i, Index);
2567 Gep2->setName("gep.indvar.idx");
2570 Ptr = Builder.Insert(Gep2);
2572 // Use the induction element ptr.
2573 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
2574 setDebugLocFromInst(Builder, Ptr);
2575 VectorParts &PtrVal = getVectorValue(Ptr);
2576 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
2579 // At this point we should vector version of GEP for Gather or Scatter
2580 assert(CreateGatherScatter && "The instruction should be scalarized");
2582 // Vectorizing GEP, across UF parts. We want to get a vector value for base
2583 // and each index that's defined inside the loop, even if it is
2584 // loop-invariant but wasn't hoisted out. Otherwise we want to keep them
2586 SmallVector<VectorParts, 4> OpsV;
2587 for (Value *Op : Gep->operands()) {
2588 Instruction *SrcInst = dyn_cast<Instruction>(Op);
2589 if (SrcInst && OrigLoop->contains(SrcInst))
2590 OpsV.push_back(getVectorValue(Op));
2592 OpsV.push_back(VectorParts(UF, Op));
2594 for (unsigned Part = 0; Part < UF; ++Part) {
2595 SmallVector<Value *, 4> Ops;
2596 Value *GEPBasePtr = OpsV[0][Part];
2597 for (unsigned i = 1; i < Gep->getNumOperands(); i++)
2598 Ops.push_back(OpsV[i][Part]);
2599 Value *NewGep = Builder.CreateGEP(GEPBasePtr, Ops, "VectorGep");
2600 cast<GetElementPtrInst>(NewGep)->setIsInBounds(Gep->isInBounds());
2601 assert(NewGep->getType()->isVectorTy() && "Expected vector GEP");
2604 Builder.CreateBitCast(NewGep, VectorType::get(Ptr->getType(), VF));
2605 VectorGep.push_back(NewGep);
2608 VectorGep = getVectorValue(Ptr);
2611 VectorParts Mask = createBlockInMask(Instr->getParent());
2614 assert(!Legal->isUniform(SI->getPointerOperand()) &&
2615 "We do not allow storing to uniform addresses");
2616 setDebugLocFromInst(Builder, SI);
2617 // We don't want to update the value in the map as it might be used in
2618 // another expression. So don't use a reference type for "StoredVal".
2619 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
2621 for (unsigned Part = 0; Part < UF; ++Part) {
2622 Instruction *NewSI = nullptr;
2623 if (CreateGatherScatter) {
2624 Value *MaskPart = Legal->isMaskRequired(SI) ? Mask[Part] : nullptr;
2625 NewSI = Builder.CreateMaskedScatter(StoredVal[Part], VectorGep[Part],
2626 Alignment, MaskPart);
2628 // Calculate the pointer for the specific unroll-part.
2630 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2633 // If we store to reverse consecutive memory locations, then we need
2634 // to reverse the order of elements in the stored value.
2635 StoredVal[Part] = reverseVector(StoredVal[Part]);
2636 // If the address is consecutive but reversed, then the
2637 // wide store needs to start at the last vector element.
2639 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2641 Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2642 Mask[Part] = reverseVector(Mask[Part]);
2646 Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace));
2648 if (Legal->isMaskRequired(SI))
2649 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
2653 Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
2655 addMetadata(NewSI, SI);
2661 assert(LI && "Must have a load instruction");
2662 setDebugLocFromInst(Builder, LI);
2663 for (unsigned Part = 0; Part < UF; ++Part) {
2665 if (CreateGatherScatter) {
2666 Value *MaskPart = Legal->isMaskRequired(LI) ? Mask[Part] : nullptr;
2667 NewLI = Builder.CreateMaskedGather(VectorGep[Part], Alignment, MaskPart,
2668 0, "wide.masked.gather");
2669 Entry[Part] = NewLI;
2671 // Calculate the pointer for the specific unroll-part.
2673 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2676 // If the address is consecutive but reversed, then the
2677 // wide load needs to start at the last vector element.
2678 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2679 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2680 Mask[Part] = reverseVector(Mask[Part]);
2684 Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace));
2685 if (Legal->isMaskRequired(LI))
2686 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
2687 UndefValue::get(DataTy),
2688 "wide.masked.load");
2690 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
2691 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
2693 addMetadata(NewLI, LI);
2697 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr,
2698 bool IfPredicateStore) {
2699 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
2700 // Holds vector parameters or scalars, in case of uniform vals.
2701 SmallVector<VectorParts, 4> Params;
2703 setDebugLocFromInst(Builder, Instr);
2705 // Find all of the vectorized parameters.
2706 for (Value *SrcOp : Instr->operands()) {
2707 // If we are accessing the old induction variable, use the new one.
2708 if (SrcOp == OldInduction) {
2709 Params.push_back(getVectorValue(SrcOp));
2713 // Try using previously calculated values.
2714 auto *SrcInst = dyn_cast<Instruction>(SrcOp);
2716 // If the src is an instruction that appeared earlier in the basic block,
2717 // then it should already be vectorized.
2718 if (SrcInst && OrigLoop->contains(SrcInst)) {
2719 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
2720 // The parameter is a vector value from earlier.
2721 Params.push_back(WidenMap.get(SrcInst));
2723 // The parameter is a scalar from outside the loop. Maybe even a constant.
2724 VectorParts Scalars;
2725 Scalars.append(UF, SrcOp);
2726 Params.push_back(Scalars);
2730 assert(Params.size() == Instr->getNumOperands() &&
2731 "Invalid number of operands");
2733 // Does this instruction return a value ?
2734 bool IsVoidRetTy = Instr->getType()->isVoidTy();
2737 IsVoidRetTy ? nullptr
2738 : UndefValue::get(VectorType::get(Instr->getType(), VF));
2739 // Create a new entry in the WidenMap and initialize it to Undef or Null.
2740 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
2743 if (IfPredicateStore) {
2744 assert(Instr->getParent()->getSinglePredecessor() &&
2745 "Only support single predecessor blocks");
2746 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
2747 Instr->getParent());
2750 // For each vector unroll 'part':
2751 for (unsigned Part = 0; Part < UF; ++Part) {
2752 // For each scalar that we create:
2753 for (unsigned Width = 0; Width < VF; ++Width) {
2756 Value *Cmp = nullptr;
2757 if (IfPredicateStore) {
2758 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
2759 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp,
2760 ConstantInt::get(Cmp->getType(), 1));
2763 Instruction *Cloned = Instr->clone();
2765 Cloned->setName(Instr->getName() + ".cloned");
2766 // Replace the operands of the cloned instructions with extracted scalars.
2767 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2769 // If the operand is an induction variable, and there's a scalarized
2770 // version of it available, use it. Otherwise, we will need to create
2771 // an extractelement instruction if vectorizing.
2772 auto *NewOp = Params[op][Part];
2773 auto *ScalarOp = Instr->getOperand(op);
2774 if (ScalarIVMap.count(ScalarOp))
2775 NewOp = ScalarIVMap[ScalarOp][VF * Part + Width];
2776 else if (NewOp->getType()->isVectorTy())
2777 NewOp = Builder.CreateExtractElement(NewOp, Builder.getInt32(Width));
2778 Cloned->setOperand(op, NewOp);
2780 addNewMetadata(Cloned, Instr);
2782 // Place the cloned scalar in the new loop.
2783 Builder.Insert(Cloned);
2785 // If we just cloned a new assumption, add it the assumption cache.
2786 if (auto *II = dyn_cast<IntrinsicInst>(Cloned))
2787 if (II->getIntrinsicID() == Intrinsic::assume)
2788 AC->registerAssumption(II);
2790 // If the original scalar returns a value we need to place it in a vector
2791 // so that future users will be able to use it.
2793 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
2794 Builder.getInt32(Width));
2796 if (IfPredicateStore)
2797 PredicatedStores.push_back(
2798 std::make_pair(cast<StoreInst>(Cloned), Cmp));
2803 PHINode *InnerLoopVectorizer::createInductionVariable(Loop *L, Value *Start,
2804 Value *End, Value *Step,
2806 BasicBlock *Header = L->getHeader();
2807 BasicBlock *Latch = L->getLoopLatch();
2808 // As we're just creating this loop, it's possible no latch exists
2809 // yet. If so, use the header as this will be a single block loop.
2813 IRBuilder<> Builder(&*Header->getFirstInsertionPt());
2814 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2815 auto *Induction = Builder.CreatePHI(Start->getType(), 2, "index");
2817 Builder.SetInsertPoint(Latch->getTerminator());
2819 // Create i+1 and fill the PHINode.
2820 Value *Next = Builder.CreateAdd(Induction, Step, "index.next");
2821 Induction->addIncoming(Start, L->getLoopPreheader());
2822 Induction->addIncoming(Next, Latch);
2823 // Create the compare.
2824 Value *ICmp = Builder.CreateICmpEQ(Next, End);
2825 Builder.CreateCondBr(ICmp, L->getExitBlock(), Header);
2827 // Now we have two terminators. Remove the old one from the block.
2828 Latch->getTerminator()->eraseFromParent();
2833 Value *InnerLoopVectorizer::getOrCreateTripCount(Loop *L) {
2837 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
2838 // Find the loop boundaries.
2839 ScalarEvolution *SE = PSE.getSE();
2840 const SCEV *BackedgeTakenCount = PSE.getBackedgeTakenCount();
2841 assert(BackedgeTakenCount != SE->getCouldNotCompute() &&
2842 "Invalid loop count");
2844 Type *IdxTy = Legal->getWidestInductionType();
2846 // The exit count might have the type of i64 while the phi is i32. This can
2847 // happen if we have an induction variable that is sign extended before the
2848 // compare. The only way that we get a backedge taken count is that the
2849 // induction variable was signed and as such will not overflow. In such a case
2850 // truncation is legal.
2851 if (BackedgeTakenCount->getType()->getPrimitiveSizeInBits() >
2852 IdxTy->getPrimitiveSizeInBits())
2853 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount, IdxTy);
2854 BackedgeTakenCount = SE->getNoopOrZeroExtend(BackedgeTakenCount, IdxTy);
2856 // Get the total trip count from the count by adding 1.
2857 const SCEV *ExitCount = SE->getAddExpr(
2858 BackedgeTakenCount, SE->getOne(BackedgeTakenCount->getType()));
2860 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
2862 // Expand the trip count and place the new instructions in the preheader.
2863 // Notice that the pre-header does not change, only the loop body.
2864 SCEVExpander Exp(*SE, DL, "induction");
2866 // Count holds the overall loop count (N).
2867 TripCount = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2868 L->getLoopPreheader()->getTerminator());
2870 if (TripCount->getType()->isPointerTy())
2872 CastInst::CreatePointerCast(TripCount, IdxTy, "exitcount.ptrcnt.to.int",
2873 L->getLoopPreheader()->getTerminator());
2878 Value *InnerLoopVectorizer::getOrCreateVectorTripCount(Loop *L) {
2879 if (VectorTripCount)
2880 return VectorTripCount;
2882 Value *TC = getOrCreateTripCount(L);
2883 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
2885 // Now we need to generate the expression for the part of the loop that the
2886 // vectorized body will execute. This is equal to N - (N % Step) if scalar
2887 // iterations are not required for correctness, or N - Step, otherwise. Step
2888 // is equal to the vectorization factor (number of SIMD elements) times the
2889 // unroll factor (number of SIMD instructions).
2890 Constant *Step = ConstantInt::get(TC->getType(), VF * UF);
2891 Value *R = Builder.CreateURem(TC, Step, "n.mod.vf");
2893 // If there is a non-reversed interleaved group that may speculatively access
2894 // memory out-of-bounds, we need to ensure that there will be at least one
2895 // iteration of the scalar epilogue loop. Thus, if the step evenly divides
2896 // the trip count, we set the remainder to be equal to the step. If the step
2897 // does not evenly divide the trip count, no adjustment is necessary since
2898 // there will already be scalar iterations. Note that the minimum iterations
2899 // check ensures that N >= Step.
2900 if (VF > 1 && Legal->requiresScalarEpilogue()) {
2901 auto *IsZero = Builder.CreateICmpEQ(R, ConstantInt::get(R->getType(), 0));
2902 R = Builder.CreateSelect(IsZero, Step, R);
2905 VectorTripCount = Builder.CreateSub(TC, R, "n.vec");
2907 return VectorTripCount;
2910 void InnerLoopVectorizer::emitMinimumIterationCountCheck(Loop *L,
2911 BasicBlock *Bypass) {
2912 Value *Count = getOrCreateTripCount(L);
2913 BasicBlock *BB = L->getLoopPreheader();
2914 IRBuilder<> Builder(BB->getTerminator());
2916 // Generate code to check that the loop's trip count that we computed by
2917 // adding one to the backedge-taken count will not overflow.
2918 Value *CheckMinIters = Builder.CreateICmpULT(
2919 Count, ConstantInt::get(Count->getType(), VF * UF), "min.iters.check");
2922 BB->splitBasicBlock(BB->getTerminator(), "min.iters.checked");
2923 // Update dominator tree immediately if the generated block is a
2924 // LoopBypassBlock because SCEV expansions to generate loop bypass
2925 // checks may query it before the current function is finished.
2926 DT->addNewBlock(NewBB, BB);
2927 if (L->getParentLoop())
2928 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
2929 ReplaceInstWithInst(BB->getTerminator(),
2930 BranchInst::Create(Bypass, NewBB, CheckMinIters));
2931 LoopBypassBlocks.push_back(BB);
2934 void InnerLoopVectorizer::emitVectorLoopEnteredCheck(Loop *L,
2935 BasicBlock *Bypass) {
2936 Value *TC = getOrCreateVectorTripCount(L);
2937 BasicBlock *BB = L->getLoopPreheader();
2938 IRBuilder<> Builder(BB->getTerminator());
2940 // Now, compare the new count to zero. If it is zero skip the vector loop and
2941 // jump to the scalar loop.
2942 Value *Cmp = Builder.CreateICmpEQ(TC, Constant::getNullValue(TC->getType()),
2945 // Generate code to check that the loop's trip count that we computed by
2946 // adding one to the backedge-taken count will not overflow.
2947 BasicBlock *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
2948 // Update dominator tree immediately if the generated block is a
2949 // LoopBypassBlock because SCEV expansions to generate loop bypass
2950 // checks may query it before the current function is finished.
2951 DT->addNewBlock(NewBB, BB);
2952 if (L->getParentLoop())
2953 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
2954 ReplaceInstWithInst(BB->getTerminator(),
2955 BranchInst::Create(Bypass, NewBB, Cmp));
2956 LoopBypassBlocks.push_back(BB);
2959 void InnerLoopVectorizer::emitSCEVChecks(Loop *L, BasicBlock *Bypass) {
2960 BasicBlock *BB = L->getLoopPreheader();
2962 // Generate the code to check that the SCEV assumptions that we made.
2963 // We want the new basic block to start at the first instruction in a
2964 // sequence of instructions that form a check.
2965 SCEVExpander Exp(*PSE.getSE(), Bypass->getModule()->getDataLayout(),
2968 Exp.expandCodeForPredicate(&PSE.getUnionPredicate(), BB->getTerminator());
2970 if (auto *C = dyn_cast<ConstantInt>(SCEVCheck))
2974 // Create a new block containing the stride check.
2975 BB->setName("vector.scevcheck");
2976 auto *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
2977 // Update dominator tree immediately if the generated block is a
2978 // LoopBypassBlock because SCEV expansions to generate loop bypass
2979 // checks may query it before the current function is finished.
2980 DT->addNewBlock(NewBB, BB);
2981 if (L->getParentLoop())
2982 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
2983 ReplaceInstWithInst(BB->getTerminator(),
2984 BranchInst::Create(Bypass, NewBB, SCEVCheck));
2985 LoopBypassBlocks.push_back(BB);
2986 AddedSafetyChecks = true;
2989 void InnerLoopVectorizer::emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass) {
2990 BasicBlock *BB = L->getLoopPreheader();
2992 // Generate the code that checks in runtime if arrays overlap. We put the
2993 // checks into a separate block to make the more common case of few elements
2995 Instruction *FirstCheckInst;
2996 Instruction *MemRuntimeCheck;
2997 std::tie(FirstCheckInst, MemRuntimeCheck) =
2998 Legal->getLAI()->addRuntimeChecks(BB->getTerminator());
2999 if (!MemRuntimeCheck)
3002 // Create a new block containing the memory check.
3003 BB->setName("vector.memcheck");
3004 auto *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
3005 // Update dominator tree immediately if the generated block is a
3006 // LoopBypassBlock because SCEV expansions to generate loop bypass
3007 // checks may query it before the current function is finished.
3008 DT->addNewBlock(NewBB, BB);
3009 if (L->getParentLoop())
3010 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
3011 ReplaceInstWithInst(BB->getTerminator(),
3012 BranchInst::Create(Bypass, NewBB, MemRuntimeCheck));
3013 LoopBypassBlocks.push_back(BB);
3014 AddedSafetyChecks = true;
3016 // We currently don't use LoopVersioning for the actual loop cloning but we
3017 // still use it to add the noalias metadata.
3018 LVer = llvm::make_unique<LoopVersioning>(*Legal->getLAI(), OrigLoop, LI, DT,
3020 LVer->prepareNoAliasMetadata();
3023 void InnerLoopVectorizer::createEmptyLoop() {
3025 In this function we generate a new loop. The new loop will contain
3026 the vectorized instructions while the old loop will continue to run the
3029 [ ] <-- loop iteration number check.
3032 | [ ] <-- vector loop bypass (may consist of multiple blocks).
3035 || [ ] <-- vector pre header.
3039 | [ ]_| <-- vector loop.
3042 | -[ ] <--- middle-block.
3045 -|- >[ ] <--- new preheader.
3049 | [ ]_| <-- old scalar loop to handle remainder.
3052 >[ ] <-- exit block.
3056 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
3057 BasicBlock *VectorPH = OrigLoop->getLoopPreheader();
3058 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
3059 assert(VectorPH && "Invalid loop structure");
3060 assert(ExitBlock && "Must have an exit block");
3062 // Some loops have a single integer induction variable, while other loops
3063 // don't. One example is c++ iterators that often have multiple pointer
3064 // induction variables. In the code below we also support a case where we
3065 // don't have a single induction variable.
3067 // We try to obtain an induction variable from the original loop as hard
3068 // as possible. However if we don't find one that:
3070 // - counts from zero, stepping by one
3071 // - is the size of the widest induction variable type
3072 // then we create a new one.
3073 OldInduction = Legal->getInduction();
3074 Type *IdxTy = Legal->getWidestInductionType();
3076 // Split the single block loop into the two loop structure described above.
3077 BasicBlock *VecBody =
3078 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
3079 BasicBlock *MiddleBlock =
3080 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
3081 BasicBlock *ScalarPH =
3082 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
3084 // Create and register the new vector loop.
3085 Loop *Lp = new Loop();
3086 Loop *ParentLoop = OrigLoop->getParentLoop();
3088 // Insert the new loop into the loop nest and register the new basic blocks
3089 // before calling any utilities such as SCEV that require valid LoopInfo.
