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 // Other ideas/concepts are from:
38 // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
40 // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of
41 // Vectorizing Compilers.
43 //===----------------------------------------------------------------------===//
45 #include "llvm/Transforms/Vectorize.h"
46 #include "llvm/ADT/DenseMap.h"
47 #include "llvm/ADT/EquivalenceClasses.h"
48 #include "llvm/ADT/Hashing.h"
49 #include "llvm/ADT/MapVector.h"
50 #include "llvm/ADT/SetVector.h"
51 #include "llvm/ADT/SmallPtrSet.h"
52 #include "llvm/ADT/SmallSet.h"
53 #include "llvm/ADT/SmallVector.h"
54 #include "llvm/ADT/Statistic.h"
55 #include "llvm/ADT/StringExtras.h"
56 #include "llvm/Analysis/AliasAnalysis.h"
57 #include "llvm/Analysis/AliasSetTracker.h"
58 #include "llvm/Analysis/AssumptionCache.h"
59 #include "llvm/Analysis/BlockFrequencyInfo.h"
60 #include "llvm/Analysis/CodeMetrics.h"
61 #include "llvm/Analysis/LoopAccessAnalysis.h"
62 #include "llvm/Analysis/LoopInfo.h"
63 #include "llvm/Analysis/LoopIterator.h"
64 #include "llvm/Analysis/LoopPass.h"
65 #include "llvm/Analysis/ScalarEvolution.h"
66 #include "llvm/Analysis/ScalarEvolutionExpander.h"
67 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
68 #include "llvm/Analysis/TargetTransformInfo.h"
69 #include "llvm/Analysis/ValueTracking.h"
70 #include "llvm/IR/Constants.h"
71 #include "llvm/IR/DataLayout.h"
72 #include "llvm/IR/DebugInfo.h"
73 #include "llvm/IR/DerivedTypes.h"
74 #include "llvm/IR/DiagnosticInfo.h"
75 #include "llvm/IR/Dominators.h"
76 #include "llvm/IR/Function.h"
77 #include "llvm/IR/IRBuilder.h"
78 #include "llvm/IR/Instructions.h"
79 #include "llvm/IR/IntrinsicInst.h"
80 #include "llvm/IR/LLVMContext.h"
81 #include "llvm/IR/Module.h"
82 #include "llvm/IR/PatternMatch.h"
83 #include "llvm/IR/Type.h"
84 #include "llvm/IR/Value.h"
85 #include "llvm/IR/ValueHandle.h"
86 #include "llvm/IR/Verifier.h"
87 #include "llvm/Pass.h"
88 #include "llvm/Support/BranchProbability.h"
89 #include "llvm/Support/CommandLine.h"
90 #include "llvm/Support/Debug.h"
91 #include "llvm/Support/raw_ostream.h"
92 #include "llvm/Transforms/Scalar.h"
93 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
94 #include "llvm/Transforms/Utils/Local.h"
95 #include "llvm/Transforms/Utils/VectorUtils.h"
100 using namespace llvm;
101 using namespace llvm::PatternMatch;
103 #define LV_NAME "loop-vectorize"
104 #define DEBUG_TYPE LV_NAME
106 STATISTIC(LoopsVectorized, "Number of loops vectorized");
107 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
110 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
111 cl::desc("Enable if-conversion during vectorization."));
113 /// We don't vectorize loops with a known constant trip count below this number.
114 static cl::opt<unsigned>
115 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
117 cl::desc("Don't vectorize loops with a constant "
118 "trip count that is smaller than this "
121 /// This enables versioning on the strides of symbolically striding memory
122 /// accesses in code like the following.
123 /// for (i = 0; i < N; ++i)
124 /// A[i * Stride1] += B[i * Stride2] ...
126 /// Will be roughly translated to
127 /// if (Stride1 == 1 && Stride2 == 1) {
128 /// for (i = 0; i < N; i+=4)
132 static cl::opt<bool> EnableMemAccessVersioning(
133 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
134 cl::desc("Enable symblic stride memory access versioning"));
136 /// We don't unroll loops with a known constant trip count below this number.
137 static const unsigned TinyTripCountUnrollThreshold = 128;
139 static cl::opt<unsigned> ForceTargetNumScalarRegs(
140 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
141 cl::desc("A flag that overrides the target's number of scalar registers."));
143 static cl::opt<unsigned> ForceTargetNumVectorRegs(
144 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
145 cl::desc("A flag that overrides the target's number of vector registers."));
147 /// Maximum vectorization interleave count.
148 static const unsigned MaxInterleaveFactor = 16;
150 static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
151 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
152 cl::desc("A flag that overrides the target's max interleave factor for "
155 static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
156 "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
157 cl::desc("A flag that overrides the target's max interleave factor for "
158 "vectorized loops."));
160 static cl::opt<unsigned> ForceTargetInstructionCost(
161 "force-target-instruction-cost", cl::init(0), cl::Hidden,
162 cl::desc("A flag that overrides the target's expected cost for "
163 "an instruction to a single constant value. Mostly "
164 "useful for getting consistent testing."));
166 static cl::opt<unsigned> SmallLoopCost(
167 "small-loop-cost", cl::init(20), cl::Hidden,
168 cl::desc("The cost of a loop that is considered 'small' by the unroller."));
170 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
171 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
172 cl::desc("Enable the use of the block frequency analysis to access PGO "
173 "heuristics minimizing code growth in cold regions and being more "
174 "aggressive in hot regions."));
176 // Runtime unroll loops for load/store throughput.
177 static cl::opt<bool> EnableLoadStoreRuntimeUnroll(
178 "enable-loadstore-runtime-unroll", cl::init(true), cl::Hidden,
179 cl::desc("Enable runtime unrolling until load/store ports are saturated"));
181 /// The number of stores in a loop that are allowed to need predication.
182 static cl::opt<unsigned> NumberOfStoresToPredicate(
183 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
184 cl::desc("Max number of stores to be predicated behind an if."));
186 static cl::opt<bool> EnableIndVarRegisterHeur(
187 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
188 cl::desc("Count the induction variable only once when unrolling"));
190 static cl::opt<bool> EnableCondStoresVectorization(
191 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
192 cl::desc("Enable if predication of stores during vectorization."));
194 static cl::opt<unsigned> MaxNestedScalarReductionUF(
195 "max-nested-scalar-reduction-unroll", cl::init(2), cl::Hidden,
196 cl::desc("The maximum unroll factor to use when unrolling a scalar "
197 "reduction in a nested loop."));
201 // Forward declarations.
202 class LoopVectorizationLegality;
203 class LoopVectorizationCostModel;
204 class LoopVectorizeHints;
206 /// \brief This modifies LoopAccessReport to initialize message with
207 /// loop-vectorizer-specific part.
208 class VectorizationReport : public LoopAccessReport {
210 VectorizationReport(Instruction *I = nullptr)
211 : LoopAccessReport("loop not vectorized: ", I) {}
213 /// \brief This allows promotion of the loop-access analysis report into the
214 /// loop-vectorizer report. It modifies the message to add the
215 /// loop-vectorizer-specific part of the message.
216 explicit VectorizationReport(const LoopAccessReport &R)
217 : LoopAccessReport(Twine("loop not vectorized: ") + R.str(),
221 /// A helper function for converting Scalar types to vector types.
222 /// If the incoming type is void, we return void. If the VF is 1, we return
224 static Type* ToVectorTy(Type *Scalar, unsigned VF) {
225 if (Scalar->isVoidTy() || VF == 1)
227 return VectorType::get(Scalar, VF);
230 /// InnerLoopVectorizer vectorizes loops which contain only one basic
231 /// block to a specified vectorization factor (VF).
232 /// This class performs the widening of scalars into vectors, or multiple
233 /// scalars. This class also implements the following features:
234 /// * It inserts an epilogue loop for handling loops that don't have iteration
235 /// counts that are known to be a multiple of the vectorization factor.
236 /// * It handles the code generation for reduction variables.
237 /// * Scalarization (implementation using scalars) of un-vectorizable
239 /// InnerLoopVectorizer does not perform any vectorization-legality
240 /// checks, and relies on the caller to check for the different legality
241 /// aspects. The InnerLoopVectorizer relies on the
242 /// LoopVectorizationLegality class to provide information about the induction
243 /// and reduction variables that were found to a given vectorization factor.
244 class InnerLoopVectorizer {
246 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
247 DominatorTree *DT, const DataLayout *DL,
248 const TargetLibraryInfo *TLI, unsigned VecWidth,
249 unsigned UnrollFactor)
250 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
251 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
252 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
253 Legal(nullptr), AddedSafetyChecks(false) {}
255 // Perform the actual loop widening (vectorization).
256 void vectorize(LoopVectorizationLegality *L) {
258 // Create a new empty loop. Unlink the old loop and connect the new one.
260 // Widen each instruction in the old loop to a new one in the new loop.
261 // Use the Legality module to find the induction and reduction variables.
263 // Register the new loop and update the analysis passes.
267 // Return true if any runtime check is added.
268 bool IsSafetyChecksAdded() {
269 return AddedSafetyChecks;
272 virtual ~InnerLoopVectorizer() {}
275 /// A small list of PHINodes.
276 typedef SmallVector<PHINode*, 4> PhiVector;
277 /// When we unroll loops we have multiple vector values for each scalar.
278 /// This data structure holds the unrolled and vectorized values that
279 /// originated from one scalar instruction.
280 typedef SmallVector<Value*, 2> VectorParts;
282 // When we if-convert we need create edge masks. We have to cache values so
283 // that we don't end up with exponential recursion/IR.
284 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
285 VectorParts> EdgeMaskCache;
287 /// \brief Add checks for strides that where assumed to be 1.
289 /// Returns the last check instruction and the first check instruction in the
290 /// pair as (first, last).
291 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
293 /// Create an empty loop, based on the loop ranges of the old loop.
294 void createEmptyLoop();
295 /// Copy and widen the instructions from the old loop.
296 virtual void vectorizeLoop();
298 /// \brief The Loop exit block may have single value PHI nodes where the
299 /// incoming value is 'Undef'. While vectorizing we only handled real values
300 /// that were defined inside the loop. Here we fix the 'undef case'.
304 /// A helper function that computes the predicate of the block BB, assuming
305 /// that the header block of the loop is set to True. It returns the *entry*
306 /// mask for the block BB.
307 VectorParts createBlockInMask(BasicBlock *BB);
308 /// A helper function that computes the predicate of the edge between SRC
310 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
312 /// A helper function to vectorize a single BB within the innermost loop.
313 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
315 /// Vectorize a single PHINode in a block. This method handles the induction
316 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
317 /// arbitrary length vectors.
318 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
319 unsigned UF, unsigned VF, PhiVector *PV);
321 /// Insert the new loop to the loop hierarchy and pass manager
322 /// and update the analysis passes.
323 void updateAnalysis();
325 /// This instruction is un-vectorizable. Implement it as a sequence
326 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
327 /// scalarized instruction behind an if block predicated on the control
328 /// dependence of the instruction.
329 virtual void scalarizeInstruction(Instruction *Instr,
330 bool IfPredicateStore=false);
332 /// Vectorize Load and Store instructions,
333 virtual void vectorizeMemoryInstruction(Instruction *Instr);
335 /// Create a broadcast instruction. This method generates a broadcast
336 /// instruction (shuffle) for loop invariant values and for the induction
337 /// value. If this is the induction variable then we extend it to N, N+1, ...
338 /// this is needed because each iteration in the loop corresponds to a SIMD
340 virtual Value *getBroadcastInstrs(Value *V);
342 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
343 /// to each vector element of Val. The sequence starts at StartIndex.
344 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step);
346 /// When we go over instructions in the basic block we rely on previous
347 /// values within the current basic block or on loop invariant values.
348 /// When we widen (vectorize) values we place them in the map. If the values
349 /// are not within the map, they have to be loop invariant, so we simply
350 /// broadcast them into a vector.
351 VectorParts &getVectorValue(Value *V);
353 /// Generate a shuffle sequence that will reverse the vector Vec.
354 virtual Value *reverseVector(Value *Vec);
356 /// This is a helper class that holds the vectorizer state. It maps scalar
357 /// instructions to vector instructions. When the code is 'unrolled' then
358 /// then a single scalar value is mapped to multiple vector parts. The parts
359 /// are stored in the VectorPart type.
361 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
363 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
365 /// \return True if 'Key' is saved in the Value Map.
366 bool has(Value *Key) const { return MapStorage.count(Key); }
368 /// Initializes a new entry in the map. Sets all of the vector parts to the
369 /// save value in 'Val'.
370 /// \return A reference to a vector with splat values.
371 VectorParts &splat(Value *Key, Value *Val) {
372 VectorParts &Entry = MapStorage[Key];
373 Entry.assign(UF, Val);
377 ///\return A reference to the value that is stored at 'Key'.
378 VectorParts &get(Value *Key) {
379 VectorParts &Entry = MapStorage[Key];
382 assert(Entry.size() == UF);
387 /// The unroll factor. Each entry in the map stores this number of vector
391 /// Map storage. We use std::map and not DenseMap because insertions to a
392 /// dense map invalidates its iterators.
393 std::map<Value *, VectorParts> MapStorage;
396 /// The original loop.
398 /// Scev analysis to use.
407 const DataLayout *DL;
408 /// Target Library Info.
409 const TargetLibraryInfo *TLI;
411 /// The vectorization SIMD factor to use. Each vector will have this many
416 /// The vectorization unroll factor to use. Each scalar is vectorized to this
417 /// many different vector instructions.
420 /// The builder that we use
423 // --- Vectorization state ---
425 /// The vector-loop preheader.
426 BasicBlock *LoopVectorPreHeader;
427 /// The scalar-loop preheader.
428 BasicBlock *LoopScalarPreHeader;
429 /// Middle Block between the vector and the scalar.
430 BasicBlock *LoopMiddleBlock;
431 ///The ExitBlock of the scalar loop.
432 BasicBlock *LoopExitBlock;
433 ///The vector loop body.
434 SmallVector<BasicBlock *, 4> LoopVectorBody;
435 ///The scalar loop body.
436 BasicBlock *LoopScalarBody;
437 /// A list of all bypass blocks. The first block is the entry of the loop.
438 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
440 /// The new Induction variable which was added to the new block.
442 /// The induction variable of the old basic block.
443 PHINode *OldInduction;
444 /// Holds the extended (to the widest induction type) start index.
446 /// Maps scalars to widened vectors.
448 EdgeMaskCache MaskCache;
450 LoopVectorizationLegality *Legal;
452 // Record whether runtime check is added.
453 bool AddedSafetyChecks;
456 class InnerLoopUnroller : public InnerLoopVectorizer {
458 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
459 DominatorTree *DT, const DataLayout *DL,
460 const TargetLibraryInfo *TLI, unsigned UnrollFactor) :
461 InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { }
464 void scalarizeInstruction(Instruction *Instr,
465 bool IfPredicateStore = false) override;
466 void vectorizeMemoryInstruction(Instruction *Instr) override;
467 Value *getBroadcastInstrs(Value *V) override;
468 Value *getStepVector(Value *Val, int StartIdx, Value *Step) override;
469 Value *reverseVector(Value *Vec) override;
472 /// \brief Look for a meaningful debug location on the instruction or it's
474 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
479 if (I->getDebugLoc() != Empty)
482 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
483 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
484 if (OpInst->getDebugLoc() != Empty)
491 /// \brief Set the debug location in the builder using the debug location in the
493 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
494 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
495 B.SetCurrentDebugLocation(Inst->getDebugLoc());
497 B.SetCurrentDebugLocation(DebugLoc());
501 /// \return string containing a file name and a line # for the given loop.
