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.h"
50 #include "llvm/ADT/DenseMap.h"
51 #include "llvm/ADT/Hashing.h"
52 #include "llvm/ADT/MapVector.h"
53 #include "llvm/ADT/SetVector.h"
54 #include "llvm/ADT/SmallPtrSet.h"
55 #include "llvm/ADT/SmallSet.h"
56 #include "llvm/ADT/SmallVector.h"
57 #include "llvm/ADT/Statistic.h"
58 #include "llvm/ADT/StringExtras.h"
59 #include "llvm/Analysis/AliasAnalysis.h"
60 #include "llvm/Analysis/BasicAliasAnalysis.h"
61 #include "llvm/Analysis/AliasSetTracker.h"
62 #include "llvm/Analysis/AssumptionCache.h"
63 #include "llvm/Analysis/BlockFrequencyInfo.h"
64 #include "llvm/Analysis/CodeMetrics.h"
65 #include "llvm/Analysis/DemandedBits.h"
66 #include "llvm/Analysis/GlobalsModRef.h"
67 #include "llvm/Analysis/LoopAccessAnalysis.h"
68 #include "llvm/Analysis/LoopInfo.h"
69 #include "llvm/Analysis/LoopIterator.h"
70 #include "llvm/Analysis/LoopPass.h"
71 #include "llvm/Analysis/ScalarEvolution.h"
72 #include "llvm/Analysis/ScalarEvolutionExpander.h"
73 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
74 #include "llvm/Analysis/TargetTransformInfo.h"
75 #include "llvm/Analysis/ValueTracking.h"
76 #include "llvm/IR/Constants.h"
77 #include "llvm/IR/DataLayout.h"
78 #include "llvm/IR/DebugInfo.h"
79 #include "llvm/IR/DerivedTypes.h"
80 #include "llvm/IR/DiagnosticInfo.h"
81 #include "llvm/IR/Dominators.h"
82 #include "llvm/IR/Function.h"
83 #include "llvm/IR/IRBuilder.h"
84 #include "llvm/IR/Instructions.h"
85 #include "llvm/IR/IntrinsicInst.h"
86 #include "llvm/IR/LLVMContext.h"
87 #include "llvm/IR/Module.h"
88 #include "llvm/IR/PatternMatch.h"
89 #include "llvm/IR/Type.h"
90 #include "llvm/IR/Value.h"
91 #include "llvm/IR/ValueHandle.h"
92 #include "llvm/IR/Verifier.h"
93 #include "llvm/Pass.h"
94 #include "llvm/Support/BranchProbability.h"
95 #include "llvm/Support/CommandLine.h"
96 #include "llvm/Support/Debug.h"
97 #include "llvm/Support/raw_ostream.h"
98 #include "llvm/Transforms/Scalar.h"
99 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
100 #include "llvm/Transforms/Utils/Local.h"
101 #include "llvm/Analysis/VectorUtils.h"
102 #include "llvm/Transforms/Utils/LoopUtils.h"
104 #include <functional>
108 using namespace llvm;
109 using namespace llvm::PatternMatch;
111 #define LV_NAME "loop-vectorize"
112 #define DEBUG_TYPE LV_NAME
114 STATISTIC(LoopsVectorized, "Number of loops vectorized");
115 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
118 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
119 cl::desc("Enable if-conversion during vectorization."));
121 /// We don't vectorize loops with a known constant trip count below this number.
122 static cl::opt<unsigned>
123 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
125 cl::desc("Don't vectorize loops with a constant "
126 "trip count that is smaller than this "
129 static cl::opt<bool> MaximizeBandwidth(
130 "vectorizer-maximize-bandwidth", cl::init(false), cl::Hidden,
131 cl::desc("Maximize bandwidth when selecting vectorization factor which "
132 "will be determined by the smallest type in loop."));
134 /// This enables versioning on the strides of symbolically striding memory
135 /// accesses in code like the following.
136 /// for (i = 0; i < N; ++i)
137 /// A[i * Stride1] += B[i * Stride2] ...
139 /// Will be roughly translated to
140 /// if (Stride1 == 1 && Stride2 == 1) {
141 /// for (i = 0; i < N; i+=4)
145 static cl::opt<bool> EnableMemAccessVersioning(
146 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
147 cl::desc("Enable symbolic stride memory access versioning"));
149 static cl::opt<bool> EnableInterleavedMemAccesses(
150 "enable-interleaved-mem-accesses", cl::init(false), cl::Hidden,
151 cl::desc("Enable vectorization on interleaved memory accesses in a loop"));
153 /// Maximum factor for an interleaved memory access.
154 static cl::opt<unsigned> MaxInterleaveGroupFactor(
155 "max-interleave-group-factor", cl::Hidden,
156 cl::desc("Maximum factor for an interleaved access group (default = 8)"),
159 /// We don't interleave loops with a known constant trip count below this
161 static const unsigned TinyTripCountInterleaveThreshold = 128;
163 static cl::opt<unsigned> ForceTargetNumScalarRegs(
164 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
165 cl::desc("A flag that overrides the target's number of scalar registers."));
167 static cl::opt<unsigned> ForceTargetNumVectorRegs(
168 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
169 cl::desc("A flag that overrides the target's number of vector registers."));
171 /// Maximum vectorization interleave count.
172 static const unsigned MaxInterleaveFactor = 16;
174 static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
175 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
176 cl::desc("A flag that overrides the target's max interleave factor for "
179 static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
180 "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
181 cl::desc("A flag that overrides the target's max interleave factor for "
182 "vectorized loops."));
184 static cl::opt<unsigned> ForceTargetInstructionCost(
185 "force-target-instruction-cost", cl::init(0), cl::Hidden,
186 cl::desc("A flag that overrides the target's expected cost for "
187 "an instruction to a single constant value. Mostly "
188 "useful for getting consistent testing."));
190 static cl::opt<unsigned> SmallLoopCost(
191 "small-loop-cost", cl::init(20), cl::Hidden,
193 "The cost of a loop that is considered 'small' by the interleaver."));
195 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
196 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
197 cl::desc("Enable the use of the block frequency analysis to access PGO "
198 "heuristics minimizing code growth in cold regions and being more "
199 "aggressive in hot regions."));
201 // Runtime interleave loops for load/store throughput.
202 static cl::opt<bool> EnableLoadStoreRuntimeInterleave(
203 "enable-loadstore-runtime-interleave", cl::init(true), cl::Hidden,
205 "Enable runtime interleaving until load/store ports are saturated"));
207 /// The number of stores in a loop that are allowed to need predication.
208 static cl::opt<unsigned> NumberOfStoresToPredicate(
209 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
210 cl::desc("Max number of stores to be predicated behind an if."));
212 static cl::opt<bool> EnableIndVarRegisterHeur(
213 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
214 cl::desc("Count the induction variable only once when interleaving"));
216 static cl::opt<bool> EnableCondStoresVectorization(
217 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
218 cl::desc("Enable if predication of stores during vectorization."));
220 static cl::opt<unsigned> MaxNestedScalarReductionIC(
221 "max-nested-scalar-reduction-interleave", cl::init(2), cl::Hidden,
222 cl::desc("The maximum interleave count to use when interleaving a scalar "
223 "reduction in a nested loop."));
225 static cl::opt<unsigned> PragmaVectorizeMemoryCheckThreshold(
226 "pragma-vectorize-memory-check-threshold", cl::init(128), cl::Hidden,
227 cl::desc("The maximum allowed number of runtime memory checks with a "
228 "vectorize(enable) pragma."));
230 static cl::opt<unsigned> VectorizeSCEVCheckThreshold(
231 "vectorize-scev-check-threshold", cl::init(16), cl::Hidden,
232 cl::desc("The maximum number of SCEV checks allowed."));
234 static cl::opt<unsigned> PragmaVectorizeSCEVCheckThreshold(
235 "pragma-vectorize-scev-check-threshold", cl::init(128), cl::Hidden,
236 cl::desc("The maximum number of SCEV checks allowed with a "
237 "vectorize(enable) pragma"));
241 // Forward declarations.
242 class LoopVectorizeHints;
243 class LoopVectorizationLegality;
244 class LoopVectorizationCostModel;
245 class LoopVectorizationRequirements;
247 /// \brief This modifies LoopAccessReport to initialize message with
248 /// loop-vectorizer-specific part.
249 class VectorizationReport : public LoopAccessReport {
251 VectorizationReport(Instruction *I = nullptr)
252 : LoopAccessReport("loop not vectorized: ", I) {}
254 /// \brief This allows promotion of the loop-access analysis report into the
255 /// loop-vectorizer report. It modifies the message to add the
256 /// loop-vectorizer-specific part of the message.
257 explicit VectorizationReport(const LoopAccessReport &R)
258 : LoopAccessReport(Twine("loop not vectorized: ") + R.str(),
262 /// A helper function for converting Scalar types to vector types.
263 /// If the incoming type is void, we return void. If the VF is 1, we return
265 static Type* ToVectorTy(Type *Scalar, unsigned VF) {
266 if (Scalar->isVoidTy() || VF == 1)
268 return VectorType::get(Scalar, VF);
271 /// A helper function that returns GEP instruction and knows to skip a
272 /// 'bitcast'. The 'bitcast' may be skipped if the source and the destination
273 /// pointee types of the 'bitcast' have the same size.
275 /// bitcast double** %var to i64* - can be skipped
276 /// bitcast double** %var to i8* - can not
277 static GetElementPtrInst *getGEPInstruction(Value *Ptr) {
279 if (isa<GetElementPtrInst>(Ptr))
280 return cast<GetElementPtrInst>(Ptr);
282 if (isa<BitCastInst>(Ptr) &&
283 isa<GetElementPtrInst>(cast<BitCastInst>(Ptr)->getOperand(0))) {
284 Type *BitcastTy = Ptr->getType();
285 Type *GEPTy = cast<BitCastInst>(Ptr)->getSrcTy();
286 if (!isa<PointerType>(BitcastTy) || !isa<PointerType>(GEPTy))
288 Type *Pointee1Ty = cast<PointerType>(BitcastTy)->getPointerElementType();
289 Type *Pointee2Ty = cast<PointerType>(GEPTy)->getPointerElementType();
290 const DataLayout &DL = cast<BitCastInst>(Ptr)->getModule()->getDataLayout();
291 if (DL.getTypeSizeInBits(Pointee1Ty) == DL.getTypeSizeInBits(Pointee2Ty))
292 return cast<GetElementPtrInst>(cast<BitCastInst>(Ptr)->getOperand(0));
297 /// InnerLoopVectorizer vectorizes loops which contain only one basic
298 /// block to a specified vectorization factor (VF).
299 /// This class performs the widening of scalars into vectors, or multiple
300 /// scalars. This class also implements the following features:
301 /// * It inserts an epilogue loop for handling loops that don't have iteration
302 /// counts that are known to be a multiple of the vectorization factor.
303 /// * It handles the code generation for reduction variables.
304 /// * Scalarization (implementation using scalars) of un-vectorizable
306 /// InnerLoopVectorizer does not perform any vectorization-legality
307 /// checks, and relies on the caller to check for the different legality
308 /// aspects. The InnerLoopVectorizer relies on the
309 /// LoopVectorizationLegality class to provide information about the induction
310 /// and reduction variables that were found to a given vectorization factor.
311 class InnerLoopVectorizer {
313 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
314 DominatorTree *DT, const TargetLibraryInfo *TLI,
315 const TargetTransformInfo *TTI, unsigned VecWidth,
316 unsigned UnrollFactor, SCEVUnionPredicate &Preds)
317 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), TLI(TLI), TTI(TTI),
318 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
319 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
320 TripCount(nullptr), VectorTripCount(nullptr), Legal(nullptr),
321 AddedSafetyChecks(false), Preds(Preds) {}
323 // Perform the actual loop widening (vectorization).
324 // MinimumBitWidths maps scalar integer values to the smallest bitwidth they
325 // can be validly truncated to. The cost model has assumed this truncation
326 // will happen when vectorizing.
327 void vectorize(LoopVectorizationLegality *L,
328 DenseMap<Instruction*,uint64_t> MinimumBitWidths) {
329 MinBWs = MinimumBitWidths;
331 // Create a new empty loop. Unlink the old loop and connect the new one.
333 // Widen each instruction in the old loop to a new one in the new loop.
334 // Use the Legality module to find the induction and reduction variables.
338 // Return true if any runtime check is added.
339 bool IsSafetyChecksAdded() {
340 return AddedSafetyChecks;
343 virtual ~InnerLoopVectorizer() {}
346 /// A small list of PHINodes.
347 typedef SmallVector<PHINode*, 4> PhiVector;
348 /// When we unroll loops we have multiple vector values for each scalar.
349 /// This data structure holds the unrolled and vectorized values that
350 /// originated from one scalar instruction.
351 typedef SmallVector<Value*, 2> VectorParts;
353 // When we if-convert we need to create edge masks. We have to cache values
354 // so that we don't end up with exponential recursion/IR.
355 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
356 VectorParts> EdgeMaskCache;
358 /// Create an empty loop, based on the loop ranges of the old loop.
359 void createEmptyLoop();
360 /// Create a new induction variable inside L.
361 PHINode *createInductionVariable(Loop *L, Value *Start, Value *End,
362 Value *Step, Instruction *DL);
363 /// Copy and widen the instructions from the old loop.
364 virtual void vectorizeLoop();
366 /// \brief The Loop exit block may have single value PHI nodes where the
367 /// incoming value is 'Undef'. While vectorizing we only handled real values
368 /// that were defined inside the loop. Here we fix the 'undef case'.
372 /// Shrinks vector element sizes based on information in "MinBWs".
373 void truncateToMinimalBitwidths();
375 /// A helper function that computes the predicate of the block BB, assuming
376 /// that the header block of the loop is set to True. It returns the *entry*
377 /// mask for the block BB.
378 VectorParts createBlockInMask(BasicBlock *BB);
379 /// A helper function that computes the predicate of the edge between SRC
381 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
383 /// A helper function to vectorize a single BB within the innermost loop.
384 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
386 /// Vectorize a single PHINode in a block. This method handles the induction
387 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
388 /// arbitrary length vectors.
389 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
390 unsigned UF, unsigned VF, PhiVector *PV);
392 /// Insert the new loop to the loop hierarchy and pass manager
393 /// and update the analysis passes.
394 void updateAnalysis();
396 /// This instruction is un-vectorizable. Implement it as a sequence
397 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
398 /// scalarized instruction behind an if block predicated on the control
399 /// dependence of the instruction.
400 virtual void scalarizeInstruction(Instruction *Instr,
401 bool IfPredicateStore=false);
403 /// Vectorize Load and Store instructions,
404 virtual void vectorizeMemoryInstruction(Instruction *Instr);
406 /// Create a broadcast instruction. This method generates a broadcast
407 /// instruction (shuffle) for loop invariant values and for the induction
408 /// value. If this is the induction variable then we extend it to N, N+1, ...
409 /// this is needed because each iteration in the loop corresponds to a SIMD
411 virtual Value *getBroadcastInstrs(Value *V);
413 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
414 /// to each vector element of Val. The sequence starts at StartIndex.
415 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step);
417 /// When we go over instructions in the basic block we rely on previous
418 /// values within the current basic block or on loop invariant values.
419 /// When we widen (vectorize) values we place them in the map. If the values
420 /// are not within the map, they have to be loop invariant, so we simply
421 /// broadcast them into a vector.
422 VectorParts &getVectorValue(Value *V);
424 /// Try to vectorize the interleaved access group that \p Instr belongs to.
425 void vectorizeInterleaveGroup(Instruction *Instr);
427 /// Generate a shuffle sequence that will reverse the vector Vec.
428 virtual Value *reverseVector(Value *Vec);
430 /// Returns (and creates if needed) the original loop trip count.
431 Value *getOrCreateTripCount(Loop *NewLoop);
433 /// Returns (and creates if needed) the trip count of the widened loop.
434 Value *getOrCreateVectorTripCount(Loop *NewLoop);
436 /// Emit a bypass check to see if the trip count would overflow, or we
437 /// wouldn't have enough iterations to execute one vector loop.
438 void emitMinimumIterationCountCheck(Loop *L, BasicBlock *Bypass);
439 /// Emit a bypass check to see if the vector trip count is nonzero.
440 void emitVectorLoopEnteredCheck(Loop *L, BasicBlock *Bypass);
441 /// Emit a bypass check to see if all of the SCEV assumptions we've
442 /// had to make are correct.
443 void emitSCEVChecks(Loop *L, BasicBlock *Bypass);
444 /// Emit bypass checks to check any memory assumptions we may have made.
445 void emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass);
447 /// This is a helper class that holds the vectorizer state. It maps scalar
448 /// instructions to vector instructions. When the code is 'unrolled' then
449 /// then a single scalar value is mapped to multiple vector parts. The parts
450 /// are stored in the VectorPart type.
452 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
454 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
456 /// \return True if 'Key' is saved in the Value Map.
457 bool has(Value *Key) const { return MapStorage.count(Key); }
459 /// Initializes a new entry in the map. Sets all of the vector parts to the
460 /// save value in 'Val'.
461 /// \return A reference to a vector with splat values.
462 VectorParts &splat(Value *Key, Value *Val) {
463 VectorParts &Entry = MapStorage[Key];
464 Entry.assign(UF, Val);
468 ///\return A reference to the value that is stored at 'Key'.
469 VectorParts &get(Value *Key) {
470 VectorParts &Entry = MapStorage[Key];
473 assert(Entry.size() == UF);
478 /// The unroll factor. Each entry in the map stores this number of vector
482 /// Map storage. We use std::map and not DenseMap because insertions to a
483 /// dense map invalidates its iterators.
484 std::map<Value *, VectorParts> MapStorage;
487 /// The original loop.
489 /// Scev analysis to use.
497 /// Target Library Info.
498 const TargetLibraryInfo *TLI;
499 /// Target Transform Info.
500 const TargetTransformInfo *TTI;
502 /// The vectorization SIMD factor to use. Each vector will have this many
507 /// The vectorization unroll factor to use. Each scalar is vectorized to this
508 /// many different vector instructions.
511 /// The builder that we use
514 // --- Vectorization state ---
516 /// The vector-loop preheader.
517 BasicBlock *LoopVectorPreHeader;
518 /// The scalar-loop preheader.
519 BasicBlock *LoopScalarPreHeader;
520 /// Middle Block between the vector and the scalar.
521 BasicBlock *LoopMiddleBlock;
522 ///The ExitBlock of the scalar loop.
523 BasicBlock *LoopExitBlock;
524 ///The vector loop body.
525 SmallVector<BasicBlock *, 4> LoopVectorBody;
526 ///The scalar loop body.
527 BasicBlock *LoopScalarBody;
528 /// A list of all bypass blocks. The first block is the entry of the loop.
529 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
531 /// The new Induction variable which was added to the new block.
533 /// The induction variable of the old basic block.
534 PHINode *OldInduction;
535 /// Maps scalars to widened vectors.
537 /// Store instructions that should be predicated, as a pair
538 /// <StoreInst, Predicate>
539 SmallVector<std::pair<StoreInst*,Value*>, 4> PredicatedStores;
540 EdgeMaskCache MaskCache;
541 /// Trip count of the original loop.
543 /// Trip count of the widened loop (TripCount - TripCount % (VF*UF))
544 Value *VectorTripCount;
546 /// Map of scalar integer values to the smallest bitwidth they can be legally
547 /// represented as. The vector equivalents of these values should be truncated
549 DenseMap<Instruction*,uint64_t> MinBWs;
550 LoopVectorizationLegality *Legal;
552 // Record whether runtime check is added.
553 bool AddedSafetyChecks;
555 /// The SCEV predicate containing all the SCEV-related assumptions.
556 /// The predicate is used to simplify existing expressions in the
557 /// context of existing SCEV assumptions. Since legality checking is
558 /// not done here, we don't need to use this predicate to record
559 /// further assumptions.
560 SCEVUnionPredicate &Preds;
563 class InnerLoopUnroller : public InnerLoopVectorizer {
565 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
566 DominatorTree *DT, const TargetLibraryInfo *TLI,
567 const TargetTransformInfo *TTI, unsigned UnrollFactor,
568 SCEVUnionPredicate &Preds)
569 : InnerLoopVectorizer(OrigLoop, SE, LI, DT, TLI, TTI, 1, UnrollFactor,
573 void scalarizeInstruction(Instruction *Instr,
574 bool IfPredicateStore = false) override;
575 void vectorizeMemoryInstruction(Instruction *Instr) override;
576 Value *getBroadcastInstrs(Value *V) override;
577 Value *getStepVector(Value *Val, int StartIdx, Value *Step) override;
578 Value *reverseVector(Value *Vec) override;
581 /// \brief Look for a meaningful debug location on the instruction or it's
583 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
588 if (I->getDebugLoc() != Empty)
591 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
592 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
593 if (OpInst->getDebugLoc() != Empty)
600 /// \brief Set the debug location in the builder using the debug location in the
602 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
603 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
604 B.SetCurrentDebugLocation(Inst->getDebugLoc());
606 B.SetCurrentDebugLocation(DebugLoc());
610 /// \return string containing a file name and a line # for the given loop.
611 static std::string getDebugLocString(const Loop *L) {
614 raw_string_ostream OS(Result);
615 if (const DebugLoc LoopDbgLoc = L->getStartLoc())
616 LoopDbgLoc.print(OS);
618 // Just print the module name.
619 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
626 /// \brief Propagate known metadata from one instruction to another.
627 static void propagateMetadata(Instruction *To, const Instruction *From) {
628 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
629 From->getAllMetadataOtherThanDebugLoc(Metadata);
631 for (auto M : Metadata) {
632 unsigned Kind = M.first;
634 // These are safe to transfer (this is safe for TBAA, even when we
635 // if-convert, because should that metadata have had a control dependency
636 // on the condition, and thus actually aliased with some other
637 // non-speculated memory access when the condition was false, this would be
638 // caught by the runtime overlap checks).
639 if (Kind != LLVMContext::MD_tbaa &&
640 Kind != LLVMContext::MD_alias_scope &&
641 Kind != LLVMContext::MD_noalias &&
642 Kind != LLVMContext::MD_fpmath &&
643 Kind != LLVMContext::MD_nontemporal)
646 To->setMetadata(Kind, M.second);
650 /// \brief Propagate known metadata from one instruction to a vector of others.
651 static void propagateMetadata(SmallVectorImpl<Value *> &To,
652 const Instruction *From) {
654 if (Instruction *I = dyn_cast<Instruction>(V))
655 propagateMetadata(I, From);
658 /// \brief The group of interleaved loads/stores sharing the same stride and
659 /// close to each other.
661 /// Each member in this group has an index starting from 0, and the largest
662 /// index should be less than interleaved factor, which is equal to the absolute
663 /// value of the access's stride.
665 /// E.g. An interleaved load group of factor 4:
666 /// for (unsigned i = 0; i < 1024; i+=4) {
667 /// a = A[i]; // Member of index 0
668 /// b = A[i+1]; // Member of index 1
669 /// d = A[i+3]; // Member of index 3
673 /// An interleaved store group of factor 4:
674 /// for (unsigned i = 0; i < 1024; i+=4) {
676 /// A[i] = a; // Member of index 0
677 /// A[i+1] = b; // Member of index 1
678 /// A[i+2] = c; // Member of index 2
679 /// A[i+3] = d; // Member of index 3
682 /// Note: the interleaved load group could have gaps (missing members), but
683 /// the interleaved store group doesn't allow gaps.
684 class InterleaveGroup {
686 InterleaveGroup(Instruction *Instr, int Stride, unsigned Align)
687 : Align(Align), SmallestKey(0), LargestKey(0), InsertPos(Instr) {
688 assert(Align && "The alignment should be non-zero");
690 Factor = std::abs(Stride);
691 assert(Factor > 1 && "Invalid interleave factor");
693 Reverse = Stride < 0;
697 bool isReverse() const { return Reverse; }
698 unsigned getFactor() const { return Factor; }
699 unsigned getAlignment() const { return Align; }
700 unsigned getNumMembers() const { return Members.size(); }
702 /// \brief Try to insert a new member \p Instr with index \p Index and
703 /// alignment \p NewAlign. The index is related to the leader and it could be
704 /// negative if it is the new leader.