3091 ParentLoop->addChildLoop(Lp);
3092 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
3093 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
3095 LI->addTopLevelLoop(Lp);
3097 Lp->addBasicBlockToLoop(VecBody, *LI);
3099 // Find the loop boundaries.
3100 Value *Count = getOrCreateTripCount(Lp);
3102 Value *StartIdx = ConstantInt::get(IdxTy, 0);
3104 // We need to test whether the backedge-taken count is uint##_max. Adding one
3105 // to it will cause overflow and an incorrect loop trip count in the vector
3106 // body. In case of overflow we want to directly jump to the scalar remainder
3108 emitMinimumIterationCountCheck(Lp, ScalarPH);
3109 // Now, compare the new count to zero. If it is zero skip the vector loop and
3110 // jump to the scalar loop.
3111 emitVectorLoopEnteredCheck(Lp, ScalarPH);
3112 // Generate the code to check any assumptions that we've made for SCEV
3114 emitSCEVChecks(Lp, ScalarPH);
3116 // Generate the code that checks in runtime if arrays overlap. We put the
3117 // checks into a separate block to make the more common case of few elements
3119 emitMemRuntimeChecks(Lp, ScalarPH);
3121 // Generate the induction variable.
3122 // The loop step is equal to the vectorization factor (num of SIMD elements)
3123 // times the unroll factor (num of SIMD instructions).
3124 Value *CountRoundDown = getOrCreateVectorTripCount(Lp);
3125 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
3127 createInductionVariable(Lp, StartIdx, CountRoundDown, Step,
3128 getDebugLocFromInstOrOperands(OldInduction));
3130 // We are going to resume the execution of the scalar loop.
3131 // Go over all of the induction variables that we found and fix the
3132 // PHIs that are left in the scalar version of the loop.
3133 // The starting values of PHI nodes depend on the counter of the last
3134 // iteration in the vectorized loop.
3135 // If we come from a bypass edge then we need to start from the original
3138 // This variable saves the new starting index for the scalar loop. It is used
3139 // to test if there are any tail iterations left once the vector loop has
3141 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
3142 for (auto &InductionEntry : *List) {
3143 PHINode *OrigPhi = InductionEntry.first;
3144 InductionDescriptor II = InductionEntry.second;
3146 // Create phi nodes to merge from the backedge-taken check block.
3147 PHINode *BCResumeVal = PHINode::Create(
3148 OrigPhi->getType(), 3, "bc.resume.val", ScalarPH->getTerminator());
3150 if (OrigPhi == OldInduction) {
3151 // We know what the end value is.
3152 EndValue = CountRoundDown;
3154 IRBuilder<> B(LoopBypassBlocks.back()->getTerminator());
3155 Value *CRD = B.CreateSExtOrTrunc(CountRoundDown,
3156 II.getStep()->getType(), "cast.crd");
3157 const DataLayout &DL = OrigLoop->getHeader()->getModule()->getDataLayout();
3158 EndValue = II.transform(B, CRD, PSE.getSE(), DL);
3159 EndValue->setName("ind.end");
3162 // The new PHI merges the original incoming value, in case of a bypass,
3163 // or the value at the end of the vectorized loop.
3164 BCResumeVal->addIncoming(EndValue, MiddleBlock);
3166 // Fix up external users of the induction variable.
3167 fixupIVUsers(OrigPhi, II, CountRoundDown, EndValue, MiddleBlock);
3169 // Fix the scalar body counter (PHI node).
3170 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
3172 // The old induction's phi node in the scalar body needs the truncated
3174 for (BasicBlock *BB : LoopBypassBlocks)
3175 BCResumeVal->addIncoming(II.getStartValue(), BB);
3176 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
3179 // Add a check in the middle block to see if we have completed
3180 // all of the iterations in the first vector loop.
3181 // If (N - N%VF) == N, then we *don't* need to run the remainder.
3183 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, Count,
3184 CountRoundDown, "cmp.n", MiddleBlock->getTerminator());
3185 ReplaceInstWithInst(MiddleBlock->getTerminator(),
3186 BranchInst::Create(ExitBlock, ScalarPH, CmpN));
3188 // Get ready to start creating new instructions into the vectorized body.
3189 Builder.SetInsertPoint(&*VecBody->getFirstInsertionPt());
3192 LoopVectorPreHeader = Lp->getLoopPreheader();
3193 LoopScalarPreHeader = ScalarPH;
3194 LoopMiddleBlock = MiddleBlock;
3195 LoopExitBlock = ExitBlock;
3196 LoopVectorBody = VecBody;
3197 LoopScalarBody = OldBasicBlock;
3199 // Keep all loop hints from the original loop on the vector loop (we'll
3200 // replace the vectorizer-specific hints below).
3201 if (MDNode *LID = OrigLoop->getLoopID())
3204 LoopVectorizeHints Hints(Lp, true);
3205 Hints.setAlreadyVectorized();
3208 // Fix up external users of the induction variable. At this point, we are
3209 // in LCSSA form, with all external PHIs that use the IV having one input value,
3210 // coming from the remainder loop. We need those PHIs to also have a correct
3211 // value for the IV when arriving directly from the middle block.
3212 void InnerLoopVectorizer::fixupIVUsers(PHINode *OrigPhi,
3213 const InductionDescriptor &II,
3214 Value *CountRoundDown, Value *EndValue,
3215 BasicBlock *MiddleBlock) {
3216 // There are two kinds of external IV usages - those that use the value
3217 // computed in the last iteration (the PHI) and those that use the penultimate
3218 // value (the value that feeds into the phi from the loop latch).
3219 // We allow both, but they, obviously, have different values.
3221 assert(OrigLoop->getExitBlock() && "Expected a single exit block");
3223 DenseMap<Value *, Value *> MissingVals;
3225 // An external user of the last iteration's value should see the value that
3226 // the remainder loop uses to initialize its own IV.
3227 Value *PostInc = OrigPhi->getIncomingValueForBlock(OrigLoop->getLoopLatch());
3228 for (User *U : PostInc->users()) {
3229 Instruction *UI = cast<Instruction>(U);
3230 if (!OrigLoop->contains(UI)) {
3231 assert(isa<PHINode>(UI) && "Expected LCSSA form");
3232 MissingVals[UI] = EndValue;
3236 // An external user of the penultimate value need to see EndValue - Step.
3237 // The simplest way to get this is to recompute it from the constituent SCEVs,
3238 // that is Start + (Step * (CRD - 1)).
3239 for (User *U : OrigPhi->users()) {
3240 auto *UI = cast<Instruction>(U);
3241 if (!OrigLoop->contains(UI)) {
3242 const DataLayout &DL =
3243 OrigLoop->getHeader()->getModule()->getDataLayout();
3244 assert(isa<PHINode>(UI) && "Expected LCSSA form");
3246 IRBuilder<> B(MiddleBlock->getTerminator());
3247 Value *CountMinusOne = B.CreateSub(
3248 CountRoundDown, ConstantInt::get(CountRoundDown->getType(), 1));
3249 Value *CMO = B.CreateSExtOrTrunc(CountMinusOne, II.getStep()->getType(),
3251 Value *Escape = II.transform(B, CMO, PSE.getSE(), DL);
3252 Escape->setName("ind.escape");
3253 MissingVals[UI] = Escape;
3257 for (auto &I : MissingVals) {
3258 PHINode *PHI = cast<PHINode>(I.first);
3259 // One corner case we have to handle is two IVs "chasing" each-other,
3260 // that is %IV2 = phi [...], [ %IV1, %latch ]
3261 // In this case, if IV1 has an external use, we need to avoid adding both
3262 // "last value of IV1" and "penultimate value of IV2". So, verify that we
3263 // don't already have an incoming value for the middle block.
3264 if (PHI->getBasicBlockIndex(MiddleBlock) == -1)
3265 PHI->addIncoming(I.second, MiddleBlock);
3270 struct CSEDenseMapInfo {
3271 static bool canHandle(Instruction *I) {
3272 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
3273 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
3275 static inline Instruction *getEmptyKey() {
3276 return DenseMapInfo<Instruction *>::getEmptyKey();
3278 static inline Instruction *getTombstoneKey() {
3279 return DenseMapInfo<Instruction *>::getTombstoneKey();
3281 static unsigned getHashValue(Instruction *I) {
3282 assert(canHandle(I) && "Unknown instruction!");
3283 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
3284 I->value_op_end()));
3286 static bool isEqual(Instruction *LHS, Instruction *RHS) {
3287 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
3288 LHS == getTombstoneKey() || RHS == getTombstoneKey())
3290 return LHS->isIdenticalTo(RHS);
3295 ///\brief Perform cse of induction variable instructions.
3296 static void cse(BasicBlock *BB) {
3297 // Perform simple cse.
3298 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
3299 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
3300 Instruction *In = &*I++;
3302 if (!CSEDenseMapInfo::canHandle(In))
3305 // Check if we can replace this instruction with any of the
3306 // visited instructions.
3307 if (Instruction *V = CSEMap.lookup(In)) {
3308 In->replaceAllUsesWith(V);
3309 In->eraseFromParent();
3317 /// \brief Adds a 'fast' flag to floating point operations.
3318 static Value *addFastMathFlag(Value *V) {
3319 if (isa<FPMathOperator>(V)) {
3320 FastMathFlags Flags;
3321 Flags.setUnsafeAlgebra();
3322 cast<Instruction>(V)->setFastMathFlags(Flags);
3327 /// Estimate the overhead of scalarizing a value. Insert and Extract are set if
3328 /// the result needs to be inserted and/or extracted from vectors.
3329 static unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract,
3330 const TargetTransformInfo &TTI) {
3334 assert(Ty->isVectorTy() && "Can only scalarize vectors");
3337 for (unsigned I = 0, E = Ty->getVectorNumElements(); I < E; ++I) {
3339 Cost += TTI.getVectorInstrCost(Instruction::InsertElement, Ty, I);
3341 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, Ty, I);
3347 // Estimate cost of a call instruction CI if it were vectorized with factor VF.
3348 // Return the cost of the instruction, including scalarization overhead if it's
3349 // needed. The flag NeedToScalarize shows if the call needs to be scalarized -
3350 // i.e. either vector version isn't available, or is too expensive.
3351 static unsigned getVectorCallCost(CallInst *CI, unsigned VF,
3352 const TargetTransformInfo &TTI,
3353 const TargetLibraryInfo *TLI,
3354 bool &NeedToScalarize) {
3355 Function *F = CI->getCalledFunction();
3356 StringRef FnName = CI->getCalledFunction()->getName();
3357 Type *ScalarRetTy = CI->getType();
3358 SmallVector<Type *, 4> Tys, ScalarTys;
3359 for (auto &ArgOp : CI->arg_operands())
3360 ScalarTys.push_back(ArgOp->getType());
3362 // Estimate cost of scalarized vector call. The source operands are assumed
3363 // to be vectors, so we need to extract individual elements from there,
3364 // execute VF scalar calls, and then gather the result into the vector return
3366 unsigned ScalarCallCost = TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys);
3368 return ScalarCallCost;
3370 // Compute corresponding vector type for return value and arguments.
3371 Type *RetTy = ToVectorTy(ScalarRetTy, VF);
3372 for (Type *ScalarTy : ScalarTys)
3373 Tys.push_back(ToVectorTy(ScalarTy, VF));
3375 // Compute costs of unpacking argument values for the scalar calls and
3376 // packing the return values to a vector.
3377 unsigned ScalarizationCost =
3378 getScalarizationOverhead(RetTy, true, false, TTI);
3379 for (Type *Ty : Tys)
3380 ScalarizationCost += getScalarizationOverhead(Ty, false, true, TTI);
3382 unsigned Cost = ScalarCallCost * VF + ScalarizationCost;
3384 // If we can't emit a vector call for this function, then the currently found
3385 // cost is the cost we need to return.
3386 NeedToScalarize = true;
3387 if (!TLI || !TLI->isFunctionVectorizable(FnName, VF) || CI->isNoBuiltin())
3390 // If the corresponding vector cost is cheaper, return its cost.
3391 unsigned VectorCallCost = TTI.getCallInstrCost(nullptr, RetTy, Tys);
3392 if (VectorCallCost < Cost) {
3393 NeedToScalarize = false;
3394 return VectorCallCost;
3399 // Estimate cost of an intrinsic call instruction CI if it were vectorized with
3400 // factor VF. Return the cost of the instruction, including scalarization
3401 // overhead if it's needed.
3402 static unsigned getVectorIntrinsicCost(CallInst *CI, unsigned VF,
3403 const TargetTransformInfo &TTI,
3404 const TargetLibraryInfo *TLI) {
3405 Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
3406 assert(ID && "Expected intrinsic call!");
3408 Type *RetTy = ToVectorTy(CI->getType(), VF);
3409 SmallVector<Type *, 4> Tys;
3410 for (Value *ArgOperand : CI->arg_operands())
3411 Tys.push_back(ToVectorTy(ArgOperand->getType(), VF));
3414 if (auto *FPMO = dyn_cast<FPMathOperator>(CI))
3415 FMF = FPMO->getFastMathFlags();
3417 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys, FMF);
3420 static Type *smallestIntegerVectorType(Type *T1, Type *T2) {
3421 auto *I1 = cast<IntegerType>(T1->getVectorElementType());
3422 auto *I2 = cast<IntegerType>(T2->getVectorElementType());
3423 return I1->getBitWidth() < I2->getBitWidth() ? T1 : T2;
3425 static Type *largestIntegerVectorType(Type *T1, Type *T2) {
3426 auto *I1 = cast<IntegerType>(T1->getVectorElementType());
3427 auto *I2 = cast<IntegerType>(T2->getVectorElementType());
3428 return I1->getBitWidth() > I2->getBitWidth() ? T1 : T2;
3431 void InnerLoopVectorizer::truncateToMinimalBitwidths() {
3432 // For every instruction `I` in MinBWs, truncate the operands, create a
3433 // truncated version of `I` and reextend its result. InstCombine runs
3434 // later and will remove any ext/trunc pairs.
3436 SmallPtrSet<Value *, 4> Erased;
3437 for (const auto &KV : *MinBWs) {
3438 VectorParts &Parts = WidenMap.get(KV.first);
3439 for (Value *&I : Parts) {
3440 if (Erased.count(I) || I->use_empty() || !isa<Instruction>(I))
3442 Type *OriginalTy = I->getType();
3443 Type *ScalarTruncatedTy =
3444 IntegerType::get(OriginalTy->getContext(), KV.second);
3445 Type *TruncatedTy = VectorType::get(ScalarTruncatedTy,
3446 OriginalTy->getVectorNumElements());
3447 if (TruncatedTy == OriginalTy)
3450 IRBuilder<> B(cast<Instruction>(I));
3451 auto ShrinkOperand = [&](Value *V) -> Value * {
3452 if (auto *ZI = dyn_cast<ZExtInst>(V))
3453 if (ZI->getSrcTy() == TruncatedTy)
3454 return ZI->getOperand(0);
3455 return B.CreateZExtOrTrunc(V, TruncatedTy);
3458 // The actual instruction modification depends on the instruction type,
3460 Value *NewI = nullptr;
3461 if (auto *BO = dyn_cast<BinaryOperator>(I)) {
3462 NewI = B.CreateBinOp(BO->getOpcode(), ShrinkOperand(BO->getOperand(0)),
3463 ShrinkOperand(BO->getOperand(1)));
3464 cast<BinaryOperator>(NewI)->copyIRFlags(I);
3465 } else if (auto *CI = dyn_cast<ICmpInst>(I)) {
3467 B.CreateICmp(CI->getPredicate(), ShrinkOperand(CI->getOperand(0)),
3468 ShrinkOperand(CI->getOperand(1)));
3469 } else if (auto *SI = dyn_cast<SelectInst>(I)) {
3470 NewI = B.CreateSelect(SI->getCondition(),
3471 ShrinkOperand(SI->getTrueValue()),
3472 ShrinkOperand(SI->getFalseValue()));
3473 } else if (auto *CI = dyn_cast<CastInst>(I)) {
3474 switch (CI->getOpcode()) {
3476 llvm_unreachable("Unhandled cast!");
3477 case Instruction::Trunc:
3478 NewI = ShrinkOperand(CI->getOperand(0));
3480 case Instruction::SExt:
3481 NewI = B.CreateSExtOrTrunc(
3483 smallestIntegerVectorType(OriginalTy, TruncatedTy));
3485 case Instruction::ZExt:
3486 NewI = B.CreateZExtOrTrunc(
3488 smallestIntegerVectorType(OriginalTy, TruncatedTy));
3491 } else if (auto *SI = dyn_cast<ShuffleVectorInst>(I)) {
3492 auto Elements0 = SI->getOperand(0)->getType()->getVectorNumElements();
3493 auto *O0 = B.CreateZExtOrTrunc(
3494 SI->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements0));
3495 auto Elements1 = SI->getOperand(1)->getType()->getVectorNumElements();
3496 auto *O1 = B.CreateZExtOrTrunc(
3497 SI->getOperand(1), VectorType::get(ScalarTruncatedTy, Elements1));
3499 NewI = B.CreateShuffleVector(O0, O1, SI->getMask());
3500 } else if (isa<LoadInst>(I)) {
3501 // Don't do anything with the operands, just extend the result.
3503 } else if (auto *IE = dyn_cast<InsertElementInst>(I)) {
3504 auto Elements = IE->getOperand(0)->getType()->getVectorNumElements();
3505 auto *O0 = B.CreateZExtOrTrunc(
3506 IE->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements));
3507 auto *O1 = B.CreateZExtOrTrunc(IE->getOperand(1), ScalarTruncatedTy);
3508 NewI = B.CreateInsertElement(O0, O1, IE->getOperand(2));
3509 } else if (auto *EE = dyn_cast<ExtractElementInst>(I)) {
3510 auto Elements = EE->getOperand(0)->getType()->getVectorNumElements();
3511 auto *O0 = B.CreateZExtOrTrunc(
3512 EE->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements));
3513 NewI = B.CreateExtractElement(O0, EE->getOperand(2));
3515 llvm_unreachable("Unhandled instruction type!");
3518 // Lastly, extend the result.
3519 NewI->takeName(cast<Instruction>(I));
3520 Value *Res = B.CreateZExtOrTrunc(NewI, OriginalTy);
3521 I->replaceAllUsesWith(Res);
3522 cast<Instruction>(I)->eraseFromParent();
3528 // We'll have created a bunch of ZExts that are now parentless. Clean up.
3529 for (const auto &KV : *MinBWs) {
3530 VectorParts &Parts = WidenMap.get(KV.first);
3531 for (Value *&I : Parts) {
3532 ZExtInst *Inst = dyn_cast<ZExtInst>(I);
3533 if (Inst && Inst->use_empty()) {
3534 Value *NewI = Inst->getOperand(0);
3535 Inst->eraseFromParent();
3542 void InnerLoopVectorizer::vectorizeLoop() {
3543 //===------------------------------------------------===//
3545 // Notice: any optimization or new instruction that go
3546 // into the code below should be also be implemented in
3549 //===------------------------------------------------===//
3550 Constant *Zero = Builder.getInt32(0);
3552 // In order to support recurrences we need to be able to vectorize Phi nodes.