502 static std::string getDebugLocString(const Loop *L) {
505 raw_string_ostream OS(Result);
506 const DebugLoc LoopDbgLoc = L->getStartLoc();
507 if (!LoopDbgLoc.isUnknown())
508 LoopDbgLoc.print(OS);
510 // Just print the module name.
511 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
518 /// \brief Propagate known metadata from one instruction to another.
519 static void propagateMetadata(Instruction *To, const Instruction *From) {
520 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
521 From->getAllMetadataOtherThanDebugLoc(Metadata);
523 for (auto M : Metadata) {
524 unsigned Kind = M.first;
526 // These are safe to transfer (this is safe for TBAA, even when we
527 // if-convert, because should that metadata have had a control dependency
528 // on the condition, and thus actually aliased with some other
529 // non-speculated memory access when the condition was false, this would be
530 // caught by the runtime overlap checks).
531 if (Kind != LLVMContext::MD_tbaa &&
532 Kind != LLVMContext::MD_alias_scope &&
533 Kind != LLVMContext::MD_noalias &&
534 Kind != LLVMContext::MD_fpmath)
537 To->setMetadata(Kind, M.second);
541 /// \brief Propagate known metadata from one instruction to a vector of others.
542 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
544 if (Instruction *I = dyn_cast<Instruction>(V))
545 propagateMetadata(I, From);
548 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
549 /// to what vectorization factor.
550 /// This class does not look at the profitability of vectorization, only the
551 /// legality. This class has two main kinds of checks:
552 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
553 /// will change the order of memory accesses in a way that will change the
554 /// correctness of the program.
555 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
556 /// checks for a number of different conditions, such as the availability of a
557 /// single induction variable, that all types are supported and vectorize-able,
558 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
559 /// This class is also used by InnerLoopVectorizer for identifying
560 /// induction variable and the different reduction variables.
561 class LoopVectorizationLegality {
563 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, const DataLayout *DL,
564 DominatorTree *DT, TargetLibraryInfo *TLI,
565 AliasAnalysis *AA, Function *F,
566 const TargetTransformInfo *TTI,
567 LoopAccessAnalysis *LAA)
568 : NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
569 TLI(TLI), TheFunction(F), TTI(TTI), DT(DT), LAA(LAA), LAI(nullptr),
570 Induction(nullptr), WidestIndTy(nullptr), HasFunNoNaNAttr(false) {}
572 /// This enum represents the kinds of reductions that we support.
574 RK_NoReduction, ///< Not a reduction.
575 RK_IntegerAdd, ///< Sum of integers.
576 RK_IntegerMult, ///< Product of integers.
577 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
578 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
579 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
580 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
581 RK_FloatAdd, ///< Sum of floats.
582 RK_FloatMult, ///< Product of floats.
583 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
586 /// This enum represents the kinds of inductions that we support.
588 IK_NoInduction, ///< Not an induction variable.
589 IK_IntInduction, ///< Integer induction variable. Step = C.
590 IK_PtrInduction ///< Pointer induction var. Step = C / sizeof(elem).
593 // This enum represents the kind of minmax reduction.
594 enum MinMaxReductionKind {
604 /// This struct holds information about reduction variables.
605 struct ReductionDescriptor {
606 ReductionDescriptor() : StartValue(nullptr), LoopExitInstr(nullptr),
607 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
609 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
610 MinMaxReductionKind MK)
611 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
613 // The starting value of the reduction.
614 // It does not have to be zero!
615 TrackingVH<Value> StartValue;
616 // The instruction who's value is used outside the loop.
617 Instruction *LoopExitInstr;
618 // The kind of the reduction.
620 // If this a min/max reduction the kind of reduction.
621 MinMaxReductionKind MinMaxKind;
624 /// This POD struct holds information about a potential reduction operation.
625 struct ReductionInstDesc {
626 ReductionInstDesc(bool IsRedux, Instruction *I) :
627 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
629 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
630 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
632 // Is this instruction a reduction candidate.
634 // The last instruction in a min/max pattern (select of the select(icmp())
635 // pattern), or the current reduction instruction otherwise.
636 Instruction *PatternLastInst;
637 // If this is a min/max pattern the comparison predicate.
638 MinMaxReductionKind MinMaxKind;
641 /// A struct for saving information about induction variables.
642 struct InductionInfo {
643 InductionInfo(Value *Start, InductionKind K, ConstantInt *Step)
644 : StartValue(Start), IK(K), StepValue(Step) {
645 assert(IK != IK_NoInduction && "Not an induction");
646 assert(StartValue && "StartValue is null");
647 assert(StepValue && !StepValue->isZero() && "StepValue is zero");
648 assert((IK != IK_PtrInduction || StartValue->getType()->isPointerTy()) &&
649 "StartValue is not a pointer for pointer induction");
650 assert((IK != IK_IntInduction || StartValue->getType()->isIntegerTy()) &&
651 "StartValue is not an integer for integer induction");
652 assert(StepValue->getType()->isIntegerTy() &&
653 "StepValue is not an integer");
656 : StartValue(nullptr), IK(IK_NoInduction), StepValue(nullptr) {}
658 /// Get the consecutive direction. Returns:
659 /// 0 - unknown or non-consecutive.
660 /// 1 - consecutive and increasing.
661 /// -1 - consecutive and decreasing.
662 int getConsecutiveDirection() const {
663 if (StepValue && (StepValue->isOne() || StepValue->isMinusOne()))
664 return StepValue->getSExtValue();
668 /// Compute the transformed value of Index at offset StartValue using step
670 /// For integer induction, returns StartValue + Index * StepValue.
671 /// For pointer induction, returns StartValue[Index * StepValue].
672 /// FIXME: The newly created binary instructions should contain nsw/nuw
673 /// flags, which can be found from the original scalar operations.
674 Value *transform(IRBuilder<> &B, Value *Index) const {
676 case IK_IntInduction:
677 assert(Index->getType() == StartValue->getType() &&
678 "Index type does not match StartValue type");
679 if (StepValue->isMinusOne())
680 return B.CreateSub(StartValue, Index);
681 if (!StepValue->isOne())
682 Index = B.CreateMul(Index, StepValue);
683 return B.CreateAdd(StartValue, Index);
685 case IK_PtrInduction:
686 if (StepValue->isMinusOne())
687 Index = B.CreateNeg(Index);
688 else if (!StepValue->isOne())
689 Index = B.CreateMul(Index, StepValue);
690 return B.CreateGEP(StartValue, Index);
695 llvm_unreachable("invalid enum");
699 TrackingVH<Value> StartValue;
703 ConstantInt *StepValue;
706 /// ReductionList contains the reduction descriptors for all
707 /// of the reductions that were found in the loop.
708 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
710 /// InductionList saves induction variables and maps them to the
711 /// induction descriptor.
712 typedef MapVector<PHINode*, InductionInfo> InductionList;
714 /// Returns true if it is legal to vectorize this loop.
715 /// This does not mean that it is profitable to vectorize this
716 /// loop, only that it is legal to do so.
719 /// Returns the Induction variable.
720 PHINode *getInduction() { return Induction; }
722 /// Returns the reduction variables found in the loop.
723 ReductionList *getReductionVars() { return &Reductions; }
725 /// Returns the induction variables found in the loop.
726 InductionList *getInductionVars() { return &Inductions; }
728 /// Returns the widest induction type.
729 Type *getWidestInductionType() { return WidestIndTy; }
731 /// Returns True if V is an induction variable in this loop.
732 bool isInductionVariable(const Value *V);
734 /// Return true if the block BB needs to be predicated in order for the loop
735 /// to be vectorized.
736 bool blockNeedsPredication(BasicBlock *BB);
738 /// Check if this pointer is consecutive when vectorizing. This happens
739 /// when the last index of the GEP is the induction variable, or that the
740 /// pointer itself is an induction variable.
741 /// This check allows us to vectorize A[idx] into a wide load/store.
743 /// 0 - Stride is unknown or non-consecutive.
744 /// 1 - Address is consecutive.
745 /// -1 - Address is consecutive, and decreasing.
746 int isConsecutivePtr(Value *Ptr);
748 /// Returns true if the value V is uniform within the loop.
749 bool isUniform(Value *V);
751 /// Returns true if this instruction will remain scalar after vectorization.
752 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
754 /// Returns the information that we collected about runtime memory check.
755 const LoopAccessInfo::RuntimePointerCheck *getRuntimePointerCheck() const {
756 return LAI->getRuntimePointerCheck();
759 const LoopAccessInfo *getLAI() const {
763 /// This function returns the identity element (or neutral element) for
765 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
767 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
769 bool hasStride(Value *V) { return StrideSet.count(V); }
770 bool mustCheckStrides() { return !StrideSet.empty(); }
771 SmallPtrSet<Value *, 8>::iterator strides_begin() {
772 return StrideSet.begin();
774 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
776 /// Returns true if the target machine supports masked store operation
777 /// for the given \p DataType and kind of access to \p Ptr.
778 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
779 return TTI->isLegalMaskedStore(DataType, isConsecutivePtr(Ptr));
781 /// Returns true if the target machine supports masked load operation
782 /// for the given \p DataType and kind of access to \p Ptr.
783 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
784 return TTI->isLegalMaskedLoad(DataType, isConsecutivePtr(Ptr));
786 /// Returns true if vector representation of the instruction \p I
788 bool isMaskRequired(const Instruction* I) {
789 return (MaskedOp.count(I) != 0);
791 unsigned getNumStores() const {
792 return LAI->getNumStores();
794 unsigned getNumLoads() const {
795 return LAI->getNumLoads();
797 unsigned getNumPredStores() const {
798 return NumPredStores;
801 /// Check if a single basic block loop is vectorizable.
802 /// At this point we know that this is a loop with a constant trip count
803 /// and we only need to check individual instructions.
804 bool canVectorizeInstrs();
806 /// When we vectorize loops we may change the order in which
807 /// we read and write from memory. This method checks if it is
808 /// legal to vectorize the code, considering only memory constrains.
809 /// Returns true if the loop is vectorizable
810 bool canVectorizeMemory();
812 /// Return true if we can vectorize this loop using the IF-conversion
814 bool canVectorizeWithIfConvert();
816 /// Collect the variables that need to stay uniform after vectorization.
817 void collectLoopUniforms();
819 /// Return true if all of the instructions in the block can be speculatively
820 /// executed. \p SafePtrs is a list of addresses that are known to be legal
821 /// and we know that we can read from them without segfault.
822 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
824 /// Returns True, if 'Phi' is the kind of reduction variable for type
825 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
826 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
827 /// Returns a struct describing if the instruction 'I' can be a reduction
828 /// variable of type 'Kind'. If the reduction is a min/max pattern of
829 /// select(icmp()) this function advances the instruction pointer 'I' from the
830 /// compare instruction to the select instruction and stores this pointer in
831 /// 'PatternLastInst' member of the returned struct.
832 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
833 ReductionInstDesc &Desc);
834 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
835 /// pattern corresponding to a min(X, Y) or max(X, Y).
836 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
837 ReductionInstDesc &Prev);
838 /// Returns the induction kind of Phi and record the step. This function may
839 /// return NoInduction if the PHI is not an induction variable.
840 InductionKind isInductionVariable(PHINode *Phi, ConstantInt *&StepValue);
842 /// \brief Collect memory access with loop invariant strides.
844 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
846 void collectStridedAccess(Value *LoadOrStoreInst);
848 /// Report an analysis message to assist the user in diagnosing loops that are
849 /// not vectorized. These are handled as LoopAccessReport rather than
850 /// VectorizationReport because the << operator of VectorizationReport returns
851 /// LoopAccessReport.
852 void emitAnalysis(const LoopAccessReport &Message) {
853 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, LV_NAME);
856 unsigned NumPredStores;
858 /// The loop that we evaluate.
862 /// DataLayout analysis.
863 const DataLayout *DL;
864 /// Target Library Info.
865 TargetLibraryInfo *TLI;
867 Function *TheFunction;
868 /// Target Transform Info
869 const TargetTransformInfo *TTI;
872 // LoopAccess analysis.
873 LoopAccessAnalysis *LAA;
874 // And the loop-accesses info corresponding to this loop. This pointer is
875 // null until canVectorizeMemory sets it up.
876 const LoopAccessInfo *LAI;
878 // --- vectorization state --- //
880 /// Holds the integer induction variable. This is the counter of the
883 /// Holds the reduction variables.
884 ReductionList Reductions;
885 /// Holds all of the induction variables that we found in the loop.
886 /// Notice that inductions don't need to start at zero and that induction
887 /// variables can be pointers.
888 InductionList Inductions;
889 /// Holds the widest induction type encountered.
892 /// Allowed outside users. This holds the reduction
893 /// vars which can be accessed from outside the loop.
894 SmallPtrSet<Value*, 4> AllowedExit;
895 /// This set holds the variables which are known to be uniform after
897 SmallPtrSet<Instruction*, 4> Uniforms;
899 /// Can we assume the absence of NaNs.
900 bool HasFunNoNaNAttr;
902 ValueToValueMap Strides;
903 SmallPtrSet<Value *, 8> StrideSet;
905 /// While vectorizing these instructions we have to generate a
906 /// call to the appropriate masked intrinsic
907 SmallPtrSet<const Instruction*, 8> MaskedOp;
910 /// LoopVectorizationCostModel - estimates the expected speedups due to
912 /// In many cases vectorization is not profitable. This can happen because of
913 /// a number of reasons. In this class we mainly attempt to predict the
914 /// expected speedup/slowdowns due to the supported instruction set. We use the
915 /// TargetTransformInfo to query the different backends for the cost of
916 /// different operations.
917 class LoopVectorizationCostModel {
919 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
920 LoopVectorizationLegality *Legal,
921 const TargetTransformInfo &TTI,
922 const DataLayout *DL, const TargetLibraryInfo *TLI,
923 AssumptionCache *AC, const Function *F,
924 const LoopVectorizeHints *Hints)
925 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI),
926 TheFunction(F), Hints(Hints) {
927 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
930 /// Information about vectorization costs
931 struct VectorizationFactor {
932 unsigned Width; // Vector width with best cost
933 unsigned Cost; // Cost of the loop with that width
935 /// \return The most profitable vectorization factor and the cost of that VF.
936 /// This method checks every power of two up to VF. If UserVF is not ZERO
937 /// then this vectorization factor will be selected if vectorization is
939 VectorizationFactor selectVectorizationFactor(bool OptForSize);
941 /// \return The size (in bits) of the widest type in the code that
942 /// needs to be vectorized. We ignore values that remain scalar such as
943 /// 64 bit loop indices.
944 unsigned getWidestType();
946 /// \return The most profitable unroll factor.
947 /// If UserUF is non-zero then this method finds the best unroll-factor
948 /// based on register pressure and other parameters.