706 /// \returns false if the instruction doesn't belong to the group.
707 bool insertMember(Instruction *Instr, int Index, unsigned NewAlign) {
708 assert(NewAlign && "The new member's alignment should be non-zero");
710 int Key = Index + SmallestKey;
712 // Skip if there is already a member with the same index.
713 if (Members.count(Key))
716 if (Key > LargestKey) {
717 // The largest index is always less than the interleave factor.
718 if (Index >= static_cast<int>(Factor))
722 } else if (Key < SmallestKey) {
723 // The largest index is always less than the interleave factor.
724 if (LargestKey - Key >= static_cast<int>(Factor))
730 // It's always safe to select the minimum alignment.
731 Align = std::min(Align, NewAlign);
732 Members[Key] = Instr;
736 /// \brief Get the member with the given index \p Index
738 /// \returns nullptr if contains no such member.
739 Instruction *getMember(unsigned Index) const {
740 int Key = SmallestKey + Index;
741 if (!Members.count(Key))
744 return Members.find(Key)->second;
747 /// \brief Get the index for the given member. Unlike the key in the member
748 /// map, the index starts from 0.
749 unsigned getIndex(Instruction *Instr) const {
750 for (auto I : Members)
751 if (I.second == Instr)
752 return I.first - SmallestKey;
754 llvm_unreachable("InterleaveGroup contains no such member");
757 Instruction *getInsertPos() const { return InsertPos; }
758 void setInsertPos(Instruction *Inst) { InsertPos = Inst; }
761 unsigned Factor; // Interleave Factor.
764 DenseMap<int, Instruction *> Members;
768 // To avoid breaking dependences, vectorized instructions of an interleave
769 // group should be inserted at either the first load or the last store in
772 // E.g. %even = load i32 // Insert Position
773 // %add = add i32 %even // Use of %even
777 // %odd = add i32 // Def of %odd
778 // store i32 %odd // Insert Position
779 Instruction *InsertPos;
782 /// \brief Drive the analysis of interleaved memory accesses in the loop.
784 /// Use this class to analyze interleaved accesses only when we can vectorize
785 /// a loop. Otherwise it's meaningless to do analysis as the vectorization
786 /// on interleaved accesses is unsafe.
788 /// The analysis collects interleave groups and records the relationships
789 /// between the member and the group in a map.
790 class InterleavedAccessInfo {
792 InterleavedAccessInfo(ScalarEvolution *SE, Loop *L, DominatorTree *DT,
793 SCEVUnionPredicate &Preds)
794 : SE(SE), TheLoop(L), DT(DT), Preds(Preds) {}
796 ~InterleavedAccessInfo() {
797 SmallSet<InterleaveGroup *, 4> DelSet;
798 // Avoid releasing a pointer twice.
799 for (auto &I : InterleaveGroupMap)
800 DelSet.insert(I.second);
801 for (auto *Ptr : DelSet)
805 /// \brief Analyze the interleaved accesses and collect them in interleave
806 /// groups. Substitute symbolic strides using \p Strides.
807 void analyzeInterleaving(const ValueToValueMap &Strides);
809 /// \brief Check if \p Instr belongs to any interleave group.
810 bool isInterleaved(Instruction *Instr) const {
811 return InterleaveGroupMap.count(Instr);
814 /// \brief Get the interleave group that \p Instr belongs to.
816 /// \returns nullptr if doesn't have such group.
817 InterleaveGroup *getInterleaveGroup(Instruction *Instr) const {
818 if (InterleaveGroupMap.count(Instr))
819 return InterleaveGroupMap.find(Instr)->second;
828 /// The SCEV predicate containing all the SCEV-related assumptions.
829 /// The predicate is used to simplify SCEV expressions in the
830 /// context of existing SCEV assumptions. The interleaved access
831 /// analysis can also add new predicates (for example by versioning
832 /// strides of pointers).
833 SCEVUnionPredicate &Preds;
835 /// Holds the relationships between the members and the interleave group.
836 DenseMap<Instruction *, InterleaveGroup *> InterleaveGroupMap;
838 /// \brief The descriptor for a strided memory access.
839 struct StrideDescriptor {
840 StrideDescriptor(int Stride, const SCEV *Scev, unsigned Size,
842 : Stride(Stride), Scev(Scev), Size(Size), Align(Align) {}
844 StrideDescriptor() : Stride(0), Scev(nullptr), Size(0), Align(0) {}
846 int Stride; // The access's stride. It is negative for a reverse access.
847 const SCEV *Scev; // The scalar expression of this access
848 unsigned Size; // The size of the memory object.
849 unsigned Align; // The alignment of this access.
852 /// \brief Create a new interleave group with the given instruction \p Instr,
853 /// stride \p Stride and alignment \p Align.
855 /// \returns the newly created interleave group.
856 InterleaveGroup *createInterleaveGroup(Instruction *Instr, int Stride,
858 assert(!InterleaveGroupMap.count(Instr) &&
859 "Already in an interleaved access group");
860 InterleaveGroupMap[Instr] = new InterleaveGroup(Instr, Stride, Align);
861 return InterleaveGroupMap[Instr];
864 /// \brief Release the group and remove all the relationships.
865 void releaseGroup(InterleaveGroup *Group) {
866 for (unsigned i = 0; i < Group->getFactor(); i++)
867 if (Instruction *Member = Group->getMember(i))
868 InterleaveGroupMap.erase(Member);
873 /// \brief Collect all the accesses with a constant stride in program order.
874 void collectConstStridedAccesses(
875 MapVector<Instruction *, StrideDescriptor> &StrideAccesses,
876 const ValueToValueMap &Strides);
879 /// Utility class for getting and setting loop vectorizer hints in the form
880 /// of loop metadata.
881 /// This class keeps a number of loop annotations locally (as member variables)
882 /// and can, upon request, write them back as metadata on the loop. It will
883 /// initially scan the loop for existing metadata, and will update the local
884 /// values based on information in the loop.
885 /// We cannot write all values to metadata, as the mere presence of some info,
886 /// for example 'force', means a decision has been made. So, we need to be
887 /// careful NOT to add them if the user hasn't specifically asked so.
888 class LoopVectorizeHints {
895 /// Hint - associates name and validation with the hint value.
898 unsigned Value; // This may have to change for non-numeric values.
901 Hint(const char * Name, unsigned Value, HintKind Kind)
902 : Name(Name), Value(Value), Kind(Kind) { }
904 bool validate(unsigned Val) {
907 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
909 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
917 /// Vectorization width.
919 /// Vectorization interleave factor.
921 /// Vectorization forced
924 /// Return the loop metadata prefix.
925 static StringRef Prefix() { return "llvm.loop."; }
929 FK_Undefined = -1, ///< Not selected.
930 FK_Disabled = 0, ///< Forcing disabled.
931 FK_Enabled = 1, ///< Forcing enabled.
934 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
935 : Width("vectorize.width", VectorizerParams::VectorizationFactor,
937 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
938 Force("vectorize.enable", FK_Undefined, HK_FORCE),
940 // Populate values with existing loop metadata.
941 getHintsFromMetadata();
943 // force-vector-interleave overrides DisableInterleaving.
944 if (VectorizerParams::isInterleaveForced())
945 Interleave.Value = VectorizerParams::VectorizationInterleave;
947 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
948 << "LV: Interleaving disabled by the pass manager\n");
951 /// Mark the loop L as already vectorized by setting the width to 1.
952 void setAlreadyVectorized() {
953 Width.Value = Interleave.Value = 1;
954 Hint Hints[] = {Width, Interleave};
955 writeHintsToMetadata(Hints);
958 bool allowVectorization(Function *F, Loop *L, bool AlwaysVectorize) const {
959 if (getForce() == LoopVectorizeHints::FK_Disabled) {
960 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
961 emitOptimizationRemarkAnalysis(F->getContext(),
962 vectorizeAnalysisPassName(), *F,
963 L->getStartLoc(), emitRemark());
967 if (!AlwaysVectorize && getForce() != LoopVectorizeHints::FK_Enabled) {
968 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
969 emitOptimizationRemarkAnalysis(F->getContext(),
970 vectorizeAnalysisPassName(), *F,
971 L->getStartLoc(), emitRemark());
975 if (getWidth() == 1 && getInterleave() == 1) {
976 // FIXME: Add a separate metadata to indicate when the loop has already
977 // been vectorized instead of setting width and count to 1.
978 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
979 // FIXME: Add interleave.disable metadata. This will allow
980 // vectorize.disable to be used without disabling the pass and errors
981 // to differentiate between disabled vectorization and a width of 1.
982 emitOptimizationRemarkAnalysis(
983 F->getContext(), vectorizeAnalysisPassName(), *F, L->getStartLoc(),
984 "loop not vectorized: vectorization and interleaving are explicitly "
985 "disabled, or vectorize width and interleave count are both set to "
993 /// Dumps all the hint information.
994 std::string emitRemark() const {
995 VectorizationReport R;
996 if (Force.Value == LoopVectorizeHints::FK_Disabled)
997 R << "vectorization is explicitly disabled";
999 R << "use -Rpass-analysis=loop-vectorize for more info";
1000 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1001 R << " (Force=true";
1002 if (Width.Value != 0)
1003 R << ", Vector Width=" << Width.Value;
1004 if (Interleave.Value != 0)
1005 R << ", Interleave Count=" << Interleave.Value;
1013 unsigned getWidth() const { return Width.Value; }
1014 unsigned getInterleave() const { return Interleave.Value; }
1015 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1016 const char *vectorizeAnalysisPassName() const {
1017 // If hints are provided that don't disable vectorization use the
1018 // AlwaysPrint pass name to force the frontend to print the diagnostic.
1019 if (getWidth() == 1)
1021 if (getForce() == LoopVectorizeHints::FK_Disabled)
1023 if (getForce() == LoopVectorizeHints::FK_Undefined && getWidth() == 0)
1025 return DiagnosticInfo::AlwaysPrint;
1028 bool allowReordering() const {
1029 // When enabling loop hints are provided we allow the vectorizer to change
1030 // the order of operations that is given by the scalar loop. This is not
1031 // enabled by default because can be unsafe or inefficient. For example,
1032 // reordering floating-point operations will change the way round-off
1033 // error accumulates in the loop.
1034 return getForce() == LoopVectorizeHints::FK_Enabled || getWidth() > 1;
1038 /// Find hints specified in the loop metadata and update local values.
1039 void getHintsFromMetadata() {
1040 MDNode *LoopID = TheLoop->getLoopID();
1044 // First operand should refer to the loop id itself.
1045 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1046 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1048 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1049 const MDString *S = nullptr;
1050 SmallVector<Metadata *, 4> Args;
1052 // The expected hint is either a MDString or a MDNode with the first
1053 // operand a MDString.
1054 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1055 if (!MD || MD->getNumOperands() == 0)
1057 S = dyn_cast<MDString>(MD->getOperand(0));
1058 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1059 Args.push_back(MD->getOperand(i));
1061 S = dyn_cast<MDString>(LoopID->getOperand(i));
1062 assert(Args.size() == 0 && "too many arguments for MDString");
1068 // Check if the hint starts with the loop metadata prefix.
1069 StringRef Name = S->getString();
1070 if (Args.size() == 1)
1071 setHint(Name, Args[0]);
1075 /// Checks string hint with one operand and set value if valid.
1076 void setHint(StringRef Name, Metadata *Arg) {
1077 if (!Name.startswith(Prefix()))
1079 Name = Name.substr(Prefix().size(), StringRef::npos);
1081 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1083 unsigned Val = C->getZExtValue();
1085 Hint *Hints[] = {&Width, &Interleave, &Force};
1086 for (auto H : Hints) {
1087 if (Name == H->Name) {
1088 if (H->validate(Val))
1091 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1097 /// Create a new hint from name / value pair.
1098 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1099 LLVMContext &Context = TheLoop->getHeader()->getContext();
1100 Metadata *MDs[] = {MDString::get(Context, Name),
1101 ConstantAsMetadata::get(
1102 ConstantInt::get(Type::getInt32Ty(Context), V))};
1103 return MDNode::get(Context, MDs);
1106 /// Matches metadata with hint name.
1107 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1108 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1112 for (auto H : HintTypes)
1113 if (Name->getString().endswith(H.Name))
1118 /// Sets current hints into loop metadata, keeping other values intact.
1119 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1120 if (HintTypes.size() == 0)
1123 // Reserve the first element to LoopID (see below).
1124 SmallVector<Metadata *, 4> MDs(1);
1125 // If the loop already has metadata, then ignore the existing operands.
1126 MDNode *LoopID = TheLoop->getLoopID();
1128 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1129 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1130 // If node in update list, ignore old value.
1131 if (!matchesHintMetadataName(Node, HintTypes))
1132 MDs.push_back(Node);
1136 // Now, add the missing hints.
1137 for (auto H : HintTypes)
1138 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1140 // Replace current metadata node with new one.
1141 LLVMContext &Context = TheLoop->getHeader()->getContext();
1142 MDNode *NewLoopID = MDNode::get(Context, MDs);
1143 // Set operand 0 to refer to the loop id itself.
1144 NewLoopID->replaceOperandWith(0, NewLoopID);
1146 TheLoop->setLoopID(NewLoopID);
1149 /// The loop these hints belong to.
1150 const Loop *TheLoop;
1153 static void emitAnalysisDiag(const Function *TheFunction, const Loop *TheLoop,
1154 const LoopVectorizeHints &Hints,
1155 const LoopAccessReport &Message) {
1156 const char *Name = Hints.vectorizeAnalysisPassName();
1157 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, Name);
1160 static void emitMissedWarning(Function *F, Loop *L,
1161 const LoopVectorizeHints &LH) {
1162 emitOptimizationRemarkMissed(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1165 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1166 if (LH.getWidth() != 1)
1167 emitLoopVectorizeWarning(
1168 F->getContext(), *F, L->getStartLoc(),
1169 "failed explicitly specified loop vectorization");
1170 else if (LH.getInterleave() != 1)
1171 emitLoopInterleaveWarning(
1172 F->getContext(), *F, L->getStartLoc(),
1173 "failed explicitly specified loop interleaving");
1177 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
1178 /// to what vectorization factor.
1179 /// This class does not look at the profitability of vectorization, only the
1180 /// legality. This class has two main kinds of checks:
1181 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
1182 /// will change the order of memory accesses in a way that will change the
1183 /// correctness of the program.
1184 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
1185 /// checks for a number of different conditions, such as the availability of a
1186 /// single induction variable, that all types are supported and vectorize-able,
1187 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
1188 /// This class is also used by InnerLoopVectorizer for identifying
1189 /// induction variable and the different reduction variables.
1190 class LoopVectorizationLegality {
1192 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DominatorTree *DT,
1193 TargetLibraryInfo *TLI, AliasAnalysis *AA,
1194 Function *F, const TargetTransformInfo *TTI,
1195 LoopAccessAnalysis *LAA,
1196 LoopVectorizationRequirements *R,
1197 const LoopVectorizeHints *H,
1198 SCEVUnionPredicate &Preds)
1199 : NumPredStores(0), TheLoop(L), SE(SE), TLI(TLI), TheFunction(F),
1200 TTI(TTI), DT(DT), LAA(LAA), LAI(nullptr),
1201 InterleaveInfo(SE, L, DT, Preds), Induction(nullptr),
1202 WidestIndTy(nullptr), HasFunNoNaNAttr(false), Requirements(R), Hints(H),
1205 /// ReductionList contains the reduction descriptors for all
1206 /// of the reductions that were found in the loop.
1207 typedef DenseMap<PHINode *, RecurrenceDescriptor> ReductionList;
1209 /// InductionList saves induction variables and maps them to the
1210 /// induction descriptor.
1211 typedef MapVector<PHINode*, InductionDescriptor> InductionList;
1213 /// Returns true if it is legal to vectorize this loop.
1214 /// This does not mean that it is profitable to vectorize this
1215 /// loop, only that it is legal to do so.
1216 bool canVectorize();
1218 /// Returns the Induction variable.
1219 PHINode *getInduction() { return Induction; }
1221 /// Returns the reduction variables found in the loop.
1222 ReductionList *getReductionVars() { return &Reductions; }
1224 /// Returns the induction variables found in the loop.
1225 InductionList *getInductionVars() { return &Inductions; }
1227 /// Returns the widest induction type.
1228 Type *getWidestInductionType() { return WidestIndTy; }
1230 /// Returns True if V is an induction variable in this loop.
1231 bool isInductionVariable(const Value *V);
1233 /// Return true if the block BB needs to be predicated in order for the loop
1234 /// to be vectorized.
1235 bool blockNeedsPredication(BasicBlock *BB);
1237 /// Check if this pointer is consecutive when vectorizing. This happens
1238 /// when the last index of the GEP is the induction variable, or that the
1239 /// pointer itself is an induction variable.
1240 /// This check allows us to vectorize A[idx] into a wide load/store.
1242 /// 0 - Stride is unknown or non-consecutive.
1243 /// 1 - Address is consecutive.
1244 /// -1 - Address is consecutive, and decreasing.
1245 int isConsecutivePtr(Value *Ptr);
1247 /// Returns true if the value V is uniform within the loop.
1248 bool isUniform(Value *V);
1250 /// Returns true if this instruction will remain scalar after vectorization.
1251 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
1253 /// Returns the information that we collected about runtime memory check.
1254 const RuntimePointerChecking *getRuntimePointerChecking() const {
1255 return LAI->getRuntimePointerChecking();
1258 const LoopAccessInfo *getLAI() const {
1262 /// \brief Check if \p Instr belongs to any interleaved access group.
1263 bool isAccessInterleaved(Instruction *Instr) {
1264 return InterleaveInfo.isInterleaved(Instr);
1267 /// \brief Get the interleaved access group that \p Instr belongs to.
1268 const InterleaveGroup *getInterleavedAccessGroup(Instruction *Instr) {
1269 return InterleaveInfo.getInterleaveGroup(Instr);
1272 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
1274 bool hasStride(Value *V) { return StrideSet.count(V); }
1275 bool mustCheckStrides() { return !StrideSet.empty(); }
1276 SmallPtrSet<Value *, 8>::iterator strides_begin() {
1277 return StrideSet.begin();
1279 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
1281 /// Returns true if the target machine supports masked store operation
1282 /// for the given \p DataType and kind of access to \p Ptr.
1283 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
1284 return isConsecutivePtr(Ptr) && TTI->isLegalMaskedStore(DataType);
1286 /// Returns true if the target machine supports masked load operation
1287 /// for the given \p DataType and kind of access to \p Ptr.
1288 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
1289 return isConsecutivePtr(Ptr) && TTI->isLegalMaskedLoad(DataType);
1291 /// Returns true if vector representation of the instruction \p I
1293 bool isMaskRequired(const Instruction* I) {
1294 return (MaskedOp.count(I) != 0);
1296 unsigned getNumStores() const {
1297 return LAI->getNumStores();
1299 unsigned getNumLoads() const {
1300 return LAI->getNumLoads();
1302 unsigned getNumPredStores() const {
1303 return NumPredStores;
1306 /// Check if a single basic block loop is vectorizable.
1307 /// At this point we know that this is a loop with a constant trip count
1308 /// and we only need to check individual instructions.
1309 bool canVectorizeInstrs();
1311 /// When we vectorize loops we may change the order in which
1312 /// we read and write from memory. This method checks if it is
1313 /// legal to vectorize the code, considering only memory constrains.
1314 /// Returns true if the loop is vectorizable
1315 bool canVectorizeMemory();
1317 /// Return true if we can vectorize this loop using the IF-conversion
1319 bool canVectorizeWithIfConvert();
1321 /// Collect the variables that need to stay uniform after vectorization.
1322 void collectLoopUniforms();
1324 /// Return true if all of the instructions in the block can be speculatively
1325 /// executed. \p SafePtrs is a list of addresses that are known to be legal
1326 /// and we know that we can read from them without segfault.
1327 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
1329 /// \brief Collect memory access with loop invariant strides.
1331 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
1333 void collectStridedAccess(Value *LoadOrStoreInst);
1335 /// Report an analysis message to assist the user in diagnosing loops that are
1336 /// not vectorized. These are handled as LoopAccessReport rather than
1337 /// VectorizationReport because the << operator of VectorizationReport returns
1338 /// LoopAccessReport.
1339 void emitAnalysis(const LoopAccessReport &Message) const {
1340 emitAnalysisDiag(TheFunction, TheLoop, *Hints, Message);
1343 unsigned NumPredStores;
1345 /// The loop that we evaluate.
1348 ScalarEvolution *SE;
1349 /// Target Library Info.
1350 TargetLibraryInfo *TLI;
1352 Function *TheFunction;
1353 /// Target Transform Info
1354 const TargetTransformInfo *TTI;
1357 // LoopAccess analysis.
1358 LoopAccessAnalysis *LAA;
1359 // And the loop-accesses info corresponding to this loop. This pointer is
1360 // null until canVectorizeMemory sets it up.
1361 const LoopAccessInfo *LAI;
1363 /// The interleave access information contains groups of interleaved accesses
1364 /// with the same stride and close to each other.
1365 InterleavedAccessInfo InterleaveInfo;
1367 // --- vectorization state --- //
1369 /// Holds the integer induction variable. This is the counter of the
1372 /// Holds the reduction variables.
1373 ReductionList Reductions;
1374 /// Holds all of the induction variables that we found in the loop.
1375 /// Notice that inductions don't need to start at zero and that induction
1376 /// variables can be pointers.
1377 InductionList Inductions;
1378 /// Holds the widest induction type encountered.
1381 /// Allowed outside users. This holds the reduction
1382 /// vars which can be accessed from outside the loop.
1383 SmallPtrSet<Value*, 4> AllowedExit;
1384 /// This set holds the variables which are known to be uniform after
1386 SmallPtrSet<Instruction*, 4> Uniforms;
1388 /// Can we assume the absence of NaNs.
1389 bool HasFunNoNaNAttr;
1391 /// Vectorization requirements that will go through late-evaluation.
1392 LoopVectorizationRequirements *Requirements;
1394 /// Used to emit an analysis of any legality issues.
1395 const LoopVectorizeHints *Hints;
1397 ValueToValueMap Strides;
1398 SmallPtrSet<Value *, 8> StrideSet;
1400 /// While vectorizing these instructions we have to generate a
1401 /// call to the appropriate masked intrinsic
1402 SmallPtrSet<const Instruction *, 8> MaskedOp;
1404 /// The SCEV predicate containing all the SCEV-related assumptions.
1405 /// The predicate is used to simplify SCEV expressions in the
1406 /// context of existing SCEV assumptions. The analysis will also
1407 /// add a minimal set of new predicates if this is required to
1408 /// enable vectorization/unrolling.
1409 SCEVUnionPredicate &Preds;
1412 /// LoopVectorizationCostModel - estimates the expected speedups due to
1414 /// In many cases vectorization is not profitable. This can happen because of
1415 /// a number of reasons. In this class we mainly attempt to predict the
1416 /// expected speedup/slowdowns due to the supported instruction set. We use the
1417 /// TargetTransformInfo to query the different backends for the cost of
1418 /// different operations.
1419 class LoopVectorizationCostModel {
1421 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
1422 LoopVectorizationLegality *Legal,
1423 const TargetTransformInfo &TTI,
1424 const TargetLibraryInfo *TLI, DemandedBits *DB,
1425 AssumptionCache *AC, const Function *F,
1426 const LoopVectorizeHints *Hints,
1427 SmallPtrSetImpl<const Value *> &ValuesToIgnore,
1428 SCEVUnionPredicate &Preds)
1429 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), TLI(TLI), DB(DB),
1430 TheFunction(F), Hints(Hints), ValuesToIgnore(ValuesToIgnore) {}
1432 /// Information about vectorization costs
1433 struct VectorizationFactor {
1434 unsigned Width; // Vector width with best cost
1435 unsigned Cost; // Cost of the loop with that width
1437 /// \return The most profitable vectorization factor and the cost of that VF.