3553 // Phi nodes have cycles, so we need to vectorize them in two stages. First,
3554 // we create a new vector PHI node with no incoming edges. We use this value
3555 // when we vectorize all of the instructions that use the PHI. Next, after
3556 // all of the instructions in the block are complete we add the new incoming
3557 // edges to the PHI. At this point all of the instructions in the basic block
3558 // are vectorized, so we can use them to construct the PHI.
3559 PhiVector PHIsToFix;
3561 // Scan the loop in a topological order to ensure that defs are vectorized
3563 LoopBlocksDFS DFS(OrigLoop);
3566 // Vectorize all of the blocks in the original loop.
3567 for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO()))
3568 vectorizeBlockInLoop(BB, &PHIsToFix);
3570 // Insert truncates and extends for any truncated instructions as hints to
3573 truncateToMinimalBitwidths();
3575 // At this point every instruction in the original loop is widened to a
3576 // vector form. Now we need to fix the recurrences in PHIsToFix. These PHI
3577 // nodes are currently empty because we did not want to introduce cycles.
3578 // This is the second stage of vectorizing recurrences.
3579 for (PHINode *Phi : PHIsToFix) {
3580 assert(Phi && "Unable to recover vectorized PHI");
3582 // Handle first-order recurrences that need to be fixed.
3583 if (Legal->isFirstOrderRecurrence(Phi)) {
3584 fixFirstOrderRecurrence(Phi);
3588 // If the phi node is not a first-order recurrence, it must be a reduction.
3589 // Get it's reduction variable descriptor.
3590 assert(Legal->isReductionVariable(Phi) &&
3591 "Unable to find the reduction variable");
3592 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[Phi];
3594 RecurrenceDescriptor::RecurrenceKind RK = RdxDesc.getRecurrenceKind();
3595 TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue();
3596 Instruction *LoopExitInst = RdxDesc.getLoopExitInstr();
3597 RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind =
3598 RdxDesc.getMinMaxRecurrenceKind();
3599 setDebugLocFromInst(Builder, ReductionStartValue);
3601 // We need to generate a reduction vector from the incoming scalar.
3602 // To do so, we need to generate the 'identity' vector and override
3603 // one of the elements with the incoming scalar reduction. We need
3604 // to do it in the vector-loop preheader.
3605 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
3607 // This is the vector-clone of the value that leaves the loop.
3608 VectorParts &VectorExit = getVectorValue(LoopExitInst);
3609 Type *VecTy = VectorExit[0]->getType();
3611 // Find the reduction identity variable. Zero for addition, or, xor,
3612 // one for multiplication, -1 for And.
3615 if (RK == RecurrenceDescriptor::RK_IntegerMinMax ||
3616 RK == RecurrenceDescriptor::RK_FloatMinMax) {
3617 // MinMax reduction have the start value as their identify.
3619 VectorStart = Identity = ReductionStartValue;
3621 VectorStart = Identity =
3622 Builder.CreateVectorSplat(VF, ReductionStartValue, "minmax.ident");
3625 // Handle other reduction kinds:
3626 Constant *Iden = RecurrenceDescriptor::getRecurrenceIdentity(
3627 RK, VecTy->getScalarType());
3630 // This vector is the Identity vector where the first element is the
3631 // incoming scalar reduction.
3632 VectorStart = ReductionStartValue;
3634 Identity = ConstantVector::getSplat(VF, Iden);
3636 // This vector is the Identity vector where the first element is the
3637 // incoming scalar reduction.
3639 Builder.CreateInsertElement(Identity, ReductionStartValue, Zero);
3643 // Fix the vector-loop phi.
3645 // Reductions do not have to start at zero. They can start with
3646 // any loop invariant values.
3647 VectorParts &VecRdxPhi = WidenMap.get(Phi);
3648 BasicBlock *Latch = OrigLoop->getLoopLatch();
3649 Value *LoopVal = Phi->getIncomingValueForBlock(Latch);
3650 VectorParts &Val = getVectorValue(LoopVal);
3651 for (unsigned part = 0; part < UF; ++part) {
3652 // Make sure to add the reduction stat value only to the
3653 // first unroll part.
3654 Value *StartVal = (part == 0) ? VectorStart : Identity;
3655 cast<PHINode>(VecRdxPhi[part])
3656 ->addIncoming(StartVal, LoopVectorPreHeader);
3657 cast<PHINode>(VecRdxPhi[part])
3658 ->addIncoming(Val[part], LoopVectorBody);
3661 // Before each round, move the insertion point right between
3662 // the PHIs and the values we are going to write.
3663 // This allows us to write both PHINodes and the extractelement
3665 Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt());
3667 VectorParts RdxParts = getVectorValue(LoopExitInst);
3668 setDebugLocFromInst(Builder, LoopExitInst);
3670 // If the vector reduction can be performed in a smaller type, we truncate
3671 // then extend the loop exit value to enable InstCombine to evaluate the
3672 // entire expression in the smaller type.
3673 if (VF > 1 && Phi->getType() != RdxDesc.getRecurrenceType()) {
3674 Type *RdxVecTy = VectorType::get(RdxDesc.getRecurrenceType(), VF);
3675 Builder.SetInsertPoint(LoopVectorBody->getTerminator());
3676 for (unsigned part = 0; part < UF; ++part) {
3677 Value *Trunc = Builder.CreateTrunc(RdxParts[part], RdxVecTy);
3678 Value *Extnd = RdxDesc.isSigned() ? Builder.CreateSExt(Trunc, VecTy)
3679 : Builder.CreateZExt(Trunc, VecTy);
3680 for (Value::user_iterator UI = RdxParts[part]->user_begin();
3681 UI != RdxParts[part]->user_end();)
3683 (*UI++)->replaceUsesOfWith(RdxParts[part], Extnd);
3684 RdxParts[part] = Extnd;
3689 Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt());
3690 for (unsigned part = 0; part < UF; ++part)
3691 RdxParts[part] = Builder.CreateTrunc(RdxParts[part], RdxVecTy);
3694 // Reduce all of the unrolled parts into a single vector.
3695 Value *ReducedPartRdx = RdxParts[0];
3696 unsigned Op = RecurrenceDescriptor::getRecurrenceBinOp(RK);
3697 setDebugLocFromInst(Builder, ReducedPartRdx);
3698 for (unsigned part = 1; part < UF; ++part) {
3699 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
3700 // Floating point operations had to be 'fast' to enable the reduction.
3701 ReducedPartRdx = addFastMathFlag(
3702 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
3703 ReducedPartRdx, "bin.rdx"));
3705 ReducedPartRdx = RecurrenceDescriptor::createMinMaxOp(
3706 Builder, MinMaxKind, ReducedPartRdx, RdxParts[part]);
3710 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
3711 // and vector ops, reducing the set of values being computed by half each
3713 assert(isPowerOf2_32(VF) &&
3714 "Reduction emission only supported for pow2 vectors!");
3715 Value *TmpVec = ReducedPartRdx;
3716 SmallVector<Constant *, 32> ShuffleMask(VF, nullptr);
3717 for (unsigned i = VF; i != 1; i >>= 1) {
3718 // Move the upper half of the vector to the lower half.
3719 for (unsigned j = 0; j != i / 2; ++j)
3720 ShuffleMask[j] = Builder.getInt32(i / 2 + j);
3722 // Fill the rest of the mask with undef.
3723 std::fill(&ShuffleMask[i / 2], ShuffleMask.end(),
3724 UndefValue::get(Builder.getInt32Ty()));
3726 Value *Shuf = Builder.CreateShuffleVector(
3727 TmpVec, UndefValue::get(TmpVec->getType()),
3728 ConstantVector::get(ShuffleMask), "rdx.shuf");
3730 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
3731 // Floating point operations had to be 'fast' to enable the reduction.
3732 TmpVec = addFastMathFlag(Builder.CreateBinOp(
3733 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
3735 TmpVec = RecurrenceDescriptor::createMinMaxOp(Builder, MinMaxKind,
3739 // The result is in the first element of the vector.
3741 Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
3743 // If the reduction can be performed in a smaller type, we need to extend
3744 // the reduction to the wider type before we branch to the original loop.
3745 if (Phi->getType() != RdxDesc.getRecurrenceType())
3748 ? Builder.CreateSExt(ReducedPartRdx, Phi->getType())
3749 : Builder.CreateZExt(ReducedPartRdx, Phi->getType());
3752 // Create a phi node that merges control-flow from the backedge-taken check
3753 // block and the middle block.
3754 PHINode *BCBlockPhi = PHINode::Create(Phi->getType(), 2, "bc.merge.rdx",
3755 LoopScalarPreHeader->getTerminator());
3756 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
3757 BCBlockPhi->addIncoming(ReductionStartValue, LoopBypassBlocks[I]);
3758 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3760 // Now, we need to fix the users of the reduction variable
3761 // inside and outside of the scalar remainder loop.
3762 // We know that the loop is in LCSSA form. We need to update the
3763 // PHI nodes in the exit blocks.
3764 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3765 LEE = LoopExitBlock->end();
3766 LEI != LEE; ++LEI) {
3767 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3771 // All PHINodes need to have a single entry edge, or two if
3772 // we already fixed them.
3773 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
3775 // We found our reduction value exit-PHI. Update it with the
3776 // incoming bypass edge.
3777 if (LCSSAPhi->getIncomingValue(0) == LoopExitInst) {
3778 // Add an edge coming from the bypass.
3779 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3782 } // end of the LCSSA phi scan.
3784 // Fix the scalar loop reduction variable with the incoming reduction sum
3785 // from the vector body and from the backedge value.
3786 int IncomingEdgeBlockIdx =
3787 Phi->getBasicBlockIndex(OrigLoop->getLoopLatch());
3788 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
3789 // Pick the other block.
3790 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
3791 Phi->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
3792 Phi->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst);
3793 } // end of for each Phi in PHIsToFix.
3797 // Make sure DomTree is updated.
3800 // Predicate any stores.
3801 for (auto KV : PredicatedStores) {
3802 BasicBlock::iterator I(KV.first);
3803 auto *BB = SplitBlock(I->getParent(), &*std::next(I), DT, LI);
3804 auto *T = SplitBlockAndInsertIfThen(KV.second, &*I, /*Unreachable=*/false,
3805 /*BranchWeights=*/nullptr, DT, LI);
3807 I->getParent()->setName("pred.store.if");
3808 BB->setName("pred.store.continue");
3810 DEBUG(DT->verifyDomTree());
3811 // Remove redundant induction instructions.
3812 cse(LoopVectorBody);
3815 void InnerLoopVectorizer::fixFirstOrderRecurrence(PHINode *Phi) {
3817 // This is the second phase of vectorizing first-order recurrences. An
3818 // overview of the transformation is described below. Suppose we have the
3821 // for (int i = 0; i < n; ++i)
3822 // b[i] = a[i] - a[i - 1];
3824 // There is a first-order recurrence on "a". For this loop, the shorthand
3825 // scalar IR looks like:
3832 // i = phi [0, scalar.ph], [i+1, scalar.body]
3833 // s1 = phi [s_init, scalar.ph], [s2, scalar.body]
3836 // br cond, scalar.body, ...
3838 // In this example, s1 is a recurrence because it's value depends on the
3839 // previous iteration. In the first phase of vectorization, we created a
3840 // temporary value for s1. We now complete the vectorization and produce the
3841 // shorthand vector IR shown below (for VF = 4, UF = 1).
3844 // v_init = vector(..., ..., ..., a[-1])
3848 // i = phi [0, vector.ph], [i+4, vector.body]
3849 // v1 = phi [v_init, vector.ph], [v2, vector.body]
3850 // v2 = a[i, i+1, i+2, i+3];
3851 // v3 = vector(v1(3), v2(0, 1, 2))
3852 // b[i, i+1, i+2, i+3] = v2 - v3
3853 // br cond, vector.body, middle.block
3860 // s_init = phi [x, middle.block], [a[-1], otherwise]
3863 // After execution completes the vector loop, we extract the next value of
3864 // the recurrence (x) to use as the initial value in the scalar loop.
3866 // Get the original loop preheader and single loop latch.
3867 auto *Preheader = OrigLoop->getLoopPreheader();
3868 auto *Latch = OrigLoop->getLoopLatch();
3870 // Get the initial and previous values of the scalar recurrence.
3871 auto *ScalarInit = Phi->getIncomingValueForBlock(Preheader);
3872 auto *Previous = Phi->getIncomingValueForBlock(Latch);
3874 // Create a vector from the initial value.
3875 auto *VectorInit = ScalarInit;
3877 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
3878 VectorInit = Builder.CreateInsertElement(
3879 UndefValue::get(VectorType::get(VectorInit->getType(), VF)), VectorInit,
3880 Builder.getInt32(VF - 1), "vector.recur.init");
3883 // We constructed a temporary phi node in the first phase of vectorization.
3884 // This phi node will eventually be deleted.
3885 auto &PhiParts = getVectorValue(Phi);
3886 Builder.SetInsertPoint(cast<Instruction>(PhiParts[0]));
3888 // Create a phi node for the new recurrence. The current value will either be
3889 // the initial value inserted into a vector or loop-varying vector value.
3890 auto *VecPhi = Builder.CreatePHI(VectorInit->getType(), 2, "vector.recur");
3891 VecPhi->addIncoming(VectorInit, LoopVectorPreHeader);
3893 // Get the vectorized previous value. We ensured the previous values was an
3894 // instruction when detecting the recurrence.
3895 auto &PreviousParts = getVectorValue(Previous);
3897 // Set the insertion point to be after this instruction. We ensured the
3898 // previous value dominated all uses of the phi when detecting the
3900 Builder.SetInsertPoint(
3901 &*++BasicBlock::iterator(cast<Instruction>(PreviousParts[UF - 1])));
3903 // We will construct a vector for the recurrence by combining the values for
3904 // the current and previous iterations. This is the required shuffle mask.
3905 SmallVector<Constant *, 8> ShuffleMask(VF);
3906 ShuffleMask[0] = Builder.getInt32(VF - 1);
3907 for (unsigned I = 1; I < VF; ++I)
3908 ShuffleMask[I] = Builder.getInt32(I + VF - 1);
3910 // The vector from which to take the initial value for the current iteration
3911 // (actual or unrolled). Initially, this is the vector phi node.
3912 Value *Incoming = VecPhi;
3914 // Shuffle the current and previous vector and update the vector parts.
3915 for (unsigned Part = 0; Part < UF; ++Part) {
3918 ? Builder.CreateShuffleVector(Incoming, PreviousParts[Part],
3919 ConstantVector::get(ShuffleMask))
3921 PhiParts[Part]->replaceAllUsesWith(Shuffle);
3922 cast<Instruction>(PhiParts[Part])->eraseFromParent();
3923 PhiParts[Part] = Shuffle;
3924 Incoming = PreviousParts[Part];
3927 // Fix the latch value of the new recurrence in the vector loop.
3928 VecPhi->addIncoming(Incoming, LI->getLoopFor(LoopVectorBody)->getLoopLatch());
3930 // Extract the last vector element in the middle block. This will be the
3931 // initial value for the recurrence when jumping to the scalar loop.
3932 auto *Extract = Incoming;
3934 Builder.SetInsertPoint(LoopMiddleBlock->getTerminator());
3935 Extract = Builder.CreateExtractElement(Extract, Builder.getInt32(VF - 1),
3936 "vector.recur.extract");
3939 // Fix the initial value of the original recurrence in the scalar loop.
3940 Builder.SetInsertPoint(&*LoopScalarPreHeader->begin());
3941 auto *Start = Builder.CreatePHI(Phi->getType(), 2, "scalar.recur.init");
3942 for (auto *BB : predecessors(LoopScalarPreHeader)) {
3943 auto *Incoming = BB == LoopMiddleBlock ? Extract : ScalarInit;
3944 Start->addIncoming(Incoming, BB);
3947 Phi->setIncomingValue(Phi->getBasicBlockIndex(LoopScalarPreHeader), Start);
3948 Phi->setName("scalar.recur");
3950 // Finally, fix users of the recurrence outside the loop. The users will need
3951 // either the last value of the scalar recurrence or the last value of the
3952 // vector recurrence we extracted in the middle block. Since the loop is in
3953 // LCSSA form, we just need to find the phi node for the original scalar
3954 // recurrence in the exit block, and then add an edge for the middle block.
3955 for (auto &I : *LoopExitBlock) {
3956 auto *LCSSAPhi = dyn_cast<PHINode>(&I);
3959 if (LCSSAPhi->getIncomingValue(0) == Phi) {
3960 LCSSAPhi->addIncoming(Extract, LoopMiddleBlock);
3966 void InnerLoopVectorizer::fixLCSSAPHIs() {
3967 for (Instruction &LEI : *LoopExitBlock) {
3968 auto *LCSSAPhi = dyn_cast<PHINode>(&LEI);
3971 if (LCSSAPhi->getNumIncomingValues() == 1)
3972 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
3977 InnerLoopVectorizer::VectorParts
3978 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
3979 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
3982 // Look for cached value.
3983 std::pair<BasicBlock *, BasicBlock *> Edge(Src, Dst);
3984 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
3985 if (ECEntryIt != MaskCache.end())
3986 return ECEntryIt->second;
3988 VectorParts SrcMask = createBlockInMask(Src);
3990 // The terminator has to be a branch inst!
3991 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
3992 assert(BI && "Unexpected terminator found");
3994 if (BI->isConditional()) {
3995 VectorParts EdgeMask = getVectorValue(BI->getCondition());
3997 if (BI->getSuccessor(0) != Dst)
3998 for (unsigned part = 0; part < UF; ++part)
3999 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
4001 for (unsigned part = 0; part < UF; ++part)
4002 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
4004 MaskCache[Edge] = EdgeMask;
4008 MaskCache[Edge] = SrcMask;
4012 InnerLoopVectorizer::VectorParts
4013 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
4014 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
4016 // Loop incoming mask is all-one.
4017 if (OrigLoop->getHeader() == BB) {
4018 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
4019 return getVectorValue(C);
4022 // This is the block mask. We OR all incoming edges, and with zero.
4023 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
4024 VectorParts BlockMask = getVectorValue(Zero);
4027 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
4028 VectorParts EM = createEdgeMask(*it, BB);
4029 for (unsigned part = 0; part < UF; ++part)
4030 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
4036 void InnerLoopVectorizer::widenPHIInstruction(
4037 Instruction *PN, InnerLoopVectorizer::VectorParts &Entry, unsigned UF,
4038 unsigned VF, PhiVector *PV) {
4039 PHINode *P = cast<PHINode>(PN);
4040 // Handle recurrences.
4041 if (Legal->isReductionVariable(P) || Legal->isFirstOrderRecurrence(P)) {
4042 for (unsigned part = 0; part < UF; ++part) {
4043 // This is phase one of vectorizing PHIs.