949 /// VF and LoopCost are the selected vectorization factor and the cost of the
951 unsigned selectUnrollFactor(bool OptForSize, unsigned VF, unsigned LoopCost);
953 /// \brief A struct that represents some properties of the register usage
955 struct RegisterUsage {
956 /// Holds the number of loop invariant values that are used in the loop.
957 unsigned LoopInvariantRegs;
958 /// Holds the maximum number of concurrent live intervals in the loop.
959 unsigned MaxLocalUsers;
960 /// Holds the number of instructions in the loop.
961 unsigned NumInstructions;
964 /// \return information about the register usage of the loop.
965 RegisterUsage calculateRegisterUsage();
968 /// Returns the expected execution cost. The unit of the cost does
969 /// not matter because we use the 'cost' units to compare different
970 /// vector widths. The cost that is returned is *not* normalized by
971 /// the factor width.
972 unsigned expectedCost(unsigned VF);
974 /// Returns the execution time cost of an instruction for a given vector
975 /// width. Vector width of one means scalar.
976 unsigned getInstructionCost(Instruction *I, unsigned VF);
978 /// Returns whether the instruction is a load or store and will be a emitted
979 /// as a vector operation.
980 bool isConsecutiveLoadOrStore(Instruction *I);
982 /// Report an analysis message to assist the user in diagnosing loops that are
983 /// not vectorized. These are handled as LoopAccessReport rather than
984 /// VectorizationReport because the << operator of VectorizationReport returns
985 /// LoopAccessReport.
986 void emitAnalysis(const LoopAccessReport &Message) {
987 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, LV_NAME);
990 /// Values used only by @llvm.assume calls.
991 SmallPtrSet<const Value *, 32> EphValues;
993 /// The loop that we evaluate.
997 /// Loop Info analysis.
999 /// Vectorization legality.
1000 LoopVectorizationLegality *Legal;
1001 /// Vector target information.
1002 const TargetTransformInfo &TTI;
1003 /// Target data layout information.
1004 const DataLayout *DL;
1005 /// Target Library Info.
1006 const TargetLibraryInfo *TLI;
1007 const Function *TheFunction;
1008 // Loop Vectorize Hint.
1009 const LoopVectorizeHints *Hints;
1012 /// Utility class for getting and setting loop vectorizer hints in the form
1013 /// of loop metadata.
1014 /// This class keeps a number of loop annotations locally (as member variables)
1015 /// and can, upon request, write them back as metadata on the loop. It will
1016 /// initially scan the loop for existing metadata, and will update the local
1017 /// values based on information in the loop.
1018 /// We cannot write all values to metadata, as the mere presence of some info,
1019 /// for example 'force', means a decision has been made. So, we need to be
1020 /// careful NOT to add them if the user hasn't specifically asked so.
1021 class LoopVectorizeHints {
1028 /// Hint - associates name and validation with the hint value.
1031 unsigned Value; // This may have to change for non-numeric values.
1034 Hint(const char * Name, unsigned Value, HintKind Kind)
1035 : Name(Name), Value(Value), Kind(Kind) { }
1037 bool validate(unsigned Val) {
1040 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
1042 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1050 /// Vectorization width.
1052 /// Vectorization interleave factor.
1054 /// Vectorization forced
1057 /// Return the loop metadata prefix.
1058 static StringRef Prefix() { return "llvm.loop."; }
1062 FK_Undefined = -1, ///< Not selected.
1063 FK_Disabled = 0, ///< Forcing disabled.
1064 FK_Enabled = 1, ///< Forcing enabled.
1067 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
1068 : Width("vectorize.width", VectorizerParams::VectorizationFactor,
1070 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1071 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1073 // Populate values with existing loop metadata.
1074 getHintsFromMetadata();
1076 // force-vector-interleave overrides DisableInterleaving.
1077 if (VectorizerParams::isInterleaveForced())
1078 Interleave.Value = VectorizerParams::VectorizationInterleave;
1080 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
1081 << "LV: Interleaving disabled by the pass manager\n");
1084 /// Mark the loop L as already vectorized by setting the width to 1.
1085 void setAlreadyVectorized() {
1086 Width.Value = Interleave.Value = 1;
1087 Hint Hints[] = {Width, Interleave};
1088 writeHintsToMetadata(Hints);
1091 /// Dumps all the hint information.
1092 std::string emitRemark() const {
1093 VectorizationReport R;
1094 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1095 R << "vectorization is explicitly disabled";
1097 R << "use -Rpass-analysis=loop-vectorize for more info";
1098 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1099 R << " (Force=true";
1100 if (Width.Value != 0)
1101 R << ", Vector Width=" << Width.Value;
1102 if (Interleave.Value != 0)
1103 R << ", Interleave Count=" << Interleave.Value;
1111 unsigned getWidth() const { return Width.Value; }
1112 unsigned getInterleave() const { return Interleave.Value; }
1113 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1116 /// Find hints specified in the loop metadata and update local values.
1117 void getHintsFromMetadata() {
1118 MDNode *LoopID = TheLoop->getLoopID();
1122 // First operand should refer to the loop id itself.
1123 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1124 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1126 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1127 const MDString *S = nullptr;
1128 SmallVector<Metadata *, 4> Args;
1130 // The expected hint is either a MDString or a MDNode with the first
1131 // operand a MDString.
1132 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1133 if (!MD || MD->getNumOperands() == 0)
1135 S = dyn_cast<MDString>(MD->getOperand(0));
1136 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1137 Args.push_back(MD->getOperand(i));
1139 S = dyn_cast<MDString>(LoopID->getOperand(i));
1140 assert(Args.size() == 0 && "too many arguments for MDString");
1146 // Check if the hint starts with the loop metadata prefix.
1147 StringRef Name = S->getString();
1148 if (Args.size() == 1)
1149 setHint(Name, Args[0]);
1153 /// Checks string hint with one operand and set value if valid.
1154 void setHint(StringRef Name, Metadata *Arg) {
1155 if (!Name.startswith(Prefix()))
1157 Name = Name.substr(Prefix().size(), StringRef::npos);
1159 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1161 unsigned Val = C->getZExtValue();
1163 Hint *Hints[] = {&Width, &Interleave, &Force};
1164 for (auto H : Hints) {
1165 if (Name == H->Name) {
1166 if (H->validate(Val))
1169 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1175 /// Create a new hint from name / value pair.
1176 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1177 LLVMContext &Context = TheLoop->getHeader()->getContext();
1178 Metadata *MDs[] = {MDString::get(Context, Name),
1179 ConstantAsMetadata::get(
1180 ConstantInt::get(Type::getInt32Ty(Context), V))};
1181 return MDNode::get(Context, MDs);
1184 /// Matches metadata with hint name.
1185 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1186 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1190 for (auto H : HintTypes)
1191 if (Name->getString().endswith(H.Name))
1196 /// Sets current hints into loop metadata, keeping other values intact.
1197 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1198 if (HintTypes.size() == 0)
1201 // Reserve the first element to LoopID (see below).
1202 SmallVector<Metadata *, 4> MDs(1);
1203 // If the loop already has metadata, then ignore the existing operands.
1204 MDNode *LoopID = TheLoop->getLoopID();
1206 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1207 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1208 // If node in update list, ignore old value.
1209 if (!matchesHintMetadataName(Node, HintTypes))
1210 MDs.push_back(Node);
1214 // Now, add the missing hints.
1215 for (auto H : HintTypes)
1216 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1218 // Replace current metadata node with new one.
1219 LLVMContext &Context = TheLoop->getHeader()->getContext();
1220 MDNode *NewLoopID = MDNode::get(Context, MDs);
1221 // Set operand 0 to refer to the loop id itself.
1222 NewLoopID->replaceOperandWith(0, NewLoopID);
1224 TheLoop->setLoopID(NewLoopID);
1227 /// The loop these hints belong to.
1228 const Loop *TheLoop;
1231 static void emitMissedWarning(Function *F, Loop *L,
1232 const LoopVectorizeHints &LH) {
1233 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1234 L->getStartLoc(), LH.emitRemark());
1236 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1237 if (LH.getWidth() != 1)
1238 emitLoopVectorizeWarning(
1239 F->getContext(), *F, L->getStartLoc(),
1240 "failed explicitly specified loop vectorization");
1241 else if (LH.getInterleave() != 1)
1242 emitLoopInterleaveWarning(
1243 F->getContext(), *F, L->getStartLoc(),
1244 "failed explicitly specified loop interleaving");
1248 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1250 return V.push_back(&L);
1252 for (Loop *InnerL : L)
1253 addInnerLoop(*InnerL, V);
1256 /// The LoopVectorize Pass.
1257 struct LoopVectorize : public FunctionPass {
1258 /// Pass identification, replacement for typeid
1261 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1263 DisableUnrolling(NoUnrolling),
1264 AlwaysVectorize(AlwaysVectorize) {
1265 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1268 ScalarEvolution *SE;
1269 const DataLayout *DL;
1271 TargetTransformInfo *TTI;
1273 BlockFrequencyInfo *BFI;
1274 TargetLibraryInfo *TLI;
1276 AssumptionCache *AC;
1277 LoopAccessAnalysis *LAA;
1278 bool DisableUnrolling;
1279 bool AlwaysVectorize;
1281 BlockFrequency ColdEntryFreq;
1283 bool runOnFunction(Function &F) override {
1284 SE = &getAnalysis<ScalarEvolution>();
1285 DL = &F.getParent()->getDataLayout();
1286 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1287 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1288 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1289 BFI = &getAnalysis<BlockFrequencyInfo>();
1290 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1291 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1292 AA = &getAnalysis<AliasAnalysis>();
1293 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1294 LAA = &getAnalysis<LoopAccessAnalysis>();
1296 // Compute some weights outside of the loop over the loops. Compute this
1297 // using a BranchProbability to re-use its scaling math.
1298 const BranchProbability ColdProb(1, 5); // 20%
1299 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1301 // If the target claims to have no vector registers don't attempt
1303 if (!TTI->getNumberOfRegisters(true))
1307 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1308 << ": Missing data layout\n");
1312 // Build up a worklist of inner-loops to vectorize. This is necessary as
1313 // the act of vectorizing or partially unrolling a loop creates new loops
1314 // and can invalidate iterators across the loops.
1315 SmallVector<Loop *, 8> Worklist;
1318 addInnerLoop(*L, Worklist);
1320 LoopsAnalyzed += Worklist.size();
1322 // Now walk the identified inner loops.
1323 bool Changed = false;
1324 while (!Worklist.empty())
1325 Changed |= processLoop(Worklist.pop_back_val());
1327 // Process each loop nest in the function.
1331 static void AddRuntimeUnrollDisableMetaData(Loop *L) {
1332 SmallVector<Metadata *, 4> MDs;
1333 // Reserve first location for self reference to the LoopID metadata node.
1334 MDs.push_back(nullptr);
1335 bool IsUnrollMetadata = false;
1336 MDNode *LoopID = L->getLoopID();
1338 // First find existing loop unrolling disable metadata.
1339 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1340 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
1342 const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
1344 S && S->getString().startswith("llvm.loop.unroll.disable");
1346 MDs.push_back(LoopID->getOperand(i));
1350 if (!IsUnrollMetadata) {
1351 // Add runtime unroll disable metadata.
1352 LLVMContext &Context = L->getHeader()->getContext();
1353 SmallVector<Metadata *, 1> DisableOperands;
1354 DisableOperands.push_back(
1355 MDString::get(Context, "llvm.loop.unroll.runtime.disable"));
1356 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
1357 MDs.push_back(DisableNode);
1358 MDNode *NewLoopID = MDNode::get(Context, MDs);
1359 // Set operand 0 to refer to the loop id itself.
1360 NewLoopID->replaceOperandWith(0, NewLoopID);
1361 L->setLoopID(NewLoopID);
1365 bool processLoop(Loop *L) {
1366 assert(L->empty() && "Only process inner loops.");
1369 const std::string DebugLocStr = getDebugLocString(L);
1372 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1373 << L->getHeader()->getParent()->getName() << "\" from "
1374 << DebugLocStr << "\n");
1376 LoopVectorizeHints Hints(L, DisableUnrolling);
1378 DEBUG(dbgs() << "LV: Loop hints:"
1380 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1382 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1384 : "?")) << " width=" << Hints.getWidth()
1385 << " unroll=" << Hints.getInterleave() << "\n");
1387 // Function containing loop
1388 Function *F = L->getHeader()->getParent();
1390 // Looking at the diagnostic output is the only way to determine if a loop
1391 // was vectorized (other than looking at the IR or machine code), so it
1392 // is important to generate an optimization remark for each loop. Most of
1393 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1394 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1395 // less verbose reporting vectorized loops and unvectorized loops that may
1396 // benefit from vectorization, respectively.
1398 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1399 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1400 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1401 L->getStartLoc(), Hints.emitRemark());
1405 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1406 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1407 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1408 L->getStartLoc(), Hints.emitRemark());
1412 if (Hints.getWidth() == 1 && Hints.getInterleave() == 1) {
1413 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1414 emitOptimizationRemarkAnalysis(
1415 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1416 "loop not vectorized: vector width and interleave count are "
1417 "explicitly set to 1");
1421 // Check the loop for a trip count threshold:
1422 // do not vectorize loops with a tiny trip count.
1423 const unsigned TC = SE->getSmallConstantTripCount(L);
1424 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1425 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1426 << "This loop is not worth vectorizing.");
1427 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1428 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1430 DEBUG(dbgs() << "\n");
1431 emitOptimizationRemarkAnalysis(
1432 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1433 "vectorization is not beneficial and is not explicitly forced");
1438 // Check if it is legal to vectorize the loop.
1439 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, AA, F, TTI, LAA);
1440 if (!LVL.canVectorize()) {
1441 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1442 emitMissedWarning(F, L, Hints);
1446 // Use the cost model.
1447 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI, AC, F,
1450 // Check the function attributes to find out if this function should be
1451 // optimized for size.
1452 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1453 F->hasFnAttribute(Attribute::OptimizeForSize);
1455 // Compute the weighted frequency of this loop being executed and see if it
1456 // is less than 20% of the function entry baseline frequency. Note that we
1457 // always have a canonical loop here because we think we *can* vectoriez.
1458 // FIXME: This is hidden behind a flag due to pervasive problems with
1459 // exactly what block frequency models.
1460 if (LoopVectorizeWithBlockFrequency) {
1461 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1462 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1463 LoopEntryFreq < ColdEntryFreq)
1467 // Check the function attributes to see if implicit floats are allowed.a
1468 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1469 // an integer loop and the vector instructions selected are purely integer
1470 // vector instructions?
1471 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1472 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1473 "attribute is used.\n");
1474 emitOptimizationRemarkAnalysis(
1475 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1476 "loop not vectorized due to NoImplicitFloat attribute");
1477 emitMissedWarning(F, L, Hints);
1481 // Select the optimal vectorization factor.
1482 const LoopVectorizationCostModel::VectorizationFactor VF =
1483 CM.selectVectorizationFactor(OptForSize);
1485 // Select the unroll factor.