1438 /// This method checks every power of two up to VF. If UserVF is not ZERO
1439 /// then this vectorization factor will be selected if vectorization is
1441 VectorizationFactor selectVectorizationFactor(bool OptForSize);
1443 /// \return The size (in bits) of the smallest and widest types in the code
1444 /// that needs to be vectorized. We ignore values that remain scalar such as
1445 /// 64 bit loop indices.
1446 std::pair<unsigned, unsigned> getSmallestAndWidestTypes();
1448 /// \return The desired interleave count.
1449 /// If interleave count has been specified by metadata it will be returned.
1450 /// Otherwise, the interleave count is computed and returned. VF and LoopCost
1451 /// are the selected vectorization factor and the cost of the selected VF.
1452 unsigned selectInterleaveCount(bool OptForSize, unsigned VF,
1455 /// \return The most profitable unroll factor.
1456 /// This method finds the best unroll-factor based on register pressure and
1457 /// other parameters. VF and LoopCost are the selected vectorization factor
1458 /// and the cost of the selected VF.
1459 unsigned computeInterleaveCount(bool OptForSize, unsigned VF,
1462 /// \brief A struct that represents some properties of the register usage
1464 struct RegisterUsage {
1465 /// Holds the number of loop invariant values that are used in the loop.
1466 unsigned LoopInvariantRegs;
1467 /// Holds the maximum number of concurrent live intervals in the loop.
1468 unsigned MaxLocalUsers;
1469 /// Holds the number of instructions in the loop.
1470 unsigned NumInstructions;
1473 /// \return Returns information about the register usages of the loop for the
1474 /// given vectorization factors.
1475 SmallVector<RegisterUsage, 8>
1476 calculateRegisterUsage(const SmallVector<unsigned, 8> &VFs);
1479 /// Returns the expected execution cost. The unit of the cost does
1480 /// not matter because we use the 'cost' units to compare different
1481 /// vector widths. The cost that is returned is *not* normalized by
1482 /// the factor width.
1483 unsigned expectedCost(unsigned VF);
1485 /// Returns the execution time cost of an instruction for a given vector
1486 /// width. Vector width of one means scalar.
1487 unsigned getInstructionCost(Instruction *I, unsigned VF);
1489 /// Returns whether the instruction is a load or store and will be a emitted
1490 /// as a vector operation.
1491 bool isConsecutiveLoadOrStore(Instruction *I);
1493 /// Report an analysis message to assist the user in diagnosing loops that are
1494 /// not vectorized. These are handled as LoopAccessReport rather than
1495 /// VectorizationReport because the << operator of VectorizationReport returns
1496 /// LoopAccessReport.
1497 void emitAnalysis(const LoopAccessReport &Message) const {
1498 emitAnalysisDiag(TheFunction, TheLoop, *Hints, Message);
1502 /// Map of scalar integer values to the smallest bitwidth they can be legally
1503 /// represented as. The vector equivalents of these values should be truncated
1505 DenseMap<Instruction*,uint64_t> MinBWs;
1507 /// The loop that we evaluate.
1510 ScalarEvolution *SE;
1511 /// Loop Info analysis.
1513 /// Vectorization legality.
1514 LoopVectorizationLegality *Legal;
1515 /// Vector target information.
1516 const TargetTransformInfo &TTI;
1517 /// Target Library Info.
1518 const TargetLibraryInfo *TLI;
1519 /// Demanded bits analysis
1521 const Function *TheFunction;
1522 // Loop Vectorize Hint.
1523 const LoopVectorizeHints *Hints;
1524 // Values to ignore in the cost model.
1525 const SmallPtrSetImpl<const Value *> &ValuesToIgnore;
1528 /// \brief This holds vectorization requirements that must be verified late in
1529 /// the process. The requirements are set by legalize and costmodel. Once
1530 /// vectorization has been determined to be possible and profitable the
1531 /// requirements can be verified by looking for metadata or compiler options.
1532 /// For example, some loops require FP commutativity which is only allowed if
1533 /// vectorization is explicitly specified or if the fast-math compiler option
1534 /// has been provided.
1535 /// Late evaluation of these requirements allows helpful diagnostics to be
1536 /// composed that tells the user what need to be done to vectorize the loop. For
1537 /// example, by specifying #pragma clang loop vectorize or -ffast-math. Late
1538 /// evaluation should be used only when diagnostics can generated that can be
1539 /// followed by a non-expert user.
1540 class LoopVectorizationRequirements {
1542 LoopVectorizationRequirements()
1543 : NumRuntimePointerChecks(0), UnsafeAlgebraInst(nullptr) {}
1545 void addUnsafeAlgebraInst(Instruction *I) {
1546 // First unsafe algebra instruction.
1547 if (!UnsafeAlgebraInst)
1548 UnsafeAlgebraInst = I;
1551 void addRuntimePointerChecks(unsigned Num) { NumRuntimePointerChecks = Num; }
1553 bool doesNotMeet(Function *F, Loop *L, const LoopVectorizeHints &Hints) {
1554 const char *Name = Hints.vectorizeAnalysisPassName();
1555 bool Failed = false;
1556 if (UnsafeAlgebraInst && !Hints.allowReordering()) {
1557 emitOptimizationRemarkAnalysisFPCommute(
1558 F->getContext(), Name, *F, UnsafeAlgebraInst->getDebugLoc(),
1559 VectorizationReport() << "cannot prove it is safe to reorder "
1560 "floating-point operations");
1564 // Test if runtime memcheck thresholds are exceeded.
1565 bool PragmaThresholdReached =
1566 NumRuntimePointerChecks > PragmaVectorizeMemoryCheckThreshold;
1567 bool ThresholdReached =
1568 NumRuntimePointerChecks > VectorizerParams::RuntimeMemoryCheckThreshold;
1569 if ((ThresholdReached && !Hints.allowReordering()) ||
1570 PragmaThresholdReached) {
1571 emitOptimizationRemarkAnalysisAliasing(
1572 F->getContext(), Name, *F, L->getStartLoc(),
1573 VectorizationReport()
1574 << "cannot prove it is safe to reorder memory operations");
1575 DEBUG(dbgs() << "LV: Too many memory checks needed.\n");
1583 unsigned NumRuntimePointerChecks;
1584 Instruction *UnsafeAlgebraInst;
1587 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1589 return V.push_back(&L);
1591 for (Loop *InnerL : L)
1592 addInnerLoop(*InnerL, V);
1595 /// The LoopVectorize Pass.
1596 struct LoopVectorize : public FunctionPass {
1597 /// Pass identification, replacement for typeid
1600 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1602 DisableUnrolling(NoUnrolling),
1603 AlwaysVectorize(AlwaysVectorize) {
1604 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1607 ScalarEvolution *SE;
1609 TargetTransformInfo *TTI;
1611 BlockFrequencyInfo *BFI;
1612 TargetLibraryInfo *TLI;
1615 AssumptionCache *AC;
1616 LoopAccessAnalysis *LAA;
1617 bool DisableUnrolling;
1618 bool AlwaysVectorize;
1620 BlockFrequency ColdEntryFreq;
1622 bool runOnFunction(Function &F) override {
1623 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1624 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1625 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1626 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1627 BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI();
1628 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1629 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1630 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
1631 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1632 LAA = &getAnalysis<LoopAccessAnalysis>();
1633 DB = &getAnalysis<DemandedBits>();
1635 // Compute some weights outside of the loop over the loops. Compute this
1636 // using a BranchProbability to re-use its scaling math.
1637 const BranchProbability ColdProb(1, 5); // 20%
1638 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1641 // 1. the target claims to have no vector registers, and
1642 // 2. interleaving won't help ILP.
1644 // The second condition is necessary because, even if the target has no
1645 // vector registers, loop vectorization may still enable scalar
1647 if (!TTI->getNumberOfRegisters(true) && TTI->getMaxInterleaveFactor(1) < 2)
1650 // Build up a worklist of inner-loops to vectorize. This is necessary as
1651 // the act of vectorizing or partially unrolling a loop creates new loops
1652 // and can invalidate iterators across the loops.
1653 SmallVector<Loop *, 8> Worklist;
1656 addInnerLoop(*L, Worklist);
1658 LoopsAnalyzed += Worklist.size();
1660 // Now walk the identified inner loops.
1661 bool Changed = false;
1662 while (!Worklist.empty())
1663 Changed |= processLoop(Worklist.pop_back_val());
1665 // Process each loop nest in the function.
1669 static void AddRuntimeUnrollDisableMetaData(Loop *L) {
1670 SmallVector<Metadata *, 4> MDs;
1671 // Reserve first location for self reference to the LoopID metadata node.
1672 MDs.push_back(nullptr);
1673 bool IsUnrollMetadata = false;
1674 MDNode *LoopID = L->getLoopID();
1676 // First find existing loop unrolling disable metadata.
1677 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1678 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
1680 const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
1682 S && S->getString().startswith("llvm.loop.unroll.disable");
1684 MDs.push_back(LoopID->getOperand(i));
1688 if (!IsUnrollMetadata) {
1689 // Add runtime unroll disable metadata.
1690 LLVMContext &Context = L->getHeader()->getContext();
1691 SmallVector<Metadata *, 1> DisableOperands;
1692 DisableOperands.push_back(
1693 MDString::get(Context, "llvm.loop.unroll.runtime.disable"));
1694 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
1695 MDs.push_back(DisableNode);
1696 MDNode *NewLoopID = MDNode::get(Context, MDs);
1697 // Set operand 0 to refer to the loop id itself.
1698 NewLoopID->replaceOperandWith(0, NewLoopID);
1699 L->setLoopID(NewLoopID);
1703 bool processLoop(Loop *L) {
1704 assert(L->empty() && "Only process inner loops.");
1707 const std::string DebugLocStr = getDebugLocString(L);
1710 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1711 << L->getHeader()->getParent()->getName() << "\" from "
1712 << DebugLocStr << "\n");
1714 LoopVectorizeHints Hints(L, DisableUnrolling);
1716 DEBUG(dbgs() << "LV: Loop hints:"
1718 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1720 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1722 : "?")) << " width=" << Hints.getWidth()
1723 << " unroll=" << Hints.getInterleave() << "\n");
1725 // Function containing loop
1726 Function *F = L->getHeader()->getParent();
1728 // Looking at the diagnostic output is the only way to determine if a loop
1729 // was vectorized (other than looking at the IR or machine code), so it
1730 // is important to generate an optimization remark for each loop. Most of
1731 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1732 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1733 // less verbose reporting vectorized loops and unvectorized loops that may
1734 // benefit from vectorization, respectively.
1736 if (!Hints.allowVectorization(F, L, AlwaysVectorize)) {
1737 DEBUG(dbgs() << "LV: Loop hints prevent vectorization.\n");
1741 // Check the loop for a trip count threshold:
1742 // do not vectorize loops with a tiny trip count.
1743 const unsigned TC = SE->getSmallConstantTripCount(L);
1744 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1745 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1746 << "This loop is not worth vectorizing.");
1747 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1748 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1750 DEBUG(dbgs() << "\n");
1751 emitAnalysisDiag(F, L, Hints, VectorizationReport()
1752 << "vectorization is not beneficial "
1753 "and is not explicitly forced");
1758 SCEVUnionPredicate Preds;
1760 // Check if it is legal to vectorize the loop.
1761 LoopVectorizationRequirements Requirements;
1762 LoopVectorizationLegality LVL(L, SE, DT, TLI, AA, F, TTI, LAA,
1763 &Requirements, &Hints, Preds);
1764 if (!LVL.canVectorize()) {
1765 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1766 emitMissedWarning(F, L, Hints);
1770 // Collect values we want to ignore in the cost model. This includes
1771 // type-promoting instructions we identified during reduction detection.
1772 SmallPtrSet<const Value *, 32> ValuesToIgnore;
1773 CodeMetrics::collectEphemeralValues(L, AC, ValuesToIgnore);
1774 for (auto &Reduction : *LVL.getReductionVars()) {
1775 RecurrenceDescriptor &RedDes = Reduction.second;
1776 SmallPtrSetImpl<Instruction *> &Casts = RedDes.getCastInsts();
1777 ValuesToIgnore.insert(Casts.begin(), Casts.end());
1780 // Use the cost model.
1781 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, TLI, DB, AC, F, &Hints,
1782 ValuesToIgnore, Preds);
1784 // Check the function attributes to find out if this function should be
1785 // optimized for size.
1786 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1789 // Compute the weighted frequency of this loop being executed and see if it
1790 // is less than 20% of the function entry baseline frequency. Note that we
1791 // always have a canonical loop here because we think we *can* vectorize.
1792 // FIXME: This is hidden behind a flag due to pervasive problems with
1793 // exactly what block frequency models.
1794 if (LoopVectorizeWithBlockFrequency) {
1795 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1796 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1797 LoopEntryFreq < ColdEntryFreq)
1801 // Check the function attributes to see if implicit floats are allowed.
1802 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1803 // an integer loop and the vector instructions selected are purely integer
1804 // vector instructions?
1805 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1806 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1807 "attribute is used.\n");
1810 VectorizationReport()
1811 << "loop not vectorized due to NoImplicitFloat attribute");
1812 emitMissedWarning(F, L, Hints);
1816 // Select the optimal vectorization factor.
1817 const LoopVectorizationCostModel::VectorizationFactor VF =
1818 CM.selectVectorizationFactor(OptForSize);
1820 // Select the interleave count.
1821 unsigned IC = CM.selectInterleaveCount(OptForSize, VF.Width, VF.Cost);
1823 // Get user interleave count.
1824 unsigned UserIC = Hints.getInterleave();
1826 // Identify the diagnostic messages that should be produced.
1827 std::string VecDiagMsg, IntDiagMsg;
1828 bool VectorizeLoop = true, InterleaveLoop = true;
1830 if (Requirements.doesNotMeet(F, L, Hints)) {
1831 DEBUG(dbgs() << "LV: Not vectorizing: loop did not meet vectorization "
1833 emitMissedWarning(F, L, Hints);
1837 if (VF.Width == 1) {
1838 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
1840 "the cost-model indicates that vectorization is not beneficial";
1841 VectorizeLoop = false;
1844 if (IC == 1 && UserIC <= 1) {
1845 // Tell the user interleaving is not beneficial.
1846 DEBUG(dbgs() << "LV: Interleaving is not beneficial.\n");
1848 "the cost-model indicates that interleaving is not beneficial";
1849 InterleaveLoop = false;
1852 " and is explicitly disabled or interleave count is set to 1";
1853 } else if (IC > 1 && UserIC == 1) {
1854 // Tell the user interleaving is beneficial, but it explicitly disabled.
1856 << "LV: Interleaving is beneficial but is explicitly disabled.");
1857 IntDiagMsg = "the cost-model indicates that interleaving is beneficial "
1858 "but is explicitly disabled or interleave count is set to 1";
1859 InterleaveLoop = false;
1862 // Override IC if user provided an interleave count.
1863 IC = UserIC > 0 ? UserIC : IC;
1865 // Emit diagnostic messages, if any.
1866 const char *VAPassName = Hints.vectorizeAnalysisPassName();
1867 if (!VectorizeLoop && !InterleaveLoop) {
1868 // Do not vectorize or interleaving the loop.
1869 emitOptimizationRemarkAnalysis(F->getContext(), VAPassName, *F,
1870 L->getStartLoc(), VecDiagMsg);
1871 emitOptimizationRemarkAnalysis(F->getContext(), LV_NAME, *F,
1872 L->getStartLoc(), IntDiagMsg);
1874 } else if (!VectorizeLoop && InterleaveLoop) {
1875 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
1876 emitOptimizationRemarkAnalysis(F->getContext(), VAPassName, *F,
1877 L->getStartLoc(), VecDiagMsg);
1878 } else if (VectorizeLoop && !InterleaveLoop) {
1879 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1880 << DebugLocStr << '\n');
1881 emitOptimizationRemarkAnalysis(F->getContext(), LV_NAME, *F,
1882 L->getStartLoc(), IntDiagMsg);
1883 } else if (VectorizeLoop && InterleaveLoop) {
1884 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1885 << DebugLocStr << '\n');
1886 DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');
1889 if (!VectorizeLoop) {
1890 assert(IC > 1 && "interleave count should not be 1 or 0");
1891 // If we decided that it is not legal to vectorize the loop then
1893 InnerLoopUnroller Unroller(L, SE, LI, DT, TLI, TTI, IC, Preds);
1894 Unroller.vectorize(&LVL, CM.MinBWs);
1896 emitOptimizationRemark(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1897 Twine("interleaved loop (interleaved count: ") +
1900 // If we decided that it is *legal* to vectorize the loop then do it.
1901 InnerLoopVectorizer LB(L, SE, LI, DT, TLI, TTI, VF.Width, IC, Preds);
1902 LB.vectorize(&LVL, CM.MinBWs);
1905 // Add metadata to disable runtime unrolling scalar loop when there's no
1906 // runtime check about strides and memory. Because at this situation,
1907 // scalar loop is rarely used not worthy to be unrolled.
1908 if (!LB.IsSafetyChecksAdded())
1909 AddRuntimeUnrollDisableMetaData(L);
1911 // Report the vectorization decision.
1912 emitOptimizationRemark(F->getContext(), LV_NAME, *F, L->getStartLoc(),
1913 Twine("vectorized loop (vectorization width: ") +
1914 Twine(VF.Width) + ", interleaved count: " +
1918 // Mark the loop as already vectorized to avoid vectorizing again.
1919 Hints.setAlreadyVectorized();
1921 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1925 void getAnalysisUsage(AnalysisUsage &AU) const override {
1926 AU.addRequired<AssumptionCacheTracker>();
1927 AU.addRequiredID(LoopSimplifyID);
1928 AU.addRequiredID(LCSSAID);
1929 AU.addRequired<BlockFrequencyInfoWrapperPass>();
1930 AU.addRequired<DominatorTreeWrapperPass>();
1931 AU.addRequired<LoopInfoWrapperPass>();
1932 AU.addRequired<ScalarEvolutionWrapperPass>();
1933 AU.addRequired<TargetTransformInfoWrapperPass>();
1934 AU.addRequired<AAResultsWrapperPass>();
1935 AU.addRequired<LoopAccessAnalysis>();
1936 AU.addRequired<DemandedBits>();
1937 AU.addPreserved<LoopInfoWrapperPass>();
1938 AU.addPreserved<DominatorTreeWrapperPass>();
1939 AU.addPreserved<BasicAAWrapperPass>();
1940 AU.addPreserved<AAResultsWrapperPass>();
1941 AU.addPreserved<GlobalsAAWrapperPass>();
1946 } // end anonymous namespace
1948 //===----------------------------------------------------------------------===//
1949 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1950 // LoopVectorizationCostModel.
1951 //===----------------------------------------------------------------------===//
1953 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1954 // We need to place the broadcast of invariant variables outside the loop.
1955 Instruction *Instr = dyn_cast<Instruction>(V);
1957 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1958 Instr->getParent()) != LoopVectorBody.end());
1959 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1961 // Place the code for broadcasting invariant variables in the new preheader.
1962 IRBuilder<>::InsertPointGuard Guard(Builder);
1964 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1966 // Broadcast the scalar into all locations in the vector.
1967 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1972 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
1974 assert(Val->getType()->isVectorTy() && "Must be a vector");
1975 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1976 "Elem must be an integer");
1977 assert(Step->getType() == Val->getType()->getScalarType() &&
1978 "Step has wrong type");
1979 // Create the types.
1980 Type *ITy = Val->getType()->getScalarType();
1981 VectorType *Ty = cast<VectorType>(Val->getType());
1982 int VLen = Ty->getNumElements();
1983 SmallVector<Constant*, 8> Indices;
1985 // Create a vector of consecutive numbers from zero to VF.
1986 for (int i = 0; i < VLen; ++i)
1987 Indices.push_back(ConstantInt::get(ITy, StartIdx + i));
1989 // Add the consecutive indices to the vector value.
1990 Constant *Cv = ConstantVector::get(Indices);
1991 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1992 Step = Builder.CreateVectorSplat(VLen, Step);
1993 assert(Step->getType() == Val->getType() && "Invalid step vec");
1994 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
1995 // which can be found from the original scalar operations.
1996 Step = Builder.CreateMul(Cv, Step);
1997 return Builder.CreateAdd(Val, Step, "induction");
2000 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
2001 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
2002 // Make sure that the pointer does not point to structs.
2003 if (Ptr->getType()->getPointerElementType()->isAggregateType())
2006 // If this value is a pointer induction variable we know it is consecutive.
2007 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
2008 if (Phi && Inductions.count(Phi)) {
2009 InductionDescriptor II = Inductions[Phi];
2010 return II.getConsecutiveDirection();
2013 GetElementPtrInst *Gep = getGEPInstruction(Ptr);
2017 unsigned NumOperands = Gep->getNumOperands();
2018 Value *GpPtr = Gep->getPointerOperand();
2019 // If this GEP value is a consecutive pointer induction variable and all of
2020 // the indices are constant then we know it is consecutive. We can
2021 Phi = dyn_cast<PHINode>(GpPtr);
2022 if (Phi && Inductions.count(Phi)) {
2024 // Make sure that the pointer does not point to structs.
2025 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
2026 if (GepPtrType->getElementType()->isAggregateType())
2029 // Make sure that all of the index operands are loop invariant.
2030 for (unsigned i = 1; i < NumOperands; ++i)
2031 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
2034 InductionDescriptor II = Inductions[Phi];
2035 return II.getConsecutiveDirection();
2038 unsigned InductionOperand = getGEPInductionOperand(Gep);
2040 // Check that all of the gep indices are uniform except for our induction
2042 for (unsigned i = 0; i != NumOperands; ++i)
2043 if (i != InductionOperand &&
2044 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
2047 // We can emit wide load/stores only if the last non-zero index is the
2048 // induction variable.
2049 const SCEV *Last = nullptr;
2050 if (!Strides.count(Gep))
2051 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
2053 // Because of the multiplication by a stride we can have a s/zext cast.
2054 // We are going to replace this stride by 1 so the cast is safe to ignore.
2056 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
2057 // %0 = trunc i64 %indvars.iv to i32
2058 // %mul = mul i32 %0, %Stride1
2059 // %idxprom = zext i32 %mul to i64 << Safe cast.
2060 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
2062 Last = replaceSymbolicStrideSCEV(SE, Strides, Preds,
2063 Gep->getOperand(InductionOperand), Gep);
2064 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
2066 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
2070 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
2071 const SCEV *Step = AR->getStepRecurrence(*SE);
2073 // The memory is consecutive because the last index is consecutive
2074 // and all other indices are loop invariant.
2077 if (Step->isAllOnesValue())
2084 bool LoopVectorizationLegality::isUniform(Value *V) {
2085 return LAI->isUniform(V);
2088 InnerLoopVectorizer::VectorParts&
2089 InnerLoopVectorizer::getVectorValue(Value *V) {
2090 assert(V != Induction && "The new induction variable should not be used.");
2091 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
2093 // If we have a stride that is replaced by one, do it here.
2094 if (Legal->hasStride(V))
2095 V = ConstantInt::get(V->getType(), 1);
2097 // If we have this scalar in the map, return it.
2098 if (WidenMap.has(V))
2099 return WidenMap.get(V);
2101 // If this scalar is unknown, assume that it is a constant or that it is
2102 // loop invariant. Broadcast V and save the value for future uses.
2103 Value *B = getBroadcastInstrs(V);
2104 return WidenMap.splat(V, B);
2107 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
2108 assert(Vec->getType()->isVectorTy() && "Invalid type");
2109 SmallVector<Constant*, 8> ShuffleMask;
2110 for (unsigned i = 0; i < VF; ++i)
2111 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
2113 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
2114 ConstantVector::get(ShuffleMask),
2118 // Get a mask to interleave \p NumVec vectors into a wide vector.