4045 (VF == 1) ? PN->getType() : VectorType::get(PN->getType(), VF);
4046 Entry[part] = PHINode::Create(
4047 VecTy, 2, "vec.phi", &*LoopVectorBody->getFirstInsertionPt());
4053 setDebugLocFromInst(Builder, P);
4054 // Check for PHI nodes that are lowered to vector selects.
4055 if (P->getParent() != OrigLoop->getHeader()) {
4056 // We know that all PHIs in non-header blocks are converted into
4057 // selects, so we don't have to worry about the insertion order and we
4058 // can just use the builder.
4059 // At this point we generate the predication tree. There may be
4060 // duplications since this is a simple recursive scan, but future
4061 // optimizations will clean it up.
4063 unsigned NumIncoming = P->getNumIncomingValues();
4065 // Generate a sequence of selects of the form:
4066 // SELECT(Mask3, In3,
4067 // SELECT(Mask2, In2,
4069 for (unsigned In = 0; In < NumIncoming; In++) {
4071 createEdgeMask(P->getIncomingBlock(In), P->getParent());
4072 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
4074 for (unsigned part = 0; part < UF; ++part) {
4075 // We might have single edge PHIs (blocks) - use an identity
4076 // 'select' for the first PHI operand.
4078 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], In0[part]);
4080 // Select between the current value and the previous incoming edge
4081 // based on the incoming mask.
4082 Entry[part] = Builder.CreateSelect(Cond[part], In0[part], Entry[part],
4089 // This PHINode must be an induction variable.
4090 // Make sure that we know about it.
4091 assert(Legal->getInductionVars()->count(P) && "Not an induction variable");
4093 InductionDescriptor II = Legal->getInductionVars()->lookup(P);
4094 const DataLayout &DL = OrigLoop->getHeader()->getModule()->getDataLayout();
4096 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
4097 // which can be found from the original scalar operations.
4098 switch (II.getKind()) {
4099 case InductionDescriptor::IK_NoInduction:
4100 llvm_unreachable("Unknown induction");
4101 case InductionDescriptor::IK_IntInduction:
4102 return widenIntInduction(P, Entry);
4103 case InductionDescriptor::IK_PtrInduction:
4104 // Handle the pointer induction variable case.
4105 assert(P->getType()->isPointerTy() && "Unexpected type.");
4106 // This is the normalized GEP that starts counting at zero.
4107 Value *PtrInd = Induction;
4108 PtrInd = Builder.CreateSExtOrTrunc(PtrInd, II.getStep()->getType());
4109 // This is the vector of results. Notice that we don't generate
4110 // vector geps because scalar geps result in better code.
4111 for (unsigned part = 0; part < UF; ++part) {
4113 int EltIndex = part;
4114 Constant *Idx = ConstantInt::get(PtrInd->getType(), EltIndex);
4115 Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx);
4116 Value *SclrGep = II.transform(Builder, GlobalIdx, PSE.getSE(), DL);
4117 SclrGep->setName("next.gep");
4118 Entry[part] = SclrGep;
4122 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
4123 for (unsigned int i = 0; i < VF; ++i) {
4124 int EltIndex = i + part * VF;
4125 Constant *Idx = ConstantInt::get(PtrInd->getType(), EltIndex);
4126 Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx);
4127 Value *SclrGep = II.transform(Builder, GlobalIdx, PSE.getSE(), DL);
4128 SclrGep->setName("next.gep");
4129 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
4130 Builder.getInt32(i), "insert.gep");
4132 Entry[part] = VecVal;
4138 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
4139 // For each instruction in the old loop.
4140 for (Instruction &I : *BB) {
4141 VectorParts &Entry = WidenMap.get(&I);
4143 switch (I.getOpcode()) {
4144 case Instruction::Br:
4145 // Nothing to do for PHIs and BR, since we already took care of the
4146 // loop control flow instructions.
4148 case Instruction::PHI: {
4149 // Vectorize PHINodes.
4150 widenPHIInstruction(&I, Entry, UF, VF, PV);
4154 case Instruction::Add:
4155 case Instruction::FAdd:
4156 case Instruction::Sub:
4157 case Instruction::FSub:
4158 case Instruction::Mul:
4159 case Instruction::FMul:
4160 case Instruction::UDiv:
4161 case Instruction::SDiv:
4162 case Instruction::FDiv:
4163 case Instruction::URem:
4164 case Instruction::SRem:
4165 case Instruction::FRem:
4166 case Instruction::Shl:
4167 case Instruction::LShr:
4168 case Instruction::AShr:
4169 case Instruction::And:
4170 case Instruction::Or:
4171 case Instruction::Xor: {
4172 // Just widen binops.
4173 auto *BinOp = cast<BinaryOperator>(&I);
4174 setDebugLocFromInst(Builder, BinOp);
4175 VectorParts &A = getVectorValue(BinOp->getOperand(0));
4176 VectorParts &B = getVectorValue(BinOp->getOperand(1));
4178 // Use this vector value for all users of the original instruction.
4179 for (unsigned Part = 0; Part < UF; ++Part) {
4180 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
4182 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
4183 VecOp->copyIRFlags(BinOp);
4188 addMetadata(Entry, BinOp);
4191 case Instruction::Select: {
4193 // If the selector is loop invariant we can create a select
4194 // instruction with a scalar condition. Otherwise, use vector-select.
4195 auto *SE = PSE.getSE();
4196 bool InvariantCond =
4197 SE->isLoopInvariant(PSE.getSCEV(I.getOperand(0)), OrigLoop);
4198 setDebugLocFromInst(Builder, &I);
4200 // The condition can be loop invariant but still defined inside the
4201 // loop. This means that we can't just use the original 'cond' value.
4202 // We have to take the 'vectorized' value and pick the first lane.
4203 // Instcombine will make this a no-op.
4204 VectorParts &Cond = getVectorValue(I.getOperand(0));
4205 VectorParts &Op0 = getVectorValue(I.getOperand(1));
4206 VectorParts &Op1 = getVectorValue(I.getOperand(2));
4211 : Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
4213 for (unsigned Part = 0; Part < UF; ++Part) {
4214 Entry[Part] = Builder.CreateSelect(
4215 InvariantCond ? ScalarCond : Cond[Part], Op0[Part], Op1[Part]);
4218 addMetadata(Entry, &I);
4222 case Instruction::ICmp:
4223 case Instruction::FCmp: {
4224 // Widen compares. Generate vector compares.
4225 bool FCmp = (I.getOpcode() == Instruction::FCmp);
4226 auto *Cmp = dyn_cast<CmpInst>(&I);
4227 setDebugLocFromInst(Builder, Cmp);
4228 VectorParts &A = getVectorValue(Cmp->getOperand(0));
4229 VectorParts &B = getVectorValue(Cmp->getOperand(1));
4230 for (unsigned Part = 0; Part < UF; ++Part) {
4233 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
4234 cast<FCmpInst>(C)->copyFastMathFlags(Cmp);
4236 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
4241 addMetadata(Entry, &I);
4245 case Instruction::Store:
4246 case Instruction::Load:
4247 vectorizeMemoryInstruction(&I);
4249 case Instruction::ZExt:
4250 case Instruction::SExt:
4251 case Instruction::FPToUI:
4252 case Instruction::FPToSI:
4253 case Instruction::FPExt:
4254 case Instruction::PtrToInt:
4255 case Instruction::IntToPtr:
4256 case Instruction::SIToFP:
4257 case Instruction::UIToFP:
4258 case Instruction::Trunc:
4259 case Instruction::FPTrunc:
4260 case Instruction::BitCast: {
4261 auto *CI = dyn_cast<CastInst>(&I);
4262 setDebugLocFromInst(Builder, CI);
4264 // Optimize the special case where the source is a constant integer
4265 // induction variable. Notice that we can only optimize the 'trunc' case
4266 // because (a) FP conversions lose precision, (b) sext/zext may wrap, and
4267 // (c) other casts depend on pointer size.
4268 auto ID = Legal->getInductionVars()->lookup(OldInduction);
4269 if (isa<TruncInst>(CI) && CI->getOperand(0) == OldInduction &&
4270 ID.getConstIntStepValue()) {
4271 widenIntInduction(OldInduction, Entry, cast<TruncInst>(CI));
4272 addMetadata(Entry, &I);
4276 /// Vectorize casts.
4278 (VF == 1) ? CI->getType() : VectorType::get(CI->getType(), VF);
4280 VectorParts &A = getVectorValue(CI->getOperand(0));
4281 for (unsigned Part = 0; Part < UF; ++Part)
4282 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
4283 addMetadata(Entry, &I);
4287 case Instruction::Call: {
4288 // Ignore dbg intrinsics.
4289 if (isa<DbgInfoIntrinsic>(I))
4291 setDebugLocFromInst(Builder, &I);
4293 Module *M = BB->getParent()->getParent();
4294 auto *CI = cast<CallInst>(&I);
4296 StringRef FnName = CI->getCalledFunction()->getName();
4297 Function *F = CI->getCalledFunction();
4298 Type *RetTy = ToVectorTy(CI->getType(), VF);
4299 SmallVector<Type *, 4> Tys;
4300 for (Value *ArgOperand : CI->arg_operands())
4301 Tys.push_back(ToVectorTy(ArgOperand->getType(), VF));
4303 Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
4304 if (ID && (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end ||
4305 ID == Intrinsic::lifetime_start)) {
4306 scalarizeInstruction(&I);
4309 // The flag shows whether we use Intrinsic or a usual Call for vectorized
4310 // version of the instruction.
4311 // Is it beneficial to perform intrinsic call compared to lib call?
4312 bool NeedToScalarize;
4313 unsigned CallCost = getVectorCallCost(CI, VF, *TTI, TLI, NeedToScalarize);
4314 bool UseVectorIntrinsic =
4315 ID && getVectorIntrinsicCost(CI, VF, *TTI, TLI) <= CallCost;
4316 if (!UseVectorIntrinsic && NeedToScalarize) {
4317 scalarizeInstruction(&I);
4321 for (unsigned Part = 0; Part < UF; ++Part) {
4322 SmallVector<Value *, 4> Args;
4323 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
4324 Value *Arg = CI->getArgOperand(i);
4325 // Some intrinsics have a scalar argument - don't replace it with a
4327 if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, i)) {
4328 VectorParts &VectorArg = getVectorValue(CI->getArgOperand(i));
4329 Arg = VectorArg[Part];
4331 Args.push_back(Arg);
4335 if (UseVectorIntrinsic) {
4336 // Use vector version of the intrinsic.
4337 Type *TysForDecl[] = {CI->getType()};
4339 TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF);
4340 VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl);
4342 // Use vector version of the library call.
4343 StringRef VFnName = TLI->getVectorizedFunction(FnName, VF);
4344 assert(!VFnName.empty() && "Vector function name is empty.");
4345 VectorF = M->getFunction(VFnName);
4347 // Generate a declaration
4348 FunctionType *FTy = FunctionType::get(RetTy, Tys, false);
4350 Function::Create(FTy, Function::ExternalLinkage, VFnName, M);
4351 VectorF->copyAttributesFrom(F);
4354 assert(VectorF && "Can't create vector function.");
4356 SmallVector<OperandBundleDef, 1> OpBundles;
4357 CI->getOperandBundlesAsDefs(OpBundles);
4358 CallInst *V = Builder.CreateCall(VectorF, Args, OpBundles);
4360 if (isa<FPMathOperator>(V))
4361 V->copyFastMathFlags(CI);
4366 addMetadata(Entry, &I);
4371 // All other instructions are unsupported. Scalarize them.
4372 scalarizeInstruction(&I);
4375 } // end of for_each instr.
4378 void InnerLoopVectorizer::updateAnalysis() {
4379 // Forget the original basic block.
4380 PSE.getSE()->forgetLoop(OrigLoop);
4382 // Update the dominator tree information.
4383 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
4384 "Entry does not dominate exit.");
4386 // We don't predicate stores by this point, so the vector body should be a
4388 DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
4390 DT->addNewBlock(LoopMiddleBlock, LoopVectorBody);
4391 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
4392 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
4393 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
4395 DEBUG(DT->verifyDomTree());
4398 /// \brief Check whether it is safe to if-convert this phi node.
4400 /// Phi nodes with constant expressions that can trap are not safe to if
4402 static bool canIfConvertPHINodes(BasicBlock *BB) {
4403 for (Instruction &I : *BB) {
4404 auto *Phi = dyn_cast<PHINode>(&I);
4407 for (Value *V : Phi->incoming_values())
4408 if (auto *C = dyn_cast<Constant>(V))
4415 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
4416 if (!EnableIfConversion) {
4417 emitAnalysis(VectorizationReport() << "if-conversion is disabled");
4421 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
4423 // A list of pointers that we can safely read and write to.
4424 SmallPtrSet<Value *, 8> SafePointes;
4426 // Collect safe addresses.
4427 for (BasicBlock *BB : TheLoop->blocks()) {
4428 if (blockNeedsPredication(BB))
4431 for (Instruction &I : *BB) {
4432 if (auto *LI = dyn_cast<LoadInst>(&I))
4433 SafePointes.insert(LI->getPointerOperand());
4434 else if (auto *SI = dyn_cast<StoreInst>(&I))
4435 SafePointes.insert(SI->getPointerOperand());
4439 // Collect the blocks that need predication.
4440 BasicBlock *Header = TheLoop->getHeader();
4441 for (BasicBlock *BB : TheLoop->blocks()) {
4442 // We don't support switch statements inside loops.
4443 if (!isa<BranchInst>(BB->getTerminator())) {
4444 emitAnalysis(VectorizationReport(BB->getTerminator())
4445 << "loop contains a switch statement");
4449 // We must be able to predicate all blocks that need to be predicated.
4450 if (blockNeedsPredication(BB)) {
4451 if (!blockCanBePredicated(BB, SafePointes)) {
4452 emitAnalysis(VectorizationReport(BB->getTerminator())
4453 << "control flow cannot be substituted for a select");
4456 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
4457 emitAnalysis(VectorizationReport(BB->getTerminator())
4458 << "control flow cannot be substituted for a select");
4463 // We can if-convert this loop.
4467 bool LoopVectorizationLegality::canVectorize() {
4468 // We must have a loop in canonical form. Loops with indirectbr in them cannot
4469 // be canonicalized.
4470 if (!TheLoop->getLoopPreheader()) {
4471 emitAnalysis(VectorizationReport()
4472 << "loop control flow is not understood by vectorizer");
4476 // FIXME: The code is currently dead, since the loop gets sent to
4477 // LoopVectorizationLegality is already an innermost loop.
4479 // We can only vectorize innermost loops.
4480 if (!TheLoop->empty()) {
4481 emitAnalysis(VectorizationReport() << "loop is not the innermost loop");
4485 // We must have a single backedge.
4486 if (TheLoop->getNumBackEdges() != 1) {
4487 emitAnalysis(VectorizationReport()
4488 << "loop control flow is not understood by vectorizer");
4492 // We must have a single exiting block.
4493 if (!TheLoop->getExitingBlock()) {
4494 emitAnalysis(VectorizationReport()
4495 << "loop control flow is not understood by vectorizer");
4499 // We only handle bottom-tested loops, i.e. loop in which the condition is
4500 // checked at the end of each iteration. With that we can assume that all
4501 // instructions in the loop are executed the same number of times.
4502 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
4503 emitAnalysis(VectorizationReport()
4504 << "loop control flow is not understood by vectorizer");
4508 // We need to have a loop header.
4509 DEBUG(dbgs() << "LV: Found a loop: " << TheLoop->getHeader()->getName()
4512 // Check if we can if-convert non-single-bb loops.
4513 unsigned NumBlocks = TheLoop->getNumBlocks();
4514 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
4515 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
4519 // ScalarEvolution needs to be able to find the exit count.
4520 const SCEV *ExitCount = PSE.getBackedgeTakenCount();
4521 if (ExitCount == PSE.getSE()->getCouldNotCompute()) {
4522 emitAnalysis(VectorizationReport()
4523 << "could not determine number of loop iterations");
4524 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
4528 // Check if we can vectorize the instructions and CFG in this loop.
4529 if (!canVectorizeInstrs()) {
4530 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
4534 // Go over each instruction and look at memory deps.
4535 if (!canVectorizeMemory()) {
4536 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
4540 // Collect all of the variables that remain uniform after vectorization.
4541 collectLoopUniforms();
4543 DEBUG(dbgs() << "LV: We can vectorize this loop"
4544 << (LAI->getRuntimePointerChecking()->Need
4545 ? " (with a runtime bound check)"
4549 bool UseInterleaved = TTI->enableInterleavedAccessVectorization();
4551 // If an override option has been passed in for interleaved accesses, use it.
4552 if (EnableInterleavedMemAccesses.getNumOccurrences() > 0)
4553 UseInterleaved = EnableInterleavedMemAccesses;
4555 // Analyze interleaved memory accesses.
4557 InterleaveInfo.analyzeInterleaving(*getSymbolicStrides());
4559 unsigned SCEVThreshold = VectorizeSCEVCheckThreshold;
4560 if (Hints->getForce() == LoopVectorizeHints::FK_Enabled)
4561 SCEVThreshold = PragmaVectorizeSCEVCheckThreshold;
4563 if (PSE.getUnionPredicate().getComplexity() > SCEVThreshold) {
4564 emitAnalysis(VectorizationReport()
4565 << "Too many SCEV assumptions need to be made and checked "
4567 DEBUG(dbgs() << "LV: Too many SCEV checks needed.\n");
4571 // Okay! We can vectorize. At this point we don't have any other mem analysis
4572 // which may limit our maximum vectorization factor, so just return true with
4577 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
4578 if (Ty->isPointerTy())
4579 return DL.getIntPtrType(Ty);
4581 // It is possible that char's or short's overflow when we ask for the loop's
4582 // trip count, work around this by changing the type size.
4583 if (Ty->getScalarSizeInBits() < 32)
4584 return Type::getInt32Ty(Ty->getContext());
4589 static Type *getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
4590 Ty0 = convertPointerToIntegerType(DL, Ty0);
4591 Ty1 = convertPointerToIntegerType(DL, Ty1);
4592 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
4597 /// \brief Check that the instruction has outside loop users and is not an
4598 /// identified reduction variable.
4599 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
4600 SmallPtrSetImpl<Value *> &AllowedExit) {
4601 // Reduction and Induction instructions are allowed to have exit users. All
4602 // other instructions must not have external users.
4603 if (!AllowedExit.count(Inst))
4604 // Check that all of the users of the loop are inside the BB.
4605 for (User *U : Inst->users()) {
4606 Instruction *UI = cast<Instruction>(U);
4607 // This user may be a reduction exit value.
4608 if (!TheLoop->contains(UI)) {
4609 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
4616 void LoopVectorizationLegality::addInductionPhi(
4617 PHINode *Phi, const InductionDescriptor &ID,
4618 SmallPtrSetImpl<Value *> &AllowedExit) {
4619 Inductions[Phi] = ID;
4620 Type *PhiTy = Phi->getType();
4621 const DataLayout &DL = Phi->getModule()->getDataLayout();
4623 // Get the widest type.