1487 CM.selectUnrollFactor(OptForSize, VF.Width, VF.Cost);
1489 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1490 << DebugLocStr << '\n');
1491 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1493 if (VF.Width == 1) {
1494 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
1497 emitOptimizationRemarkAnalysis(
1498 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1499 "not beneficial to vectorize and user disabled interleaving");
1502 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1504 // Report the unrolling decision.
1505 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1506 Twine("unrolled with interleaving factor " +
1508 " (vectorization not beneficial)"));
1510 // We decided not to vectorize, but we may want to unroll.
1512 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1513 Unroller.vectorize(&LVL);
1515 // If we decided that it is *legal* to vectorize the loop then do it.
1516 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1520 // Add metadata to disable runtime unrolling scalar loop when there's no
1521 // runtime check about strides and memory. Because at this situation,
1522 // scalar loop is rarely used not worthy to be unrolled.
1523 if (!LB.IsSafetyChecksAdded())
1524 AddRuntimeUnrollDisableMetaData(L);
1526 // Report the vectorization decision.
1527 emitOptimizationRemark(
1528 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1529 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1530 ", unrolling interleave factor: " + Twine(UF) + ")");
1533 // Mark the loop as already vectorized to avoid vectorizing again.
1534 Hints.setAlreadyVectorized();
1536 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1540 void getAnalysisUsage(AnalysisUsage &AU) const override {
1541 AU.addRequired<AssumptionCacheTracker>();
1542 AU.addRequiredID(LoopSimplifyID);
1543 AU.addRequiredID(LCSSAID);
1544 AU.addRequired<BlockFrequencyInfo>();
1545 AU.addRequired<DominatorTreeWrapperPass>();
1546 AU.addRequired<LoopInfoWrapperPass>();
1547 AU.addRequired<ScalarEvolution>();
1548 AU.addRequired<TargetTransformInfoWrapperPass>();
1549 AU.addRequired<AliasAnalysis>();
1550 AU.addRequired<LoopAccessAnalysis>();
1551 AU.addPreserved<LoopInfoWrapperPass>();
1552 AU.addPreserved<DominatorTreeWrapperPass>();
1553 AU.addPreserved<AliasAnalysis>();
1558 } // end anonymous namespace
1560 //===----------------------------------------------------------------------===//
1561 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1562 // LoopVectorizationCostModel.
1563 //===----------------------------------------------------------------------===//
1565 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1566 // We need to place the broadcast of invariant variables outside the loop.
1567 Instruction *Instr = dyn_cast<Instruction>(V);
1569 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1570 Instr->getParent()) != LoopVectorBody.end());
1571 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1573 // Place the code for broadcasting invariant variables in the new preheader.
1574 IRBuilder<>::InsertPointGuard Guard(Builder);
1576 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1578 // Broadcast the scalar into all locations in the vector.
1579 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1584 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
1586 assert(Val->getType()->isVectorTy() && "Must be a vector");
1587 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1588 "Elem must be an integer");
1589 assert(Step->getType() == Val->getType()->getScalarType() &&
1590 "Step has wrong type");
1591 // Create the types.
1592 Type *ITy = Val->getType()->getScalarType();
1593 VectorType *Ty = cast<VectorType>(Val->getType());
1594 int VLen = Ty->getNumElements();
1595 SmallVector<Constant*, 8> Indices;
1597 // Create a vector of consecutive numbers from zero to VF.
1598 for (int i = 0; i < VLen; ++i)
1599 Indices.push_back(ConstantInt::get(ITy, StartIdx + i));
1601 // Add the consecutive indices to the vector value.
1602 Constant *Cv = ConstantVector::get(Indices);
1603 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1604 Step = Builder.CreateVectorSplat(VLen, Step);
1605 assert(Step->getType() == Val->getType() && "Invalid step vec");
1606 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
1607 // which can be found from the original scalar operations.
1608 Step = Builder.CreateMul(Cv, Step);
1609 return Builder.CreateAdd(Val, Step, "induction");
1612 /// \brief Find the operand of the GEP that should be checked for consecutive
1613 /// stores. This ignores trailing indices that have no effect on the final
1615 static unsigned getGEPInductionOperand(const DataLayout *DL,
1616 const GetElementPtrInst *Gep) {
1617 unsigned LastOperand = Gep->getNumOperands() - 1;
1618 unsigned GEPAllocSize = DL->getTypeAllocSize(
1619 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1621 // Walk backwards and try to peel off zeros.
1622 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1623 // Find the type we're currently indexing into.
1624 gep_type_iterator GEPTI = gep_type_begin(Gep);
1625 std::advance(GEPTI, LastOperand - 1);
1627 // If it's a type with the same allocation size as the result of the GEP we
1628 // can peel off the zero index.
1629 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1637 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1638 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1639 // Make sure that the pointer does not point to structs.
1640 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1643 // If this value is a pointer induction variable we know it is consecutive.
1644 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1645 if (Phi && Inductions.count(Phi)) {
1646 InductionInfo II = Inductions[Phi];
1647 return II.getConsecutiveDirection();
1650 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1654 unsigned NumOperands = Gep->getNumOperands();
1655 Value *GpPtr = Gep->getPointerOperand();
1656 // If this GEP value is a consecutive pointer induction variable and all of
1657 // the indices are constant then we know it is consecutive. We can
1658 Phi = dyn_cast<PHINode>(GpPtr);
1659 if (Phi && Inductions.count(Phi)) {
1661 // Make sure that the pointer does not point to structs.
1662 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1663 if (GepPtrType->getElementType()->isAggregateType())
1666 // Make sure that all of the index operands are loop invariant.
1667 for (unsigned i = 1; i < NumOperands; ++i)
1668 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1671 InductionInfo II = Inductions[Phi];
1672 return II.getConsecutiveDirection();
1675 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1677 // Check that all of the gep indices are uniform except for our induction
1679 for (unsigned i = 0; i != NumOperands; ++i)
1680 if (i != InductionOperand &&
1681 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1684 // We can emit wide load/stores only if the last non-zero index is the
1685 // induction variable.
1686 const SCEV *Last = nullptr;
1687 if (!Strides.count(Gep))
1688 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1690 // Because of the multiplication by a stride we can have a s/zext cast.
1691 // We are going to replace this stride by 1 so the cast is safe to ignore.
1693 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1694 // %0 = trunc i64 %indvars.iv to i32
1695 // %mul = mul i32 %0, %Stride1
1696 // %idxprom = zext i32 %mul to i64 << Safe cast.
1697 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1699 Last = replaceSymbolicStrideSCEV(SE, Strides,
1700 Gep->getOperand(InductionOperand), Gep);
1701 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1703 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1707 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1708 const SCEV *Step = AR->getStepRecurrence(*SE);
1710 // The memory is consecutive because the last index is consecutive
1711 // and all other indices are loop invariant.
1714 if (Step->isAllOnesValue())
1721 bool LoopVectorizationLegality::isUniform(Value *V) {
1722 return LAI->isUniform(V);
1725 InnerLoopVectorizer::VectorParts&
1726 InnerLoopVectorizer::getVectorValue(Value *V) {
1727 assert(V != Induction && "The new induction variable should not be used.");
1728 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1730 // If we have a stride that is replaced by one, do it here.
1731 if (Legal->hasStride(V))
1732 V = ConstantInt::get(V->getType(), 1);
1734 // If we have this scalar in the map, return it.
1735 if (WidenMap.has(V))
1736 return WidenMap.get(V);
1738 // If this scalar is unknown, assume that it is a constant or that it is
1739 // loop invariant. Broadcast V and save the value for future uses.
1740 Value *B = getBroadcastInstrs(V);
1741 return WidenMap.splat(V, B);
1744 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1745 assert(Vec->getType()->isVectorTy() && "Invalid type");
1746 SmallVector<Constant*, 8> ShuffleMask;
1747 for (unsigned i = 0; i < VF; ++i)
1748 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1750 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1751 ConstantVector::get(ShuffleMask),
1755 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1756 // Attempt to issue a wide load.
1757 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1758 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1760 assert((LI || SI) && "Invalid Load/Store instruction");
1762 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1763 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1764 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1765 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1766 // An alignment of 0 means target abi alignment. We need to use the scalar's
1767 // target abi alignment in such a case.
1769 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1770 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1771 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1772 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1774 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
1775 !Legal->isMaskRequired(SI))
1776 return scalarizeInstruction(Instr, true);
1778 if (ScalarAllocatedSize != VectorElementSize)
1779 return scalarizeInstruction(Instr);
1781 // If the pointer is loop invariant or if it is non-consecutive,
1782 // scalarize the load.
1783 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1784 bool Reverse = ConsecutiveStride < 0;
1785 bool UniformLoad = LI && Legal->isUniform(Ptr);
1786 if (!ConsecutiveStride || UniformLoad)
1787 return scalarizeInstruction(Instr);
1789 Constant *Zero = Builder.getInt32(0);
1790 VectorParts &Entry = WidenMap.get(Instr);
1792 // Handle consecutive loads/stores.
1793 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1794 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1795 setDebugLocFromInst(Builder, Gep);
1796 Value *PtrOperand = Gep->getPointerOperand();
1797 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1798 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1800 // Create the new GEP with the new induction variable.
1801 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1802 Gep2->setOperand(0, FirstBasePtr);
1803 Gep2->setName("gep.indvar.base");
1804 Ptr = Builder.Insert(Gep2);
1806 setDebugLocFromInst(Builder, Gep);
1807 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1808 OrigLoop) && "Base ptr must be invariant");
1810 // The last index does not have to be the induction. It can be
1811 // consecutive and be a function of the index. For example A[I+1];
1812 unsigned NumOperands = Gep->getNumOperands();
1813 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1814 // Create the new GEP with the new induction variable.
1815 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1817 for (unsigned i = 0; i < NumOperands; ++i) {
1818 Value *GepOperand = Gep->getOperand(i);
1819 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1821 // Update last index or loop invariant instruction anchored in loop.
1822 if (i == InductionOperand ||
1823 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1824 assert((i == InductionOperand ||
1825 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1826 "Must be last index or loop invariant");
1828 VectorParts &GEPParts = getVectorValue(GepOperand);
1829 Value *Index = GEPParts[0];
1830 Index = Builder.CreateExtractElement(Index, Zero);
1831 Gep2->setOperand(i, Index);
1832 Gep2->setName("gep.indvar.idx");
1835 Ptr = Builder.Insert(Gep2);
1837 // Use the induction element ptr.
1838 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1839 setDebugLocFromInst(Builder, Ptr);
1840 VectorParts &PtrVal = getVectorValue(Ptr);
1841 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1844 VectorParts Mask = createBlockInMask(Instr->getParent());
1847 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1848 "We do not allow storing to uniform addresses");
1849 setDebugLocFromInst(Builder, SI);
1850 // We don't want to update the value in the map as it might be used in
1851 // another expression. So don't use a reference type for "StoredVal".
1852 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1854 for (unsigned Part = 0; Part < UF; ++Part) {
1855 // Calculate the pointer for the specific unroll-part.
1856 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1859 // If we store to reverse consecutive memory locations then we need
1860 // to reverse the order of elements in the stored value.
1861 StoredVal[Part] = reverseVector(StoredVal[Part]);
1862 // If the address is consecutive but reversed, then the
1863 // wide store needs to start at the last vector element.
1864 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1865 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1866 Mask[Part] = reverseVector(Mask[Part]);
1869 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1870 DataTy->getPointerTo(AddressSpace));
1873 if (Legal->isMaskRequired(SI))
1874 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
1877 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
1878 propagateMetadata(NewSI, SI);
1884 assert(LI && "Must have a load instruction");
1885 setDebugLocFromInst(Builder, LI);
1886 for (unsigned Part = 0; Part < UF; ++Part) {
1887 // Calculate the pointer for the specific unroll-part.
1888 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1891 // If the address is consecutive but reversed, then the
1892 // wide load needs to start at the last vector element.
1893 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1894 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1895 Mask[Part] = reverseVector(Mask[Part]);
1899 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1900 DataTy->getPointerTo(AddressSpace));
1901 if (Legal->isMaskRequired(LI))
1902 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
1903 UndefValue::get(DataTy),
1904 "wide.masked.load");
1906 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
1907 propagateMetadata(NewLI, LI);
1908 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
1912 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1913 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1914 // Holds vector parameters or scalars, in case of uniform vals.
1915 SmallVector<VectorParts, 4> Params;
1917 setDebugLocFromInst(Builder, Instr);
1919 // Find all of the vectorized parameters.
1920 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1921 Value *SrcOp = Instr->getOperand(op);
1923 // If we are accessing the old induction variable, use the new one.
1924 if (SrcOp == OldInduction) {
1925 Params.push_back(getVectorValue(SrcOp));
1929 // Try using previously calculated values.
1930 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1932 // If the src is an instruction that appeared earlier in the basic block
1933 // then it should already be vectorized.
1934 if (SrcInst && OrigLoop->contains(SrcInst)) {
1935 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1936 // The parameter is a vector value from earlier.
1937 Params.push_back(WidenMap.get(SrcInst));
1939 // The parameter is a scalar from outside the loop. Maybe even a constant.
1940 VectorParts Scalars;
1941 Scalars.append(UF, SrcOp);
1942 Params.push_back(Scalars);
1946 assert(Params.size() == Instr->getNumOperands() &&
1947 "Invalid number of operands");
1949 // Does this instruction return a value ?
1950 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1952 Value *UndefVec = IsVoidRetTy ? nullptr :
1953 UndefValue::get(VectorType::get(Instr->getType(), VF));
1954 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1955 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1957 Instruction *InsertPt = Builder.GetInsertPoint();
1958 BasicBlock *IfBlock = Builder.GetInsertBlock();
1959 BasicBlock *CondBlock = nullptr;
1962 Loop *VectorLp = nullptr;
1963 if (IfPredicateStore) {
1964 assert(Instr->getParent()->getSinglePredecessor() &&
1965 "Only support single predecessor blocks");
1966 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1967 Instr->getParent());
1968 VectorLp = LI->getLoopFor(IfBlock);
1969 assert(VectorLp && "Must have a loop for this block");
1972 // For each vector unroll 'part':
1973 for (unsigned Part = 0; Part < UF; ++Part) {
1974 // For each scalar that we create:
1975 for (unsigned Width = 0; Width < VF; ++Width) {
1978 Value *Cmp = nullptr;
1979 if (IfPredicateStore) {
1980 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1981 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1982 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1983 LoopVectorBody.push_back(CondBlock);
1984 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
1985 // Update Builder with newly created basic block.
1986 Builder.SetInsertPoint(InsertPt);
1989 Instruction *Cloned = Instr->clone();
1991 Cloned->setName(Instr->getName() + ".cloned");
1992 // Replace the operands of the cloned instructions with extracted scalars.
1993 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1994 Value *Op = Params[op][Part];
1995 // Param is a vector. Need to extract the right lane.
1996 if (Op->getType()->isVectorTy())
1997 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1998 Cloned->setOperand(op, Op);
2001 // Place the cloned scalar in the new loop.
2002 Builder.Insert(Cloned);
2004 // If the original scalar returns a value we need to place it in a vector
2005 // so that future users will be able to use it.