2119 // I.e. <0, VF, VF*2, ..., VF*(NumVec-1), 1, VF+1, VF*2+1, ...>
2120 // E.g. For 2 interleaved vectors, if VF is 4, the mask is:
2121 // <0, 4, 1, 5, 2, 6, 3, 7>
2122 static Constant *getInterleavedMask(IRBuilder<> &Builder, unsigned VF,
2124 SmallVector<Constant *, 16> Mask;
2125 for (unsigned i = 0; i < VF; i++)
2126 for (unsigned j = 0; j < NumVec; j++)
2127 Mask.push_back(Builder.getInt32(j * VF + i));
2129 return ConstantVector::get(Mask);
2132 // Get the strided mask starting from index \p Start.
2133 // I.e. <Start, Start + Stride, ..., Start + Stride*(VF-1)>
2134 static Constant *getStridedMask(IRBuilder<> &Builder, unsigned Start,
2135 unsigned Stride, unsigned VF) {
2136 SmallVector<Constant *, 16> Mask;
2137 for (unsigned i = 0; i < VF; i++)
2138 Mask.push_back(Builder.getInt32(Start + i * Stride));
2140 return ConstantVector::get(Mask);
2143 // Get a mask of two parts: The first part consists of sequential integers
2144 // starting from 0, The second part consists of UNDEFs.
2145 // I.e. <0, 1, 2, ..., NumInt - 1, undef, ..., undef>
2146 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned NumInt,
2147 unsigned NumUndef) {
2148 SmallVector<Constant *, 16> Mask;
2149 for (unsigned i = 0; i < NumInt; i++)
2150 Mask.push_back(Builder.getInt32(i));
2152 Constant *Undef = UndefValue::get(Builder.getInt32Ty());
2153 for (unsigned i = 0; i < NumUndef; i++)
2154 Mask.push_back(Undef);
2156 return ConstantVector::get(Mask);
2159 // Concatenate two vectors with the same element type. The 2nd vector should
2160 // not have more elements than the 1st vector. If the 2nd vector has less
2161 // elements, extend it with UNDEFs.
2162 static Value *ConcatenateTwoVectors(IRBuilder<> &Builder, Value *V1,
2164 VectorType *VecTy1 = dyn_cast<VectorType>(V1->getType());
2165 VectorType *VecTy2 = dyn_cast<VectorType>(V2->getType());
2166 assert(VecTy1 && VecTy2 &&
2167 VecTy1->getScalarType() == VecTy2->getScalarType() &&
2168 "Expect two vectors with the same element type");
2170 unsigned NumElts1 = VecTy1->getNumElements();
2171 unsigned NumElts2 = VecTy2->getNumElements();
2172 assert(NumElts1 >= NumElts2 && "Unexpect the first vector has less elements");
2174 if (NumElts1 > NumElts2) {
2175 // Extend with UNDEFs.
2177 getSequentialMask(Builder, NumElts2, NumElts1 - NumElts2);
2178 V2 = Builder.CreateShuffleVector(V2, UndefValue::get(VecTy2), ExtMask);
2181 Constant *Mask = getSequentialMask(Builder, NumElts1 + NumElts2, 0);
2182 return Builder.CreateShuffleVector(V1, V2, Mask);
2185 // Concatenate vectors in the given list. All vectors have the same type.
2186 static Value *ConcatenateVectors(IRBuilder<> &Builder,
2187 ArrayRef<Value *> InputList) {
2188 unsigned NumVec = InputList.size();
2189 assert(NumVec > 1 && "Should be at least two vectors");
2191 SmallVector<Value *, 8> ResList;
2192 ResList.append(InputList.begin(), InputList.end());
2194 SmallVector<Value *, 8> TmpList;
2195 for (unsigned i = 0; i < NumVec - 1; i += 2) {
2196 Value *V0 = ResList[i], *V1 = ResList[i + 1];
2197 assert((V0->getType() == V1->getType() || i == NumVec - 2) &&
2198 "Only the last vector may have a different type");
2200 TmpList.push_back(ConcatenateTwoVectors(Builder, V0, V1));
2203 // Push the last vector if the total number of vectors is odd.
2204 if (NumVec % 2 != 0)
2205 TmpList.push_back(ResList[NumVec - 1]);
2208 NumVec = ResList.size();
2209 } while (NumVec > 1);
2214 // Try to vectorize the interleave group that \p Instr belongs to.
2216 // E.g. Translate following interleaved load group (factor = 3):
2217 // for (i = 0; i < N; i+=3) {
2218 // R = Pic[i]; // Member of index 0
2219 // G = Pic[i+1]; // Member of index 1
2220 // B = Pic[i+2]; // Member of index 2
2221 // ... // do something to R, G, B
2224 // %wide.vec = load <12 x i32> ; Read 4 tuples of R,G,B
2225 // %R.vec = shuffle %wide.vec, undef, <0, 3, 6, 9> ; R elements
2226 // %G.vec = shuffle %wide.vec, undef, <1, 4, 7, 10> ; G elements
2227 // %B.vec = shuffle %wide.vec, undef, <2, 5, 8, 11> ; B elements
2229 // Or translate following interleaved store group (factor = 3):
2230 // for (i = 0; i < N; i+=3) {
2231 // ... do something to R, G, B
2232 // Pic[i] = R; // Member of index 0
2233 // Pic[i+1] = G; // Member of index 1
2234 // Pic[i+2] = B; // Member of index 2
2237 // %R_G.vec = shuffle %R.vec, %G.vec, <0, 1, 2, ..., 7>
2238 // %B_U.vec = shuffle %B.vec, undef, <0, 1, 2, 3, u, u, u, u>
2239 // %interleaved.vec = shuffle %R_G.vec, %B_U.vec,
2240 // <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> ; Interleave R,G,B elements
2241 // store <12 x i32> %interleaved.vec ; Write 4 tuples of R,G,B
2242 void InnerLoopVectorizer::vectorizeInterleaveGroup(Instruction *Instr) {
2243 const InterleaveGroup *Group = Legal->getInterleavedAccessGroup(Instr);
2244 assert(Group && "Fail to get an interleaved access group.");
2246 // Skip if current instruction is not the insert position.
2247 if (Instr != Group->getInsertPos())
2250 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2251 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2252 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
2254 // Prepare for the vector type of the interleaved load/store.
2255 Type *ScalarTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2256 unsigned InterleaveFactor = Group->getFactor();
2257 Type *VecTy = VectorType::get(ScalarTy, InterleaveFactor * VF);
2258 Type *PtrTy = VecTy->getPointerTo(Ptr->getType()->getPointerAddressSpace());
2260 // Prepare for the new pointers.
2261 setDebugLocFromInst(Builder, Ptr);
2262 VectorParts &PtrParts = getVectorValue(Ptr);
2263 SmallVector<Value *, 2> NewPtrs;
2264 unsigned Index = Group->getIndex(Instr);
2265 for (unsigned Part = 0; Part < UF; Part++) {
2266 // Extract the pointer for current instruction from the pointer vector. A
2267 // reverse access uses the pointer in the last lane.
2268 Value *NewPtr = Builder.CreateExtractElement(
2270 Group->isReverse() ? Builder.getInt32(VF - 1) : Builder.getInt32(0));
2272 // Notice current instruction could be any index. Need to adjust the address
2273 // to the member of index 0.
2275 // E.g. a = A[i+1]; // Member of index 1 (Current instruction)
2276 // b = A[i]; // Member of index 0
2277 // Current pointer is pointed to A[i+1], adjust it to A[i].
2279 // E.g. A[i+1] = a; // Member of index 1
2280 // A[i] = b; // Member of index 0
2281 // A[i+2] = c; // Member of index 2 (Current instruction)
2282 // Current pointer is pointed to A[i+2], adjust it to A[i].
2283 NewPtr = Builder.CreateGEP(NewPtr, Builder.getInt32(-Index));
2285 // Cast to the vector pointer type.
2286 NewPtrs.push_back(Builder.CreateBitCast(NewPtr, PtrTy));
2289 setDebugLocFromInst(Builder, Instr);
2290 Value *UndefVec = UndefValue::get(VecTy);
2292 // Vectorize the interleaved load group.
2294 for (unsigned Part = 0; Part < UF; Part++) {
2295 Instruction *NewLoadInstr = Builder.CreateAlignedLoad(
2296 NewPtrs[Part], Group->getAlignment(), "wide.vec");
2298 for (unsigned i = 0; i < InterleaveFactor; i++) {
2299 Instruction *Member = Group->getMember(i);
2301 // Skip the gaps in the group.
2305 Constant *StrideMask = getStridedMask(Builder, i, InterleaveFactor, VF);
2306 Value *StridedVec = Builder.CreateShuffleVector(
2307 NewLoadInstr, UndefVec, StrideMask, "strided.vec");
2309 // If this member has different type, cast the result type.
2310 if (Member->getType() != ScalarTy) {
2311 VectorType *OtherVTy = VectorType::get(Member->getType(), VF);
2312 StridedVec = Builder.CreateBitOrPointerCast(StridedVec, OtherVTy);
2315 VectorParts &Entry = WidenMap.get(Member);
2317 Group->isReverse() ? reverseVector(StridedVec) : StridedVec;
2320 propagateMetadata(NewLoadInstr, Instr);
2325 // The sub vector type for current instruction.
2326 VectorType *SubVT = VectorType::get(ScalarTy, VF);
2328 // Vectorize the interleaved store group.
2329 for (unsigned Part = 0; Part < UF; Part++) {
2330 // Collect the stored vector from each member.
2331 SmallVector<Value *, 4> StoredVecs;
2332 for (unsigned i = 0; i < InterleaveFactor; i++) {
2333 // Interleaved store group doesn't allow a gap, so each index has a member
2334 Instruction *Member = Group->getMember(i);
2335 assert(Member && "Fail to get a member from an interleaved store group");
2338 getVectorValue(dyn_cast<StoreInst>(Member)->getValueOperand())[Part];
2339 if (Group->isReverse())
2340 StoredVec = reverseVector(StoredVec);
2342 // If this member has different type, cast it to an unified type.
2343 if (StoredVec->getType() != SubVT)
2344 StoredVec = Builder.CreateBitOrPointerCast(StoredVec, SubVT);
2346 StoredVecs.push_back(StoredVec);
2349 // Concatenate all vectors into a wide vector.
2350 Value *WideVec = ConcatenateVectors(Builder, StoredVecs);
2352 // Interleave the elements in the wide vector.
2353 Constant *IMask = getInterleavedMask(Builder, VF, InterleaveFactor);
2354 Value *IVec = Builder.CreateShuffleVector(WideVec, UndefVec, IMask,
2357 Instruction *NewStoreInstr =
2358 Builder.CreateAlignedStore(IVec, NewPtrs[Part], Group->getAlignment());
2359 propagateMetadata(NewStoreInstr, Instr);
2363 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
2364 // Attempt to issue a wide load.
2365 LoadInst *LI = dyn_cast<LoadInst>(Instr);
2366 StoreInst *SI = dyn_cast<StoreInst>(Instr);
2368 assert((LI || SI) && "Invalid Load/Store instruction");
2370 // Try to vectorize the interleave group if this access is interleaved.
2371 if (Legal->isAccessInterleaved(Instr))
2372 return vectorizeInterleaveGroup(Instr);
2374 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
2375 Type *DataTy = VectorType::get(ScalarDataTy, VF);
2376 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
2377 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
2378 // An alignment of 0 means target abi alignment. We need to use the scalar's
2379 // target abi alignment in such a case.
2380 const DataLayout &DL = Instr->getModule()->getDataLayout();
2382 Alignment = DL.getABITypeAlignment(ScalarDataTy);
2383 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
2384 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ScalarDataTy);
2385 unsigned VectorElementSize = DL.getTypeStoreSize(DataTy) / VF;
2387 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
2388 !Legal->isMaskRequired(SI))
2389 return scalarizeInstruction(Instr, true);
2391 if (ScalarAllocatedSize != VectorElementSize)
2392 return scalarizeInstruction(Instr);
2394 // If the pointer is loop invariant or if it is non-consecutive,
2395 // scalarize the load.
2396 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
2397 bool Reverse = ConsecutiveStride < 0;
2398 bool UniformLoad = LI && Legal->isUniform(Ptr);
2399 if (!ConsecutiveStride || UniformLoad)
2400 return scalarizeInstruction(Instr);
2402 Constant *Zero = Builder.getInt32(0);
2403 VectorParts &Entry = WidenMap.get(Instr);
2405 // Handle consecutive loads/stores.
2406 GetElementPtrInst *Gep = getGEPInstruction(Ptr);
2407 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
2408 setDebugLocFromInst(Builder, Gep);
2409 Value *PtrOperand = Gep->getPointerOperand();
2410 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
2411 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
2413 // Create the new GEP with the new induction variable.
2414 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2415 Gep2->setOperand(0, FirstBasePtr);
2416 Gep2->setName("gep.indvar.base");
2417 Ptr = Builder.Insert(Gep2);
2419 setDebugLocFromInst(Builder, Gep);
2420 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
2421 OrigLoop) && "Base ptr must be invariant");
2423 // The last index does not have to be the induction. It can be
2424 // consecutive and be a function of the index. For example A[I+1];
2425 unsigned NumOperands = Gep->getNumOperands();
2426 unsigned InductionOperand = getGEPInductionOperand(Gep);
2427 // Create the new GEP with the new induction variable.
2428 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
2430 for (unsigned i = 0; i < NumOperands; ++i) {
2431 Value *GepOperand = Gep->getOperand(i);
2432 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
2434 // Update last index or loop invariant instruction anchored in loop.
2435 if (i == InductionOperand ||
2436 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
2437 assert((i == InductionOperand ||
2438 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
2439 "Must be last index or loop invariant");
2441 VectorParts &GEPParts = getVectorValue(GepOperand);
2442 Value *Index = GEPParts[0];
2443 Index = Builder.CreateExtractElement(Index, Zero);
2444 Gep2->setOperand(i, Index);
2445 Gep2->setName("gep.indvar.idx");
2448 Ptr = Builder.Insert(Gep2);
2450 // Use the induction element ptr.
2451 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
2452 setDebugLocFromInst(Builder, Ptr);
2453 VectorParts &PtrVal = getVectorValue(Ptr);
2454 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
2457 VectorParts Mask = createBlockInMask(Instr->getParent());
2460 assert(!Legal->isUniform(SI->getPointerOperand()) &&
2461 "We do not allow storing to uniform addresses");
2462 setDebugLocFromInst(Builder, SI);
2463 // We don't want to update the value in the map as it might be used in
2464 // another expression. So don't use a reference type for "StoredVal".
2465 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
2467 for (unsigned Part = 0; Part < UF; ++Part) {
2468 // Calculate the pointer for the specific unroll-part.
2470 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2473 // If we store to reverse consecutive memory locations, then we need
2474 // to reverse the order of elements in the stored value.
2475 StoredVal[Part] = reverseVector(StoredVal[Part]);
2476 // If the address is consecutive but reversed, then the
2477 // wide store needs to start at the last vector element.
2478 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2479 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2480 Mask[Part] = reverseVector(Mask[Part]);
2483 Value *VecPtr = Builder.CreateBitCast(PartPtr,
2484 DataTy->getPointerTo(AddressSpace));
2487 if (Legal->isMaskRequired(SI))
2488 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
2491 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
2492 propagateMetadata(NewSI, SI);
2498 assert(LI && "Must have a load instruction");
2499 setDebugLocFromInst(Builder, LI);
2500 for (unsigned Part = 0; Part < UF; ++Part) {
2501 // Calculate the pointer for the specific unroll-part.
2503 Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
2506 // If the address is consecutive but reversed, then the
2507 // wide load needs to start at the last vector element.
2508 PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
2509 PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
2510 Mask[Part] = reverseVector(Mask[Part]);
2514 Value *VecPtr = Builder.CreateBitCast(PartPtr,
2515 DataTy->getPointerTo(AddressSpace));
2516 if (Legal->isMaskRequired(LI))
2517 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
2518 UndefValue::get(DataTy),
2519 "wide.masked.load");
2521 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
2522 propagateMetadata(NewLI, LI);
2523 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
2527 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr,
2528 bool IfPredicateStore) {
2529 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
2530 // Holds vector parameters or scalars, in case of uniform vals.
2531 SmallVector<VectorParts, 4> Params;
2533 setDebugLocFromInst(Builder, Instr);
2535 // Find all of the vectorized parameters.
2536 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2537 Value *SrcOp = Instr->getOperand(op);
2539 // If we are accessing the old induction variable, use the new one.
2540 if (SrcOp == OldInduction) {
2541 Params.push_back(getVectorValue(SrcOp));
2545 // Try using previously calculated values.
2546 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
2548 // If the src is an instruction that appeared earlier in the basic block,
2549 // then it should already be vectorized.
2550 if (SrcInst && OrigLoop->contains(SrcInst)) {
2551 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
2552 // The parameter is a vector value from earlier.
2553 Params.push_back(WidenMap.get(SrcInst));
2555 // The parameter is a scalar from outside the loop. Maybe even a constant.
2556 VectorParts Scalars;
2557 Scalars.append(UF, SrcOp);
2558 Params.push_back(Scalars);
2562 assert(Params.size() == Instr->getNumOperands() &&
2563 "Invalid number of operands");
2565 // Does this instruction return a value ?
2566 bool IsVoidRetTy = Instr->getType()->isVoidTy();
2568 Value *UndefVec = IsVoidRetTy ? nullptr :
2569 UndefValue::get(VectorType::get(Instr->getType(), VF));
2570 // Create a new entry in the WidenMap and initialize it to Undef or Null.
2571 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
2574 if (IfPredicateStore) {
2575 assert(Instr->getParent()->getSinglePredecessor() &&
2576 "Only support single predecessor blocks");
2577 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
2578 Instr->getParent());
2581 // For each vector unroll 'part':
2582 for (unsigned Part = 0; Part < UF; ++Part) {
2583 // For each scalar that we create:
2584 for (unsigned Width = 0; Width < VF; ++Width) {
2587 Value *Cmp = nullptr;
2588 if (IfPredicateStore) {
2589 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
2590 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp,
2591 ConstantInt::get(Cmp->getType(), 1));
2594 Instruction *Cloned = Instr->clone();
2596 Cloned->setName(Instr->getName() + ".cloned");
2597 // Replace the operands of the cloned instructions with extracted scalars.
2598 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2599 Value *Op = Params[op][Part];
2600 // Param is a vector. Need to extract the right lane.
2601 if (Op->getType()->isVectorTy())
2602 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
2603 Cloned->setOperand(op, Op);
2606 // Place the cloned scalar in the new loop.
2607 Builder.Insert(Cloned);
2609 // If the original scalar returns a value we need to place it in a vector
2610 // so that future users will be able to use it.
2612 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
2613 Builder.getInt32(Width));
2615 if (IfPredicateStore)
2616 PredicatedStores.push_back(std::make_pair(cast<StoreInst>(Cloned),
2622 PHINode *InnerLoopVectorizer::createInductionVariable(Loop *L, Value *Start,
2623 Value *End, Value *Step,
2625 BasicBlock *Header = L->getHeader();
2626 BasicBlock *Latch = L->getLoopLatch();
2627 // As we're just creating this loop, it's possible no latch exists
2628 // yet. If so, use the header as this will be a single block loop.
2632 IRBuilder<> Builder(&*Header->getFirstInsertionPt());
2633 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2634 auto *Induction = Builder.CreatePHI(Start->getType(), 2, "index");
2636 Builder.SetInsertPoint(Latch->getTerminator());
2638 // Create i+1 and fill the PHINode.
2639 Value *Next = Builder.CreateAdd(Induction, Step, "index.next");
2640 Induction->addIncoming(Start, L->getLoopPreheader());
2641 Induction->addIncoming(Next, Latch);
2642 // Create the compare.
2643 Value *ICmp = Builder.CreateICmpEQ(Next, End);
2644 Builder.CreateCondBr(ICmp, L->getExitBlock(), Header);
2646 // Now we have two terminators. Remove the old one from the block.
2647 Latch->getTerminator()->eraseFromParent();
2652 Value *InnerLoopVectorizer::getOrCreateTripCount(Loop *L) {
2656 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
2657 // Find the loop boundaries.
2658 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(OrigLoop);
2659 assert(BackedgeTakenCount != SE->getCouldNotCompute() &&
2660 "Invalid loop count");
2662 Type *IdxTy = Legal->getWidestInductionType();
2664 // The exit count might have the type of i64 while the phi is i32. This can
2665 // happen if we have an induction variable that is sign extended before the
2666 // compare. The only way that we get a backedge taken count is that the
2667 // induction variable was signed and as such will not overflow. In such a case
2668 // truncation is legal.
2669 if (BackedgeTakenCount->getType()->getPrimitiveSizeInBits() >
2670 IdxTy->getPrimitiveSizeInBits())
2671 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount, IdxTy);
2672 BackedgeTakenCount = SE->getNoopOrZeroExtend(BackedgeTakenCount, IdxTy);
2674 // Get the total trip count from the count by adding 1.
2675 const SCEV *ExitCount = SE->getAddExpr(
2676 BackedgeTakenCount, SE->getOne(BackedgeTakenCount->getType()));
2678 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
2680 // Expand the trip count and place the new instructions in the preheader.
2681 // Notice that the pre-header does not change, only the loop body.
2682 SCEVExpander Exp(*SE, DL, "induction");
2684 // Count holds the overall loop count (N).
2685 TripCount = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2686 L->getLoopPreheader()->getTerminator());
2688 if (TripCount->getType()->isPointerTy())
2690 CastInst::CreatePointerCast(TripCount, IdxTy,
2691 "exitcount.ptrcnt.to.int",
2692 L->getLoopPreheader()->getTerminator());
2697 Value *InnerLoopVectorizer::getOrCreateVectorTripCount(Loop *L) {
2698 if (VectorTripCount)
2699 return VectorTripCount;
2701 Value *TC = getOrCreateTripCount(L);
2702 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
2704 // Now we need to generate the expression for N - (N % VF), which is
2705 // the part that the vectorized body will execute.
2706 // The loop step is equal to the vectorization factor (num of SIMD elements)
2707 // times the unroll factor (num of SIMD instructions).
2708 Constant *Step = ConstantInt::get(TC->getType(), VF * UF);
2709 Value *R = Builder.CreateURem(TC, Step, "n.mod.vf");
2710 VectorTripCount = Builder.CreateSub(TC, R, "n.vec");
2712 return VectorTripCount;
2715 void InnerLoopVectorizer::emitMinimumIterationCountCheck(Loop *L,
2716 BasicBlock *Bypass) {
2717 Value *Count = getOrCreateTripCount(L);
2718 BasicBlock *BB = L->getLoopPreheader();
2719 IRBuilder<> Builder(BB->getTerminator());
2721 // Generate code to check that the loop's trip count that we computed by
2722 // adding one to the backedge-taken count will not overflow.
2723 Value *CheckMinIters =
2724 Builder.CreateICmpULT(Count,
2725 ConstantInt::get(Count->getType(), VF * UF),
2728 BasicBlock *NewBB = BB->splitBasicBlock(BB->getTerminator(),
2729 "min.iters.checked");
2730 if (L->getParentLoop())
2731 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
2732 ReplaceInstWithInst(BB->getTerminator(),
2733 BranchInst::Create(Bypass, NewBB, CheckMinIters));
2734 LoopBypassBlocks.push_back(BB);
2737 void InnerLoopVectorizer::emitVectorLoopEnteredCheck(Loop *L,
2738 BasicBlock *Bypass) {
2739 Value *TC = getOrCreateVectorTripCount(L);
2740 BasicBlock *BB = L->getLoopPreheader();
2741 IRBuilder<> Builder(BB->getTerminator());
2743 // Now, compare the new count to zero. If it is zero skip the vector loop and
2744 // jump to the scalar loop.
2745 Value *Cmp = Builder.CreateICmpEQ(TC, Constant::getNullValue(TC->getType()),
2748 // Generate code to check that the loop's trip count that we computed by
2749 // adding one to the backedge-taken count will not overflow.