4625 WidestIndTy = convertPointerToIntegerType(DL, PhiTy);
4627 WidestIndTy = getWiderType(DL, PhiTy, WidestIndTy);
4629 // Int inductions are special because we only allow one IV.
4630 if (ID.getKind() == InductionDescriptor::IK_IntInduction &&
4631 ID.getConstIntStepValue() &&
4632 ID.getConstIntStepValue()->isOne() &&
4633 isa<Constant>(ID.getStartValue()) &&
4634 cast<Constant>(ID.getStartValue())->isNullValue()) {
4636 // Use the phi node with the widest type as induction. Use the last
4637 // one if there are multiple (no good reason for doing this other
4638 // than it is expedient). We've checked that it begins at zero and
4639 // steps by one, so this is a canonical induction variable.
4640 if (!Induction || PhiTy == WidestIndTy)
4644 // Both the PHI node itself, and the "post-increment" value feeding
4645 // back into the PHI node may have external users.
4646 AllowedExit.insert(Phi);
4647 AllowedExit.insert(Phi->getIncomingValueForBlock(TheLoop->getLoopLatch()));
4649 DEBUG(dbgs() << "LV: Found an induction variable.\n");
4653 bool LoopVectorizationLegality::canVectorizeInstrs() {
4654 BasicBlock *Header = TheLoop->getHeader();
4656 // Look for the attribute signaling the absence of NaNs.
4657 Function &F = *Header->getParent();
4659 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
4661 // For each block in the loop.
4662 for (BasicBlock *BB : TheLoop->blocks()) {
4663 // Scan the instructions in the block and look for hazards.
4664 for (Instruction &I : *BB) {
4665 if (auto *Phi = dyn_cast<PHINode>(&I)) {
4666 Type *PhiTy = Phi->getType();
4667 // Check that this PHI type is allowed.
4668 if (!PhiTy->isIntegerTy() && !PhiTy->isFloatingPointTy() &&
4669 !PhiTy->isPointerTy()) {
4670 emitAnalysis(VectorizationReport(Phi)
4671 << "loop control flow is not understood by vectorizer");
4672 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
4676 // If this PHINode is not in the header block, then we know that we
4677 // can convert it to select during if-conversion. No need to check if
4678 // the PHIs in this block are induction or reduction variables.
4680 // Check that this instruction has no outside users or is an
4681 // identified reduction value with an outside user.
4682 if (!hasOutsideLoopUser(TheLoop, Phi, AllowedExit))
4684 emitAnalysis(VectorizationReport(Phi)
4685 << "value could not be identified as "
4686 "an induction or reduction variable");
4690 // We only allow if-converted PHIs with exactly two incoming values.
4691 if (Phi->getNumIncomingValues() != 2) {
4692 emitAnalysis(VectorizationReport(Phi)
4693 << "control flow not understood by vectorizer");
4694 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
4698 RecurrenceDescriptor RedDes;
4699 if (RecurrenceDescriptor::isReductionPHI(Phi, TheLoop, RedDes)) {
4700 if (RedDes.hasUnsafeAlgebra())
4701 Requirements->addUnsafeAlgebraInst(RedDes.getUnsafeAlgebraInst());
4702 AllowedExit.insert(RedDes.getLoopExitInstr());
4703 Reductions[Phi] = RedDes;
4707 InductionDescriptor ID;
4708 if (InductionDescriptor::isInductionPHI(Phi, PSE, ID)) {
4709 addInductionPhi(Phi, ID, AllowedExit);
4713 if (RecurrenceDescriptor::isFirstOrderRecurrence(Phi, TheLoop, DT)) {
4714 FirstOrderRecurrences.insert(Phi);
4718 // As a last resort, coerce the PHI to a AddRec expression
4719 // and re-try classifying it a an induction PHI.
4720 if (InductionDescriptor::isInductionPHI(Phi, PSE, ID, true)) {
4721 addInductionPhi(Phi, ID, AllowedExit);
4725 emitAnalysis(VectorizationReport(Phi)
4726 << "value that could not be identified as "
4727 "reduction is used outside the loop");
4728 DEBUG(dbgs() << "LV: Found an unidentified PHI." << *Phi << "\n");
4730 } // end of PHI handling
4732 // We handle calls that:
4733 // * Are debug info intrinsics.
4734 // * Have a mapping to an IR intrinsic.
4735 // * Have a vector version available.
4736 auto *CI = dyn_cast<CallInst>(&I);
4737 if (CI && !getVectorIntrinsicIDForCall(CI, TLI) &&
4738 !isa<DbgInfoIntrinsic>(CI) &&
4739 !(CI->getCalledFunction() && TLI &&
4740 TLI->isFunctionVectorizable(CI->getCalledFunction()->getName()))) {
4741 emitAnalysis(VectorizationReport(CI)
4742 << "call instruction cannot be vectorized");
4743 DEBUG(dbgs() << "LV: Found a non-intrinsic, non-libfunc callsite.\n");
4747 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
4748 // second argument is the same (i.e. loop invariant)
4749 if (CI && hasVectorInstrinsicScalarOpd(
4750 getVectorIntrinsicIDForCall(CI, TLI), 1)) {
4751 auto *SE = PSE.getSE();
4752 if (!SE->isLoopInvariant(PSE.getSCEV(CI->getOperand(1)), TheLoop)) {
4753 emitAnalysis(VectorizationReport(CI)
4754 << "intrinsic instruction cannot be vectorized");
4755 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
4760 // Check that the instruction return type is vectorizable.
4761 // Also, we can't vectorize extractelement instructions.
4762 if ((!VectorType::isValidElementType(I.getType()) &&
4763 !I.getType()->isVoidTy()) ||
4764 isa<ExtractElementInst>(I)) {
4765 emitAnalysis(VectorizationReport(&I)
4766 << "instruction return type cannot be vectorized");
4767 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
4771 // Check that the stored type is vectorizable.
4772 if (auto *ST = dyn_cast<StoreInst>(&I)) {
4773 Type *T = ST->getValueOperand()->getType();
4774 if (!VectorType::isValidElementType(T)) {
4775 emitAnalysis(VectorizationReport(ST)
4776 << "store instruction cannot be vectorized");
4780 // FP instructions can allow unsafe algebra, thus vectorizable by
4781 // non-IEEE-754 compliant SIMD units.
4782 // This applies to floating-point math operations and calls, not memory
4783 // operations, shuffles, or casts, as they don't change precision or
4785 } else if (I.getType()->isFloatingPointTy() && (CI || I.isBinaryOp()) &&
4786 !I.hasUnsafeAlgebra()) {
4787 DEBUG(dbgs() << "LV: Found FP op with unsafe algebra.\n");
4788 Hints->setPotentiallyUnsafe();
4791 // Reduction instructions are allowed to have exit users.
4792 // All other instructions must not have external users.
4793 if (hasOutsideLoopUser(TheLoop, &I, AllowedExit)) {
4794 emitAnalysis(VectorizationReport(&I)
4795 << "value cannot be used outside the loop");
4803 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
4804 if (Inductions.empty()) {
4805 emitAnalysis(VectorizationReport()
4806 << "loop induction variable could not be identified");
4811 // Now we know the widest induction type, check if our found induction
4812 // is the same size. If it's not, unset it here and InnerLoopVectorizer
4813 // will create another.
4814 if (Induction && WidestIndTy != Induction->getType())
4815 Induction = nullptr;
4820 void LoopVectorizationLegality::collectLoopUniforms() {
4821 // We now know that the loop is vectorizable!
4822 // Collect variables that will remain uniform after vectorization.
4824 // If V is not an instruction inside the current loop, it is a Value
4825 // outside of the scope which we are interesting in.
4826 auto isOutOfScope = [&](Value *V) -> bool {
4827 Instruction *I = dyn_cast<Instruction>(V);
4828 return (!I || !TheLoop->contains(I));
4831 SetVector<Instruction *> Worklist;
4832 BasicBlock *Latch = TheLoop->getLoopLatch();
4833 // Start with the conditional branch.
4834 if (!isOutOfScope(Latch->getTerminator()->getOperand(0))) {
4835 Instruction *Cmp = cast<Instruction>(Latch->getTerminator()->getOperand(0));
4836 Worklist.insert(Cmp);
4837 DEBUG(dbgs() << "LV: Found uniform instruction: " << *Cmp << "\n");
4840 // Also add all consecutive pointer values; these values will be uniform
4841 // after vectorization (and subsequent cleanup).
4842 for (auto *BB : TheLoop->blocks()) {
4843 for (auto &I : *BB) {
4844 if (I.getType()->isPointerTy() && isConsecutivePtr(&I)) {
4845 Worklist.insert(&I);
4846 DEBUG(dbgs() << "LV: Found uniform instruction: " << I << "\n");
4851 // Expand Worklist in topological order: whenever a new instruction
4852 // is added , its users should be either already inside Worklist, or
4853 // out of scope. It ensures a uniform instruction will only be used
4854 // by uniform instructions or out of scope instructions.
4857 Instruction *I = Worklist[idx++];
4859 for (auto OV : I->operand_values()) {
4860 if (isOutOfScope(OV))
4862 auto *OI = cast<Instruction>(OV);
4863 if (all_of(OI->users(), [&](User *U) -> bool {
4864 return isOutOfScope(U) || Worklist.count(cast<Instruction>(U));
4866 Worklist.insert(OI);
4867 DEBUG(dbgs() << "LV: Found uniform instruction: " << *OI << "\n");
4870 } while (idx != Worklist.size());
4872 // For an instruction to be added into Worklist above, all its users inside
4873 // the current loop should be already added into Worklist. This condition
4874 // cannot be true for phi instructions which is always in a dependence loop.
4875 // Because any instruction in the dependence cycle always depends on others
4876 // in the cycle to be added into Worklist first, the result is no ones in
4877 // the cycle will be added into Worklist in the end.
4878 // That is why we process PHI separately.
4879 for (auto &Induction : *getInductionVars()) {
4880 auto *PN = Induction.first;
4881 auto *UpdateV = PN->getIncomingValueForBlock(TheLoop->getLoopLatch());
4882 if (all_of(PN->users(),
4883 [&](User *U) -> bool {
4884 return U == UpdateV || isOutOfScope(U) ||
4885 Worklist.count(cast<Instruction>(U));
4887 all_of(UpdateV->users(), [&](User *U) -> bool {
4888 return U == PN || isOutOfScope(U) ||
4889 Worklist.count(cast<Instruction>(U));
4891 Worklist.insert(cast<Instruction>(PN));
4892 Worklist.insert(cast<Instruction>(UpdateV));
4893 DEBUG(dbgs() << "LV: Found uniform instruction: " << *PN << "\n");
4894 DEBUG(dbgs() << "LV: Found uniform instruction: " << *UpdateV << "\n");
4898 Uniforms.insert(Worklist.begin(), Worklist.end());
4901 bool LoopVectorizationLegality::canVectorizeMemory() {
4902 LAI = &(*GetLAA)(*TheLoop);
4903 InterleaveInfo.setLAI(LAI);
4904 auto &OptionalReport = LAI->getReport();
4906 emitAnalysis(VectorizationReport(*OptionalReport));
4907 if (!LAI->canVectorizeMemory())
4910 if (LAI->hasStoreToLoopInvariantAddress()) {
4912 VectorizationReport()
4913 << "write to a loop invariant address could not be vectorized");
4914 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4918 Requirements->addRuntimePointerChecks(LAI->getNumRuntimePointerChecks());
4919 PSE.addPredicate(LAI->getPSE().getUnionPredicate());
4924 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4925 Value *In0 = const_cast<Value *>(V);
4926 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4930 return Inductions.count(PN);
4933 bool LoopVectorizationLegality::isFirstOrderRecurrence(const PHINode *Phi) {
4934 return FirstOrderRecurrences.count(Phi);
4937 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4938 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4941 bool LoopVectorizationLegality::blockCanBePredicated(
4942 BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs) {
4943 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
4945 for (Instruction &I : *BB) {
4946 // Check that we don't have a constant expression that can trap as operand.
4947 for (Value *Operand : I.operands()) {
4948 if (auto *C = dyn_cast<Constant>(Operand))
4952 // We might be able to hoist the load.
4953 if (I.mayReadFromMemory()) {
4954 auto *LI = dyn_cast<LoadInst>(&I);
4957 if (!SafePtrs.count(LI->getPointerOperand())) {
4958 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand()) ||
4959 isLegalMaskedGather(LI->getType())) {
4960 MaskedOp.insert(LI);
4963 // !llvm.mem.parallel_loop_access implies if-conversion safety.
4964 if (IsAnnotatedParallel)
4970 // We don't predicate stores at the moment.
4971 if (I.mayWriteToMemory()) {
4972 auto *SI = dyn_cast<StoreInst>(&I);
4973 // We only support predication of stores in basic blocks with one
4978 // Build a masked store if it is legal for the target.
4979 if (isLegalMaskedStore(SI->getValueOperand()->getType(),
4980 SI->getPointerOperand()) ||
4981 isLegalMaskedScatter(SI->getValueOperand()->getType())) {
4982 MaskedOp.insert(SI);
4986 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
4987 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
4989 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
4990 !isSinglePredecessor)
4996 // The instructions below can trap.
4997 switch (I.getOpcode()) {
5000 case Instruction::UDiv:
5001 case Instruction::SDiv:
5002 case Instruction::URem:
5003 case Instruction::SRem:
5011 void InterleavedAccessInfo::collectConstStrideAccesses(
5012 MapVector<Instruction *, StrideDescriptor> &AccessStrideInfo,
5013 const ValueToValueMap &Strides) {
5015 auto &DL = TheLoop->getHeader()->getModule()->getDataLayout();
5017 // Since it's desired that the load/store instructions be maintained in
5018 // "program order" for the interleaved access analysis, we have to visit the
5019 // blocks in the loop in reverse postorder (i.e., in a topological order).
5020 // Such an ordering will ensure that any load/store that may be executed
5021 // before a second load/store will precede the second load/store in
5022 // AccessStrideInfo.
5023 LoopBlocksDFS DFS(TheLoop);
5025 for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO()))
5026 for (auto &I : *BB) {
5027 auto *LI = dyn_cast<LoadInst>(&I);
5028 auto *SI = dyn_cast<StoreInst>(&I);
5032 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
5033 int64_t Stride = getPtrStride(PSE, Ptr, TheLoop, Strides);
5035 const SCEV *Scev = replaceSymbolicStrideSCEV(PSE, Strides, Ptr);
5036 PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
5037 uint64_t Size = DL.getTypeAllocSize(PtrTy->getElementType());
5039 // An alignment of 0 means target ABI alignment.
5040 unsigned Align = LI ? LI->getAlignment() : SI->getAlignment();
5042 Align = DL.getABITypeAlignment(PtrTy->getElementType());
5044 AccessStrideInfo[&I] = StrideDescriptor(Stride, Scev, Size, Align);
5048 // Analyze interleaved accesses and collect them into interleaved load and
5051 // When generating code for an interleaved load group, we effectively hoist all
5052 // loads in the group to the location of the first load in program order. When
5053 // generating code for an interleaved store group, we sink all stores to the
5054 // location of the last store. This code motion can change the order of load
5055 // and store instructions and may break dependences.
5057 // The code generation strategy mentioned above ensures that we won't violate
5058 // any write-after-read (WAR) dependences.
5060 // E.g., for the WAR dependence: a = A[i]; // (1)
5063 // The store group of (2) is always inserted at or below (2), and the load
5064 // group of (1) is always inserted at or above (1). Thus, the instructions will
5065 // never be reordered. All other dependences are checked to ensure the
5066 // correctness of the instruction reordering.
5068 // The algorithm visits all memory accesses in the loop in bottom-up program
5069 // order. Program order is established by traversing the blocks in the loop in
5070 // reverse postorder when collecting the accesses.
5072 // We visit the memory accesses in bottom-up order because it can simplify the
5073 // construction of store groups in the presence of write-after-write (WAW)
5076 // E.g., for the WAW dependence: A[i] = a; // (1)
5078 // A[i + 1] = c; // (3)
5080 // We will first create a store group with (3) and (2). (1) can't be added to
5081 // this group because it and (2) are dependent. However, (1) can be grouped
5082 // with other accesses that may precede it in program order. Note that a
5083 // bottom-up order does not imply that WAW dependences should not be checked.
5084 void InterleavedAccessInfo::analyzeInterleaving(
5085 const ValueToValueMap &Strides) {
5086 DEBUG(dbgs() << "LV: Analyzing interleaved accesses...\n");
5088 // Holds all accesses with a constant stride.
5089 MapVector<Instruction *, StrideDescriptor> AccessStrideInfo;
5090 collectConstStrideAccesses(AccessStrideInfo, Strides);
5092 if (AccessStrideInfo.empty())
5095 // Collect the dependences in the loop.
5096 collectDependences();
5098 // Holds all interleaved store groups temporarily.
5099 SmallSetVector<InterleaveGroup *, 4> StoreGroups;
5100 // Holds all interleaved load groups temporarily.
5101 SmallSetVector<InterleaveGroup *, 4> LoadGroups;
5103 // Search in bottom-up program order for pairs of accesses (A and B) that can
5104 // form interleaved load or store groups. In the algorithm below, access A
5105 // precedes access B in program order. We initialize a group for B in the
5106 // outer loop of the algorithm, and then in the inner loop, we attempt to
5107 // insert each A into B's group if:
5109 // 1. A and B have the same stride,
5110 // 2. A and B have the same memory object size, and
5111 // 3. A belongs in B's group according to its distance from B.
5113 // Special care is taken to ensure group formation will not break any
5115 for (auto BI = AccessStrideInfo.rbegin(), E = AccessStrideInfo.rend();
5117 Instruction *B = BI->first;
5118 StrideDescriptor DesB = BI->second;
5120 // Initialize a group for B if it has an allowable stride. Even if we don't
5121 // create a group for B, we continue with the bottom-up algorithm to ensure
5122 // we don't break any of B's dependences.
5123 InterleaveGroup *Group = nullptr;
5124 if (isStrided(DesB.Stride)) {
5125 Group = getInterleaveGroup(B);
5127 DEBUG(dbgs() << "LV: Creating an interleave group with:" << *B << '\n');
5128 Group = createInterleaveGroup(B, DesB.Stride, DesB.Align);
5130 if (B->mayWriteToMemory())
5131 StoreGroups.insert(Group);
5133 LoadGroups.insert(Group);
5136 for (auto AI = std::next(BI); AI != E; ++AI) {
5137 Instruction *A = AI->first;
5138 StrideDescriptor DesA = AI->second;
5140 // Our code motion strategy implies that we can't have dependences
5141 // between accesses in an interleaved group and other accesses located
5142 // between the first and last member of the group. Note that this also
5143 // means that a group can't have more than one member at a given offset.
5144 // The accesses in a group can have dependences with other accesses, but
5145 // we must ensure we don't extend the boundaries of the group such that
5146 // we encompass those dependent accesses.