2007 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
2008 Builder.getInt32(Width));
2010 if (IfPredicateStore) {
2011 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2012 LoopVectorBody.push_back(NewIfBlock);
2013 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
2014 Builder.SetInsertPoint(InsertPt);
2015 Instruction *OldBr = IfBlock->getTerminator();
2016 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2017 OldBr->eraseFromParent();
2018 IfBlock = NewIfBlock;
2024 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
2028 if (Instruction *I = dyn_cast<Instruction>(V))
2029 return I->getParent() == Loc->getParent() ? I : nullptr;
2033 std::pair<Instruction *, Instruction *>
2034 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
2035 Instruction *tnullptr = nullptr;
2036 if (!Legal->mustCheckStrides())
2037 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2039 IRBuilder<> ChkBuilder(Loc);
2042 Value *Check = nullptr;
2043 Instruction *FirstInst = nullptr;
2044 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
2045 SE = Legal->strides_end();
2047 Value *Ptr = stripIntegerCast(*SI);
2048 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
2050 // Store the first instruction we create.
2051 FirstInst = getFirstInst(FirstInst, C, Loc);
2053 Check = ChkBuilder.CreateOr(Check, C);
2058 // We have to do this trickery because the IRBuilder might fold the check to a
2059 // constant expression in which case there is no Instruction anchored in a
2061 LLVMContext &Ctx = Loc->getContext();
2062 Instruction *TheCheck =
2063 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2064 ChkBuilder.Insert(TheCheck, "stride.not.one");
2065 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2067 return std::make_pair(FirstInst, TheCheck);
2070 void InnerLoopVectorizer::createEmptyLoop() {
2072 In this function we generate a new loop. The new loop will contain
2073 the vectorized instructions while the old loop will continue to run the
2076 [ ] <-- Back-edge taken count overflow check.
2079 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2082 || [ ] <-- vector pre header.
2086 || [ ]_| <-- vector loop.
2089 | >[ ] <--- middle-block.
2092 -|- >[ ] <--- new preheader.
2096 | [ ]_| <-- old scalar loop to handle remainder.
2099 >[ ] <-- exit block.
2103 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2104 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2105 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2106 assert(BypassBlock && "Invalid loop structure");
2107 assert(ExitBlock && "Must have an exit block");
2109 // Some loops have a single integer induction variable, while other loops
2110 // don't. One example is c++ iterators that often have multiple pointer
2111 // induction variables. In the code below we also support a case where we
2112 // don't have a single induction variable.
2113 OldInduction = Legal->getInduction();
2114 Type *IdxTy = Legal->getWidestInductionType();
2116 // Find the loop boundaries.
2117 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2118 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2120 // The exit count might have the type of i64 while the phi is i32. This can
2121 // happen if we have an induction variable that is sign extended before the
2122 // compare. The only way that we get a backedge taken count is that the
2123 // induction variable was signed and as such will not overflow. In such a case
2124 // truncation is legal.
2125 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2126 IdxTy->getPrimitiveSizeInBits())
2127 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2129 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2130 // Get the total trip count from the count by adding 1.
2131 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2132 SE->getConstant(BackedgeTakeCount->getType(), 1));
2134 // Expand the trip count and place the new instructions in the preheader.
2135 // Notice that the pre-header does not change, only the loop body.
2136 SCEVExpander Exp(*SE, "induction");
2138 // We need to test whether the backedge-taken count is uint##_max. Adding one
2139 // to it will cause overflow and an incorrect loop trip count in the vector
2140 // body. In case of overflow we want to directly jump to the scalar remainder
2142 Value *BackedgeCount =
2143 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2144 BypassBlock->getTerminator());
2145 if (BackedgeCount->getType()->isPointerTy())
2146 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2147 "backedge.ptrcnt.to.int",
2148 BypassBlock->getTerminator());
2149 Instruction *CheckBCOverflow =
2150 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2151 Constant::getAllOnesValue(BackedgeCount->getType()),
2152 "backedge.overflow", BypassBlock->getTerminator());
2154 // The loop index does not have to start at Zero. Find the original start
2155 // value from the induction PHI node. If we don't have an induction variable
2156 // then we know that it starts at zero.
2157 Builder.SetInsertPoint(BypassBlock->getTerminator());
2158 Value *StartIdx = ExtendedIdx = OldInduction ?
2159 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2161 ConstantInt::get(IdxTy, 0);
2163 // We need an instruction to anchor the overflow check on. StartIdx needs to
2164 // be defined before the overflow check branch. Because the scalar preheader
2165 // is going to merge the start index and so the overflow branch block needs to
2166 // contain a definition of the start index.
2167 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2168 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2169 BypassBlock->getTerminator());
2171 // Count holds the overall loop count (N).
2172 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2173 BypassBlock->getTerminator());
2175 LoopBypassBlocks.push_back(BypassBlock);
2177 // Split the single block loop into the two loop structure described above.
2178 BasicBlock *VectorPH =
2179 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2180 BasicBlock *VecBody =
2181 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2182 BasicBlock *MiddleBlock =
2183 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2184 BasicBlock *ScalarPH =
2185 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2187 // Create and register the new vector loop.
2188 Loop* Lp = new Loop();
2189 Loop *ParentLoop = OrigLoop->getParentLoop();
2191 // Insert the new loop into the loop nest and register the new basic blocks
2192 // before calling any utilities such as SCEV that require valid LoopInfo.
2194 ParentLoop->addChildLoop(Lp);
2195 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
2196 ParentLoop->addBasicBlockToLoop(VectorPH, *LI);
2197 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
2199 LI->addTopLevelLoop(Lp);
2201 Lp->addBasicBlockToLoop(VecBody, *LI);
2203 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2205 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2207 // Generate the induction variable.
2208 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2209 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2210 // The loop step is equal to the vectorization factor (num of SIMD elements)
2211 // times the unroll factor (num of SIMD instructions).
2212 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2214 // This is the IR builder that we use to add all of the logic for bypassing
2215 // the new vector loop.
2216 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2217 setDebugLocFromInst(BypassBuilder,
2218 getDebugLocFromInstOrOperands(OldInduction));
2220 // We may need to extend the index in case there is a type mismatch.
2221 // We know that the count starts at zero and does not overflow.
2222 if (Count->getType() != IdxTy) {
2223 // The exit count can be of pointer type. Convert it to the correct
2225 if (ExitCount->getType()->isPointerTy())
2226 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2228 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2231 // Add the start index to the loop count to get the new end index.
2232 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2234 // Now we need to generate the expression for N - (N % VF), which is
2235 // the part that the vectorized body will execute.
2236 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2237 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2238 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2239 "end.idx.rnd.down");
2241 // Now, compare the new count to zero. If it is zero skip the vector loop and
2242 // jump to the scalar loop.
2244 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2246 BasicBlock *LastBypassBlock = BypassBlock;
2248 // Generate code to check that the loops trip count that we computed by adding
2249 // one to the backedge-taken count will not overflow.
2251 auto PastOverflowCheck =
2252 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2253 BasicBlock *CheckBlock =
2254 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2256 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2257 LoopBypassBlocks.push_back(CheckBlock);
2258 Instruction *OldTerm = LastBypassBlock->getTerminator();
2259 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2260 OldTerm->eraseFromParent();
2261 LastBypassBlock = CheckBlock;
2264 // Generate the code to check that the strides we assumed to be one are really
2265 // one. We want the new basic block to start at the first instruction in a
2266 // sequence of instructions that form a check.
2267 Instruction *StrideCheck;
2268 Instruction *FirstCheckInst;
2269 std::tie(FirstCheckInst, StrideCheck) =
2270 addStrideCheck(LastBypassBlock->getTerminator());
2272 AddedSafetyChecks = true;
2273 // Create a new block containing the stride check.
2274 BasicBlock *CheckBlock =
2275 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2277 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2278 LoopBypassBlocks.push_back(CheckBlock);
2280 // Replace the branch into the memory check block with a conditional branch
2281 // for the "few elements case".
2282 Instruction *OldTerm = LastBypassBlock->getTerminator();
2283 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2284 OldTerm->eraseFromParent();
2287 LastBypassBlock = CheckBlock;
2290 // Generate the code that checks in runtime if arrays overlap. We put the
2291 // checks into a separate block to make the more common case of few elements
2293 Instruction *MemRuntimeCheck;
2294 std::tie(FirstCheckInst, MemRuntimeCheck) =
2295 Legal->getLAI()->addRuntimeCheck(LastBypassBlock->getTerminator());
2296 if (MemRuntimeCheck) {
2297 AddedSafetyChecks = true;
2298 // Create a new block containing the memory check.
2299 BasicBlock *CheckBlock =
2300 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.memcheck");
2302 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2303 LoopBypassBlocks.push_back(CheckBlock);
2305 // Replace the branch into the memory check block with a conditional branch
2306 // for the "few elements case".
2307 Instruction *OldTerm = LastBypassBlock->getTerminator();
2308 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2309 OldTerm->eraseFromParent();
2311 Cmp = MemRuntimeCheck;
2312 LastBypassBlock = CheckBlock;
2315 LastBypassBlock->getTerminator()->eraseFromParent();
2316 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2319 // We are going to resume the execution of the scalar loop.
2320 // Go over all of the induction variables that we found and fix the
2321 // PHIs that are left in the scalar version of the loop.
2322 // The starting values of PHI nodes depend on the counter of the last
2323 // iteration in the vectorized loop.
2324 // If we come from a bypass edge then we need to start from the original
2327 // This variable saves the new starting index for the scalar loop.
2328 PHINode *ResumeIndex = nullptr;
2329 LoopVectorizationLegality::InductionList::iterator I, E;
2330 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2331 // Set builder to point to last bypass block.
2332 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2333 for (I = List->begin(), E = List->end(); I != E; ++I) {
2334 PHINode *OrigPhi = I->first;
2335 LoopVectorizationLegality::InductionInfo II = I->second;
2337 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2338 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2339 MiddleBlock->getTerminator());
2340 // We might have extended the type of the induction variable but we need a
2341 // truncated version for the scalar loop.
2342 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2343 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2344 MiddleBlock->getTerminator()) : nullptr;
2346 // Create phi nodes to merge from the backedge-taken check block.
2347 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2348 ScalarPH->getTerminator());
2349 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2351 PHINode *BCTruncResumeVal = nullptr;
2352 if (OrigPhi == OldInduction) {
2354 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2355 ScalarPH->getTerminator());
2356 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2359 Value *EndValue = nullptr;
2361 case LoopVectorizationLegality::IK_NoInduction:
2362 llvm_unreachable("Unknown induction");
2363 case LoopVectorizationLegality::IK_IntInduction: {
2364 // Handle the integer induction counter.
2365 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2367 // We have the canonical induction variable.
2368 if (OrigPhi == OldInduction) {
2369 // Create a truncated version of the resume value for the scalar loop,
2370 // we might have promoted the type to a larger width.
2372 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2373 // The new PHI merges the original incoming value, in case of a bypass,
2374 // or the value at the end of the vectorized loop.
2375 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2376 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2377 TruncResumeVal->addIncoming(EndValue, VecBody);
2379 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2381 // We know what the end value is.
2382 EndValue = IdxEndRoundDown;
2383 // We also know which PHI node holds it.
2384 ResumeIndex = ResumeVal;
2388 // Not the canonical induction variable - add the vector loop count to the
2390 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2391 II.StartValue->getType(),
2393 EndValue = II.transform(BypassBuilder, CRD);
2394 EndValue->setName("ind.end");
2397 case LoopVectorizationLegality::IK_PtrInduction: {
2398 EndValue = II.transform(BypassBuilder, CountRoundDown);
2399 EndValue->setName("ptr.ind.end");
2404 // The new PHI merges the original incoming value, in case of a bypass,
2405 // or the value at the end of the vectorized loop.
2406 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2407 if (OrigPhi == OldInduction)
2408 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2410 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2412 ResumeVal->addIncoming(EndValue, VecBody);
2414 // Fix the scalar body counter (PHI node).
2415 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2417 // The old induction's phi node in the scalar body needs the truncated
2419 if (OrigPhi == OldInduction) {
2420 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2421 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2423 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2424 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2428 // If we are generating a new induction variable then we also need to
2429 // generate the code that calculates the exit value. This value is not
2430 // simply the end of the counter because we may skip the vectorized body
2431 // in case of a runtime check.
2433 assert(!ResumeIndex && "Unexpected resume value found");
2434 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2435 MiddleBlock->getTerminator());
2436 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2437 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2438 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2441 // Make sure that we found the index where scalar loop needs to continue.
2442 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2443 "Invalid resume Index");
2445 // Add a check in the middle block to see if we have completed
2446 // all of the iterations in the first vector loop.
2447 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2448 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2449 ResumeIndex, "cmp.n",
2450 MiddleBlock->getTerminator());
2452 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2453 // Remove the old terminator.
2454 MiddleBlock->getTerminator()->eraseFromParent();
2456 // Create i+1 and fill the PHINode.
2457 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2458 Induction->addIncoming(StartIdx, VectorPH);
2459 Induction->addIncoming(NextIdx, VecBody);
2460 // Create the compare.
2461 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2462 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2464 // Now we have two terminators. Remove the old one from the block.
2465 VecBody->getTerminator()->eraseFromParent();
2467 // Get ready to start creating new instructions into the vectorized body.
2468 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2471 LoopVectorPreHeader = VectorPH;
2472 LoopScalarPreHeader = ScalarPH;
2473 LoopMiddleBlock = MiddleBlock;
2474 LoopExitBlock = ExitBlock;
2475 LoopVectorBody.push_back(VecBody);
2476 LoopScalarBody = OldBasicBlock;
2478 LoopVectorizeHints Hints(Lp, true);
2479 Hints.setAlreadyVectorized();
2482 /// This function returns the identity element (or neutral element) for
2483 /// the operation K.
2485 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2490 // Adding, Xoring, Oring zero to a number does not change it.
2491 return ConstantInt::get(Tp, 0);
2492 case RK_IntegerMult:
2493 // Multiplying a number by 1 does not change it.
2494 return ConstantInt::get(Tp, 1);
2496 // AND-ing a number with an all-1 value does not change it.
2497 return ConstantInt::get(Tp, -1, true);
2499 // Multiplying a number by 1 does not change it.
2500 return ConstantFP::get(Tp, 1.0L);
2502 // Adding zero to a number does not change it.
2503 return ConstantFP::get(Tp, 0.0L);
2505 llvm_unreachable("Unknown reduction kind");
2509 /// This function translates the reduction kind to an LLVM binary operator.