2750 BasicBlock *NewBB = BB->splitBasicBlock(BB->getTerminator(),
2752 if (L->getParentLoop())
2753 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
2754 ReplaceInstWithInst(BB->getTerminator(),
2755 BranchInst::Create(Bypass, NewBB, Cmp));
2756 LoopBypassBlocks.push_back(BB);
2759 void InnerLoopVectorizer::emitSCEVChecks(Loop *L, BasicBlock *Bypass) {
2760 BasicBlock *BB = L->getLoopPreheader();
2762 // Generate the code to check that the SCEV assumptions that we made.
2763 // We want the new basic block to start at the first instruction in a
2764 // sequence of instructions that form a check.
2765 SCEVExpander Exp(*SE, Bypass->getModule()->getDataLayout(), "scev.check");
2766 Value *SCEVCheck = Exp.expandCodeForPredicate(&Preds, BB->getTerminator());
2768 if (auto *C = dyn_cast<ConstantInt>(SCEVCheck))
2772 // Create a new block containing the stride check.
2773 BB->setName("vector.scevcheck");
2774 auto *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
2775 if (L->getParentLoop())
2776 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
2777 ReplaceInstWithInst(BB->getTerminator(),
2778 BranchInst::Create(Bypass, NewBB, SCEVCheck));
2779 LoopBypassBlocks.push_back(BB);
2780 AddedSafetyChecks = true;
2783 void InnerLoopVectorizer::emitMemRuntimeChecks(Loop *L,
2784 BasicBlock *Bypass) {
2785 BasicBlock *BB = L->getLoopPreheader();
2787 // Generate the code that checks in runtime if arrays overlap. We put the
2788 // checks into a separate block to make the more common case of few elements
2790 Instruction *FirstCheckInst;
2791 Instruction *MemRuntimeCheck;
2792 std::tie(FirstCheckInst, MemRuntimeCheck) =
2793 Legal->getLAI()->addRuntimeChecks(BB->getTerminator());
2794 if (!MemRuntimeCheck)
2797 // Create a new block containing the memory check.
2798 BB->setName("vector.memcheck");
2799 auto *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
2800 if (L->getParentLoop())
2801 L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
2802 ReplaceInstWithInst(BB->getTerminator(),
2803 BranchInst::Create(Bypass, NewBB, MemRuntimeCheck));
2804 LoopBypassBlocks.push_back(BB);
2805 AddedSafetyChecks = true;
2809 void InnerLoopVectorizer::createEmptyLoop() {
2811 In this function we generate a new loop. The new loop will contain
2812 the vectorized instructions while the old loop will continue to run the
2815 [ ] <-- loop iteration number check.
2818 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2821 || [ ] <-- vector pre header.
2825 | [ ]_| <-- vector loop.
2828 | -[ ] <--- middle-block.
2831 -|- >[ ] <--- new preheader.
2835 | [ ]_| <-- old scalar loop to handle remainder.
2838 >[ ] <-- exit block.
2842 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2843 BasicBlock *VectorPH = OrigLoop->getLoopPreheader();
2844 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2845 assert(VectorPH && "Invalid loop structure");
2846 assert(ExitBlock && "Must have an exit block");
2848 // Some loops have a single integer induction variable, while other loops
2849 // don't. One example is c++ iterators that often have multiple pointer
2850 // induction variables. In the code below we also support a case where we
2851 // don't have a single induction variable.
2853 // We try to obtain an induction variable from the original loop as hard
2854 // as possible. However if we don't find one that:
2856 // - counts from zero, stepping by one
2857 // - is the size of the widest induction variable type
2858 // then we create a new one.
2859 OldInduction = Legal->getInduction();
2860 Type *IdxTy = Legal->getWidestInductionType();
2862 // Split the single block loop into the two loop structure described above.
2863 BasicBlock *VecBody =
2864 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2865 BasicBlock *MiddleBlock =
2866 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2867 BasicBlock *ScalarPH =
2868 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2870 // Create and register the new vector loop.
2871 Loop* Lp = new Loop();
2872 Loop *ParentLoop = OrigLoop->getParentLoop();
2874 // Insert the new loop into the loop nest and register the new basic blocks
2875 // before calling any utilities such as SCEV that require valid LoopInfo.
2877 ParentLoop->addChildLoop(Lp);
2878 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
2879 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
2881 LI->addTopLevelLoop(Lp);
2883 Lp->addBasicBlockToLoop(VecBody, *LI);
2885 // Find the loop boundaries.
2886 Value *Count = getOrCreateTripCount(Lp);
2888 Value *StartIdx = ConstantInt::get(IdxTy, 0);
2890 // We need to test whether the backedge-taken count is uint##_max. Adding one
2891 // to it will cause overflow and an incorrect loop trip count in the vector
2892 // body. In case of overflow we want to directly jump to the scalar remainder
2894 emitMinimumIterationCountCheck(Lp, ScalarPH);
2895 // Now, compare the new count to zero. If it is zero skip the vector loop and
2896 // jump to the scalar loop.
2897 emitVectorLoopEnteredCheck(Lp, ScalarPH);
2898 // Generate the code to check any assumptions that we've made for SCEV
2900 emitSCEVChecks(Lp, ScalarPH);
2902 // Generate the code that checks in runtime if arrays overlap. We put the
2903 // checks into a separate block to make the more common case of few elements
2905 emitMemRuntimeChecks(Lp, ScalarPH);
2907 // Generate the induction variable.
2908 // The loop step is equal to the vectorization factor (num of SIMD elements)
2909 // times the unroll factor (num of SIMD instructions).
2910 Value *CountRoundDown = getOrCreateVectorTripCount(Lp);
2911 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2913 createInductionVariable(Lp, StartIdx, CountRoundDown, Step,
2914 getDebugLocFromInstOrOperands(OldInduction));
2916 // We are going to resume the execution of the scalar loop.
2917 // Go over all of the induction variables that we found and fix the
2918 // PHIs that are left in the scalar version of the loop.
2919 // The starting values of PHI nodes depend on the counter of the last
2920 // iteration in the vectorized loop.
2921 // If we come from a bypass edge then we need to start from the original
2924 // This variable saves the new starting index for the scalar loop. It is used
2925 // to test if there are any tail iterations left once the vector loop has
2927 LoopVectorizationLegality::InductionList::iterator I, E;
2928 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2929 for (I = List->begin(), E = List->end(); I != E; ++I) {
2930 PHINode *OrigPhi = I->first;
2931 InductionDescriptor II = I->second;
2933 // Create phi nodes to merge from the backedge-taken check block.
2934 PHINode *BCResumeVal = PHINode::Create(OrigPhi->getType(), 3,
2936 ScalarPH->getTerminator());
2938 if (OrigPhi == OldInduction) {
2939 // We know what the end value is.
2940 EndValue = CountRoundDown;
2942 IRBuilder<> B(LoopBypassBlocks.back()->getTerminator());
2943 Value *CRD = B.CreateSExtOrTrunc(CountRoundDown,
2944 II.getStepValue()->getType(),
2946 EndValue = II.transform(B, CRD);
2947 EndValue->setName("ind.end");
2950 // The new PHI merges the original incoming value, in case of a bypass,
2951 // or the value at the end of the vectorized loop.
2952 BCResumeVal->addIncoming(EndValue, MiddleBlock);
2954 // Fix the scalar body counter (PHI node).
2955 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2957 // The old induction's phi node in the scalar body needs the truncated
2959 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
2960 BCResumeVal->addIncoming(II.getStartValue(), LoopBypassBlocks[I]);
2961 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2964 // Add a check in the middle block to see if we have completed
2965 // all of the iterations in the first vector loop.
2966 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2967 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, Count,
2968 CountRoundDown, "cmp.n",
2969 MiddleBlock->getTerminator());
2970 ReplaceInstWithInst(MiddleBlock->getTerminator(),
2971 BranchInst::Create(ExitBlock, ScalarPH, CmpN));
2973 // Get ready to start creating new instructions into the vectorized body.
2974 Builder.SetInsertPoint(&*VecBody->getFirstInsertionPt());
2977 LoopVectorPreHeader = Lp->getLoopPreheader();
2978 LoopScalarPreHeader = ScalarPH;
2979 LoopMiddleBlock = MiddleBlock;
2980 LoopExitBlock = ExitBlock;
2981 LoopVectorBody.push_back(VecBody);
2982 LoopScalarBody = OldBasicBlock;
2984 LoopVectorizeHints Hints(Lp, true);
2985 Hints.setAlreadyVectorized();
2989 struct CSEDenseMapInfo {
2990 static bool canHandle(Instruction *I) {
2991 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2992 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2994 static inline Instruction *getEmptyKey() {
2995 return DenseMapInfo<Instruction *>::getEmptyKey();
2997 static inline Instruction *getTombstoneKey() {
2998 return DenseMapInfo<Instruction *>::getTombstoneKey();
3000 static unsigned getHashValue(Instruction *I) {
3001 assert(canHandle(I) && "Unknown instruction!");
3002 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
3003 I->value_op_end()));
3005 static bool isEqual(Instruction *LHS, Instruction *RHS) {
3006 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
3007 LHS == getTombstoneKey() || RHS == getTombstoneKey())
3009 return LHS->isIdenticalTo(RHS);
3014 /// \brief Check whether this block is a predicated block.
3015 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
3016 /// = ...; " blocks. We start with one vectorized basic block. For every
3017 /// conditional block we split this vectorized block. Therefore, every second
3018 /// block will be a predicated one.
3019 static bool isPredicatedBlock(unsigned BlockNum) {
3020 return BlockNum % 2;
3023 ///\brief Perform cse of induction variable instructions.
3024 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
3025 // Perform simple cse.
3026 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
3027 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
3028 BasicBlock *BB = BBs[i];
3029 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
3030 Instruction *In = &*I++;
3032 if (!CSEDenseMapInfo::canHandle(In))
3035 // Check if we can replace this instruction with any of the
3036 // visited instructions.
3037 if (Instruction *V = CSEMap.lookup(In)) {
3038 In->replaceAllUsesWith(V);
3039 In->eraseFromParent();
3042 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
3043 // ...;" blocks for predicated stores. Every second block is a predicated
3045 if (isPredicatedBlock(i))
3053 /// \brief Adds a 'fast' flag to floating point operations.
3054 static Value *addFastMathFlag(Value *V) {
3055 if (isa<FPMathOperator>(V)){
3056 FastMathFlags Flags;
3057 Flags.setUnsafeAlgebra();
3058 cast<Instruction>(V)->setFastMathFlags(Flags);
3063 /// Estimate the overhead of scalarizing a value. Insert and Extract are set if
3064 /// the result needs to be inserted and/or extracted from vectors.
3065 static unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract,
3066 const TargetTransformInfo &TTI) {
3070 assert(Ty->isVectorTy() && "Can only scalarize vectors");
3073 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
3075 Cost += TTI.getVectorInstrCost(Instruction::InsertElement, Ty, i);
3077 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, Ty, i);
3083 // Estimate cost of a call instruction CI if it were vectorized with factor VF.
3084 // Return the cost of the instruction, including scalarization overhead if it's
3085 // needed. The flag NeedToScalarize shows if the call needs to be scalarized -
3086 // i.e. either vector version isn't available, or is too expensive.
3087 static unsigned getVectorCallCost(CallInst *CI, unsigned VF,
3088 const TargetTransformInfo &TTI,
3089 const TargetLibraryInfo *TLI,
3090 bool &NeedToScalarize) {
3091 Function *F = CI->getCalledFunction();
3092 StringRef FnName = CI->getCalledFunction()->getName();
3093 Type *ScalarRetTy = CI->getType();
3094 SmallVector<Type *, 4> Tys, ScalarTys;
3095 for (auto &ArgOp : CI->arg_operands())
3096 ScalarTys.push_back(ArgOp->getType());
3098 // Estimate cost of scalarized vector call. The source operands are assumed
3099 // to be vectors, so we need to extract individual elements from there,
3100 // execute VF scalar calls, and then gather the result into the vector return
3102 unsigned ScalarCallCost = TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys);
3104 return ScalarCallCost;
3106 // Compute corresponding vector type for return value and arguments.
3107 Type *RetTy = ToVectorTy(ScalarRetTy, VF);
3108 for (unsigned i = 0, ie = ScalarTys.size(); i != ie; ++i)
3109 Tys.push_back(ToVectorTy(ScalarTys[i], VF));
3111 // Compute costs of unpacking argument values for the scalar calls and
3112 // packing the return values to a vector.
3113 unsigned ScalarizationCost =
3114 getScalarizationOverhead(RetTy, true, false, TTI);
3115 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i)
3116 ScalarizationCost += getScalarizationOverhead(Tys[i], false, true, TTI);
3118 unsigned Cost = ScalarCallCost * VF + ScalarizationCost;
3120 // If we can't emit a vector call for this function, then the currently found
3121 // cost is the cost we need to return.
3122 NeedToScalarize = true;
3123 if (!TLI || !TLI->isFunctionVectorizable(FnName, VF) || CI->isNoBuiltin())
3126 // If the corresponding vector cost is cheaper, return its cost.
3127 unsigned VectorCallCost = TTI.getCallInstrCost(nullptr, RetTy, Tys);
3128 if (VectorCallCost < Cost) {
3129 NeedToScalarize = false;
3130 return VectorCallCost;
3135 // Estimate cost of an intrinsic call instruction CI if it were vectorized with
3136 // factor VF. Return the cost of the instruction, including scalarization
3137 // overhead if it's needed.
3138 static unsigned getVectorIntrinsicCost(CallInst *CI, unsigned VF,
3139 const TargetTransformInfo &TTI,
3140 const TargetLibraryInfo *TLI) {
3141 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3142 assert(ID && "Expected intrinsic call!");
3144 Type *RetTy = ToVectorTy(CI->getType(), VF);
3145 SmallVector<Type *, 4> Tys;
3146 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3147 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3149 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
3152 static Type *smallestIntegerVectorType(Type *T1, Type *T2) {
3153 IntegerType *I1 = cast<IntegerType>(T1->getVectorElementType());
3154 IntegerType *I2 = cast<IntegerType>(T2->getVectorElementType());
3155 return I1->getBitWidth() < I2->getBitWidth() ? T1 : T2;
3157 static Type *largestIntegerVectorType(Type *T1, Type *T2) {
3158 IntegerType *I1 = cast<IntegerType>(T1->getVectorElementType());
3159 IntegerType *I2 = cast<IntegerType>(T2->getVectorElementType());
3160 return I1->getBitWidth() > I2->getBitWidth() ? T1 : T2;
3163 void InnerLoopVectorizer::truncateToMinimalBitwidths() {
3164 // For every instruction `I` in MinBWs, truncate the operands, create a
3165 // truncated version of `I` and reextend its result. InstCombine runs
3166 // later and will remove any ext/trunc pairs.
3168 for (auto &KV : MinBWs) {
3169 VectorParts &Parts = WidenMap.get(KV.first);
3170 for (Value *&I : Parts) {
3173 Type *OriginalTy = I->getType();
3174 Type *ScalarTruncatedTy = IntegerType::get(OriginalTy->getContext(),
3176 Type *TruncatedTy = VectorType::get(ScalarTruncatedTy,
3177 OriginalTy->getVectorNumElements());
3178 if (TruncatedTy == OriginalTy)
3181 IRBuilder<> B(cast<Instruction>(I));
3182 auto ShrinkOperand = [&](Value *V) -> Value* {
3183 if (auto *ZI = dyn_cast<ZExtInst>(V))
3184 if (ZI->getSrcTy() == TruncatedTy)
3185 return ZI->getOperand(0);
3186 return B.CreateZExtOrTrunc(V, TruncatedTy);
3189 // The actual instruction modification depends on the instruction type,
3191 Value *NewI = nullptr;
3192 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
3193 NewI = B.CreateBinOp(BO->getOpcode(),
3194 ShrinkOperand(BO->getOperand(0)),
3195 ShrinkOperand(BO->getOperand(1)));
3196 cast<BinaryOperator>(NewI)->copyIRFlags(I);
3197 } else if (ICmpInst *CI = dyn_cast<ICmpInst>(I)) {
3198 NewI = B.CreateICmp(CI->getPredicate(),
3199 ShrinkOperand(CI->getOperand(0)),
3200 ShrinkOperand(CI->getOperand(1)));
3201 } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
3202 NewI = B.CreateSelect(SI->getCondition(),
3203 ShrinkOperand(SI->getTrueValue()),
3204 ShrinkOperand(SI->getFalseValue()));
3205 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
3206 switch (CI->getOpcode()) {
3207 default: llvm_unreachable("Unhandled cast!");
3208 case Instruction::Trunc:
3209 NewI = ShrinkOperand(CI->getOperand(0));
3211 case Instruction::SExt:
3212 NewI = B.CreateSExtOrTrunc(CI->getOperand(0),
3213 smallestIntegerVectorType(OriginalTy,
3216 case Instruction::ZExt:
3217 NewI = B.CreateZExtOrTrunc(CI->getOperand(0),
3218 smallestIntegerVectorType(OriginalTy,
3222 } else if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(I)) {
3223 auto Elements0 = SI->getOperand(0)->getType()->getVectorNumElements();
3225 B.CreateZExtOrTrunc(SI->getOperand(0),
3226 VectorType::get(ScalarTruncatedTy, Elements0));
3227 auto Elements1 = SI->getOperand(1)->getType()->getVectorNumElements();
3229 B.CreateZExtOrTrunc(SI->getOperand(1),
3230 VectorType::get(ScalarTruncatedTy, Elements1));
3232 NewI = B.CreateShuffleVector(O0, O1, SI->getMask());
3233 } else if (isa<LoadInst>(I)) {
3234 // Don't do anything with the operands, just extend the result.
3237 llvm_unreachable("Unhandled instruction type!");
3240 // Lastly, extend the result.
3241 NewI->takeName(cast<Instruction>(I));
3242 Value *Res = B.CreateZExtOrTrunc(NewI, OriginalTy);
3243 I->replaceAllUsesWith(Res);
3244 cast<Instruction>(I)->eraseFromParent();
3249 // We'll have created a bunch of ZExts that are now parentless. Clean up.
3250 for (auto &KV : MinBWs) {
3251 VectorParts &Parts = WidenMap.get(KV.first);
3252 for (Value *&I : Parts) {
3253 ZExtInst *Inst = dyn_cast<ZExtInst>(I);
3254 if (Inst && Inst->use_empty()) {
3255 Value *NewI = Inst->getOperand(0);
3256 Inst->eraseFromParent();
3263 void InnerLoopVectorizer::vectorizeLoop() {
3264 //===------------------------------------------------===//
3266 // Notice: any optimization or new instruction that go
3267 // into the code below should be also be implemented in
3270 //===------------------------------------------------===//
3271 Constant *Zero = Builder.getInt32(0);
3273 // In order to support reduction variables we need to be able to vectorize
3274 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
3275 // stages. First, we create a new vector PHI node with no incoming edges.
3276 // We use this value when we vectorize all of the instructions that use the
3277 // PHI. Next, after all of the instructions in the block are complete we
3278 // add the new incoming edges to the PHI. At this point all of the
3279 // instructions in the basic block are vectorized, so we can use them to
3280 // construct the PHI.
3281 PhiVector RdxPHIsToFix;
3283 // Scan the loop in a topological order to ensure that defs are vectorized
3285 LoopBlocksDFS DFS(OrigLoop);
3288 // Vectorize all of the blocks in the original loop.
3289 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
3290 be = DFS.endRPO(); bb != be; ++bb)
3291 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
3293 // Insert truncates and extends for any truncated instructions as hints to
3296 truncateToMinimalBitwidths();
3298 // At this point every instruction in the original loop is widened to
3299 // a vector form. We are almost done. Now, we need to fix the PHI nodes
3300 // that we vectorized. The PHI nodes are currently empty because we did
3301 // not want to introduce cycles. Notice that the remaining PHI nodes
3302 // that we need to fix are reduction variables.
3304 // Create the 'reduced' values for each of the induction vars.
3305 // The reduced values are the vector values that we scalarize and combine
3306 // after the loop is finished.
3307 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
3309 PHINode *RdxPhi = *it;
3310 assert(RdxPhi && "Unable to recover vectorized PHI");
3312 // Find the reduction variable descriptor.
3313 assert(Legal->getReductionVars()->count(RdxPhi) &&
3314 "Unable to find the reduction variable");
3315 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[RdxPhi];
3317 RecurrenceDescriptor::RecurrenceKind RK = RdxDesc.getRecurrenceKind();
3318 TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue();
3319 Instruction *LoopExitInst = RdxDesc.getLoopExitInstr();
3320 RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind =
3321 RdxDesc.getMinMaxRecurrenceKind();
3322 setDebugLocFromInst(Builder, ReductionStartValue);
3324 // We need to generate a reduction vector from the incoming scalar.
3325 // To do so, we need to generate the 'identity' vector and override
3326 // one of the elements with the incoming scalar reduction. We need
3327 // to do it in the vector-loop preheader.
3328 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
3330 // This is the vector-clone of the value that leaves the loop.
3331 VectorParts &VectorExit = getVectorValue(LoopExitInst);
3332 Type *VecTy = VectorExit[0]->getType();
3334 // Find the reduction identity variable. Zero for addition, or, xor,
3335 // one for multiplication, -1 for And.
3338 if (RK == RecurrenceDescriptor::RK_IntegerMinMax ||
3339 RK == RecurrenceDescriptor::RK_FloatMinMax) {
3340 // MinMax reduction have the start value as their identify.
3342 VectorStart = Identity = ReductionStartValue;
3344 VectorStart = Identity =
3345 Builder.CreateVectorSplat(VF, ReductionStartValue, "minmax.ident");
3348 // Handle other reduction kinds:
3349 Constant *Iden = RecurrenceDescriptor::getRecurrenceIdentity(
3350 RK, VecTy->getScalarType());
3353 // This vector is the Identity vector where the first element is the
3354 // incoming scalar reduction.
3355 VectorStart = ReductionStartValue;
3357 Identity = ConstantVector::getSplat(VF, Iden);
3359 // This vector is the Identity vector where the first element is the
3360 // incoming scalar reduction.
3362 Builder.CreateInsertElement(Identity, ReductionStartValue, Zero);
3366 // Fix the vector-loop phi.
3368 // Reductions do not have to start at zero. They can start with
3369 // any loop invariant values.
3370 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
3371 BasicBlock *Latch = OrigLoop->getLoopLatch();
3372 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
3373 VectorParts &Val = getVectorValue(LoopVal);
3374 for (unsigned part = 0; part < UF; ++part) {
3375 // Make sure to add the reduction stat value only to the
3376 // first unroll part.
3377 Value *StartVal = (part == 0) ? VectorStart : Identity;
3378 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
3379 LoopVectorPreHeader);
3380 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
3381 LoopVectorBody.back());
3384 // Before each round, move the insertion point right between
3385 // the PHIs and the values we are going to write.
3386 // This allows us to write both PHINodes and the extractelement
3388 Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt());
3390 VectorParts RdxParts = getVectorValue(LoopExitInst);
3391 setDebugLocFromInst(Builder, LoopExitInst);
3393 // If the vector reduction can be performed in a smaller type, we truncate
3394 // then extend the loop exit value to enable InstCombine to evaluate the
3395 // entire expression in the smaller type.
3396 if (VF > 1 && RdxPhi->getType() != RdxDesc.getRecurrenceType()) {
3397 Type *RdxVecTy = VectorType::get(RdxDesc.getRecurrenceType(), VF);
3398 Builder.SetInsertPoint(LoopVectorBody.back()->getTerminator());
3399 for (unsigned part = 0; part < UF; ++part) {
3400 Value *Trunc = Builder.CreateTrunc(RdxParts[part], RdxVecTy);
3401 Value *Extnd = RdxDesc.isSigned() ? Builder.CreateSExt(Trunc, VecTy)
3402 : Builder.CreateZExt(Trunc, VecTy);
3403 for (Value::user_iterator UI = RdxParts[part]->user_begin();
3404 UI != RdxParts[part]->user_end();)
3406 (*UI++)->replaceUsesOfWith(RdxParts[part], Extnd);
3407 RdxParts[part] = Extnd;
3412 Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt());
3413 for (unsigned part = 0; part < UF; ++part)
3414 RdxParts[part] = Builder.CreateTrunc(RdxParts[part], RdxVecTy);
3417 // Reduce all of the unrolled parts into a single vector.