5148 // For example, assume we have the sequence of accesses shown below in a
5151 // (1, 2) is a group | A[i] = a; // (1)
5152 // | A[i-1] = b; // (2) |
5153 // A[i-3] = c; // (3)
5154 // A[i] = d; // (4) | (2, 4) is not a group
5156 // Because accesses (2) and (3) are dependent, we can group (2) with (1)
5157 // but not with (4). If we did, the dependent access (3) would be within
5158 // the boundaries of the (2, 4) group.
5159 if (!canReorderMemAccessesForInterleavedGroups(&*AI, &*BI)) {
5161 // If a dependence exists and A is already in a group, we know that A
5162 // must be a store since A precedes B and WAR dependences are allowed.
5163 // Thus, A would be sunk below B. We release A's group to prevent this
5164 // illegal code motion. A will then be free to form another group with
5165 // instructions that precede it.
5166 if (isInterleaved(A)) {
5167 InterleaveGroup *StoreGroup = getInterleaveGroup(A);
5168 StoreGroups.remove(StoreGroup);
5169 releaseGroup(StoreGroup);
5172 // If a dependence exists and A is not already in a group (or it was
5173 // and we just released it), B might be hoisted above A (if B is a
5174 // load) or another store might be sunk below A (if B is a store). In
5175 // either case, we can't add additional instructions to B's group. B
5176 // will only form a group with instructions that it precedes.
5180 // At this point, we've checked for illegal code motion. If either A or B
5181 // isn't strided, there's nothing left to do.
5182 if (!isStrided(DesA.Stride) || !isStrided(DesB.Stride))
5185 // Ignore A if it's already in a group or isn't the same kind of memory
5187 if (isInterleaved(A) || A->mayReadFromMemory() != B->mayReadFromMemory())
5190 // Check rules 1 and 2. Ignore A if its stride or size is different from
5192 if (DesA.Stride != DesB.Stride || DesA.Size != DesB.Size)
5195 // Calculate the distance from A to B.
5196 const SCEVConstant *DistToB = dyn_cast<SCEVConstant>(
5197 PSE.getSE()->getMinusSCEV(DesA.Scev, DesB.Scev));
5200 int64_t DistanceToB = DistToB->getAPInt().getSExtValue();
5202 // Check rule 3. Ignore A if its distance to B is not a multiple of the
5204 if (DistanceToB % static_cast<int64_t>(DesB.Size))
5207 // Ignore A if either A or B is in a predicated block. Although we
5208 // currently prevent group formation for predicated accesses, we may be
5209 // able to relax this limitation in the future once we handle more
5210 // complicated blocks.
5211 if (isPredicated(A->getParent()) || isPredicated(B->getParent()))
5214 // The index of A is the index of B plus A's distance to B in multiples
5217 Group->getIndex(B) + DistanceToB / static_cast<int64_t>(DesB.Size);
5219 // Try to insert A into B's group.
5220 if (Group->insertMember(A, IndexA, DesA.Align)) {
5221 DEBUG(dbgs() << "LV: Inserted:" << *A << '\n'
5222 << " into the interleave group with" << *B << '\n');
5223 InterleaveGroupMap[A] = Group;
5225 // Set the first load in program order as the insert position.
5226 if (A->mayReadFromMemory())
5227 Group->setInsertPos(A);
5229 } // Iteration over A accesses.
5230 } // Iteration over B accesses.
5232 // Remove interleaved store groups with gaps.
5233 for (InterleaveGroup *Group : StoreGroups)
5234 if (Group->getNumMembers() != Group->getFactor())
5235 releaseGroup(Group);
5237 // If there is a non-reversed interleaved load group with gaps, we will need
5238 // to execute at least one scalar epilogue iteration. This will ensure that
5239 // we don't speculatively access memory out-of-bounds. Note that we only need
5240 // to look for a member at index factor - 1, since every group must have a
5241 // member at index zero.
5242 for (InterleaveGroup *Group : LoadGroups)
5243 if (!Group->getMember(Group->getFactor() - 1)) {
5244 if (Group->isReverse()) {
5245 releaseGroup(Group);
5247 DEBUG(dbgs() << "LV: Interleaved group requires epilogue iteration.\n");
5248 RequiresScalarEpilogue = true;
5253 LoopVectorizationCostModel::VectorizationFactor
5254 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
5255 // Width 1 means no vectorize
5256 VectorizationFactor Factor = {1U, 0U};
5257 if (OptForSize && Legal->getRuntimePointerChecking()->Need) {
5259 VectorizationReport()
5260 << "runtime pointer checks needed. Enable vectorization of this "
5261 "loop with '#pragma clang loop vectorize(enable)' when "
5262 "compiling with -Os/-Oz");
5264 << "LV: Aborting. Runtime ptr check is required with -Os/-Oz.\n");
5268 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
5270 VectorizationReport()
5271 << "store that is conditionally executed prevents vectorization");
5272 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
5276 // Find the trip count.
5277 unsigned TC = PSE.getSE()->getSmallConstantTripCount(TheLoop);
5278 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
5280 MinBWs = computeMinimumValueSizes(TheLoop->getBlocks(), *DB, &TTI);
5281 unsigned SmallestType, WidestType;
5282 std::tie(SmallestType, WidestType) = getSmallestAndWidestTypes();
5283 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
5284 unsigned MaxSafeDepDist = -1U;
5286 // Get the maximum safe dependence distance in bits computed by LAA. If the
5287 // loop contains any interleaved accesses, we divide the dependence distance
5288 // by the maximum interleave factor of all interleaved groups. Note that
5289 // although the division ensures correctness, this is a fairly conservative
5290 // computation because the maximum distance computed by LAA may not involve
5291 // any of the interleaved accesses.
5292 if (Legal->getMaxSafeDepDistBytes() != -1U)
5294 Legal->getMaxSafeDepDistBytes() * 8 / Legal->getMaxInterleaveFactor();
5297 ((WidestRegister < MaxSafeDepDist) ? WidestRegister : MaxSafeDepDist);
5298 unsigned MaxVectorSize = WidestRegister / WidestType;
5300 DEBUG(dbgs() << "LV: The Smallest and Widest types: " << SmallestType << " / "
5301 << WidestType << " bits.\n");
5302 DEBUG(dbgs() << "LV: The Widest register is: " << WidestRegister
5305 if (MaxVectorSize == 0) {
5306 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
5310 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
5311 " into one vector!");
5313 unsigned VF = MaxVectorSize;
5314 if (MaximizeBandwidth && !OptForSize) {
5315 // Collect all viable vectorization factors.
5316 SmallVector<unsigned, 8> VFs;
5317 unsigned NewMaxVectorSize = WidestRegister / SmallestType;
5318 for (unsigned VS = MaxVectorSize; VS <= NewMaxVectorSize; VS *= 2)
5321 // For each VF calculate its register usage.
5322 auto RUs = calculateRegisterUsage(VFs);
5324 // Select the largest VF which doesn't require more registers than existing
5326 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(true);
5327 for (int i = RUs.size() - 1; i >= 0; --i) {
5328 if (RUs[i].MaxLocalUsers <= TargetNumRegisters) {
5335 // If we optimize the program for size, avoid creating the tail loop.
5337 // If we are unable to calculate the trip count then don't try to vectorize.
5340 VectorizationReport()
5341 << "unable to calculate the loop count due to complex control flow");
5342 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n");
5346 // Find the maximum SIMD width that can fit within the trip count.
5347 VF = TC % MaxVectorSize;
5352 // If the trip count that we found modulo the vectorization factor is not
5353 // zero then we require a tail.
5354 emitAnalysis(VectorizationReport()
5355 << "cannot optimize for size and vectorize at the "
5356 "same time. Enable vectorization of this loop "
5357 "with '#pragma clang loop vectorize(enable)' "
5358 "when compiling with -Os/-Oz");
5359 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n");
5364 int UserVF = Hints->getWidth();
5366 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
5367 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
5369 Factor.Width = UserVF;
5373 float Cost = expectedCost(1).first;
5375 const float ScalarCost = Cost;
5378 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
5380 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
5381 // Ignore scalar width, because the user explicitly wants vectorization.
5382 if (ForceVectorization && VF > 1) {
5384 Cost = expectedCost(Width).first / (float)Width;
5387 for (unsigned i = 2; i <= VF; i *= 2) {
5388 // Notice that the vector loop needs to be executed less times, so
5389 // we need to divide the cost of the vector loops by the width of
5390 // the vector elements.
5391 VectorizationCostTy C = expectedCost(i);
5392 float VectorCost = C.first / (float)i;
5393 DEBUG(dbgs() << "LV: Vector loop of width " << i
5394 << " costs: " << (int)VectorCost << ".\n");
5395 if (!C.second && !ForceVectorization) {
5397 dbgs() << "LV: Not considering vector loop of width " << i
5398 << " because it will not generate any vector instructions.\n");
5401 if (VectorCost < Cost) {
5407 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
5408 << "LV: Vectorization seems to be not beneficial, "
5409 << "but was forced by a user.\n");
5410 DEBUG(dbgs() << "LV: Selecting VF: " << Width << ".\n");
5411 Factor.Width = Width;
5412 Factor.Cost = Width * Cost;
5416 std::pair<unsigned, unsigned>
5417 LoopVectorizationCostModel::getSmallestAndWidestTypes() {
5418 unsigned MinWidth = -1U;
5419 unsigned MaxWidth = 8;
5420 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
5423 for (BasicBlock *BB : TheLoop->blocks()) {
5424 // For each instruction in the loop.
5425 for (Instruction &I : *BB) {
5426 Type *T = I.getType();
5428 // Skip ignored values.
5429 if (ValuesToIgnore.count(&I))
5432 // Only examine Loads, Stores and PHINodes.
5433 if (!isa<LoadInst>(I) && !isa<StoreInst>(I) && !isa<PHINode>(I))
5436 // Examine PHI nodes that are reduction variables. Update the type to
5437 // account for the recurrence type.
5438 if (auto *PN = dyn_cast<PHINode>(&I)) {
5439 if (!Legal->isReductionVariable(PN))
5441 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[PN];
5442 T = RdxDesc.getRecurrenceType();
5445 // Examine the stored values.
5446 if (auto *ST = dyn_cast<StoreInst>(&I))
5447 T = ST->getValueOperand()->getType();
5449 // Ignore loaded pointer types and stored pointer types that are not
5450 // consecutive. However, we do want to take consecutive stores/loads of
5451 // pointer vectors into account.
5452 if (T->isPointerTy() && !isConsecutiveLoadOrStore(&I))
5455 MinWidth = std::min(MinWidth,
5456 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
5457 MaxWidth = std::max(MaxWidth,
5458 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
5462 return {MinWidth, MaxWidth};
5465 unsigned LoopVectorizationCostModel::selectInterleaveCount(bool OptForSize,
5467 unsigned LoopCost) {
5469 // -- The interleave heuristics --
5470 // We interleave the loop in order to expose ILP and reduce the loop overhead.
5471 // There are many micro-architectural considerations that we can't predict
5472 // at this level. For example, frontend pressure (on decode or fetch) due to
5473 // code size, or the number and capabilities of the execution ports.
5475 // We use the following heuristics to select the interleave count:
5476 // 1. If the code has reductions, then we interleave to break the cross
5477 // iteration dependency.
5478 // 2. If the loop is really small, then we interleave to reduce the loop
5480 // 3. We don't interleave if we think that we will spill registers to memory
5481 // due to the increased register pressure.
5483 // When we optimize for size, we don't interleave.
5487 // We used the distance for the interleave count.
5488 if (Legal->getMaxSafeDepDistBytes() != -1U)
5491 // Do not interleave loops with a relatively small trip count.
5492 unsigned TC = PSE.getSE()->getSmallConstantTripCount(TheLoop);
5493 if (TC > 1 && TC < TinyTripCountInterleaveThreshold)
5496 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
5497 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters
5501 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
5502 TargetNumRegisters = ForceTargetNumScalarRegs;
5504 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
5505 TargetNumRegisters = ForceTargetNumVectorRegs;
5508 RegisterUsage R = calculateRegisterUsage({VF})[0];
5509 // We divide by these constants so assume that we have at least one
5510 // instruction that uses at least one register.
5511 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
5512 R.NumInstructions = std::max(R.NumInstructions, 1U);
5514 // We calculate the interleave count using the following formula.
5515 // Subtract the number of loop invariants from the number of available
5516 // registers. These registers are used by all of the interleaved instances.
5517 // Next, divide the remaining registers by the number of registers that is
5518 // required by the loop, in order to estimate how many parallel instances
5519 // fit without causing spills. All of this is rounded down if necessary to be
5520 // a power of two. We want power of two interleave count to simplify any
5521 // addressing operations or alignment considerations.
5522 unsigned IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
5525 // Don't count the induction variable as interleaved.
5526 if (EnableIndVarRegisterHeur)
5527 IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
5528 std::max(1U, (R.MaxLocalUsers - 1)));
5530 // Clamp the interleave ranges to reasonable counts.
5531 unsigned MaxInterleaveCount = TTI.getMaxInterleaveFactor(VF);
5533 // Check if the user has overridden the max.
5535 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
5536 MaxInterleaveCount = ForceTargetMaxScalarInterleaveFactor;
5538 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
5539 MaxInterleaveCount = ForceTargetMaxVectorInterleaveFactor;
5542 // If we did not calculate the cost for VF (because the user selected the VF)
5543 // then we calculate the cost of VF here.
5545 LoopCost = expectedCost(VF).first;
5547 // Clamp the calculated IC to be between the 1 and the max interleave count
5548 // that the target allows.
5549 if (IC > MaxInterleaveCount)
5550 IC = MaxInterleaveCount;
5554 // Interleave if we vectorized this loop and there is a reduction that could
5555 // benefit from interleaving.
5556 if (VF > 1 && Legal->getReductionVars()->size()) {
5557 DEBUG(dbgs() << "LV: Interleaving because of reductions.\n");
5561 // Note that if we've already vectorized the loop we will have done the
5562 // runtime check and so interleaving won't require further checks.
5563 bool InterleavingRequiresRuntimePointerCheck =
5564 (VF == 1 && Legal->getRuntimePointerChecking()->Need);
5566 // We want to interleave small loops in order to reduce the loop overhead and
5567 // potentially expose ILP opportunities.
5568 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
5569 if (!InterleavingRequiresRuntimePointerCheck && LoopCost < SmallLoopCost) {
5570 // We assume that the cost overhead is 1 and we use the cost model
5571 // to estimate the cost of the loop and interleave until the cost of the
5572 // loop overhead is about 5% of the cost of the loop.
5574 std::min(IC, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
5576 // Interleave until store/load ports (estimated by max interleave count) are
5578 unsigned NumStores = Legal->getNumStores();
5579 unsigned NumLoads = Legal->getNumLoads();
5580 unsigned StoresIC = IC / (NumStores ? NumStores : 1);
5581 unsigned LoadsIC = IC / (NumLoads ? NumLoads : 1);
5583 // If we have a scalar reduction (vector reductions are already dealt with
5584 // by this point), we can increase the critical path length if the loop
5585 // we're interleaving is inside another loop. Limit, by default to 2, so the
5586 // critical path only gets increased by one reduction operation.
5587 if (Legal->getReductionVars()->size() && TheLoop->getLoopDepth() > 1) {
5588 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionIC);
5589 SmallIC = std::min(SmallIC, F);
5590 StoresIC = std::min(StoresIC, F);
5591 LoadsIC = std::min(LoadsIC, F);
5594 if (EnableLoadStoreRuntimeInterleave &&
5595 std::max(StoresIC, LoadsIC) > SmallIC) {
5596 DEBUG(dbgs() << "LV: Interleaving to saturate store or load ports.\n");
5597 return std::max(StoresIC, LoadsIC);
5600 DEBUG(dbgs() << "LV: Interleaving to reduce branch cost.\n");
5604 // Interleave if this is a large loop (small loops are already dealt with by
5605 // this point) that could benefit from interleaving.
5606 bool HasReductions = (Legal->getReductionVars()->size() > 0);
5607 if (TTI.enableAggressiveInterleaving(HasReductions)) {
5608 DEBUG(dbgs() << "LV: Interleaving to expose ILP.\n");
5612 DEBUG(dbgs() << "LV: Not Interleaving.\n");
5616 SmallVector<LoopVectorizationCostModel::RegisterUsage, 8>
5617 LoopVectorizationCostModel::calculateRegisterUsage(ArrayRef<unsigned> VFs) {
5618 // This function calculates the register usage by measuring the highest number
5619 // of values that are alive at a single location. Obviously, this is a very
5620 // rough estimation. We scan the loop in a topological order in order and
5621 // assign a number to each instruction. We use RPO to ensure that defs are
5622 // met before their users. We assume that each instruction that has in-loop
5623 // users starts an interval. We record every time that an in-loop value is
5624 // used, so we have a list of the first and last occurrences of each
5625 // instruction. Next, we transpose this data structure into a multi map that
5626 // holds the list of intervals that *end* at a specific location. This multi
5627 // map allows us to perform a linear search. We scan the instructions linearly
5628 // and record each time that a new interval starts, by placing it in a set.
5629 // If we find this value in the multi-map then we remove it from the set.
5630 // The max register usage is the maximum size of the set.
5631 // We also search for instructions that are defined outside the loop, but are
5632 // used inside the loop. We need this number separately from the max-interval
5633 // usage number because when we unroll, loop-invariant values do not take
5635 LoopBlocksDFS DFS(TheLoop);
5639 RU.NumInstructions = 0;
5641 // Each 'key' in the map opens a new interval. The values
5642 // of the map are the index of the 'last seen' usage of the
5643 // instruction that is the key.
5644 typedef DenseMap<Instruction *, unsigned> IntervalMap;
5645 // Maps instruction to its index.
5646 DenseMap<unsigned, Instruction *> IdxToInstr;
5647 // Marks the end of each interval.
5648 IntervalMap EndPoint;
5649 // Saves the list of instruction indices that are used in the loop.
5650 SmallSet<Instruction *, 8> Ends;
5651 // Saves the list of values that are used in the loop but are
5652 // defined outside the loop, such as arguments and constants.
5653 SmallPtrSet<Value *, 8> LoopInvariants;
5656 for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO())) {
5657 RU.NumInstructions += BB->size();
5658 for (Instruction &I : *BB) {
5659 IdxToInstr[Index++] = &I;
5661 // Save the end location of each USE.
5662 for (Value *U : I.operands()) {
5663 auto *Instr = dyn_cast<Instruction>(U);
5665 // Ignore non-instruction values such as arguments, constants, etc.
5669 // If this instruction is outside the loop then record it and continue.
5670 if (!TheLoop->contains(Instr)) {
5671 LoopInvariants.insert(Instr);
5675 // Overwrite previous end points.
5676 EndPoint[Instr] = Index;
5682 // Saves the list of intervals that end with the index in 'key'.
5683 typedef SmallVector<Instruction *, 2> InstrList;
5684 DenseMap<unsigned, InstrList> TransposeEnds;
5686 // Transpose the EndPoints to a list of values that end at each index.