2511 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2513 case LoopVectorizationLegality::RK_IntegerAdd:
2514 return Instruction::Add;
2515 case LoopVectorizationLegality::RK_IntegerMult:
2516 return Instruction::Mul;
2517 case LoopVectorizationLegality::RK_IntegerOr:
2518 return Instruction::Or;
2519 case LoopVectorizationLegality::RK_IntegerAnd:
2520 return Instruction::And;
2521 case LoopVectorizationLegality::RK_IntegerXor:
2522 return Instruction::Xor;
2523 case LoopVectorizationLegality::RK_FloatMult:
2524 return Instruction::FMul;
2525 case LoopVectorizationLegality::RK_FloatAdd:
2526 return Instruction::FAdd;
2527 case LoopVectorizationLegality::RK_IntegerMinMax:
2528 return Instruction::ICmp;
2529 case LoopVectorizationLegality::RK_FloatMinMax:
2530 return Instruction::FCmp;
2532 llvm_unreachable("Unknown reduction operation");
2536 static Value *createMinMaxOp(IRBuilder<> &Builder,
2537 LoopVectorizationLegality::MinMaxReductionKind RK,
2538 Value *Left, Value *Right) {
2539 CmpInst::Predicate P = CmpInst::ICMP_NE;
2542 llvm_unreachable("Unknown min/max reduction kind");
2543 case LoopVectorizationLegality::MRK_UIntMin:
2544 P = CmpInst::ICMP_ULT;
2546 case LoopVectorizationLegality::MRK_UIntMax:
2547 P = CmpInst::ICMP_UGT;
2549 case LoopVectorizationLegality::MRK_SIntMin:
2550 P = CmpInst::ICMP_SLT;
2552 case LoopVectorizationLegality::MRK_SIntMax:
2553 P = CmpInst::ICMP_SGT;
2555 case LoopVectorizationLegality::MRK_FloatMin:
2556 P = CmpInst::FCMP_OLT;
2558 case LoopVectorizationLegality::MRK_FloatMax:
2559 P = CmpInst::FCMP_OGT;
2564 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2565 RK == LoopVectorizationLegality::MRK_FloatMax)
2566 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2568 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2570 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2575 struct CSEDenseMapInfo {
2576 static bool canHandle(Instruction *I) {
2577 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2578 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2580 static inline Instruction *getEmptyKey() {
2581 return DenseMapInfo<Instruction *>::getEmptyKey();
2583 static inline Instruction *getTombstoneKey() {
2584 return DenseMapInfo<Instruction *>::getTombstoneKey();
2586 static unsigned getHashValue(Instruction *I) {
2587 assert(canHandle(I) && "Unknown instruction!");
2588 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2589 I->value_op_end()));
2591 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2592 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2593 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2595 return LHS->isIdenticalTo(RHS);
2600 /// \brief Check whether this block is a predicated block.
2601 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2602 /// = ...; " blocks. We start with one vectorized basic block. For every
2603 /// conditional block we split this vectorized block. Therefore, every second
2604 /// block will be a predicated one.
2605 static bool isPredicatedBlock(unsigned BlockNum) {
2606 return BlockNum % 2;
2609 ///\brief Perform cse of induction variable instructions.
2610 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2611 // Perform simple cse.
2612 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2613 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2614 BasicBlock *BB = BBs[i];
2615 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2616 Instruction *In = I++;
2618 if (!CSEDenseMapInfo::canHandle(In))
2621 // Check if we can replace this instruction with any of the
2622 // visited instructions.
2623 if (Instruction *V = CSEMap.lookup(In)) {
2624 In->replaceAllUsesWith(V);
2625 In->eraseFromParent();
2628 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2629 // ...;" blocks for predicated stores. Every second block is a predicated
2631 if (isPredicatedBlock(i))
2639 /// \brief Adds a 'fast' flag to floating point operations.
2640 static Value *addFastMathFlag(Value *V) {
2641 if (isa<FPMathOperator>(V)){
2642 FastMathFlags Flags;
2643 Flags.setUnsafeAlgebra();
2644 cast<Instruction>(V)->setFastMathFlags(Flags);
2649 void InnerLoopVectorizer::vectorizeLoop() {
2650 //===------------------------------------------------===//
2652 // Notice: any optimization or new instruction that go
2653 // into the code below should be also be implemented in
2656 //===------------------------------------------------===//
2657 Constant *Zero = Builder.getInt32(0);
2659 // In order to support reduction variables we need to be able to vectorize
2660 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2661 // stages. First, we create a new vector PHI node with no incoming edges.
2662 // We use this value when we vectorize all of the instructions that use the
2663 // PHI. Next, after all of the instructions in the block are complete we
2664 // add the new incoming edges to the PHI. At this point all of the
2665 // instructions in the basic block are vectorized, so we can use them to
2666 // construct the PHI.
2667 PhiVector RdxPHIsToFix;
2669 // Scan the loop in a topological order to ensure that defs are vectorized
2671 LoopBlocksDFS DFS(OrigLoop);
2674 // Vectorize all of the blocks in the original loop.
2675 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2676 be = DFS.endRPO(); bb != be; ++bb)
2677 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2679 // At this point every instruction in the original loop is widened to
2680 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2681 // that we vectorized. The PHI nodes are currently empty because we did
2682 // not want to introduce cycles. Notice that the remaining PHI nodes
2683 // that we need to fix are reduction variables.
2685 // Create the 'reduced' values for each of the induction vars.
2686 // The reduced values are the vector values that we scalarize and combine
2687 // after the loop is finished.
2688 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2690 PHINode *RdxPhi = *it;
2691 assert(RdxPhi && "Unable to recover vectorized PHI");
2693 // Find the reduction variable descriptor.
2694 assert(Legal->getReductionVars()->count(RdxPhi) &&
2695 "Unable to find the reduction variable");
2696 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2697 (*Legal->getReductionVars())[RdxPhi];
2699 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2701 // We need to generate a reduction vector from the incoming scalar.
2702 // To do so, we need to generate the 'identity' vector and override
2703 // one of the elements with the incoming scalar reduction. We need
2704 // to do it in the vector-loop preheader.
2705 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
2707 // This is the vector-clone of the value that leaves the loop.
2708 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2709 Type *VecTy = VectorExit[0]->getType();
2711 // Find the reduction identity variable. Zero for addition, or, xor,
2712 // one for multiplication, -1 for And.
2715 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2716 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2717 // MinMax reduction have the start value as their identify.
2719 VectorStart = Identity = RdxDesc.StartValue;
2721 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2726 // Handle other reduction kinds:
2728 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2729 VecTy->getScalarType());
2732 // This vector is the Identity vector where the first element is the
2733 // incoming scalar reduction.
2734 VectorStart = RdxDesc.StartValue;
2736 Identity = ConstantVector::getSplat(VF, Iden);
2738 // This vector is the Identity vector where the first element is the
2739 // incoming scalar reduction.
2740 VectorStart = Builder.CreateInsertElement(Identity,
2741 RdxDesc.StartValue, Zero);
2745 // Fix the vector-loop phi.
2747 // Reductions do not have to start at zero. They can start with
2748 // any loop invariant values.
2749 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2750 BasicBlock *Latch = OrigLoop->getLoopLatch();
2751 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2752 VectorParts &Val = getVectorValue(LoopVal);
2753 for (unsigned part = 0; part < UF; ++part) {
2754 // Make sure to add the reduction stat value only to the
2755 // first unroll part.
2756 Value *StartVal = (part == 0) ? VectorStart : Identity;
2757 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
2758 LoopVectorPreHeader);
2759 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2760 LoopVectorBody.back());
2763 // Before each round, move the insertion point right between
2764 // the PHIs and the values we are going to write.
2765 // This allows us to write both PHINodes and the extractelement
2767 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2769 VectorParts RdxParts;
2770 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2771 for (unsigned part = 0; part < UF; ++part) {
2772 // This PHINode contains the vectorized reduction variable, or
2773 // the initial value vector, if we bypass the vector loop.
2774 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2775 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2776 Value *StartVal = (part == 0) ? VectorStart : Identity;
2777 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2778 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2779 NewPhi->addIncoming(RdxExitVal[part],
2780 LoopVectorBody.back());
2781 RdxParts.push_back(NewPhi);
2784 // Reduce all of the unrolled parts into a single vector.
2785 Value *ReducedPartRdx = RdxParts[0];
2786 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2787 setDebugLocFromInst(Builder, ReducedPartRdx);
2788 for (unsigned part = 1; part < UF; ++part) {
2789 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2790 // Floating point operations had to be 'fast' to enable the reduction.
2791 ReducedPartRdx = addFastMathFlag(
2792 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2793 ReducedPartRdx, "bin.rdx"));
2795 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2796 ReducedPartRdx, RdxParts[part]);
2800 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2801 // and vector ops, reducing the set of values being computed by half each
2803 assert(isPowerOf2_32(VF) &&
2804 "Reduction emission only supported for pow2 vectors!");
2805 Value *TmpVec = ReducedPartRdx;
2806 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2807 for (unsigned i = VF; i != 1; i >>= 1) {
2808 // Move the upper half of the vector to the lower half.
2809 for (unsigned j = 0; j != i/2; ++j)
2810 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2812 // Fill the rest of the mask with undef.
2813 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2814 UndefValue::get(Builder.getInt32Ty()));
2817 Builder.CreateShuffleVector(TmpVec,
2818 UndefValue::get(TmpVec->getType()),
2819 ConstantVector::get(ShuffleMask),
2822 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2823 // Floating point operations had to be 'fast' to enable the reduction.
2824 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2825 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2827 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2830 // The result is in the first element of the vector.
2831 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2832 Builder.getInt32(0));
2835 // Create a phi node that merges control-flow from the backedge-taken check
2836 // block and the middle block.
2837 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
2838 LoopScalarPreHeader->getTerminator());
2839 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
2840 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2842 // Now, we need to fix the users of the reduction variable
2843 // inside and outside of the scalar remainder loop.
2844 // We know that the loop is in LCSSA form. We need to update the
2845 // PHI nodes in the exit blocks.
2846 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2847 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2848 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2849 if (!LCSSAPhi) break;
2851 // All PHINodes need to have a single entry edge, or two if
2852 // we already fixed them.
2853 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2855 // We found our reduction value exit-PHI. Update it with the
2856 // incoming bypass edge.
2857 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2858 // Add an edge coming from the bypass.
2859 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2862 }// end of the LCSSA phi scan.
2864 // Fix the scalar loop reduction variable with the incoming reduction sum
2865 // from the vector body and from the backedge value.
2866 int IncomingEdgeBlockIdx =
2867 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2868 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2869 // Pick the other block.
2870 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2871 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
2872 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2873 }// end of for each redux variable.
2877 // Remove redundant induction instructions.
2878 cse(LoopVectorBody);
2881 void InnerLoopVectorizer::fixLCSSAPHIs() {
2882 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2883 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2884 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2885 if (!LCSSAPhi) break;
2886 if (LCSSAPhi->getNumIncomingValues() == 1)
2887 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2892 InnerLoopVectorizer::VectorParts
2893 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2894 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2897 // Look for cached value.
2898 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2899 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2900 if (ECEntryIt != MaskCache.end())
2901 return ECEntryIt->second;
2903 VectorParts SrcMask = createBlockInMask(Src);
2905 // The terminator has to be a branch inst!
2906 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2907 assert(BI && "Unexpected terminator found");
2909 if (BI->isConditional()) {
2910 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2912 if (BI->getSuccessor(0) != Dst)
2913 for (unsigned part = 0; part < UF; ++part)
2914 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2916 for (unsigned part = 0; part < UF; ++part)
2917 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2919 MaskCache[Edge] = EdgeMask;
2923 MaskCache[Edge] = SrcMask;
2927 InnerLoopVectorizer::VectorParts
2928 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2929 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2931 // Loop incoming mask is all-one.
2932 if (OrigLoop->getHeader() == BB) {
2933 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2934 return getVectorValue(C);
2937 // This is the block mask. We OR all incoming edges, and with zero.
2938 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2939 VectorParts BlockMask = getVectorValue(Zero);
2942 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2943 VectorParts EM = createEdgeMask(*it, BB);
2944 for (unsigned part = 0; part < UF; ++part)
2945 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2951 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2952 InnerLoopVectorizer::VectorParts &Entry,
2953 unsigned UF, unsigned VF, PhiVector *PV) {
2954 PHINode* P = cast<PHINode>(PN);
2955 // Handle reduction variables:
2956 if (Legal->getReductionVars()->count(P)) {
2957 for (unsigned part = 0; part < UF; ++part) {
2958 // This is phase one of vectorizing PHIs.
2959 Type *VecTy = (VF == 1) ? PN->getType() :
2960 VectorType::get(PN->getType(), VF);
2961 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2962 LoopVectorBody.back()-> getFirstInsertionPt());
2968 setDebugLocFromInst(Builder, P);
2969 // Check for PHI nodes that are lowered to vector selects.
2970 if (P->getParent() != OrigLoop->getHeader()) {
2971 // We know that all PHIs in non-header blocks are converted into
2972 // selects, so we don't have to worry about the insertion order and we
2973 // can just use the builder.
2974 // At this point we generate the predication tree. There may be
2975 // duplications since this is a simple recursive scan, but future
2976 // optimizations will clean it up.
2978 unsigned NumIncoming = P->getNumIncomingValues();
2980 // Generate a sequence of selects of the form:
2981 // SELECT(Mask3, In3,
2982 // SELECT(Mask2, In2,
2984 for (unsigned In = 0; In < NumIncoming; In++) {
2985 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2987 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2989 for (unsigned part = 0; part < UF; ++part) {
2990 // We might have single edge PHIs (blocks) - use an identity
2991 // 'select' for the first PHI operand.
2993 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2996 // Select between the current value and the previous incoming edge
2997 // based on the incoming mask.
2998 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2999 Entry[part], "predphi");
3005 // This PHINode must be an induction variable.
3006 // Make sure that we know about it.
3007 assert(Legal->getInductionVars()->count(P) &&
3008 "Not an induction variable");
3010 LoopVectorizationLegality::InductionInfo II =
3011 Legal->getInductionVars()->lookup(P);
3013 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
3014 // which can be found from the original scalar operations.
3016 case LoopVectorizationLegality::IK_NoInduction:
3017 llvm_unreachable("Unknown induction");
3018 case LoopVectorizationLegality::IK_IntInduction: {
3019 assert(P->getType() == II.StartValue->getType() && "Types must match");
3020 Type *PhiTy = P->getType();
3022 if (P == OldInduction) {
3023 // Handle the canonical induction variable. We might have had to
3025 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
3027 // Handle other induction variables that are now based on the
3029 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
3031 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
3032 Broadcasted = II.transform(Builder, NormalizedIdx);
3033 Broadcasted->setName("offset.idx");
3035 Broadcasted = getBroadcastInstrs(Broadcasted);
3036 // After broadcasting the induction variable we need to make the vector
3037 // consecutive by adding 0, 1, 2, etc.
3038 for (unsigned part = 0; part < UF; ++part)
3039 Entry[part] = getStepVector(Broadcasted, VF * part, II.StepValue);
3042 case LoopVectorizationLegality::IK_PtrInduction:
3043 // Handle the pointer induction variable case.
3044 assert(P->getType()->isPointerTy() && "Unexpected type.");
3045 // This is the normalized GEP that starts counting at zero.
3046 Value *NormalizedIdx =
3047 Builder.CreateSub(Induction, ExtendedIdx, "normalized.idx");
3048 // This is the vector of results. Notice that we don't generate
3049 // vector geps because scalar geps result in better code.
3050 for (unsigned part = 0; part < UF; ++part) {
3052 int EltIndex = part;
3053 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3054 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3055 Value *SclrGep = II.transform(Builder, GlobalIdx);
3056 SclrGep->setName("next.gep");
3057 Entry[part] = SclrGep;
3061 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3062 for (unsigned int i = 0; i < VF; ++i) {
3063 int EltIndex = i + part * VF;
3064 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3065 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3066 Value *SclrGep = II.transform(Builder, GlobalIdx);
3067 SclrGep->setName("next.gep");
3068 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3069 Builder.getInt32(i),
3072 Entry[part] = VecVal;
3078 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3079 // For each instruction in the old loop.