3418 Value *ReducedPartRdx = RdxParts[0];
3419 unsigned Op = RecurrenceDescriptor::getRecurrenceBinOp(RK);
3420 setDebugLocFromInst(Builder, ReducedPartRdx);
3421 for (unsigned part = 1; part < UF; ++part) {
3422 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
3423 // Floating point operations had to be 'fast' to enable the reduction.
3424 ReducedPartRdx = addFastMathFlag(
3425 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
3426 ReducedPartRdx, "bin.rdx"));
3428 ReducedPartRdx = RecurrenceDescriptor::createMinMaxOp(
3429 Builder, MinMaxKind, ReducedPartRdx, RdxParts[part]);
3433 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
3434 // and vector ops, reducing the set of values being computed by half each
3436 assert(isPowerOf2_32(VF) &&
3437 "Reduction emission only supported for pow2 vectors!");
3438 Value *TmpVec = ReducedPartRdx;
3439 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
3440 for (unsigned i = VF; i != 1; i >>= 1) {
3441 // Move the upper half of the vector to the lower half.
3442 for (unsigned j = 0; j != i/2; ++j)
3443 ShuffleMask[j] = Builder.getInt32(i/2 + j);
3445 // Fill the rest of the mask with undef.
3446 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
3447 UndefValue::get(Builder.getInt32Ty()));
3450 Builder.CreateShuffleVector(TmpVec,
3451 UndefValue::get(TmpVec->getType()),
3452 ConstantVector::get(ShuffleMask),
3455 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
3456 // Floating point operations had to be 'fast' to enable the reduction.
3457 TmpVec = addFastMathFlag(Builder.CreateBinOp(
3458 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
3460 TmpVec = RecurrenceDescriptor::createMinMaxOp(Builder, MinMaxKind,
3464 // The result is in the first element of the vector.
3465 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
3466 Builder.getInt32(0));
3468 // If the reduction can be performed in a smaller type, we need to extend
3469 // the reduction to the wider type before we branch to the original loop.
3470 if (RdxPhi->getType() != RdxDesc.getRecurrenceType())
3473 ? Builder.CreateSExt(ReducedPartRdx, RdxPhi->getType())
3474 : Builder.CreateZExt(ReducedPartRdx, RdxPhi->getType());
3477 // Create a phi node that merges control-flow from the backedge-taken check
3478 // block and the middle block.
3479 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
3480 LoopScalarPreHeader->getTerminator());
3481 for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
3482 BCBlockPhi->addIncoming(ReductionStartValue, LoopBypassBlocks[I]);
3483 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3485 // Now, we need to fix the users of the reduction variable
3486 // inside and outside of the scalar remainder loop.
3487 // We know that the loop is in LCSSA form. We need to update the
3488 // PHI nodes in the exit blocks.
3489 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3490 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3491 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3492 if (!LCSSAPhi) break;
3494 // All PHINodes need to have a single entry edge, or two if
3495 // we already fixed them.
3496 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
3498 // We found our reduction value exit-PHI. Update it with the
3499 // incoming bypass edge.
3500 if (LCSSAPhi->getIncomingValue(0) == LoopExitInst) {
3501 // Add an edge coming from the bypass.
3502 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3505 }// end of the LCSSA phi scan.
3507 // Fix the scalar loop reduction variable with the incoming reduction sum
3508 // from the vector body and from the backedge value.
3509 int IncomingEdgeBlockIdx =
3510 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
3511 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
3512 // Pick the other block.
3513 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
3514 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
3515 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst);
3516 }// end of for each redux variable.
3520 // Make sure DomTree is updated.
3523 // Predicate any stores.
3524 for (auto KV : PredicatedStores) {
3525 BasicBlock::iterator I(KV.first);
3526 auto *BB = SplitBlock(I->getParent(), &*std::next(I), DT, LI);
3527 auto *T = SplitBlockAndInsertIfThen(KV.second, &*I, /*Unreachable=*/false,
3528 /*BranchWeights=*/nullptr, DT);
3530 I->getParent()->setName("pred.store.if");
3531 BB->setName("pred.store.continue");
3533 DEBUG(DT->verifyDomTree());
3534 // Remove redundant induction instructions.
3535 cse(LoopVectorBody);
3538 void InnerLoopVectorizer::fixLCSSAPHIs() {
3539 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3540 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3541 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3542 if (!LCSSAPhi) break;
3543 if (LCSSAPhi->getNumIncomingValues() == 1)
3544 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
3549 InnerLoopVectorizer::VectorParts
3550 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
3551 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
3554 // Look for cached value.
3555 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
3556 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
3557 if (ECEntryIt != MaskCache.end())
3558 return ECEntryIt->second;
3560 VectorParts SrcMask = createBlockInMask(Src);
3562 // The terminator has to be a branch inst!
3563 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
3564 assert(BI && "Unexpected terminator found");
3566 if (BI->isConditional()) {
3567 VectorParts EdgeMask = getVectorValue(BI->getCondition());
3569 if (BI->getSuccessor(0) != Dst)
3570 for (unsigned part = 0; part < UF; ++part)
3571 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
3573 for (unsigned part = 0; part < UF; ++part)
3574 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
3576 MaskCache[Edge] = EdgeMask;
3580 MaskCache[Edge] = SrcMask;
3584 InnerLoopVectorizer::VectorParts
3585 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
3586 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
3588 // Loop incoming mask is all-one.
3589 if (OrigLoop->getHeader() == BB) {
3590 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
3591 return getVectorValue(C);
3594 // This is the block mask. We OR all incoming edges, and with zero.
3595 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
3596 VectorParts BlockMask = getVectorValue(Zero);
3599 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
3600 VectorParts EM = createEdgeMask(*it, BB);
3601 for (unsigned part = 0; part < UF; ++part)
3602 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
3608 void InnerLoopVectorizer::widenPHIInstruction(
3609 Instruction *PN, InnerLoopVectorizer::VectorParts &Entry, unsigned UF,
3610 unsigned VF, PhiVector *PV) {
3611 PHINode* P = cast<PHINode>(PN);
3612 // Handle reduction variables:
3613 if (Legal->getReductionVars()->count(P)) {
3614 for (unsigned part = 0; part < UF; ++part) {
3615 // This is phase one of vectorizing PHIs.
3616 Type *VecTy = (VF == 1) ? PN->getType() :
3617 VectorType::get(PN->getType(), VF);
3618 Entry[part] = PHINode::Create(
3619 VecTy, 2, "vec.phi", &*LoopVectorBody.back()->getFirstInsertionPt());
3625 setDebugLocFromInst(Builder, P);
3626 // Check for PHI nodes that are lowered to vector selects.
3627 if (P->getParent() != OrigLoop->getHeader()) {
3628 // We know that all PHIs in non-header blocks are converted into
3629 // selects, so we don't have to worry about the insertion order and we
3630 // can just use the builder.
3631 // At this point we generate the predication tree. There may be
3632 // duplications since this is a simple recursive scan, but future
3633 // optimizations will clean it up.
3635 unsigned NumIncoming = P->getNumIncomingValues();
3637 // Generate a sequence of selects of the form:
3638 // SELECT(Mask3, In3,
3639 // SELECT(Mask2, In2,
3641 for (unsigned In = 0; In < NumIncoming; In++) {
3642 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
3644 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3646 for (unsigned part = 0; part < UF; ++part) {
3647 // We might have single edge PHIs (blocks) - use an identity
3648 // 'select' for the first PHI operand.
3650 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3653 // Select between the current value and the previous incoming edge
3654 // based on the incoming mask.
3655 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3656 Entry[part], "predphi");
3662 // This PHINode must be an induction variable.
3663 // Make sure that we know about it.
3664 assert(Legal->getInductionVars()->count(P) &&
3665 "Not an induction variable");
3667 InductionDescriptor II = Legal->getInductionVars()->lookup(P);
3669 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
3670 // which can be found from the original scalar operations.
3671 switch (II.getKind()) {
3672 case InductionDescriptor::IK_NoInduction:
3673 llvm_unreachable("Unknown induction");
3674 case InductionDescriptor::IK_IntInduction: {
3675 assert(P->getType() == II.getStartValue()->getType() &&
3676 "Types must match");
3677 // Handle other induction variables that are now based on the
3679 Value *V = Induction;
3680 if (P != OldInduction) {
3681 V = Builder.CreateSExtOrTrunc(Induction, P->getType());
3682 V = II.transform(Builder, V);
3683 V->setName("offset.idx");
3685 Value *Broadcasted = getBroadcastInstrs(V);
3686 // After broadcasting the induction variable we need to make the vector
3687 // consecutive by adding 0, 1, 2, etc.
3688 for (unsigned part = 0; part < UF; ++part)
3689 Entry[part] = getStepVector(Broadcasted, VF * part, II.getStepValue());
3692 case InductionDescriptor::IK_PtrInduction:
3693 // Handle the pointer induction variable case.
3694 assert(P->getType()->isPointerTy() && "Unexpected type.");
3695 // This is the normalized GEP that starts counting at zero.
3696 Value *PtrInd = Induction;
3697 PtrInd = Builder.CreateSExtOrTrunc(PtrInd, II.getStepValue()->getType());
3698 // This is the vector of results. Notice that we don't generate
3699 // vector geps because scalar geps result in better code.
3700 for (unsigned part = 0; part < UF; ++part) {
3702 int EltIndex = part;
3703 Constant *Idx = ConstantInt::get(PtrInd->getType(), EltIndex);
3704 Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx);
3705 Value *SclrGep = II.transform(Builder, GlobalIdx);
3706 SclrGep->setName("next.gep");
3707 Entry[part] = SclrGep;
3711 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3712 for (unsigned int i = 0; i < VF; ++i) {
3713 int EltIndex = i + part * VF;
3714 Constant *Idx = ConstantInt::get(PtrInd->getType(), EltIndex);
3715 Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx);
3716 Value *SclrGep = II.transform(Builder, GlobalIdx);
3717 SclrGep->setName("next.gep");
3718 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3719 Builder.getInt32(i),
3722 Entry[part] = VecVal;
3728 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3729 // For each instruction in the old loop.
3730 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3731 VectorParts &Entry = WidenMap.get(&*it);
3733 switch (it->getOpcode()) {
3734 case Instruction::Br:
3735 // Nothing to do for PHIs and BR, since we already took care of the
3736 // loop control flow instructions.
3738 case Instruction::PHI: {
3739 // Vectorize PHINodes.
3740 widenPHIInstruction(&*it, Entry, UF, VF, PV);
3744 case Instruction::Add:
3745 case Instruction::FAdd:
3746 case Instruction::Sub:
3747 case Instruction::FSub:
3748 case Instruction::Mul:
3749 case Instruction::FMul:
3750 case Instruction::UDiv:
3751 case Instruction::SDiv:
3752 case Instruction::FDiv:
3753 case Instruction::URem:
3754 case Instruction::SRem:
3755 case Instruction::FRem:
3756 case Instruction::Shl:
3757 case Instruction::LShr:
3758 case Instruction::AShr:
3759 case Instruction::And:
3760 case Instruction::Or:
3761 case Instruction::Xor: {
3762 // Just widen binops.
3763 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3764 setDebugLocFromInst(Builder, BinOp);
3765 VectorParts &A = getVectorValue(it->getOperand(0));
3766 VectorParts &B = getVectorValue(it->getOperand(1));
3768 // Use this vector value for all users of the original instruction.
3769 for (unsigned Part = 0; Part < UF; ++Part) {
3770 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3772 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3773 VecOp->copyIRFlags(BinOp);
3778 propagateMetadata(Entry, &*it);
3781 case Instruction::Select: {
3783 // If the selector is loop invariant we can create a select
3784 // instruction with a scalar condition. Otherwise, use vector-select.
3785 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3787 setDebugLocFromInst(Builder, &*it);
3789 // The condition can be loop invariant but still defined inside the
3790 // loop. This means that we can't just use the original 'cond' value.
3791 // We have to take the 'vectorized' value and pick the first lane.
3792 // Instcombine will make this a no-op.
3793 VectorParts &Cond = getVectorValue(it->getOperand(0));
3794 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3795 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3797 Value *ScalarCond = (VF == 1) ? Cond[0] :
3798 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3800 for (unsigned Part = 0; Part < UF; ++Part) {
3801 Entry[Part] = Builder.CreateSelect(
3802 InvariantCond ? ScalarCond : Cond[Part],
3807 propagateMetadata(Entry, &*it);
3811 case Instruction::ICmp:
3812 case Instruction::FCmp: {
3813 // Widen compares. Generate vector compares.
3814 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3815 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3816 setDebugLocFromInst(Builder, &*it);
3817 VectorParts &A = getVectorValue(it->getOperand(0));
3818 VectorParts &B = getVectorValue(it->getOperand(1));
3819 for (unsigned Part = 0; Part < UF; ++Part) {
3822 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3823 cast<FCmpInst>(C)->copyFastMathFlags(&*it);
3825 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3830 propagateMetadata(Entry, &*it);
3834 case Instruction::Store:
3835 case Instruction::Load:
3836 vectorizeMemoryInstruction(&*it);
3838 case Instruction::ZExt:
3839 case Instruction::SExt:
3840 case Instruction::FPToUI:
3841 case Instruction::FPToSI:
3842 case Instruction::FPExt:
3843 case Instruction::PtrToInt:
3844 case Instruction::IntToPtr:
3845 case Instruction::SIToFP:
3846 case Instruction::UIToFP:
3847 case Instruction::Trunc:
3848 case Instruction::FPTrunc:
3849 case Instruction::BitCast: {
3850 CastInst *CI = dyn_cast<CastInst>(it);
3851 setDebugLocFromInst(Builder, &*it);
3852 /// Optimize the special case where the source is the induction
3853 /// variable. Notice that we can only optimize the 'trunc' case
3854 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3855 /// c. other casts depend on pointer size.
3856 if (CI->getOperand(0) == OldInduction &&
3857 it->getOpcode() == Instruction::Trunc) {
3858 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3860 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3861 InductionDescriptor II =
3862 Legal->getInductionVars()->lookup(OldInduction);
3863 Constant *Step = ConstantInt::getSigned(
3864 CI->getType(), II.getStepValue()->getSExtValue());
3865 for (unsigned Part = 0; Part < UF; ++Part)
3866 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
3867 propagateMetadata(Entry, &*it);
3870 /// Vectorize casts.
3871 Type *DestTy = (VF == 1) ? CI->getType() :
3872 VectorType::get(CI->getType(), VF);
3874 VectorParts &A = getVectorValue(it->getOperand(0));
3875 for (unsigned Part = 0; Part < UF; ++Part)
3876 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3877 propagateMetadata(Entry, &*it);
3881 case Instruction::Call: {
3882 // Ignore dbg intrinsics.
3883 if (isa<DbgInfoIntrinsic>(it))
3885 setDebugLocFromInst(Builder, &*it);
3887 Module *M = BB->getParent()->getParent();
3888 CallInst *CI = cast<CallInst>(it);
3890 StringRef FnName = CI->getCalledFunction()->getName();
3891 Function *F = CI->getCalledFunction();
3892 Type *RetTy = ToVectorTy(CI->getType(), VF);
3893 SmallVector<Type *, 4> Tys;
3894 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3895 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3897 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3899 (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end ||
3900 ID == Intrinsic::lifetime_start)) {
3901 scalarizeInstruction(&*it);
3904 // The flag shows whether we use Intrinsic or a usual Call for vectorized
3905 // version of the instruction.
3906 // Is it beneficial to perform intrinsic call compared to lib call?
3907 bool NeedToScalarize;
3908 unsigned CallCost = getVectorCallCost(CI, VF, *TTI, TLI, NeedToScalarize);
3909 bool UseVectorIntrinsic =
3910 ID && getVectorIntrinsicCost(CI, VF, *TTI, TLI) <= CallCost;
3911 if (!UseVectorIntrinsic && NeedToScalarize) {
3912 scalarizeInstruction(&*it);
3916 for (unsigned Part = 0; Part < UF; ++Part) {
3917 SmallVector<Value *, 4> Args;
3918 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3919 Value *Arg = CI->getArgOperand(i);
3920 // Some intrinsics have a scalar argument - don't replace it with a
3922 if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, i)) {
3923 VectorParts &VectorArg = getVectorValue(CI->getArgOperand(i));
3924 Arg = VectorArg[Part];
3926 Args.push_back(Arg);
3930 if (UseVectorIntrinsic) {
3931 // Use vector version of the intrinsic.
3932 Type *TysForDecl[] = {CI->getType()};
3934 TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3935 VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl);
3937 // Use vector version of the library call.
3938 StringRef VFnName = TLI->getVectorizedFunction(FnName, VF);
3939 assert(!VFnName.empty() && "Vector function name is empty.");
3940 VectorF = M->getFunction(VFnName);
3942 // Generate a declaration
3943 FunctionType *FTy = FunctionType::get(RetTy, Tys, false);
3945 Function::Create(FTy, Function::ExternalLinkage, VFnName, M);
3946 VectorF->copyAttributesFrom(F);
3949 assert(VectorF && "Can't create vector function.");
3950 Entry[Part] = Builder.CreateCall(VectorF, Args);
3953 propagateMetadata(Entry, &*it);
3958 // All other instructions are unsupported. Scalarize them.
3959 scalarizeInstruction(&*it);
3962 }// end of for_each instr.
3965 void InnerLoopVectorizer::updateAnalysis() {
3966 // Forget the original basic block.
3967 SE->forgetLoop(OrigLoop);
3969 // Update the dominator tree information.
3970 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3971 "Entry does not dominate exit.");
3973 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3974 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3975 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3977 // We don't predicate stores by this point, so the vector body should be a
3979 assert(LoopVectorBody.size() == 1 && "Expected single block loop!");
3980 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3982 DT->addNewBlock(LoopMiddleBlock, LoopVectorBody.back());
3983 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3984 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3985 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3987 DEBUG(DT->verifyDomTree());
3990 /// \brief Check whether it is safe to if-convert this phi node.
3992 /// Phi nodes with constant expressions that can trap are not safe to if
3994 static bool canIfConvertPHINodes(BasicBlock *BB) {
3995 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3996 PHINode *Phi = dyn_cast<PHINode>(I);
3999 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
4000 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
4007 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
4008 if (!EnableIfConversion) {
4009 emitAnalysis(VectorizationReport() << "if-conversion is disabled");
4013 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
4015 // A list of pointers that we can safely read and write to.
4016 SmallPtrSet<Value *, 8> SafePointes;
4018 // Collect safe addresses.
4019 for (Loop::block_iterator BI = TheLoop->block_begin(),
4020 BE = TheLoop->block_end(); BI != BE; ++BI) {
4021 BasicBlock *BB = *BI;
4023 if (blockNeedsPredication(BB))
4026 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
4027 if (LoadInst *LI = dyn_cast<LoadInst>(I))
4028 SafePointes.insert(LI->getPointerOperand());
4029 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
4030 SafePointes.insert(SI->getPointerOperand());
4034 // Collect the blocks that need predication.
4035 BasicBlock *Header = TheLoop->getHeader();
4036 for (Loop::block_iterator BI = TheLoop->block_begin(),
4037 BE = TheLoop->block_end(); BI != BE; ++BI) {
4038 BasicBlock *BB = *BI;
4040 // We don't support switch statements inside loops.
4041 if (!isa<BranchInst>(BB->getTerminator())) {
4042 emitAnalysis(VectorizationReport(BB->getTerminator())
4043 << "loop contains a switch statement");
4047 // We must be able to predicate all blocks that need to be predicated.
4048 if (blockNeedsPredication(BB)) {
4049 if (!blockCanBePredicated(BB, SafePointes)) {
4050 emitAnalysis(VectorizationReport(BB->getTerminator())
4051 << "control flow cannot be substituted for a select");
4054 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
4055 emitAnalysis(VectorizationReport(BB->getTerminator())
4056 << "control flow cannot be substituted for a select");
4061 // We can if-convert this loop.
4065 bool LoopVectorizationLegality::canVectorize() {
4066 // We must have a loop in canonical form. Loops with indirectbr in them cannot
4067 // be canonicalized.
4068 if (!TheLoop->getLoopPreheader()) {
4070 VectorizationReport() <<
4071 "loop control flow is not understood by vectorizer");
4075 // We can only vectorize innermost loops.
4076 if (!TheLoop->empty()) {
4077 emitAnalysis(VectorizationReport() << "loop is not the innermost loop");
4081 // We must have a single backedge.
4082 if (TheLoop->getNumBackEdges() != 1) {
4084 VectorizationReport() <<
4085 "loop control flow is not understood by vectorizer");
4089 // We must have a single exiting block.
4090 if (!TheLoop->getExitingBlock()) {
4092 VectorizationReport() <<
4093 "loop control flow is not understood by vectorizer");
4097 // We only handle bottom-tested loops, i.e. loop in which the condition is
4098 // checked at the end of each iteration. With that we can assume that all
4099 // instructions in the loop are executed the same number of times.
4100 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
4102 VectorizationReport() <<
4103 "loop control flow is not understood by vectorizer");
4107 // We need to have a loop header.
4108 DEBUG(dbgs() << "LV: Found a loop: " <<
4109 TheLoop->getHeader()->getName() << '\n');
4111 // Check if we can if-convert non-single-bb loops.
4112 unsigned NumBlocks = TheLoop->getNumBlocks();
4113 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
4114 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
4118 // ScalarEvolution needs to be able to find the exit count.
4119 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
4120 if (ExitCount == SE->getCouldNotCompute()) {
4121 emitAnalysis(VectorizationReport() <<
4122 "could not determine number of loop iterations");
4123 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
4127 // Check if we can vectorize the instructions and CFG in this loop.
4128 if (!canVectorizeInstrs()) {
4129 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
4133 // Go over each instruction and look at memory deps.
4134 if (!canVectorizeMemory()) {
4135 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
4139 // Collect all of the variables that remain uniform after vectorization.
4140 collectLoopUniforms();
4142 DEBUG(dbgs() << "LV: We can vectorize this loop"
4143 << (LAI->getRuntimePointerChecking()->Need
4144 ? " (with a runtime bound check)"
4148 bool UseInterleaved = TTI->enableInterleavedAccessVectorization();
4150 // If an override option has been passed in for interleaved accesses, use it.
4151 if (EnableInterleavedMemAccesses.getNumOccurrences() > 0)
4152 UseInterleaved = EnableInterleavedMemAccesses;
4154 // Analyze interleaved memory accesses.
4156 InterleaveInfo.analyzeInterleaving(Strides);
4158 unsigned SCEVThreshold = VectorizeSCEVCheckThreshold;
4159 if (Hints->getForce() == LoopVectorizeHints::FK_Enabled)
4160 SCEVThreshold = PragmaVectorizeSCEVCheckThreshold;
4162 if (Preds.getComplexity() > SCEVThreshold) {
4163 emitAnalysis(VectorizationReport()
4164 << "Too many SCEV assumptions need to be made and checked "
4166 DEBUG(dbgs() << "LV: Too many SCEV checks needed.\n");
4170 // Okay! We can vectorize. At this point we don't have any other mem analysis
4171 // which may limit our maximum vectorization factor, so just return true with
4176 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
4177 if (Ty->isPointerTy())
4178 return DL.getIntPtrType(Ty);
4180 // It is possible that char's or short's overflow when we ask for the loop's
4181 // trip count, work around this by changing the type size.
4182 if (Ty->getScalarSizeInBits() < 32)
4183 return Type::getInt32Ty(Ty->getContext());
4188 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
4189 Ty0 = convertPointerToIntegerType(DL, Ty0);
4190 Ty1 = convertPointerToIntegerType(DL, Ty1);
4191 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
4196 /// \brief Check that the instruction has outside loop users and is not an
4197 /// identified reduction variable.
4198 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
4199 SmallPtrSetImpl<Value *> &Reductions) {
4200 // Reduction instructions are allowed to have exit users. All other
4201 // instructions must not have external users.