5687 for (auto &Interval : EndPoint)
5688 TransposeEnds[Interval.second].push_back(Interval.first);
5690 SmallSet<Instruction *, 8> OpenIntervals;
5692 // Get the size of the widest register.
5693 unsigned MaxSafeDepDist = -1U;
5694 if (Legal->getMaxSafeDepDistBytes() != -1U)
5695 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
5696 unsigned WidestRegister =
5697 std::min(TTI.getRegisterBitWidth(true), MaxSafeDepDist);
5698 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
5700 SmallVector<RegisterUsage, 8> RUs(VFs.size());
5701 SmallVector<unsigned, 8> MaxUsages(VFs.size(), 0);
5703 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5705 // A lambda that gets the register usage for the given type and VF.
5706 auto GetRegUsage = [&DL, WidestRegister](Type *Ty, unsigned VF) {
5707 if (Ty->isTokenTy())
5709 unsigned TypeSize = DL.getTypeSizeInBits(Ty->getScalarType());
5710 return std::max<unsigned>(1, VF * TypeSize / WidestRegister);
5713 for (unsigned int i = 0; i < Index; ++i) {
5714 Instruction *I = IdxToInstr[i];
5715 // Ignore instructions that are never used within the loop.
5719 // Remove all of the instructions that end at this location.
5720 InstrList &List = TransposeEnds[i];
5721 for (Instruction *ToRemove : List)
5722 OpenIntervals.erase(ToRemove);
5724 // Skip ignored values.
5725 if (ValuesToIgnore.count(I))
5728 // For each VF find the maximum usage of registers.
5729 for (unsigned j = 0, e = VFs.size(); j < e; ++j) {
5731 MaxUsages[j] = std::max(MaxUsages[j], OpenIntervals.size());
5735 // Count the number of live intervals.
5736 unsigned RegUsage = 0;
5737 for (auto Inst : OpenIntervals) {
5738 // Skip ignored values for VF > 1.
5739 if (VecValuesToIgnore.count(Inst))
5741 RegUsage += GetRegUsage(Inst->getType(), VFs[j]);
5743 MaxUsages[j] = std::max(MaxUsages[j], RegUsage);
5746 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # "
5747 << OpenIntervals.size() << '\n');
5749 // Add the current instruction to the list of open intervals.
5750 OpenIntervals.insert(I);
5753 for (unsigned i = 0, e = VFs.size(); i < e; ++i) {
5754 unsigned Invariant = 0;
5756 Invariant = LoopInvariants.size();
5758 for (auto Inst : LoopInvariants)
5759 Invariant += GetRegUsage(Inst->getType(), VFs[i]);
5762 DEBUG(dbgs() << "LV(REG): VF = " << VFs[i] << '\n');
5763 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsages[i] << '\n');
5764 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5765 DEBUG(dbgs() << "LV(REG): LoopSize: " << RU.NumInstructions << '\n');
5767 RU.LoopInvariantRegs = Invariant;
5768 RU.MaxLocalUsers = MaxUsages[i];
5775 LoopVectorizationCostModel::VectorizationCostTy
5776 LoopVectorizationCostModel::expectedCost(unsigned VF) {
5777 VectorizationCostTy Cost;
5780 for (BasicBlock *BB : TheLoop->blocks()) {
5781 VectorizationCostTy BlockCost;
5783 // For each instruction in the old loop.
5784 for (Instruction &I : *BB) {
5785 // Skip dbg intrinsics.
5786 if (isa<DbgInfoIntrinsic>(I))
5789 // Skip ignored values.
5790 if (ValuesToIgnore.count(&I))
5793 VectorizationCostTy C = getInstructionCost(&I, VF);
5795 // Check if we should override the cost.
5796 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5797 C.first = ForceTargetInstructionCost;
5799 BlockCost.first += C.first;
5800 BlockCost.second |= C.second;
5801 DEBUG(dbgs() << "LV: Found an estimated cost of " << C.first << " for VF "
5802 << VF << " For instruction: " << I << '\n');
5805 // We assume that if-converted blocks have a 50% chance of being executed.
5806 // When the code is scalar then some of the blocks are avoided due to CF.
5807 // When the code is vectorized we execute all code paths.
5808 if (VF == 1 && Legal->blockNeedsPredication(BB))
5809 BlockCost.first /= 2;
5811 Cost.first += BlockCost.first;
5812 Cost.second |= BlockCost.second;
5818 /// \brief Check if the load/store instruction \p I may be translated into
5819 /// gather/scatter during vectorization.
5821 /// Pointer \p Ptr specifies address in memory for the given scalar memory
5822 /// instruction. We need it to retrieve data type.
5823 /// Using gather/scatter is possible when it is supported by target.
5824 static bool isGatherOrScatterLegal(Instruction *I, Value *Ptr,
5825 LoopVectorizationLegality *Legal) {
5826 auto *DataTy = cast<PointerType>(Ptr->getType())->getElementType();
5827 return (isa<LoadInst>(I) && Legal->isLegalMaskedGather(DataTy)) ||
5828 (isa<StoreInst>(I) && Legal->isLegalMaskedScatter(DataTy));
5831 /// \brief Check whether the address computation for a non-consecutive memory
5832 /// access looks like an unlikely candidate for being merged into the indexing
5835 /// We look for a GEP which has one index that is an induction variable and all
5836 /// other indices are loop invariant. If the stride of this access is also
5837 /// within a small bound we decide that this address computation can likely be
5838 /// merged into the addressing mode.
5839 /// In all other cases, we identify the address computation as complex.
5840 static bool isLikelyComplexAddressComputation(Value *Ptr,
5841 LoopVectorizationLegality *Legal,
5842 ScalarEvolution *SE,
5843 const Loop *TheLoop) {
5844 auto *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5848 // We are looking for a gep with all loop invariant indices except for one
5849 // which should be an induction variable.
5850 unsigned NumOperands = Gep->getNumOperands();
5851 for (unsigned i = 1; i < NumOperands; ++i) {
5852 Value *Opd = Gep->getOperand(i);
5853 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5854 !Legal->isInductionVariable(Opd))
5858 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5859 // can likely be merged into the address computation.
5860 unsigned MaxMergeDistance = 64;
5862 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5866 // Check the step is constant.
5867 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5868 // Calculate the pointer stride and check if it is consecutive.
5869 const auto *C = dyn_cast<SCEVConstant>(Step);
5873 const APInt &APStepVal = C->getAPInt();
5875 // Huge step value - give up.
5876 if (APStepVal.getBitWidth() > 64)
5879 int64_t StepVal = APStepVal.getSExtValue();
5881 return StepVal > MaxMergeDistance;
5884 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5885 return Legal->hasStride(I->getOperand(0)) ||
5886 Legal->hasStride(I->getOperand(1));
5889 LoopVectorizationCostModel::VectorizationCostTy
5890 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5891 // If we know that this instruction will remain uniform, check the cost of
5892 // the scalar version.
5893 if (Legal->isUniformAfterVectorization(I))
5897 unsigned C = getInstructionCost(I, VF, VectorTy);
5899 bool TypeNotScalarized =
5900 VF > 1 && !VectorTy->isVoidTy() && TTI.getNumberOfParts(VectorTy) < VF;
5901 return VectorizationCostTy(C, TypeNotScalarized);
5904 unsigned LoopVectorizationCostModel::getInstructionCost(Instruction *I,
5907 Type *RetTy = I->getType();
5908 if (VF > 1 && MinBWs.count(I))
5909 RetTy = IntegerType::get(RetTy->getContext(), MinBWs[I]);
5910 VectorTy = ToVectorTy(RetTy, VF);
5911 auto SE = PSE.getSE();
5913 // TODO: We need to estimate the cost of intrinsic calls.
5914 switch (I->getOpcode()) {
5915 case Instruction::GetElementPtr:
5916 // We mark this instruction as zero-cost because the cost of GEPs in
5917 // vectorized code depends on whether the corresponding memory instruction
5918 // is scalarized or not. Therefore, we handle GEPs with the memory
5919 // instruction cost.
5921 case Instruction::Br: {
5922 return TTI.getCFInstrCost(I->getOpcode());
5924 case Instruction::PHI: {
5925 auto *Phi = cast<PHINode>(I);
5927 // First-order recurrences are replaced by vector shuffles inside the loop.
5928 if (VF > 1 && Legal->isFirstOrderRecurrence(Phi))
5929 return TTI.getShuffleCost(TargetTransformInfo::SK_ExtractSubvector,
5930 VectorTy, VF - 1, VectorTy);
5932 // TODO: IF-converted IFs become selects.
5935 case Instruction::Add:
5936 case Instruction::FAdd:
5937 case Instruction::Sub:
5938 case Instruction::FSub:
5939 case Instruction::Mul:
5940 case Instruction::FMul:
5941 case Instruction::UDiv:
5942 case Instruction::SDiv:
5943 case Instruction::FDiv:
5944 case Instruction::URem:
5945 case Instruction::SRem:
5946 case Instruction::FRem:
5947 case Instruction::Shl:
5948 case Instruction::LShr:
5949 case Instruction::AShr:
5950 case Instruction::And:
5951 case Instruction::Or:
5952 case Instruction::Xor: {
5953 // Since we will replace the stride by 1 the multiplication should go away.
5954 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5956 // Certain instructions can be cheaper to vectorize if they have a constant
5957 // second vector operand. One example of this are shifts on x86.
5958 TargetTransformInfo::OperandValueKind Op1VK =
5959 TargetTransformInfo::OK_AnyValue;
5960 TargetTransformInfo::OperandValueKind Op2VK =
5961 TargetTransformInfo::OK_AnyValue;
5962 TargetTransformInfo::OperandValueProperties Op1VP =
5963 TargetTransformInfo::OP_None;
5964 TargetTransformInfo::OperandValueProperties Op2VP =
5965 TargetTransformInfo::OP_None;
5966 Value *Op2 = I->getOperand(1);
5968 // Check for a splat of a constant or for a non uniform vector of constants.
5969 if (isa<ConstantInt>(Op2)) {
5970 ConstantInt *CInt = cast<ConstantInt>(Op2);
5971 if (CInt && CInt->getValue().isPowerOf2())
5972 Op2VP = TargetTransformInfo::OP_PowerOf2;
5973 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5974 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5975 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5976 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
5978 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
5979 if (CInt && CInt->getValue().isPowerOf2())
5980 Op2VP = TargetTransformInfo::OP_PowerOf2;
5981 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5985 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
5988 case Instruction::Select: {
5989 SelectInst *SI = cast<SelectInst>(I);
5990 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5991 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5992 Type *CondTy = SI->getCondition()->getType();
5994 CondTy = VectorType::get(CondTy, VF);
5996 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5998 case Instruction::ICmp:
5999 case Instruction::FCmp: {
6000 Type *ValTy = I->getOperand(0)->getType();
6001 Instruction *Op0AsInstruction = dyn_cast<Instruction>(I->getOperand(0));
6002 auto It = MinBWs.find(Op0AsInstruction);
6003 if (VF > 1 && It != MinBWs.end())
6004 ValTy = IntegerType::get(ValTy->getContext(), It->second);
6005 VectorTy = ToVectorTy(ValTy, VF);
6006 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
6008 case Instruction::Store:
6009 case Instruction::Load: {
6010 StoreInst *SI = dyn_cast<StoreInst>(I);
6011 LoadInst *LI = dyn_cast<LoadInst>(I);
6012 Type *ValTy = (SI ? SI->getValueOperand()->getType() : LI->getType());
6013 VectorTy = ToVectorTy(ValTy, VF);
6015 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
6017 SI ? SI->getPointerAddressSpace() : LI->getPointerAddressSpace();
6018 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
6019 // We add the cost of address computation here instead of with the gep
6020 // instruction because only here we know whether the operation is
6023 return TTI.getAddressComputationCost(VectorTy) +
6024 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
6026 if (LI && Legal->isUniform(Ptr)) {
6027 // Scalar load + broadcast
6028 unsigned Cost = TTI.getAddressComputationCost(ValTy->getScalarType());
6029 Cost += TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
6032 TTI.getShuffleCost(TargetTransformInfo::SK_Broadcast, ValTy);
6035 // For an interleaved access, calculate the total cost of the whole
6036 // interleave group.
6037 if (Legal->isAccessInterleaved(I)) {
6038 auto Group = Legal->getInterleavedAccessGroup(I);
6039 assert(Group && "Fail to get an interleaved access group.");
6041 // Only calculate the cost once at the insert position.
6042 if (Group->getInsertPos() != I)
6045 unsigned InterleaveFactor = Group->getFactor();
6047 VectorType::get(VectorTy->getVectorElementType(),
6048 VectorTy->getVectorNumElements() * InterleaveFactor);
6050 // Holds the indices of existing members in an interleaved load group.
6051 // An interleaved store group doesn't need this as it doesn't allow gaps.
6052 SmallVector<unsigned, 4> Indices;
6054 for (unsigned i = 0; i < InterleaveFactor; i++)
6055 if (Group->getMember(i))
6056 Indices.push_back(i);
6059 // Calculate the cost of the whole interleaved group.
6060 unsigned Cost = TTI.getInterleavedMemoryOpCost(
6061 I->getOpcode(), WideVecTy, Group->getFactor(), Indices,
6062 Group->getAlignment(), AS);
6064 if (Group->isReverse())
6066 Group->getNumMembers() *
6067 TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
6069 // FIXME: The interleaved load group with a huge gap could be even more
6070 // expensive than scalar operations. Then we could ignore such group and
6071 // use scalar operations instead.
6075 // Scalarized loads/stores.
6076 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
6077 bool UseGatherOrScatter =
6078 (ConsecutiveStride == 0) && isGatherOrScatterLegal(I, Ptr, Legal);
6080 bool Reverse = ConsecutiveStride < 0;
6081 const DataLayout &DL = I->getModule()->getDataLayout();
6082 uint64_t ScalarAllocatedSize = DL.getTypeAllocSize(ValTy);
6083 uint64_t VectorElementSize = DL.getTypeStoreSize(VectorTy) / VF;
6084 if ((!ConsecutiveStride && !UseGatherOrScatter) ||
6085 ScalarAllocatedSize != VectorElementSize) {
6086 bool IsComplexComputation =
6087 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
6089 // The cost of extracting from the value vector and pointer vector.
6090 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
6091 for (unsigned i = 0; i < VF; ++i) {
6092 // The cost of extracting the pointer operand.
6093 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
6094 // In case of STORE, the cost of ExtractElement from the vector.
6095 // In case of LOAD, the cost of InsertElement into the returned
6097 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement
6098 : Instruction::InsertElement,
6102 // The cost of the scalar loads/stores.
6103 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
6105 TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
6110 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
6111 if (UseGatherOrScatter) {
6112 assert(ConsecutiveStride == 0 &&
6113 "Gather/Scatter are not used for consecutive stride");
6115 TTI.getGatherScatterOpCost(I->getOpcode(), VectorTy, Ptr,
6116 Legal->isMaskRequired(I), Alignment);
6118 // Wide load/stores.
6119 if (Legal->isMaskRequired(I))
6121 TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
6123 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
6126 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
6129 case Instruction::ZExt:
6130 case Instruction::SExt:
6131 case Instruction::FPToUI:
6132 case Instruction::FPToSI:
6133 case Instruction::FPExt:
6134 case Instruction::PtrToInt:
6135 case Instruction::IntToPtr:
6136 case Instruction::SIToFP:
6137 case Instruction::UIToFP:
6138 case Instruction::Trunc:
6139 case Instruction::FPTrunc:
6140 case Instruction::BitCast: {
6141 // We optimize the truncation of induction variable.
6142 // The cost of these is the same as the scalar operation.
6143 if (I->getOpcode() == Instruction::Trunc &&
6144 Legal->isInductionVariable(I->getOperand(0)))
6145 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
6146 I->getOperand(0)->getType());
6148 Type *SrcScalarTy = I->getOperand(0)->getType();
6149 Type *SrcVecTy = ToVectorTy(SrcScalarTy, VF);
6150 if (VF > 1 && MinBWs.count(I)) {
6151 // This cast is going to be shrunk. This may remove the cast or it might
6152 // turn it into slightly different cast. For example, if MinBW == 16,
6153 // "zext i8 %1 to i32" becomes "zext i8 %1 to i16".
6155 // Calculate the modified src and dest types.
6156 Type *MinVecTy = VectorTy;
6157 if (I->getOpcode() == Instruction::Trunc) {
6158 SrcVecTy = smallestIntegerVectorType(SrcVecTy, MinVecTy);
6160 largestIntegerVectorType(ToVectorTy(I->getType(), VF), MinVecTy);
6161 } else if (I->getOpcode() == Instruction::ZExt ||
6162 I->getOpcode() == Instruction::SExt) {
6163 SrcVecTy = largestIntegerVectorType(SrcVecTy, MinVecTy);
6165 smallestIntegerVectorType(ToVectorTy(I->getType(), VF), MinVecTy);
6169 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
6171 case Instruction::Call: {
6172 bool NeedToScalarize;
6173 CallInst *CI = cast<CallInst>(I);
6174 unsigned CallCost = getVectorCallCost(CI, VF, TTI, TLI, NeedToScalarize);
6175 if (getVectorIntrinsicIDForCall(CI, TLI))
6176 return std::min(CallCost, getVectorIntrinsicCost(CI, VF, TTI, TLI));
6180 // We are scalarizing the instruction. Return the cost of the scalar
6181 // instruction, plus the cost of insert and extract into vector
6182 // elements, times the vector width.
6185 if (!RetTy->isVoidTy() && VF != 1) {
6187 TTI.getVectorInstrCost(Instruction::InsertElement, VectorTy);
6189 TTI.getVectorInstrCost(Instruction::ExtractElement, VectorTy);
6191 // The cost of inserting the results plus extracting each one of the
6193 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
6196 // The cost of executing VF copies of the scalar instruction. This opcode
6197 // is unknown. Assume that it is the same as 'mul'.
6198 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
6204 char LoopVectorize::ID = 0;
6205 static const char lv_name[] = "Loop Vectorization";
6206 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
6207 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
6208 INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass)
6209 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
6210 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
6211 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
6212 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
6213 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
6214 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
6215 INITIALIZE_PASS_DEPENDENCY(LCSSAWrapperPass)
6216 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
6217 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
6218 INITIALIZE_PASS_DEPENDENCY(LoopAccessLegacyAnalysis)
6219 INITIALIZE_PASS_DEPENDENCY(DemandedBitsWrapperPass)
6220 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
6223 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
6224 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
6228 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
6229 // Check for a store.
6230 if (auto *ST = dyn_cast<StoreInst>(Inst))
6231 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
6233 // Check for a load.
6234 if (auto *LI = dyn_cast<LoadInst>(Inst))
6235 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
6240 void LoopVectorizationCostModel::collectValuesToIgnore() {
6241 // Ignore ephemeral values.