3080 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3081 VectorParts &Entry = WidenMap.get(it);
3082 switch (it->getOpcode()) {
3083 case Instruction::Br:
3084 // Nothing to do for PHIs and BR, since we already took care of the
3085 // loop control flow instructions.
3087 case Instruction::PHI: {
3088 // Vectorize PHINodes.
3089 widenPHIInstruction(it, Entry, UF, VF, PV);
3093 case Instruction::Add:
3094 case Instruction::FAdd:
3095 case Instruction::Sub:
3096 case Instruction::FSub:
3097 case Instruction::Mul:
3098 case Instruction::FMul:
3099 case Instruction::UDiv:
3100 case Instruction::SDiv:
3101 case Instruction::FDiv:
3102 case Instruction::URem:
3103 case Instruction::SRem:
3104 case Instruction::FRem:
3105 case Instruction::Shl:
3106 case Instruction::LShr:
3107 case Instruction::AShr:
3108 case Instruction::And:
3109 case Instruction::Or:
3110 case Instruction::Xor: {
3111 // Just widen binops.
3112 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3113 setDebugLocFromInst(Builder, BinOp);
3114 VectorParts &A = getVectorValue(it->getOperand(0));
3115 VectorParts &B = getVectorValue(it->getOperand(1));
3117 // Use this vector value for all users of the original instruction.
3118 for (unsigned Part = 0; Part < UF; ++Part) {
3119 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3121 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3122 VecOp->copyIRFlags(BinOp);
3127 propagateMetadata(Entry, it);
3130 case Instruction::Select: {
3132 // If the selector is loop invariant we can create a select
3133 // instruction with a scalar condition. Otherwise, use vector-select.
3134 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3136 setDebugLocFromInst(Builder, it);
3138 // The condition can be loop invariant but still defined inside the
3139 // loop. This means that we can't just use the original 'cond' value.
3140 // We have to take the 'vectorized' value and pick the first lane.
3141 // Instcombine will make this a no-op.
3142 VectorParts &Cond = getVectorValue(it->getOperand(0));
3143 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3144 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3146 Value *ScalarCond = (VF == 1) ? Cond[0] :
3147 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3149 for (unsigned Part = 0; Part < UF; ++Part) {
3150 Entry[Part] = Builder.CreateSelect(
3151 InvariantCond ? ScalarCond : Cond[Part],
3156 propagateMetadata(Entry, it);
3160 case Instruction::ICmp:
3161 case Instruction::FCmp: {
3162 // Widen compares. Generate vector compares.
3163 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3164 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3165 setDebugLocFromInst(Builder, it);
3166 VectorParts &A = getVectorValue(it->getOperand(0));
3167 VectorParts &B = getVectorValue(it->getOperand(1));
3168 for (unsigned Part = 0; Part < UF; ++Part) {
3171 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3173 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3177 propagateMetadata(Entry, it);
3181 case Instruction::Store:
3182 case Instruction::Load:
3183 vectorizeMemoryInstruction(it);
3185 case Instruction::ZExt:
3186 case Instruction::SExt:
3187 case Instruction::FPToUI:
3188 case Instruction::FPToSI:
3189 case Instruction::FPExt:
3190 case Instruction::PtrToInt:
3191 case Instruction::IntToPtr:
3192 case Instruction::SIToFP:
3193 case Instruction::UIToFP:
3194 case Instruction::Trunc:
3195 case Instruction::FPTrunc:
3196 case Instruction::BitCast: {
3197 CastInst *CI = dyn_cast<CastInst>(it);
3198 setDebugLocFromInst(Builder, it);
3199 /// Optimize the special case where the source is the induction
3200 /// variable. Notice that we can only optimize the 'trunc' case
3201 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3202 /// c. other casts depend on pointer size.
3203 if (CI->getOperand(0) == OldInduction &&
3204 it->getOpcode() == Instruction::Trunc) {
3205 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3207 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3208 LoopVectorizationLegality::InductionInfo II =
3209 Legal->getInductionVars()->lookup(OldInduction);
3211 ConstantInt::getSigned(CI->getType(), II.StepValue->getSExtValue());
3212 for (unsigned Part = 0; Part < UF; ++Part)
3213 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
3214 propagateMetadata(Entry, it);
3217 /// Vectorize casts.
3218 Type *DestTy = (VF == 1) ? CI->getType() :
3219 VectorType::get(CI->getType(), VF);
3221 VectorParts &A = getVectorValue(it->getOperand(0));
3222 for (unsigned Part = 0; Part < UF; ++Part)
3223 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3224 propagateMetadata(Entry, it);
3228 case Instruction::Call: {
3229 // Ignore dbg intrinsics.
3230 if (isa<DbgInfoIntrinsic>(it))
3232 setDebugLocFromInst(Builder, it);
3234 Module *M = BB->getParent()->getParent();
3235 CallInst *CI = cast<CallInst>(it);
3236 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3237 assert(ID && "Not an intrinsic call!");
3239 case Intrinsic::assume:
3240 case Intrinsic::lifetime_end:
3241 case Intrinsic::lifetime_start:
3242 scalarizeInstruction(it);
3245 bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
3246 for (unsigned Part = 0; Part < UF; ++Part) {
3247 SmallVector<Value *, 4> Args;
3248 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3249 if (HasScalarOpd && i == 1) {
3250 Args.push_back(CI->getArgOperand(i));
3253 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3254 Args.push_back(Arg[Part]);
3256 Type *Tys[] = {CI->getType()};
3258 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3260 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3261 Entry[Part] = Builder.CreateCall(F, Args);
3264 propagateMetadata(Entry, it);
3271 // All other instructions are unsupported. Scalarize them.
3272 scalarizeInstruction(it);
3275 }// end of for_each instr.
3278 void InnerLoopVectorizer::updateAnalysis() {
3279 // Forget the original basic block.
3280 SE->forgetLoop(OrigLoop);
3282 // Update the dominator tree information.
3283 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3284 "Entry does not dominate exit.");
3286 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3287 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3288 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3290 // Due to if predication of stores we might create a sequence of "if(pred)
3291 // a[i] = ...; " blocks.
3292 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3294 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3295 else if (isPredicatedBlock(i)) {
3296 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3298 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3302 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3303 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3304 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3305 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3307 DEBUG(DT->verifyDomTree());
3310 /// \brief Check whether it is safe to if-convert this phi node.
3312 /// Phi nodes with constant expressions that can trap are not safe to if
3314 static bool canIfConvertPHINodes(BasicBlock *BB) {
3315 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3316 PHINode *Phi = dyn_cast<PHINode>(I);
3319 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3320 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3327 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3328 if (!EnableIfConversion) {
3329 emitAnalysis(VectorizationReport() << "if-conversion is disabled");
3333 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3335 // A list of pointers that we can safely read and write to.
3336 SmallPtrSet<Value *, 8> SafePointes;
3338 // Collect safe addresses.
3339 for (Loop::block_iterator BI = TheLoop->block_begin(),
3340 BE = TheLoop->block_end(); BI != BE; ++BI) {
3341 BasicBlock *BB = *BI;
3343 if (blockNeedsPredication(BB))
3346 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3347 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3348 SafePointes.insert(LI->getPointerOperand());
3349 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3350 SafePointes.insert(SI->getPointerOperand());
3354 // Collect the blocks that need predication.
3355 BasicBlock *Header = TheLoop->getHeader();
3356 for (Loop::block_iterator BI = TheLoop->block_begin(),
3357 BE = TheLoop->block_end(); BI != BE; ++BI) {
3358 BasicBlock *BB = *BI;
3360 // We don't support switch statements inside loops.
3361 if (!isa<BranchInst>(BB->getTerminator())) {
3362 emitAnalysis(VectorizationReport(BB->getTerminator())
3363 << "loop contains a switch statement");
3367 // We must be able to predicate all blocks that need to be predicated.
3368 if (blockNeedsPredication(BB)) {
3369 if (!blockCanBePredicated(BB, SafePointes)) {
3370 emitAnalysis(VectorizationReport(BB->getTerminator())
3371 << "control flow cannot be substituted for a select");
3374 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3375 emitAnalysis(VectorizationReport(BB->getTerminator())
3376 << "control flow cannot be substituted for a select");
3381 // We can if-convert this loop.
3385 bool LoopVectorizationLegality::canVectorize() {
3386 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3387 // be canonicalized.
3388 if (!TheLoop->getLoopPreheader()) {
3390 VectorizationReport() <<
3391 "loop control flow is not understood by vectorizer");
3395 // We can only vectorize innermost loops.
3396 if (!TheLoop->getSubLoopsVector().empty()) {
3397 emitAnalysis(VectorizationReport() << "loop is not the innermost loop");
3401 // We must have a single backedge.
3402 if (TheLoop->getNumBackEdges() != 1) {
3404 VectorizationReport() <<
3405 "loop control flow is not understood by vectorizer");
3409 // We must have a single exiting block.
3410 if (!TheLoop->getExitingBlock()) {
3412 VectorizationReport() <<
3413 "loop control flow is not understood by vectorizer");
3417 // We only handle bottom-tested loops, i.e. loop in which the condition is
3418 // checked at the end of each iteration. With that we can assume that all
3419 // instructions in the loop are executed the same number of times.
3420 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
3422 VectorizationReport() <<
3423 "loop control flow is not understood by vectorizer");
3427 // We need to have a loop header.
3428 DEBUG(dbgs() << "LV: Found a loop: " <<
3429 TheLoop->getHeader()->getName() << '\n');
3431 // Check if we can if-convert non-single-bb loops.
3432 unsigned NumBlocks = TheLoop->getNumBlocks();
3433 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3434 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3438 // ScalarEvolution needs to be able to find the exit count.
3439 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3440 if (ExitCount == SE->getCouldNotCompute()) {
3441 emitAnalysis(VectorizationReport() <<
3442 "could not determine number of loop iterations");
3443 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3447 // Check if we can vectorize the instructions and CFG in this loop.
3448 if (!canVectorizeInstrs()) {
3449 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3453 // Go over each instruction and look at memory deps.
3454 if (!canVectorizeMemory()) {
3455 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3459 // Collect all of the variables that remain uniform after vectorization.
3460 collectLoopUniforms();
3462 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3463 (LAI->getRuntimePointerCheck()->Need ? " (with a runtime bound check)" :
3467 // Okay! We can vectorize. At this point we don't have any other mem analysis
3468 // which may limit our maximum vectorization factor, so just return true with
3473 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3474 if (Ty->isPointerTy())
3475 return DL.getIntPtrType(Ty);
3477 // It is possible that char's or short's overflow when we ask for the loop's
3478 // trip count, work around this by changing the type size.
3479 if (Ty->getScalarSizeInBits() < 32)
3480 return Type::getInt32Ty(Ty->getContext());
3485 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3486 Ty0 = convertPointerToIntegerType(DL, Ty0);
3487 Ty1 = convertPointerToIntegerType(DL, Ty1);
3488 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3493 /// \brief Check that the instruction has outside loop users and is not an
3494 /// identified reduction variable.
3495 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3496 SmallPtrSetImpl<Value *> &Reductions) {
3497 // Reduction instructions are allowed to have exit users. All other
3498 // instructions must not have external users.
3499 if (!Reductions.count(Inst))
3500 //Check that all of the users of the loop are inside the BB.
3501 for (User *U : Inst->users()) {
3502 Instruction *UI = cast<Instruction>(U);
3503 // This user may be a reduction exit value.
3504 if (!TheLoop->contains(UI)) {
3505 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3512 bool LoopVectorizationLegality::canVectorizeInstrs() {
3513 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3514 BasicBlock *Header = TheLoop->getHeader();
3516 // Look for the attribute signaling the absence of NaNs.
3517 Function &F = *Header->getParent();
3518 if (F.hasFnAttribute("no-nans-fp-math"))
3520 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
3522 // For each block in the loop.
3523 for (Loop::block_iterator bb = TheLoop->block_begin(),
3524 be = TheLoop->block_end(); bb != be; ++bb) {
3526 // Scan the instructions in the block and look for hazards.
3527 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3530 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3531 Type *PhiTy = Phi->getType();
3532 // Check that this PHI type is allowed.
3533 if (!PhiTy->isIntegerTy() &&
3534 !PhiTy->isFloatingPointTy() &&
3535 !PhiTy->isPointerTy()) {
3536 emitAnalysis(VectorizationReport(it)
3537 << "loop control flow is not understood by vectorizer");
3538 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3542 // If this PHINode is not in the header block, then we know that we
3543 // can convert it to select during if-conversion. No need to check if
3544 // the PHIs in this block are induction or reduction variables.
3545 if (*bb != Header) {
3546 // Check that this instruction has no outside users or is an
3547 // identified reduction value with an outside user.
3548 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3550 emitAnalysis(VectorizationReport(it) <<
3551 "value could not be identified as "
3552 "an induction or reduction variable");
3556 // We only allow if-converted PHIs with exactly two incoming values.
3557 if (Phi->getNumIncomingValues() != 2) {
3558 emitAnalysis(VectorizationReport(it)
3559 << "control flow not understood by vectorizer");
3560 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3564 // This is the value coming from the preheader.
3565 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3566 ConstantInt *StepValue = nullptr;
3567 // Check if this is an induction variable.
3568 InductionKind IK = isInductionVariable(Phi, StepValue);
3570 if (IK_NoInduction != IK) {
3571 // Get the widest type.
3573 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3575 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3577 // Int inductions are special because we only allow one IV.
3578 if (IK == IK_IntInduction && StepValue->isOne()) {
3579 // Use the phi node with the widest type as induction. Use the last
3580 // one if there are multiple (no good reason for doing this other
3581 // than it is expedient).
3582 if (!Induction || PhiTy == WidestIndTy)
3586 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3587 Inductions[Phi] = InductionInfo(StartValue, IK, StepValue);
3589 // Until we explicitly handle the case of an induction variable with
3590 // an outside loop user we have to give up vectorizing this loop.
3591 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3592 emitAnalysis(VectorizationReport(it) <<
3593 "use of induction value outside of the "
3594 "loop is not handled by vectorizer");
3601 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3602 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3605 if (AddReductionVar(Phi, RK_IntegerMult)) {
3606 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3609 if (AddReductionVar(Phi, RK_IntegerOr)) {
3610 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3613 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3614 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3617 if (AddReductionVar(Phi, RK_IntegerXor)) {
3618 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3621 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3622 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3625 if (AddReductionVar(Phi, RK_FloatMult)) {
3626 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3629 if (AddReductionVar(Phi, RK_FloatAdd)) {
3630 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3633 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3634 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3639 emitAnalysis(VectorizationReport(it) <<
3640 "value that could not be identified as "
3641 "reduction is used outside the loop");
3642 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3644 }// end of PHI handling
3646 // We still don't handle functions. However, we can ignore dbg intrinsic
3647 // calls and we do handle certain intrinsic and libm functions.