4202 if (!Reductions.count(Inst))
4203 //Check that all of the users of the loop are inside the BB.
4204 for (User *U : Inst->users()) {
4205 Instruction *UI = cast<Instruction>(U);
4206 // This user may be a reduction exit value.
4207 if (!TheLoop->contains(UI)) {
4208 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
4215 bool LoopVectorizationLegality::canVectorizeInstrs() {
4216 BasicBlock *Header = TheLoop->getHeader();
4218 // Look for the attribute signaling the absence of NaNs.
4219 Function &F = *Header->getParent();
4220 const DataLayout &DL = F.getParent()->getDataLayout();
4221 if (F.hasFnAttribute("no-nans-fp-math"))
4223 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
4225 // For each block in the loop.
4226 for (Loop::block_iterator bb = TheLoop->block_begin(),
4227 be = TheLoop->block_end(); bb != be; ++bb) {
4229 // Scan the instructions in the block and look for hazards.
4230 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4233 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
4234 Type *PhiTy = Phi->getType();
4235 // Check that this PHI type is allowed.
4236 if (!PhiTy->isIntegerTy() &&
4237 !PhiTy->isFloatingPointTy() &&
4238 !PhiTy->isPointerTy()) {
4239 emitAnalysis(VectorizationReport(&*it)
4240 << "loop control flow is not understood by vectorizer");
4241 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
4245 // If this PHINode is not in the header block, then we know that we
4246 // can convert it to select during if-conversion. No need to check if
4247 // the PHIs in this block are induction or reduction variables.
4248 if (*bb != Header) {
4249 // Check that this instruction has no outside users or is an
4250 // identified reduction value with an outside user.
4251 if (!hasOutsideLoopUser(TheLoop, &*it, AllowedExit))
4253 emitAnalysis(VectorizationReport(&*it) <<
4254 "value could not be identified as "
4255 "an induction or reduction variable");
4259 // We only allow if-converted PHIs with exactly two incoming values.
4260 if (Phi->getNumIncomingValues() != 2) {
4261 emitAnalysis(VectorizationReport(&*it)
4262 << "control flow not understood by vectorizer");
4263 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
4267 InductionDescriptor ID;
4268 if (InductionDescriptor::isInductionPHI(Phi, SE, ID)) {
4269 Inductions[Phi] = ID;
4270 // Get the widest type.
4272 WidestIndTy = convertPointerToIntegerType(DL, PhiTy);
4274 WidestIndTy = getWiderType(DL, PhiTy, WidestIndTy);
4276 // Int inductions are special because we only allow one IV.
4277 if (ID.getKind() == InductionDescriptor::IK_IntInduction &&
4278 ID.getStepValue()->isOne() &&
4279 isa<Constant>(ID.getStartValue()) &&
4280 cast<Constant>(ID.getStartValue())->isNullValue()) {
4281 // Use the phi node with the widest type as induction. Use the last
4282 // one if there are multiple (no good reason for doing this other
4283 // than it is expedient). We've checked that it begins at zero and
4284 // steps by one, so this is a canonical induction variable.
4285 if (!Induction || PhiTy == WidestIndTy)
4289 DEBUG(dbgs() << "LV: Found an induction variable.\n");
4291 // Until we explicitly handle the case of an induction variable with
4292 // an outside loop user we have to give up vectorizing this loop.
4293 if (hasOutsideLoopUser(TheLoop, &*it, AllowedExit)) {
4294 emitAnalysis(VectorizationReport(&*it) <<
4295 "use of induction value outside of the "
4296 "loop is not handled by vectorizer");
4303 if (RecurrenceDescriptor::isReductionPHI(Phi, TheLoop,
4305 if (Reductions[Phi].hasUnsafeAlgebra())
4306 Requirements->addUnsafeAlgebraInst(
4307 Reductions[Phi].getUnsafeAlgebraInst());
4308 AllowedExit.insert(Reductions[Phi].getLoopExitInstr());
4312 emitAnalysis(VectorizationReport(&*it) <<
4313 "value that could not be identified as "
4314 "reduction is used outside the loop");
4315 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
4317 }// end of PHI handling
4319 // We handle calls that:
4320 // * Are debug info intrinsics.
4321 // * Have a mapping to an IR intrinsic.
4322 // * Have a vector version available.
4323 CallInst *CI = dyn_cast<CallInst>(it);
4324 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI) &&
4325 !(CI->getCalledFunction() && TLI &&
4326 TLI->isFunctionVectorizable(CI->getCalledFunction()->getName()))) {
4327 emitAnalysis(VectorizationReport(&*it)
4328 << "call instruction cannot be vectorized");
4329 DEBUG(dbgs() << "LV: Found a non-intrinsic, non-libfunc callsite.\n");
4333 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
4334 // second argument is the same (i.e. loop invariant)
4336 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
4337 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
4338 emitAnalysis(VectorizationReport(&*it)
4339 << "intrinsic instruction cannot be vectorized");
4340 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
4345 // Check that the instruction return type is vectorizable.
4346 // Also, we can't vectorize extractelement instructions.
4347 if ((!VectorType::isValidElementType(it->getType()) &&
4348 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
4349 emitAnalysis(VectorizationReport(&*it)
4350 << "instruction return type cannot be vectorized");
4351 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
4355 // Check that the stored type is vectorizable.
4356 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
4357 Type *T = ST->getValueOperand()->getType();
4358 if (!VectorType::isValidElementType(T)) {
4359 emitAnalysis(VectorizationReport(ST) <<
4360 "store instruction cannot be vectorized");
4363 if (EnableMemAccessVersioning)
4364 collectStridedAccess(ST);
4367 if (EnableMemAccessVersioning)
4368 if (LoadInst *LI = dyn_cast<LoadInst>(it))
4369 collectStridedAccess(LI);
4371 // Reduction instructions are allowed to have exit users.
4372 // All other instructions must not have external users.
4373 if (hasOutsideLoopUser(TheLoop, &*it, AllowedExit)) {
4374 emitAnalysis(VectorizationReport(&*it) <<
4375 "value cannot be used outside the loop");
4384 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
4385 if (Inductions.empty()) {
4386 emitAnalysis(VectorizationReport()
4387 << "loop induction variable could not be identified");
4392 // Now we know the widest induction type, check if our found induction
4393 // is the same size. If it's not, unset it here and InnerLoopVectorizer
4394 // will create another.
4395 if (Induction && WidestIndTy != Induction->getType())
4396 Induction = nullptr;
4401 void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) {
4402 Value *Ptr = nullptr;
4403 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
4404 Ptr = LI->getPointerOperand();
4405 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
4406 Ptr = SI->getPointerOperand();
4410 Value *Stride = getStrideFromPointer(Ptr, SE, TheLoop);
4414 DEBUG(dbgs() << "LV: Found a strided access that we can version");
4415 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
4416 Strides[Ptr] = Stride;
4417 StrideSet.insert(Stride);
4420 void LoopVectorizationLegality::collectLoopUniforms() {
4421 // We now know that the loop is vectorizable!
4422 // Collect variables that will remain uniform after vectorization.
4423 std::vector<Value*> Worklist;
4424 BasicBlock *Latch = TheLoop->getLoopLatch();
4426 // Start with the conditional branch and walk up the block.
4427 Worklist.push_back(Latch->getTerminator()->getOperand(0));
4429 // Also add all consecutive pointer values; these values will be uniform
4430 // after vectorization (and subsequent cleanup) and, until revectorization is
4431 // supported, all dependencies must also be uniform.
4432 for (Loop::block_iterator B = TheLoop->block_begin(),
4433 BE = TheLoop->block_end(); B != BE; ++B)
4434 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
4436 if (I->getType()->isPointerTy() && isConsecutivePtr(&*I))
4437 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4439 while (!Worklist.empty()) {
4440 Instruction *I = dyn_cast<Instruction>(Worklist.back());
4441 Worklist.pop_back();
4443 // Look at instructions inside this loop.
4444 // Stop when reaching PHI nodes.
4445 // TODO: we need to follow values all over the loop, not only in this block.
4446 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
4449 // This is a known uniform.
4452 // Insert all operands.
4453 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4457 bool LoopVectorizationLegality::canVectorizeMemory() {
4458 LAI = &LAA->getInfo(TheLoop, Strides);
4459 auto &OptionalReport = LAI->getReport();
4461 emitAnalysis(VectorizationReport(*OptionalReport));
4462 if (!LAI->canVectorizeMemory())
4465 if (LAI->hasStoreToLoopInvariantAddress()) {
4467 VectorizationReport()
4468 << "write to a loop invariant address could not be vectorized");
4469 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4473 Requirements->addRuntimePointerChecks(LAI->getNumRuntimePointerChecks());
4474 Preds.add(&LAI->Preds);
4479 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4480 Value *In0 = const_cast<Value*>(V);
4481 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4485 return Inductions.count(PN);
4488 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4489 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4492 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4493 SmallPtrSetImpl<Value *> &SafePtrs) {
4495 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4496 // Check that we don't have a constant expression that can trap as operand.
4497 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4499 if (Constant *C = dyn_cast<Constant>(*OI))
4503 // We might be able to hoist the load.
4504 if (it->mayReadFromMemory()) {
4505 LoadInst *LI = dyn_cast<LoadInst>(it);
4508 if (!SafePtrs.count(LI->getPointerOperand())) {
4509 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
4510 MaskedOp.insert(LI);
4517 // We don't predicate stores at the moment.
4518 if (it->mayWriteToMemory()) {
4519 StoreInst *SI = dyn_cast<StoreInst>(it);
4520 // We only support predication of stores in basic blocks with one
4525 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
4526 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
4528 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
4529 !isSinglePredecessor) {
4530 // Build a masked store if it is legal for the target, otherwise
4531 // scalarize the block.
4532 bool isLegalMaskedOp =
4533 isLegalMaskedStore(SI->getValueOperand()->getType(),
4534 SI->getPointerOperand());
4535 if (isLegalMaskedOp) {
4537 MaskedOp.insert(SI);
4546 // The instructions below can trap.
4547 switch (it->getOpcode()) {
4549 case Instruction::UDiv:
4550 case Instruction::SDiv:
4551 case Instruction::URem:
4552 case Instruction::SRem:
4560 void InterleavedAccessInfo::collectConstStridedAccesses(
4561 MapVector<Instruction *, StrideDescriptor> &StrideAccesses,
4562 const ValueToValueMap &Strides) {
4563 // Holds load/store instructions in program order.
4564 SmallVector<Instruction *, 16> AccessList;
4566 for (auto *BB : TheLoop->getBlocks()) {
4567 bool IsPred = LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4569 for (auto &I : *BB) {
4570 if (!isa<LoadInst>(&I) && !isa<StoreInst>(&I))
4572 // FIXME: Currently we can't handle mixed accesses and predicated accesses
4576 AccessList.push_back(&I);
4580 if (AccessList.empty())
4583 auto &DL = TheLoop->getHeader()->getModule()->getDataLayout();
4584 for (auto I : AccessList) {
4585 LoadInst *LI = dyn_cast<LoadInst>(I);
4586 StoreInst *SI = dyn_cast<StoreInst>(I);
4588 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
4589 int Stride = isStridedPtr(SE, Ptr, TheLoop, Strides, Preds);
4591 // The factor of the corresponding interleave group.
4592 unsigned Factor = std::abs(Stride);
4594 // Ignore the access if the factor is too small or too large.
4595 if (Factor < 2 || Factor > MaxInterleaveGroupFactor)
4598 const SCEV *Scev = replaceSymbolicStrideSCEV(SE, Strides, Preds, Ptr);
4599 PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
4600 unsigned Size = DL.getTypeAllocSize(PtrTy->getElementType());
4602 // An alignment of 0 means target ABI alignment.
4603 unsigned Align = LI ? LI->getAlignment() : SI->getAlignment();
4605 Align = DL.getABITypeAlignment(PtrTy->getElementType());
4607 StrideAccesses[I] = StrideDescriptor(Stride, Scev, Size, Align);
4611 // Analyze interleaved accesses and collect them into interleave groups.
4613 // Notice that the vectorization on interleaved groups will change instruction
4614 // orders and may break dependences. But the memory dependence check guarantees
4615 // that there is no overlap between two pointers of different strides, element
4616 // sizes or underlying bases.
4618 // For pointers sharing the same stride, element size and underlying base, no
4619 // need to worry about Read-After-Write dependences and Write-After-Read
4622 // E.g. The RAW dependence: A[i] = a;
4624 // This won't exist as it is a store-load forwarding conflict, which has
4625 // already been checked and forbidden in the dependence check.
4627 // E.g. The WAR dependence: a = A[i]; // (1)
4629 // The store group of (2) is always inserted at or below (2), and the load group
4630 // of (1) is always inserted at or above (1). The dependence is safe.
4631 void InterleavedAccessInfo::analyzeInterleaving(
4632 const ValueToValueMap &Strides) {
4633 DEBUG(dbgs() << "LV: Analyzing interleaved accesses...\n");
4635 // Holds all the stride accesses.
4636 MapVector<Instruction *, StrideDescriptor> StrideAccesses;
4637 collectConstStridedAccesses(StrideAccesses, Strides);
4639 if (StrideAccesses.empty())
4642 // Holds all interleaved store groups temporarily.
4643 SmallSetVector<InterleaveGroup *, 4> StoreGroups;
4645 // Search the load-load/write-write pair B-A in bottom-up order and try to
4646 // insert B into the interleave group of A according to 3 rules:
4647 // 1. A and B have the same stride.
4648 // 2. A and B have the same memory object size.
4649 // 3. B belongs to the group according to the distance.
4651 // The bottom-up order can avoid breaking the Write-After-Write dependences
4652 // between two pointers of the same base.
4653 // E.g. A[i] = a; (1)
4656 // We form the group (2)+(3) in front, so (1) has to form groups with accesses
4657 // above (1), which guarantees that (1) is always above (2).
4658 for (auto I = StrideAccesses.rbegin(), E = StrideAccesses.rend(); I != E;
4660 Instruction *A = I->first;
4661 StrideDescriptor DesA = I->second;
4663 InterleaveGroup *Group = getInterleaveGroup(A);
4665 DEBUG(dbgs() << "LV: Creating an interleave group with:" << *A << '\n');
4666 Group = createInterleaveGroup(A, DesA.Stride, DesA.Align);
4669 if (A->mayWriteToMemory())
4670 StoreGroups.insert(Group);
4672 for (auto II = std::next(I); II != E; ++II) {
4673 Instruction *B = II->first;
4674 StrideDescriptor DesB = II->second;
4676 // Ignore if B is already in a group or B is a different memory operation.
4677 if (isInterleaved(B) || A->mayReadFromMemory() != B->mayReadFromMemory())
4680 // Check the rule 1 and 2.
4681 if (DesB.Stride != DesA.Stride || DesB.Size != DesA.Size)
4684 // Calculate the distance and prepare for the rule 3.
4685 const SCEVConstant *DistToA =
4686 dyn_cast<SCEVConstant>(SE->getMinusSCEV(DesB.Scev, DesA.Scev));
4690 int DistanceToA = DistToA->getValue()->getValue().getSExtValue();
4692 // Skip if the distance is not multiple of size as they are not in the
4694 if (DistanceToA % static_cast<int>(DesA.Size))
4697 // The index of B is the index of A plus the related index to A.
4699 Group->getIndex(A) + DistanceToA / static_cast<int>(DesA.Size);
4701 // Try to insert B into the group.
4702 if (Group->insertMember(B, IndexB, DesB.Align)) {
4703 DEBUG(dbgs() << "LV: Inserted:" << *B << '\n'
4704 << " into the interleave group with" << *A << '\n');
4705 InterleaveGroupMap[B] = Group;
4707 // Set the first load in program order as the insert position.
4708 if (B->mayReadFromMemory())
4709 Group->setInsertPos(B);
4711 } // Iteration on instruction B
4712 } // Iteration on instruction A
4714 // Remove interleaved store groups with gaps.
4715 for (InterleaveGroup *Group : StoreGroups)
4716 if (Group->getNumMembers() != Group->getFactor())
4717 releaseGroup(Group);
4720 LoopVectorizationCostModel::VectorizationFactor
4721 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
4722 // Width 1 means no vectorize
4723 VectorizationFactor Factor = { 1U, 0U };
4724 if (OptForSize && Legal->getRuntimePointerChecking()->Need) {
4725 emitAnalysis(VectorizationReport() <<
4726 "runtime pointer checks needed. Enable vectorization of this "
4727 "loop with '#pragma clang loop vectorize(enable)' when "
4728 "compiling with -Os/-Oz");
4730 "LV: Aborting. Runtime ptr check is required with -Os/-Oz.\n");
4734 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
4735 emitAnalysis(VectorizationReport() <<
4736 "store that is conditionally executed prevents vectorization");
4737 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
4741 // Find the trip count.
4742 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4743 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4745 MinBWs = computeMinimumValueSizes(TheLoop->getBlocks(), *DB, &TTI);
4746 unsigned SmallestType, WidestType;
4747 std::tie(SmallestType, WidestType) = getSmallestAndWidestTypes();
4748 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4749 unsigned MaxSafeDepDist = -1U;
4750 if (Legal->getMaxSafeDepDistBytes() != -1U)
4751 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4752 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4753 WidestRegister : MaxSafeDepDist);
4754 unsigned MaxVectorSize = WidestRegister / WidestType;
4756 DEBUG(dbgs() << "LV: The Smallest and Widest types: " << SmallestType << " / "
4757 << WidestType << " bits.\n");
4758 DEBUG(dbgs() << "LV: The Widest register is: "
4759 << WidestRegister << " bits.\n");
4761 if (MaxVectorSize == 0) {
4762 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4766 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
4767 " into one vector!");
4769 unsigned VF = MaxVectorSize;
4770 if (MaximizeBandwidth && !OptForSize) {
4771 // Collect all viable vectorization factors.
4772 SmallVector<unsigned, 8> VFs;
4773 unsigned NewMaxVectorSize = WidestRegister / SmallestType;
4774 for (unsigned VS = MaxVectorSize; VS <= NewMaxVectorSize; VS *= 2)
4777 // For each VF calculate its register usage.
4778 auto RUs = calculateRegisterUsage(VFs);
4780 // Select the largest VF which doesn't require more registers than existing
4782 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(true);
4783 for (int i = RUs.size() - 1; i >= 0; --i) {
4784 if (RUs[i].MaxLocalUsers <= TargetNumRegisters) {
4791 // If we optimize the program for size, avoid creating the tail loop.
4793 // If we are unable to calculate the trip count then don't try to vectorize.
4796 (VectorizationReport() <<
4797 "unable to calculate the loop count due to complex control flow");
4798 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n");
4802 // Find the maximum SIMD width that can fit within the trip count.
4803 VF = TC % MaxVectorSize;
4808 // If the trip count that we found modulo the vectorization factor is not
4809 // zero then we require a tail.
4810 emitAnalysis(VectorizationReport() <<
4811 "cannot optimize for size and vectorize at the "
4812 "same time. Enable vectorization of this loop "
4813 "with '#pragma clang loop vectorize(enable)' "
4814 "when compiling with -Os/-Oz");
4815 DEBUG(dbgs() << "LV: Aborting. A tail loop is required with -Os/-Oz.\n");
4820 int UserVF = Hints->getWidth();
4822 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4823 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4825 Factor.Width = UserVF;
4829 float Cost = expectedCost(1);
4831 const float ScalarCost = Cost;
4834 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
4836 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
4837 // Ignore scalar width, because the user explicitly wants vectorization.
4838 if (ForceVectorization && VF > 1) {
4840 Cost = expectedCost(Width) / (float)Width;
4843 for (unsigned i=2; i <= VF; i*=2) {
4844 // Notice that the vector loop needs to be executed less times, so
4845 // we need to divide the cost of the vector loops by the width of
4846 // the vector elements.
4847 float VectorCost = expectedCost(i) / (float)i;
4848 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
4849 (int)VectorCost << ".\n");
4850 if (VectorCost < Cost) {
4856 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
4857 << "LV: Vectorization seems to be not beneficial, "
4858 << "but was forced by a user.\n");
4859 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
4860 Factor.Width = Width;
4861 Factor.Cost = Width * Cost;
4865 std::pair<unsigned, unsigned>
4866 LoopVectorizationCostModel::getSmallestAndWidestTypes() {
4867 unsigned MinWidth = -1U;
4868 unsigned MaxWidth = 8;
4869 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
4872 for (Loop::block_iterator bb = TheLoop->block_begin(),
4873 be = TheLoop->block_end(); bb != be; ++bb) {
4874 BasicBlock *BB = *bb;
4876 // For each instruction in the loop.
4877 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4878 Type *T = it->getType();
4880 // Skip ignored values.
4881 if (ValuesToIgnore.count(&*it))
4884 // Only examine Loads, Stores and PHINodes.
4885 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4888 // Examine PHI nodes that are reduction variables. Update the type to
4889 // account for the recurrence type.
4890 if (PHINode *PN = dyn_cast<PHINode>(it)) {
4891 if (!Legal->getReductionVars()->count(PN))
4893 RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[PN];
4894 T = RdxDesc.getRecurrenceType();
4897 // Examine the stored values.
4898 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4899 T = ST->getValueOperand()->getType();
4901 // Ignore loaded pointer types and stored pointer types that are not
4902 // consecutive. However, we do want to take consecutive stores/loads of
4903 // pointer vectors into account.
4904 if (T->isPointerTy() && !isConsecutiveLoadOrStore(&*it))
4907 MinWidth = std::min(MinWidth,
4908 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
4909 MaxWidth = std::max(MaxWidth,
4910 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
4914 return {MinWidth, MaxWidth};
4917 unsigned LoopVectorizationCostModel::selectInterleaveCount(bool OptForSize,
4919 unsigned LoopCost) {
4921 // -- The interleave heuristics --
4922 // We interleave the loop in order to expose ILP and reduce the loop overhead.
4923 // There are many micro-architectural considerations that we can't predict
4924 // at this level. For example, frontend pressure (on decode or fetch) due to
4925 // code size, or the number and capabilities of the execution ports.
4927 // We use the following heuristics to select the interleave count:
4928 // 1. If the code has reductions, then we interleave to break the cross
4929 // iteration dependency.
4930 // 2. If the loop is really small, then we interleave to reduce the loop
4932 // 3. We don't interleave if we think that we will spill registers to memory
4933 // due to the increased register pressure.
4935 // When we optimize for size, we don't interleave.
4939 // We used the distance for the interleave count.
4940 if (Legal->getMaxSafeDepDistBytes() != -1U)
4943 // Do not interleave loops with a relatively small trip count.
4944 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4945 if (TC > 1 && TC < TinyTripCountInterleaveThreshold)
4948 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
4949 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
4953 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
4954 TargetNumRegisters = ForceTargetNumScalarRegs;
4956 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
4957 TargetNumRegisters = ForceTargetNumVectorRegs;
4960 RegisterUsage R = calculateRegisterUsage({VF})[0];
4961 // We divide by these constants so assume that we have at least one
4962 // instruction that uses at least one register.
4963 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4964 R.NumInstructions = std::max(R.NumInstructions, 1U);
4966 // We calculate the interleave count using the following formula.
4967 // Subtract the number of loop invariants from the number of available
4968 // registers. These registers are used by all of the interleaved instances.
4969 // Next, divide the remaining registers by the number of registers that is
4970 // required by the loop, in order to estimate how many parallel instances
4971 // fit without causing spills. All of this is rounded down if necessary to be
4972 // a power of two. We want power of two interleave count to simplify any
4973 // addressing operations or alignment considerations.
4974 unsigned IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
4977 // Don't count the induction variable as interleaved.
4978 if (EnableIndVarRegisterHeur)
4979 IC = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
4980 std::max(1U, (R.MaxLocalUsers - 1)));
4982 // Clamp the interleave ranges to reasonable counts.
4983 unsigned MaxInterleaveCount = TTI.getMaxInterleaveFactor(VF);
4985 // Check if the user has overridden the max.