6242 CodeMetrics::collectEphemeralValues(TheLoop, AC, ValuesToIgnore);
6244 // Ignore type-promoting instructions we identified during reduction
6246 for (auto &Reduction : *Legal->getReductionVars()) {
6247 RecurrenceDescriptor &RedDes = Reduction.second;
6248 SmallPtrSetImpl<Instruction *> &Casts = RedDes.getCastInsts();
6249 VecValuesToIgnore.insert(Casts.begin(), Casts.end());
6252 // Ignore induction phis that are only used in either GetElementPtr or ICmp
6253 // instruction to exit loop. Induction variables usually have large types and
6254 // can have big impact when estimating register usage.
6255 // This is for when VF > 1.
6256 for (auto &Induction : *Legal->getInductionVars()) {
6257 auto *PN = Induction.first;
6258 auto *UpdateV = PN->getIncomingValueForBlock(TheLoop->getLoopLatch());
6260 // Check that the PHI is only used by the induction increment (UpdateV) or
6261 // by GEPs. Then check that UpdateV is only used by a compare instruction,
6262 // the loop header PHI, or by GEPs.
6263 // FIXME: Need precise def-use analysis to determine if this instruction
6264 // variable will be vectorized.
6265 if (all_of(PN->users(),
6266 [&](const User *U) -> bool {
6267 return U == UpdateV || isa<GetElementPtrInst>(U);
6269 all_of(UpdateV->users(), [&](const User *U) -> bool {
6270 return U == PN || isa<ICmpInst>(U) || isa<GetElementPtrInst>(U);
6272 VecValuesToIgnore.insert(PN);
6273 VecValuesToIgnore.insert(UpdateV);
6277 // Ignore instructions that will not be vectorized.
6278 // This is for when VF > 1.
6279 for (BasicBlock *BB : TheLoop->blocks()) {
6280 for (auto &Inst : *BB) {
6281 switch (Inst.getOpcode())
6282 case Instruction::GetElementPtr: {
6283 // Ignore GEP if its last operand is an induction variable so that it is
6284 // a consecutive load/store and won't be vectorized as scatter/gather
6287 GetElementPtrInst *Gep = cast<GetElementPtrInst>(&Inst);
6288 unsigned NumOperands = Gep->getNumOperands();
6289 unsigned InductionOperand = getGEPInductionOperand(Gep);
6290 bool GepToIgnore = true;
6292 // Check that all of the gep indices are uniform except for the
6293 // induction operand.
6294 for (unsigned i = 0; i != NumOperands; ++i) {
6295 if (i != InductionOperand &&
6296 !PSE.getSE()->isLoopInvariant(PSE.getSCEV(Gep->getOperand(i)),
6298 GepToIgnore = false;
6304 VecValuesToIgnore.insert(&Inst);
6311 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
6312 bool IfPredicateStore) {
6313 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
6314 // Holds vector parameters or scalars, in case of uniform vals.
6315 SmallVector<VectorParts, 4> Params;
6317 setDebugLocFromInst(Builder, Instr);
6319 // Find all of the vectorized parameters.
6320 for (Value *SrcOp : Instr->operands()) {
6321 // If we are accessing the old induction variable, use the new one.
6322 if (SrcOp == OldInduction) {
6323 Params.push_back(getVectorValue(SrcOp));
6327 // Try using previously calculated values.
6328 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
6330 // If the src is an instruction that appeared earlier in the basic block
6331 // then it should already be vectorized.
6332 if (SrcInst && OrigLoop->contains(SrcInst)) {
6333 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
6334 // The parameter is a vector value from earlier.
6335 Params.push_back(WidenMap.get(SrcInst));
6337 // The parameter is a scalar from outside the loop. Maybe even a constant.
6338 VectorParts Scalars;
6339 Scalars.append(UF, SrcOp);
6340 Params.push_back(Scalars);
6344 assert(Params.size() == Instr->getNumOperands() &&
6345 "Invalid number of operands");
6347 // Does this instruction return a value ?
6348 bool IsVoidRetTy = Instr->getType()->isVoidTy();
6350 Value *UndefVec = IsVoidRetTy ? nullptr : UndefValue::get(Instr->getType());
6351 // Create a new entry in the WidenMap and initialize it to Undef or Null.
6352 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
6355 if (IfPredicateStore) {
6356 assert(Instr->getParent()->getSinglePredecessor() &&
6357 "Only support single predecessor blocks");
6358 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
6359 Instr->getParent());
6362 // For each vector unroll 'part':
6363 for (unsigned Part = 0; Part < UF; ++Part) {
6364 // For each scalar that we create:
6366 // Start an "if (pred) a[i] = ..." block.
6367 Value *Cmp = nullptr;
6368 if (IfPredicateStore) {
6369 if (Cond[Part]->getType()->isVectorTy())
6371 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
6372 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
6373 ConstantInt::get(Cond[Part]->getType(), 1));
6376 Instruction *Cloned = Instr->clone();
6378 Cloned->setName(Instr->getName() + ".cloned");
6379 // Replace the operands of the cloned instructions with extracted scalars.
6380 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
6381 Value *Op = Params[op][Part];
6382 Cloned->setOperand(op, Op);
6385 // Place the cloned scalar in the new loop.
6386 Builder.Insert(Cloned);
6388 // If we just cloned a new assumption, add it the assumption cache.
6389 if (auto *II = dyn_cast<IntrinsicInst>(Cloned))
6390 if (II->getIntrinsicID() == Intrinsic::assume)
6391 AC->registerAssumption(II);
6393 // If the original scalar returns a value we need to place it in a vector
6394 // so that future users will be able to use it.
6396 VecResults[Part] = Cloned;
6399 if (IfPredicateStore)
6400 PredicatedStores.push_back(std::make_pair(cast<StoreInst>(Cloned), Cmp));
6404 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
6405 auto *SI = dyn_cast<StoreInst>(Instr);
6406 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
6408 return scalarizeInstruction(Instr, IfPredicateStore);
6411 Value *InnerLoopUnroller::reverseVector(Value *Vec) { return Vec; }
6413 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) { return V; }
6415 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) {
6416 // When unrolling and the VF is 1, we only need to add a simple scalar.
6417 Type *ITy = Val->getType();
6418 assert(!ITy->isVectorTy() && "Val must be a scalar");
6419 Constant *C = ConstantInt::get(ITy, StartIdx);
6420 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");
6423 static void AddRuntimeUnrollDisableMetaData(Loop *L) {
6424 SmallVector<Metadata *, 4> MDs;
6425 // Reserve first location for self reference to the LoopID metadata node.
6426 MDs.push_back(nullptr);
6427 bool IsUnrollMetadata = false;
6428 MDNode *LoopID = L->getLoopID();
6430 // First find existing loop unrolling disable metadata.
6431 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
6432 auto *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
6434 const auto *S = dyn_cast<MDString>(MD->getOperand(0));
6436 S && S->getString().startswith("llvm.loop.unroll.disable");
6438 MDs.push_back(LoopID->getOperand(i));
6442 if (!IsUnrollMetadata) {
6443 // Add runtime unroll disable metadata.
6444 LLVMContext &Context = L->getHeader()->getContext();
6445 SmallVector<Metadata *, 1> DisableOperands;
6446 DisableOperands.push_back(
6447 MDString::get(Context, "llvm.loop.unroll.runtime.disable"));
6448 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
6449 MDs.push_back(DisableNode);
6450 MDNode *NewLoopID = MDNode::get(Context, MDs);
6451 // Set operand 0 to refer to the loop id itself.
6452 NewLoopID->replaceOperandWith(0, NewLoopID);
6453 L->setLoopID(NewLoopID);
6457 bool LoopVectorizePass::processLoop(Loop *L) {
6458 assert(L->empty() && "Only process inner loops.");
6461 const std::string DebugLocStr = getDebugLocString(L);
6464 DEBUG(dbgs() << "\nLV: Checking a loop in \""
6465 << L->getHeader()->getParent()->getName() << "\" from "
6466 << DebugLocStr << "\n");
6468 LoopVectorizeHints Hints(L, DisableUnrolling);
6470 DEBUG(dbgs() << "LV: Loop hints:"
6472 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
6474 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
6477 << " width=" << Hints.getWidth()
6478 << " unroll=" << Hints.getInterleave() << "\n");
6480 // Function containing loop
6481 Function *F = L->getHeader()->getParent();
6483 // Looking at the diagnostic output is the only way to determine if a loop
6484 // was vectorized (other than looking at the IR or machine code), so it
6485 // is important to generate an optimization remark for each loop. Most of
6486 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
6487 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
6488 // less verbose reporting vectorized loops and unvectorized loops that may
6489 // benefit from vectorization, respectively.
6491 if (!Hints.allowVectorization(F, L, AlwaysVectorize)) {
6492 DEBUG(dbgs() << "LV: Loop hints prevent vectorization.\n");
6496 // Check the loop for a trip count threshold:
6497 // do not vectorize loops with a tiny trip count.
6498 const unsigned TC = SE->getSmallConstantTripCount(L);
6499 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
6500 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
6501 << "This loop is not worth vectorizing.");
6502 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
6503 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
6505 DEBUG(dbgs() << "\n");
6506 emitAnalysisDiag(F, L, Hints, VectorizationReport()
6507 << "vectorization is not beneficial "
6508 "and is not explicitly forced");
6513 PredicatedScalarEvolution PSE(*SE, *L);
6515 // Check if it is legal to vectorize the loop.
6516 LoopVectorizationRequirements Requirements;
6517 LoopVectorizationLegality LVL(L, PSE, DT, TLI, AA, F, TTI, GetLAA, LI,
6518 &Requirements, &Hints);
6519 if (!LVL.canVectorize()) {
6520 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
6521 emitMissedWarning(F, L, Hints);
6525 // Use the cost model.
6526 LoopVectorizationCostModel CM(L, PSE, LI, &LVL, *TTI, TLI, DB, AC, F,
6528 CM.collectValuesToIgnore();
6530 // Check the function attributes to find out if this function should be
6531 // optimized for size.
6533 Hints.getForce() != LoopVectorizeHints::FK_Enabled && F->optForSize();
6535 // Compute the weighted frequency of this loop being executed and see if it
6536 // is less than 20% of the function entry baseline frequency. Note that we
6537 // always have a canonical loop here because we think we *can* vectorize.
6538 // FIXME: This is hidden behind a flag due to pervasive problems with
6539 // exactly what block frequency models.
6540 if (LoopVectorizeWithBlockFrequency) {
6541 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
6542 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
6543 LoopEntryFreq < ColdEntryFreq)
6547 // Check the function attributes to see if implicit floats are allowed.
6548 // FIXME: This check doesn't seem possibly correct -- what if the loop is
6549 // an integer loop and the vector instructions selected are purely integer
6550 // vector instructions?
6551 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
6552 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
6553 "attribute is used.\n");
6556 VectorizationReport()
6557 << "loop not vectorized due to NoImplicitFloat attribute");
6558 emitMissedWarning(F, L, Hints);
6562 // Check if the target supports potentially unsafe FP vectorization.
6563 // FIXME: Add a check for the type of safety issue (denormal, signaling)
6564 // for the target we're vectorizing for, to make sure none of the
6565 // additional fp-math flags can help.
6566 if (Hints.isPotentiallyUnsafe() &&
6567 TTI->isFPVectorizationPotentiallyUnsafe()) {
6568 DEBUG(dbgs() << "LV: Potentially unsafe FP op prevents vectorization.\n");
6569 emitAnalysisDiag(F, L, Hints,
6570 VectorizationReport()
6571 << "loop not vectorized due to unsafe FP support.");
6572 emitMissedWarning(F, L, Hints);
6576 // Select the optimal vectorization factor.
6577 const LoopVectorizationCostModel::VectorizationFactor VF =
6578 CM.selectVectorizationFactor(OptForSize);
6580 // Select the interleave count.
6581 unsigned IC = CM.selectInterleaveCount(OptForSize, VF.Width, VF.Cost);
6583 // Get user interleave count.
6584 unsigned UserIC = Hints.getInterleave();
6586 // Identify the diagnostic messages that should be produced.
6587 std::string VecDiagMsg, IntDiagMsg;
6588 bool VectorizeLoop = true, InterleaveLoop = true;
6590 if (Requirements.doesNotMeet(F, L, Hints)) {
6591 DEBUG(dbgs() << "LV: Not vectorizing: loop did not meet vectorization "
6593 emitMissedWarning(F, L, Hints);
6597 if (VF.Width == 1) {
6598 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
6600 "the cost-model indicates that vectorization is not beneficial";
6601 VectorizeLoop = false;
6604 if (IC == 1 && UserIC <= 1) {
6605 // Tell the user interleaving is not beneficial.
6606 DEBUG(dbgs() << "LV: Interleaving is not beneficial.\n");
6608 "the cost-model indicates that interleaving is not beneficial";
6609 InterleaveLoop = false;
6612 " and is explicitly disabled or interleave count is set to 1";
6613 } else if (IC > 1 && UserIC == 1) {
6614 // Tell the user interleaving is beneficial, but it explicitly disabled.
6616 << "LV: Interleaving is beneficial but is explicitly disabled.");
6617 IntDiagMsg = "the cost-model indicates that interleaving is beneficial "
6618 "but is explicitly disabled or interleave count is set to 1";
6619 InterleaveLoop = false;
6622 // Override IC if user provided an interleave count.
6623 IC = UserIC > 0 ? UserIC : IC;
6625 // Emit diagnostic messages, if any.
6626 const char *VAPassName = Hints.vectorizeAnalysisPassName();
6627 if (!VectorizeLoop && !InterleaveLoop) {
6628 // Do not vectorize or interleaving the loop.
6629 emitOptimizationRemarkAnalysis(F->getContext(), VAPassName, *F,
6630 L->getStartLoc(), VecDiagMsg);
6631 emitOptimizationRemarkAnalysis(F->getContext(), LV_NAME, *F,
6632 L->getStartLoc(), IntDiagMsg);
6634 } else if (!VectorizeLoop && InterleaveLoop) {
6635 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
6636 emitOptimizationRemarkAnalysis(F->getContext(), VAPassName, *F,
6637 L->getStartLoc(), VecDiagMsg);
6638 } else if (VectorizeLoop && !InterleaveLoop) {
6639 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
6640 << DebugLocStr << '\n');
6641 emitOptimizationRemarkAnalysis(F->getContext(), LV_NAME, *F,
6642 L->getStartLoc(), IntDiagMsg);
6643 } else if (VectorizeLoop && InterleaveLoop) {
6644 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
6645 << DebugLocStr << '\n');
6646 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
6649 if (!VectorizeLoop) {
6650 assert(IC > 1 && "interleave count should not be 1 or 0");
6651 // If we decided that it is not legal to vectorize the loop, then
6653 InnerLoopUnroller Unroller(L, PSE, LI, DT, TLI, TTI, AC, IC);
6654 Unroller.vectorize(&LVL, CM.MinBWs, CM.VecValuesToIgnore);
6656 emitOptimizationRemark(F->getContext(), LV_NAME, *F, L->getStartLoc(),
6657 Twine("interleaved loop (interleaved count: ") +
6660 // If we decided that it is *legal* to vectorize the loop, then do it.
6661 InnerLoopVectorizer LB(L, PSE, LI, DT, TLI, TTI, AC, VF.Width, IC);
6662 LB.vectorize(&LVL, CM.MinBWs, CM.VecValuesToIgnore);
6665 // Add metadata to disable runtime unrolling a scalar loop when there are
6666 // no runtime checks about strides and memory. A scalar loop that is
6667 // rarely used is not worth unrolling.
6668 if (!LB.areSafetyChecksAdded())
6669 AddRuntimeUnrollDisableMetaData(L);
6671 // Report the vectorization decision.
6672 emitOptimizationRemark(F->getContext(), LV_NAME, *F, L->getStartLoc(),
6673 Twine("vectorized loop (vectorization width: ") +
6674 Twine(VF.Width) + ", interleaved count: " +
6678 // Mark the loop as already vectorized to avoid vectorizing again.
6679 Hints.setAlreadyVectorized();
6681 DEBUG(verifyFunction(*L->getHeader()->getParent()));
6685 bool LoopVectorizePass::runImpl(
6686 Function &F, ScalarEvolution &SE_, LoopInfo &LI_, TargetTransformInfo &TTI_,
6687 DominatorTree &DT_, BlockFrequencyInfo &BFI_, TargetLibraryInfo *TLI_,
6688 DemandedBits &DB_, AliasAnalysis &AA_, AssumptionCache &AC_,
6689 std::function<const LoopAccessInfo &(Loop &)> &GetLAA_) {
6702 // Compute some weights outside of the loop over the loops. Compute this
6703 // using a BranchProbability to re-use its scaling math.
6704 const BranchProbability ColdProb(1, 5); // 20%
6705 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
6708 // 1. the target claims to have no vector registers, and
6709 // 2. interleaving won't help ILP.
6711 // The second condition is necessary because, even if the target has no
6712 // vector registers, loop vectorization may still enable scalar
6714 if (!TTI->getNumberOfRegisters(true) && TTI->getMaxInterleaveFactor(1) < 2)
6717 // Build up a worklist of inner-loops to vectorize. This is necessary as
6718 // the act of vectorizing or partially unrolling a loop creates new loops
6719 // and can invalidate iterators across the loops.
6720 SmallVector<Loop *, 8> Worklist;
6723 addAcyclicInnerLoop(*L, Worklist);
6725 LoopsAnalyzed += Worklist.size();
6727 // Now walk the identified inner loops.
6728 bool Changed = false;
6729 while (!Worklist.empty())
6730 Changed |= processLoop(Worklist.pop_back_val());
6732 // Process each loop nest in the function.
6738 PreservedAnalyses LoopVectorizePass::run(Function &F,
6739 FunctionAnalysisManager &AM) {
6740 auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F);
6741 auto &LI = AM.getResult<LoopAnalysis>(F);
6742 auto &TTI = AM.getResult<TargetIRAnalysis>(F);
6743 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
6744 auto &BFI = AM.getResult<BlockFrequencyAnalysis>(F);
6745 auto *TLI = AM.getCachedResult<TargetLibraryAnalysis>(F);
6746 auto &AA = AM.getResult<AAManager>(F);
6747 auto &AC = AM.getResult<AssumptionAnalysis>(F);
6748 auto &DB = AM.getResult<DemandedBitsAnalysis>(F);
6750 auto &LAM = AM.getResult<LoopAnalysisManagerFunctionProxy>(F).getManager();
6751 std::function<const LoopAccessInfo &(Loop &)> GetLAA =
6752 [&](Loop &L) -> const LoopAccessInfo & {
6753 return LAM.getResult<LoopAccessAnalysis>(L);
6755 bool Changed = runImpl(F, SE, LI, TTI, DT, BFI, TLI, DB, AA, AC, GetLAA);
6757 return PreservedAnalyses::all();
6758 PreservedAnalyses PA;
6759 PA.preserve<LoopAnalysis>();
6760 PA.preserve<DominatorTreeAnalysis>();
6761 PA.preserve<BasicAA>();
6762 PA.preserve<GlobalsAA>();