4987 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
4988 MaxInterleaveCount = ForceTargetMaxScalarInterleaveFactor;
4990 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
4991 MaxInterleaveCount = ForceTargetMaxVectorInterleaveFactor;
4994 // If we did not calculate the cost for VF (because the user selected the VF)
4995 // then we calculate the cost of VF here.
4997 LoopCost = expectedCost(VF);
4999 // Clamp the calculated IC to be between the 1 and the max interleave count
5000 // that the target allows.
5001 if (IC > MaxInterleaveCount)
5002 IC = MaxInterleaveCount;
5006 // Interleave if we vectorized this loop and there is a reduction that could
5007 // benefit from interleaving.
5008 if (VF > 1 && Legal->getReductionVars()->size()) {
5009 DEBUG(dbgs() << "LV: Interleaving because of reductions.\n");
5013 // Note that if we've already vectorized the loop we will have done the
5014 // runtime check and so interleaving won't require further checks.
5015 bool InterleavingRequiresRuntimePointerCheck =
5016 (VF == 1 && Legal->getRuntimePointerChecking()->Need);
5018 // We want to interleave small loops in order to reduce the loop overhead and
5019 // potentially expose ILP opportunities.
5020 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
5021 if (!InterleavingRequiresRuntimePointerCheck && LoopCost < SmallLoopCost) {
5022 // We assume that the cost overhead is 1 and we use the cost model
5023 // to estimate the cost of the loop and interleave until the cost of the
5024 // loop overhead is about 5% of the cost of the loop.
5026 std::min(IC, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
5028 // Interleave until store/load ports (estimated by max interleave count) are
5030 unsigned NumStores = Legal->getNumStores();
5031 unsigned NumLoads = Legal->getNumLoads();
5032 unsigned StoresIC = IC / (NumStores ? NumStores : 1);
5033 unsigned LoadsIC = IC / (NumLoads ? NumLoads : 1);
5035 // If we have a scalar reduction (vector reductions are already dealt with
5036 // by this point), we can increase the critical path length if the loop
5037 // we're interleaving is inside another loop. Limit, by default to 2, so the
5038 // critical path only gets increased by one reduction operation.
5039 if (Legal->getReductionVars()->size() &&
5040 TheLoop->getLoopDepth() > 1) {
5041 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionIC);
5042 SmallIC = std::min(SmallIC, F);
5043 StoresIC = std::min(StoresIC, F);
5044 LoadsIC = std::min(LoadsIC, F);
5047 if (EnableLoadStoreRuntimeInterleave &&
5048 std::max(StoresIC, LoadsIC) > SmallIC) {
5049 DEBUG(dbgs() << "LV: Interleaving to saturate store or load ports.\n");
5050 return std::max(StoresIC, LoadsIC);
5053 DEBUG(dbgs() << "LV: Interleaving to reduce branch cost.\n");
5057 // Interleave if this is a large loop (small loops are already dealt with by
5059 // point) that could benefit from interleaving.
5060 bool HasReductions = (Legal->getReductionVars()->size() > 0);
5061 if (TTI.enableAggressiveInterleaving(HasReductions)) {
5062 DEBUG(dbgs() << "LV: Interleaving to expose ILP.\n");
5066 DEBUG(dbgs() << "LV: Not Interleaving.\n");
5070 SmallVector<LoopVectorizationCostModel::RegisterUsage, 8>
5071 LoopVectorizationCostModel::calculateRegisterUsage(
5072 const SmallVector<unsigned, 8> &VFs) {
5073 // This function calculates the register usage by measuring the highest number
5074 // of values that are alive at a single location. Obviously, this is a very
5075 // rough estimation. We scan the loop in a topological order in order and
5076 // assign a number to each instruction. We use RPO to ensure that defs are
5077 // met before their users. We assume that each instruction that has in-loop
5078 // users starts an interval. We record every time that an in-loop value is
5079 // used, so we have a list of the first and last occurrences of each
5080 // instruction. Next, we transpose this data structure into a multi map that
5081 // holds the list of intervals that *end* at a specific location. This multi
5082 // map allows us to perform a linear search. We scan the instructions linearly
5083 // and record each time that a new interval starts, by placing it in a set.
5084 // If we find this value in the multi-map then we remove it from the set.
5085 // The max register usage is the maximum size of the set.
5086 // We also search for instructions that are defined outside the loop, but are
5087 // used inside the loop. We need this number separately from the max-interval
5088 // usage number because when we unroll, loop-invariant values do not take
5090 LoopBlocksDFS DFS(TheLoop);
5094 RU.NumInstructions = 0;
5096 // Each 'key' in the map opens a new interval. The values
5097 // of the map are the index of the 'last seen' usage of the
5098 // instruction that is the key.
5099 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5100 // Maps instruction to its index.
5101 DenseMap<unsigned, Instruction*> IdxToInstr;
5102 // Marks the end of each interval.
5103 IntervalMap EndPoint;
5104 // Saves the list of instruction indices that are used in the loop.
5105 SmallSet<Instruction*, 8> Ends;
5106 // Saves the list of values that are used in the loop but are
5107 // defined outside the loop, such as arguments and constants.
5108 SmallPtrSet<Value*, 8> LoopInvariants;
5111 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5112 be = DFS.endRPO(); bb != be; ++bb) {
5113 RU.NumInstructions += (*bb)->size();
5114 for (Instruction &I : **bb) {
5115 IdxToInstr[Index++] = &I;
5117 // Save the end location of each USE.
5118 for (unsigned i = 0; i < I.getNumOperands(); ++i) {
5119 Value *U = I.getOperand(i);
5120 Instruction *Instr = dyn_cast<Instruction>(U);
5122 // Ignore non-instruction values such as arguments, constants, etc.
5123 if (!Instr) continue;
5125 // If this instruction is outside the loop then record it and continue.
5126 if (!TheLoop->contains(Instr)) {
5127 LoopInvariants.insert(Instr);
5131 // Overwrite previous end points.
5132 EndPoint[Instr] = Index;
5138 // Saves the list of intervals that end with the index in 'key'.
5139 typedef SmallVector<Instruction*, 2> InstrList;
5140 DenseMap<unsigned, InstrList> TransposeEnds;
5142 // Transpose the EndPoints to a list of values that end at each index.
5143 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5145 TransposeEnds[it->second].push_back(it->first);
5147 SmallSet<Instruction*, 8> OpenIntervals;
5149 // Get the size of the widest register.
5150 unsigned MaxSafeDepDist = -1U;
5151 if (Legal->getMaxSafeDepDistBytes() != -1U)
5152 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
5153 unsigned WidestRegister =
5154 std::min(TTI.getRegisterBitWidth(true), MaxSafeDepDist);
5155 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
5157 SmallVector<RegisterUsage, 8> RUs(VFs.size());
5158 SmallVector<unsigned, 8> MaxUsages(VFs.size(), 0);
5160 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5162 // A lambda that gets the register usage for the given type and VF.
5163 auto GetRegUsage = [&DL, WidestRegister](Type *Ty, unsigned VF) {
5164 unsigned TypeSize = DL.getTypeSizeInBits(Ty->getScalarType());
5165 return std::max<unsigned>(1, VF * TypeSize / WidestRegister);
5168 for (unsigned int i = 0; i < Index; ++i) {
5169 Instruction *I = IdxToInstr[i];
5170 // Ignore instructions that are never used within the loop.
5171 if (!Ends.count(I)) continue;
5173 // Skip ignored values.
5174 if (ValuesToIgnore.count(I))
5177 // Remove all of the instructions that end at this location.
5178 InstrList &List = TransposeEnds[i];
5179 for (unsigned int j = 0, e = List.size(); j < e; ++j)
5180 OpenIntervals.erase(List[j]);
5182 // For each VF find the maximum usage of registers.
5183 for (unsigned j = 0, e = VFs.size(); j < e; ++j) {
5185 MaxUsages[j] = std::max(MaxUsages[j], OpenIntervals.size());
5189 // Count the number of live interals.
5190 unsigned RegUsage = 0;
5191 for (auto Inst : OpenIntervals)
5192 RegUsage += GetRegUsage(Inst->getType(), VFs[j]);
5193 MaxUsages[j] = std::max(MaxUsages[j], RegUsage);
5196 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # "
5197 << OpenIntervals.size() << '\n');
5199 // Add the current instruction to the list of open intervals.
5200 OpenIntervals.insert(I);
5203 for (unsigned i = 0, e = VFs.size(); i < e; ++i) {
5204 unsigned Invariant = 0;
5206 Invariant = LoopInvariants.size();
5208 for (auto Inst : LoopInvariants)
5209 Invariant += GetRegUsage(Inst->getType(), VFs[i]);
5212 DEBUG(dbgs() << "LV(REG): VF = " << VFs[i] << '\n');
5213 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsages[i] << '\n');
5214 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5215 DEBUG(dbgs() << "LV(REG): LoopSize: " << RU.NumInstructions << '\n');
5217 RU.LoopInvariantRegs = Invariant;
5218 RU.MaxLocalUsers = MaxUsages[i];
5225 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5229 for (Loop::block_iterator bb = TheLoop->block_begin(),
5230 be = TheLoop->block_end(); bb != be; ++bb) {
5231 unsigned BlockCost = 0;
5232 BasicBlock *BB = *bb;
5234 // For each instruction in the old loop.
5235 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5236 // Skip dbg intrinsics.
5237 if (isa<DbgInfoIntrinsic>(it))
5240 // Skip ignored values.
5241 if (ValuesToIgnore.count(&*it))
5244 unsigned C = getInstructionCost(&*it, VF);
5246 // Check if we should override the cost.
5247 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5248 C = ForceTargetInstructionCost;
5251 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5252 VF << " For instruction: " << *it << '\n');
5255 // We assume that if-converted blocks have a 50% chance of being executed.
5256 // When the code is scalar then some of the blocks are avoided due to CF.
5257 // When the code is vectorized we execute all code paths.
5258 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5267 /// \brief Check whether the address computation for a non-consecutive memory
5268 /// access looks like an unlikely candidate for being merged into the indexing
5271 /// We look for a GEP which has one index that is an induction variable and all
5272 /// other indices are loop invariant. If the stride of this access is also
5273 /// within a small bound we decide that this address computation can likely be
5274 /// merged into the addressing mode.
5275 /// In all other cases, we identify the address computation as complex.
5276 static bool isLikelyComplexAddressComputation(Value *Ptr,
5277 LoopVectorizationLegality *Legal,
5278 ScalarEvolution *SE,
5279 const Loop *TheLoop) {
5280 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5284 // We are looking for a gep with all loop invariant indices except for one
5285 // which should be an induction variable.
5286 unsigned NumOperands = Gep->getNumOperands();
5287 for (unsigned i = 1; i < NumOperands; ++i) {
5288 Value *Opd = Gep->getOperand(i);
5289 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5290 !Legal->isInductionVariable(Opd))
5294 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5295 // can likely be merged into the address computation.
5296 unsigned MaxMergeDistance = 64;
5298 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5302 // Check the step is constant.
5303 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5304 // Calculate the pointer stride and check if it is consecutive.
5305 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5309 const APInt &APStepVal = C->getValue()->getValue();
5311 // Huge step value - give up.
5312 if (APStepVal.getBitWidth() > 64)
5315 int64_t StepVal = APStepVal.getSExtValue();
5317 return StepVal > MaxMergeDistance;
5320 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5321 return Legal->hasStride(I->getOperand(0)) ||
5322 Legal->hasStride(I->getOperand(1));
5326 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5327 // If we know that this instruction will remain uniform, check the cost of
5328 // the scalar version.
5329 if (Legal->isUniformAfterVectorization(I))
5332 Type *RetTy = I->getType();
5333 if (VF > 1 && MinBWs.count(I))
5334 RetTy = IntegerType::get(RetTy->getContext(), MinBWs[I]);
5335 Type *VectorTy = ToVectorTy(RetTy, VF);
5337 // TODO: We need to estimate the cost of intrinsic calls.
5338 switch (I->getOpcode()) {
5339 case Instruction::GetElementPtr:
5340 // We mark this instruction as zero-cost because the cost of GEPs in
5341 // vectorized code depends on whether the corresponding memory instruction
5342 // is scalarized or not. Therefore, we handle GEPs with the memory
5343 // instruction cost.
5345 case Instruction::Br: {
5346 return TTI.getCFInstrCost(I->getOpcode());
5348 case Instruction::PHI:
5349 //TODO: IF-converted IFs become selects.
5351 case Instruction::Add:
5352 case Instruction::FAdd:
5353 case Instruction::Sub:
5354 case Instruction::FSub:
5355 case Instruction::Mul:
5356 case Instruction::FMul:
5357 case Instruction::UDiv:
5358 case Instruction::SDiv:
5359 case Instruction::FDiv:
5360 case Instruction::URem:
5361 case Instruction::SRem:
5362 case Instruction::FRem:
5363 case Instruction::Shl:
5364 case Instruction::LShr:
5365 case Instruction::AShr:
5366 case Instruction::And:
5367 case Instruction::Or:
5368 case Instruction::Xor: {
5369 // Since we will replace the stride by 1 the multiplication should go away.
5370 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5372 // Certain instructions can be cheaper to vectorize if they have a constant
5373 // second vector operand. One example of this are shifts on x86.
5374 TargetTransformInfo::OperandValueKind Op1VK =
5375 TargetTransformInfo::OK_AnyValue;
5376 TargetTransformInfo::OperandValueKind Op2VK =
5377 TargetTransformInfo::OK_AnyValue;
5378 TargetTransformInfo::OperandValueProperties Op1VP =
5379 TargetTransformInfo::OP_None;
5380 TargetTransformInfo::OperandValueProperties Op2VP =
5381 TargetTransformInfo::OP_None;
5382 Value *Op2 = I->getOperand(1);
5384 // Check for a splat of a constant or for a non uniform vector of constants.
5385 if (isa<ConstantInt>(Op2)) {
5386 ConstantInt *CInt = cast<ConstantInt>(Op2);
5387 if (CInt && CInt->getValue().isPowerOf2())
5388 Op2VP = TargetTransformInfo::OP_PowerOf2;
5389 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5390 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5391 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5392 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
5394 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
5395 if (CInt && CInt->getValue().isPowerOf2())
5396 Op2VP = TargetTransformInfo::OP_PowerOf2;
5397 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5401 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
5404 case Instruction::Select: {
5405 SelectInst *SI = cast<SelectInst>(I);
5406 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5407 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5408 Type *CondTy = SI->getCondition()->getType();
5410 CondTy = VectorType::get(CondTy, VF);
5412 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5414 case Instruction::ICmp:
5415 case Instruction::FCmp: {
5416 Type *ValTy = I->getOperand(0)->getType();
5417 Instruction *Op0AsInstruction = dyn_cast<Instruction>(I->getOperand(0));
5418 auto It = MinBWs.find(Op0AsInstruction);
5419 if (VF > 1 && It != MinBWs.end())
5420 ValTy = IntegerType::get(ValTy->getContext(), It->second);
5421 VectorTy = ToVectorTy(ValTy, VF);
5422 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5424 case Instruction::Store:
5425 case Instruction::Load: {
5426 StoreInst *SI = dyn_cast<StoreInst>(I);
5427 LoadInst *LI = dyn_cast<LoadInst>(I);
5428 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5430 VectorTy = ToVectorTy(ValTy, VF);
5432 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5433 unsigned AS = SI ? SI->getPointerAddressSpace() :
5434 LI->getPointerAddressSpace();
5435 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5436 // We add the cost of address computation here instead of with the gep
5437 // instruction because only here we know whether the operation is
5440 return TTI.getAddressComputationCost(VectorTy) +
5441 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5443 // For an interleaved access, calculate the total cost of the whole
5444 // interleave group.
5445 if (Legal->isAccessInterleaved(I)) {
5446 auto Group = Legal->getInterleavedAccessGroup(I);
5447 assert(Group && "Fail to get an interleaved access group.");
5449 // Only calculate the cost once at the insert position.
5450 if (Group->getInsertPos() != I)
5453 unsigned InterleaveFactor = Group->getFactor();
5455 VectorType::get(VectorTy->getVectorElementType(),
5456 VectorTy->getVectorNumElements() * InterleaveFactor);
5458 // Holds the indices of existing members in an interleaved load group.
5459 // An interleaved store group doesn't need this as it dones't allow gaps.
5460 SmallVector<unsigned, 4> Indices;
5462 for (unsigned i = 0; i < InterleaveFactor; i++)
5463 if (Group->getMember(i))
5464 Indices.push_back(i);
5467 // Calculate the cost of the whole interleaved group.
5468 unsigned Cost = TTI.getInterleavedMemoryOpCost(
5469 I->getOpcode(), WideVecTy, Group->getFactor(), Indices,
5470 Group->getAlignment(), AS);
5472 if (Group->isReverse())
5474 Group->getNumMembers() *
5475 TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
5477 // FIXME: The interleaved load group with a huge gap could be even more
5478 // expensive than scalar operations. Then we could ignore such group and
5479 // use scalar operations instead.
5483 // Scalarized loads/stores.
5484 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5485 bool Reverse = ConsecutiveStride < 0;
5486 const DataLayout &DL = I->getModule()->getDataLayout();
5487 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ValTy);
5488 unsigned VectorElementSize = DL.getTypeStoreSize(VectorTy) / VF;
5489 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5490 bool IsComplexComputation =
5491 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5493 // The cost of extracting from the value vector and pointer vector.
5494 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5495 for (unsigned i = 0; i < VF; ++i) {
5496 // The cost of extracting the pointer operand.
5497 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5498 // In case of STORE, the cost of ExtractElement from the vector.
5499 // In case of LOAD, the cost of InsertElement into the returned
5501 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5502 Instruction::InsertElement,
5506 // The cost of the scalar loads/stores.
5507 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5508 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5513 // Wide load/stores.
5514 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5515 if (Legal->isMaskRequired(I))
5516 Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment,
5519 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5522 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5526 case Instruction::ZExt:
5527 case Instruction::SExt:
5528 case Instruction::FPToUI:
5529 case Instruction::FPToSI:
5530 case Instruction::FPExt:
5531 case Instruction::PtrToInt:
5532 case Instruction::IntToPtr:
5533 case Instruction::SIToFP:
5534 case Instruction::UIToFP:
5535 case Instruction::Trunc:
5536 case Instruction::FPTrunc:
5537 case Instruction::BitCast: {
5538 // We optimize the truncation of induction variable.
5539 // The cost of these is the same as the scalar operation.
5540 if (I->getOpcode() == Instruction::Trunc &&
5541 Legal->isInductionVariable(I->getOperand(0)))
5542 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5543 I->getOperand(0)->getType());
5545 Type *SrcScalarTy = I->getOperand(0)->getType();
5546 Type *SrcVecTy = ToVectorTy(SrcScalarTy, VF);
5547 if (VF > 1 && MinBWs.count(I)) {
5548 // This cast is going to be shrunk. This may remove the cast or it might
5549 // turn it into slightly different cast. For example, if MinBW == 16,
5550 // "zext i8 %1 to i32" becomes "zext i8 %1 to i16".
5552 // Calculate the modified src and dest types.
5553 Type *MinVecTy = VectorTy;
5554 if (I->getOpcode() == Instruction::Trunc) {
5555 SrcVecTy = smallestIntegerVectorType(SrcVecTy, MinVecTy);
5556 VectorTy = largestIntegerVectorType(ToVectorTy(I->getType(), VF),
5558 } else if (I->getOpcode() == Instruction::ZExt ||
5559 I->getOpcode() == Instruction::SExt) {
5560 SrcVecTy = largestIntegerVectorType(SrcVecTy, MinVecTy);
5561 VectorTy = smallestIntegerVectorType(ToVectorTy(I->getType(), VF),
5566 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5568 case Instruction::Call: {
5569 bool NeedToScalarize;
5570 CallInst *CI = cast<CallInst>(I);
5571 unsigned CallCost = getVectorCallCost(CI, VF, TTI, TLI, NeedToScalarize);
5572 if (getIntrinsicIDForCall(CI, TLI))
5573 return std::min(CallCost, getVectorIntrinsicCost(CI, VF, TTI, TLI));
5577 // We are scalarizing the instruction. Return the cost of the scalar
5578 // instruction, plus the cost of insert and extract into vector
5579 // elements, times the vector width.
5582 if (!RetTy->isVoidTy() && VF != 1) {
5583 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5585 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5588 // The cost of inserting the results plus extracting each one of the
5590 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5593 // The cost of executing VF copies of the scalar instruction. This opcode
5594 // is unknown. Assume that it is the same as 'mul'.
5595 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5601 char LoopVectorize::ID = 0;
5602 static const char lv_name[] = "Loop Vectorization";
5603 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5604 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
5605 INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass)
5606 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
5607 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
5608 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
5609 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
5610 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5611 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
5612 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5613 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
5614 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5615 INITIALIZE_PASS_DEPENDENCY(LoopAccessAnalysis)
5616 INITIALIZE_PASS_DEPENDENCY(DemandedBits)
5617 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5620 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5621 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5625 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5626 // Check for a store.
5627 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5628 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5630 // Check for a load.
5631 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5632 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5638 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5639 bool IfPredicateStore) {
5640 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5641 // Holds vector parameters or scalars, in case of uniform vals.
5642 SmallVector<VectorParts, 4> Params;
5644 setDebugLocFromInst(Builder, Instr);
5646 // Find all of the vectorized parameters.
5647 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5648 Value *SrcOp = Instr->getOperand(op);
5650 // If we are accessing the old induction variable, use the new one.
5651 if (SrcOp == OldInduction) {
5652 Params.push_back(getVectorValue(SrcOp));
5656 // Try using previously calculated values.
5657 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5659 // If the src is an instruction that appeared earlier in the basic block
5660 // then it should already be vectorized.
5661 if (SrcInst && OrigLoop->contains(SrcInst)) {
5662 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5663 // The parameter is a vector value from earlier.
5664 Params.push_back(WidenMap.get(SrcInst));
5666 // The parameter is a scalar from outside the loop. Maybe even a constant.
5667 VectorParts Scalars;
5668 Scalars.append(UF, SrcOp);
5669 Params.push_back(Scalars);
5673 assert(Params.size() == Instr->getNumOperands() &&
5674 "Invalid number of operands");
5676 // Does this instruction return a value ?
5677 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5679 Value *UndefVec = IsVoidRetTy ? nullptr :
5680 UndefValue::get(Instr->getType());
5681 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5682 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5685 if (IfPredicateStore) {
5686 assert(Instr->getParent()->getSinglePredecessor() &&
5687 "Only support single predecessor blocks");
5688 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5689 Instr->getParent());
5692 // For each vector unroll 'part':
5693 for (unsigned Part = 0; Part < UF; ++Part) {
5694 // For each scalar that we create:
5696 // Start an "if (pred) a[i] = ..." block.
5697 Value *Cmp = nullptr;
5698 if (IfPredicateStore) {
5699 if (Cond[Part]->getType()->isVectorTy())
5701 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5702 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5703 ConstantInt::get(Cond[Part]->getType(), 1));
5706 Instruction *Cloned = Instr->clone();
5708 Cloned->setName(Instr->getName() + ".cloned");
5709 // Replace the operands of the cloned instructions with extracted scalars.
5710 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5711 Value *Op = Params[op][Part];
5712 Cloned->setOperand(op, Op);
5715 // Place the cloned scalar in the new loop.
5716 Builder.Insert(Cloned);
5718 // If the original scalar returns a value we need to place it in a vector
5719 // so that future users will be able to use it.
5721 VecResults[Part] = Cloned;
5724 if (IfPredicateStore)
5725 PredicatedStores.push_back(std::make_pair(cast<StoreInst>(Cloned),
5730 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5731 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5732 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5734 return scalarizeInstruction(Instr, IfPredicateStore);
5737 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5741 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5745 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) {
5746 // When unrolling and the VF is 1, we only need to add a simple scalar.
5747 Type *ITy = Val->getType();
5748 assert(!ITy->isVectorTy() && "Val must be a scalar");
5749 Constant *C = ConstantInt::get(ITy, StartIdx);
5750 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");