3648 CallInst *CI = dyn_cast<CallInst>(it);
3649 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3650 emitAnalysis(VectorizationReport(it) <<
3651 "call instruction cannot be vectorized");
3652 DEBUG(dbgs() << "LV: Found a call site.\n");
3656 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
3657 // second argument is the same (i.e. loop invariant)
3659 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
3660 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
3661 emitAnalysis(VectorizationReport(it)
3662 << "intrinsic instruction cannot be vectorized");
3663 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
3668 // Check that the instruction return type is vectorizable.
3669 // Also, we can't vectorize extractelement instructions.
3670 if ((!VectorType::isValidElementType(it->getType()) &&
3671 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3672 emitAnalysis(VectorizationReport(it)
3673 << "instruction return type cannot be vectorized");
3674 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3678 // Check that the stored type is vectorizable.
3679 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3680 Type *T = ST->getValueOperand()->getType();
3681 if (!VectorType::isValidElementType(T)) {
3682 emitAnalysis(VectorizationReport(ST) <<
3683 "store instruction cannot be vectorized");
3686 if (EnableMemAccessVersioning)
3687 collectStridedAccess(ST);
3690 if (EnableMemAccessVersioning)
3691 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3692 collectStridedAccess(LI);
3694 // Reduction instructions are allowed to have exit users.
3695 // All other instructions must not have external users.
3696 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3697 emitAnalysis(VectorizationReport(it) <<
3698 "value cannot be used outside the loop");
3707 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3708 if (Inductions.empty()) {
3709 emitAnalysis(VectorizationReport()
3710 << "loop induction variable could not be identified");
3718 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3719 /// return the induction operand of the gep pointer.
3720 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3721 const DataLayout *DL, Loop *Lp) {
3722 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3726 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3728 // Check that all of the gep indices are uniform except for our induction
3730 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3731 if (i != InductionOperand &&
3732 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3734 return GEP->getOperand(InductionOperand);
3737 ///\brief Look for a cast use of the passed value.
3738 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3739 Value *UniqueCast = nullptr;
3740 for (User *U : Ptr->users()) {
3741 CastInst *CI = dyn_cast<CastInst>(U);
3742 if (CI && CI->getType() == Ty) {
3752 ///\brief Get the stride of a pointer access in a loop.
3753 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3754 /// pointer to the Value, or null otherwise.
3755 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3756 const DataLayout *DL, Loop *Lp) {
3757 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3758 if (!PtrTy || PtrTy->isAggregateType())
3761 // Try to remove a gep instruction to make the pointer (actually index at this
3762 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3763 // pointer, otherwise, we are analyzing the index.
3764 Value *OrigPtr = Ptr;
3766 // The size of the pointer access.
3767 int64_t PtrAccessSize = 1;
3769 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3770 const SCEV *V = SE->getSCEV(Ptr);
3774 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3775 V = C->getOperand();
3777 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3781 V = S->getStepRecurrence(*SE);
3785 // Strip off the size of access multiplication if we are still analyzing the
3787 if (OrigPtr == Ptr) {
3788 DL->getTypeAllocSize(PtrTy->getElementType());
3789 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3790 if (M->getOperand(0)->getSCEVType() != scConstant)
3793 const APInt &APStepVal =
3794 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3796 // Huge step value - give up.
3797 if (APStepVal.getBitWidth() > 64)
3800 int64_t StepVal = APStepVal.getSExtValue();
3801 if (PtrAccessSize != StepVal)
3803 V = M->getOperand(1);
3808 Type *StripedOffRecurrenceCast = nullptr;
3809 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3810 StripedOffRecurrenceCast = C->getType();
3811 V = C->getOperand();
3814 // Look for the loop invariant symbolic value.
3815 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3819 Value *Stride = U->getValue();
3820 if (!Lp->isLoopInvariant(Stride))
3823 // If we have stripped off the recurrence cast we have to make sure that we
3824 // return the value that is used in this loop so that we can replace it later.
3825 if (StripedOffRecurrenceCast)
3826 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3831 void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) {
3832 Value *Ptr = nullptr;
3833 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3834 Ptr = LI->getPointerOperand();
3835 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3836 Ptr = SI->getPointerOperand();
3840 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3844 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3845 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3846 Strides[Ptr] = Stride;
3847 StrideSet.insert(Stride);
3850 void LoopVectorizationLegality::collectLoopUniforms() {
3851 // We now know that the loop is vectorizable!
3852 // Collect variables that will remain uniform after vectorization.
3853 std::vector<Value*> Worklist;
3854 BasicBlock *Latch = TheLoop->getLoopLatch();
3856 // Start with the conditional branch and walk up the block.
3857 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3859 // Also add all consecutive pointer values; these values will be uniform
3860 // after vectorization (and subsequent cleanup) and, until revectorization is
3861 // supported, all dependencies must also be uniform.
3862 for (Loop::block_iterator B = TheLoop->block_begin(),
3863 BE = TheLoop->block_end(); B != BE; ++B)
3864 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
3866 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
3867 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3869 while (!Worklist.empty()) {
3870 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3871 Worklist.pop_back();
3873 // Look at instructions inside this loop.
3874 // Stop when reaching PHI nodes.
3875 // TODO: we need to follow values all over the loop, not only in this block.
3876 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3879 // This is a known uniform.
3882 // Insert all operands.
3883 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3887 bool LoopVectorizationLegality::canVectorizeMemory() {
3888 LAI = &LAA->getInfo(TheLoop, Strides);
3889 auto &OptionalReport = LAI->getReport();
3891 emitAnalysis(VectorizationReport(*OptionalReport));
3892 return LAI->canVectorizeMemory();
3895 static bool hasMultipleUsesOf(Instruction *I,
3896 SmallPtrSetImpl<Instruction *> &Insts) {
3897 unsigned NumUses = 0;
3898 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
3899 if (Insts.count(dyn_cast<Instruction>(*Use)))
3908 static bool areAllUsesIn(Instruction *I, SmallPtrSetImpl<Instruction *> &Set) {
3909 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
3910 if (!Set.count(dyn_cast<Instruction>(*Use)))
3915 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
3916 ReductionKind Kind) {
3917 if (Phi->getNumIncomingValues() != 2)
3920 // Reduction variables are only found in the loop header block.
3921 if (Phi->getParent() != TheLoop->getHeader())
3924 // Obtain the reduction start value from the value that comes from the loop
3926 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
3928 // ExitInstruction is the single value which is used outside the loop.
3929 // We only allow for a single reduction value to be used outside the loop.
3930 // This includes users of the reduction, variables (which form a cycle
3931 // which ends in the phi node).
3932 Instruction *ExitInstruction = nullptr;
3933 // Indicates that we found a reduction operation in our scan.
3934 bool FoundReduxOp = false;
3936 // We start with the PHI node and scan for all of the users of this
3937 // instruction. All users must be instructions that can be used as reduction
3938 // variables (such as ADD). We must have a single out-of-block user. The cycle
3939 // must include the original PHI.
3940 bool FoundStartPHI = false;
3942 // To recognize min/max patterns formed by a icmp select sequence, we store
3943 // the number of instruction we saw from the recognized min/max pattern,
3944 // to make sure we only see exactly the two instructions.
3945 unsigned NumCmpSelectPatternInst = 0;
3946 ReductionInstDesc ReduxDesc(false, nullptr);
3948 SmallPtrSet<Instruction *, 8> VisitedInsts;
3949 SmallVector<Instruction *, 8> Worklist;
3950 Worklist.push_back(Phi);
3951 VisitedInsts.insert(Phi);
3953 // A value in the reduction can be used:
3954 // - By the reduction:
3955 // - Reduction operation:
3956 // - One use of reduction value (safe).
3957 // - Multiple use of reduction value (not safe).
3959 // - All uses of the PHI must be the reduction (safe).
3960 // - Otherwise, not safe.
3961 // - By one instruction outside of the loop (safe).
3962 // - By further instructions outside of the loop (not safe).
3963 // - By an instruction that is not part of the reduction (not safe).
3965 // * An instruction type other than PHI or the reduction operation.
3966 // * A PHI in the header other than the initial PHI.
3967 while (!Worklist.empty()) {
3968 Instruction *Cur = Worklist.back();
3969 Worklist.pop_back();
3972 // If the instruction has no users then this is a broken chain and can't be
3973 // a reduction variable.
3974 if (Cur->use_empty())
3977 bool IsAPhi = isa<PHINode>(Cur);
3979 // A header PHI use other than the original PHI.
3980 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
3983 // Reductions of instructions such as Div, and Sub is only possible if the
3984 // LHS is the reduction variable.
3985 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
3986 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
3987 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
3990 // Any reduction instruction must be of one of the allowed kinds.
3991 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
3992 if (!ReduxDesc.IsReduction)
3995 // A reduction operation must only have one use of the reduction value.
3996 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
3997 hasMultipleUsesOf(Cur, VisitedInsts))
4000 // All inputs to a PHI node must be a reduction value.
4001 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
4004 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
4005 isa<SelectInst>(Cur)))
4006 ++NumCmpSelectPatternInst;
4007 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
4008 isa<SelectInst>(Cur)))
4009 ++NumCmpSelectPatternInst;
4011 // Check whether we found a reduction operator.
4012 FoundReduxOp |= !IsAPhi;
4014 // Process users of current instruction. Push non-PHI nodes after PHI nodes
4015 // onto the stack. This way we are going to have seen all inputs to PHI
4016 // nodes once we get to them.
4017 SmallVector<Instruction *, 8> NonPHIs;
4018 SmallVector<Instruction *, 8> PHIs;
4019 for (User *U : Cur->users()) {
4020 Instruction *UI = cast<Instruction>(U);
4022 // Check if we found the exit user.
4023 BasicBlock *Parent = UI->getParent();
4024 if (!TheLoop->contains(Parent)) {
4025 // Exit if you find multiple outside users or if the header phi node is
4026 // being used. In this case the user uses the value of the previous
4027 // iteration, in which case we would loose "VF-1" iterations of the
4028 // reduction operation if we vectorize.
4029 if (ExitInstruction != nullptr || Cur == Phi)
4032 // The instruction used by an outside user must be the last instruction
4033 // before we feed back to the reduction phi. Otherwise, we loose VF-1
4034 // operations on the value.
4035 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
4038 ExitInstruction = Cur;
4042 // Process instructions only once (termination). Each reduction cycle
4043 // value must only be used once, except by phi nodes and min/max
4044 // reductions which are represented as a cmp followed by a select.
4045 ReductionInstDesc IgnoredVal(false, nullptr);
4046 if (VisitedInsts.insert(UI).second) {
4047 if (isa<PHINode>(UI))
4050 NonPHIs.push_back(UI);
4051 } else if (!isa<PHINode>(UI) &&
4052 ((!isa<FCmpInst>(UI) &&
4053 !isa<ICmpInst>(UI) &&
4054 !isa<SelectInst>(UI)) ||
4055 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
4058 // Remember that we completed the cycle.
4060 FoundStartPHI = true;
4062 Worklist.append(PHIs.begin(), PHIs.end());
4063 Worklist.append(NonPHIs.begin(), NonPHIs.end());
4066 // This means we have seen one but not the other instruction of the
4067 // pattern or more than just a select and cmp.
4068 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
4069 NumCmpSelectPatternInst != 2)
4072 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
4075 // We found a reduction var if we have reached the original phi node and we
4076 // only have a single instruction with out-of-loop users.
4078 // This instruction is allowed to have out-of-loop users.
4079 AllowedExit.insert(ExitInstruction);
4081 // Save the description of this reduction variable.
4082 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
4083 ReduxDesc.MinMaxKind);
4084 Reductions[Phi] = RD;
4085 // We've ended the cycle. This is a reduction variable if we have an
4086 // outside user and it has a binary op.
4091 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
4092 /// pattern corresponding to a min(X, Y) or max(X, Y).
4093 LoopVectorizationLegality::ReductionInstDesc
4094 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
4095 ReductionInstDesc &Prev) {
4097 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
4098 "Expect a select instruction");
4099 Instruction *Cmp = nullptr;
4100 SelectInst *Select = nullptr;
4102 // We must handle the select(cmp()) as a single instruction. Advance to the
4104 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
4105 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
4106 return ReductionInstDesc(false, I);
4107 return ReductionInstDesc(Select, Prev.MinMaxKind);
4110 // Only handle single use cases for now.
4111 if (!(Select = dyn_cast<SelectInst>(I)))
4112 return ReductionInstDesc(false, I);
4113 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
4114 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
4115 return ReductionInstDesc(false, I);
4116 if (!Cmp->hasOneUse())
4117 return ReductionInstDesc(false, I);
4122 // Look for a min/max pattern.
4123 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4124 return ReductionInstDesc(Select, MRK_UIntMin);
4125 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4126 return ReductionInstDesc(Select, MRK_UIntMax);
4127 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4128 return ReductionInstDesc(Select, MRK_SIntMax);
4129 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4130 return ReductionInstDesc(Select, MRK_SIntMin);
4131 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4132 return ReductionInstDesc(Select, MRK_FloatMin);
4133 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4134 return ReductionInstDesc(Select, MRK_FloatMax);
4135 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4136 return ReductionInstDesc(Select, MRK_FloatMin);
4137 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4138 return ReductionInstDesc(Select, MRK_FloatMax);
4140 return ReductionInstDesc(false, I);
4143 LoopVectorizationLegality::ReductionInstDesc
4144 LoopVectorizationLegality::isReductionInstr(Instruction *I,
4146 ReductionInstDesc &Prev) {
4147 bool FP = I->getType()->isFloatingPointTy();
4148 bool FastMath = FP && I->hasUnsafeAlgebra();
4149 switch (I->getOpcode()) {
4151 return ReductionInstDesc(false, I);
4152 case Instruction::PHI:
4153 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
4154 Kind != RK_FloatMinMax))
4155 return ReductionInstDesc(false, I);
4156 return ReductionInstDesc(I, Prev.MinMaxKind);
4157 case Instruction::Sub:
4158 case Instruction::Add:
4159 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4160 case Instruction::Mul:
4161 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4162 case Instruction::And:
4163 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4164 case Instruction::Or:
4165 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4166 case Instruction::Xor:
4167 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4168 case Instruction::FMul:
4169 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4170 case Instruction::FSub:
4171 case Instruction::FAdd:
4172 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4173 case Instruction::FCmp:
4174 case Instruction::ICmp:
4175 case Instruction::Select:
4176 if (Kind != RK_IntegerMinMax &&
4177 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4178 return ReductionInstDesc(false, I);
4179 return isMinMaxSelectCmpPattern(I, Prev);
4183 LoopVectorizationLegality::InductionKind
4184 LoopVectorizationLegality::isInductionVariable(PHINode *Phi,
4185 ConstantInt *&StepValue) {
4186 Type *PhiTy = Phi->getType();
4187 // We only handle integer and pointer inductions variables.
4188 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4189 return IK_NoInduction;
4191 // Check that the PHI is consecutive.
4192 const SCEV *PhiScev = SE->getSCEV(Phi);
4193 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4195 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4196 return IK_NoInduction;
4199 const SCEV *Step = AR->getStepRecurrence(*SE);
4200 // Calculate the pointer stride and check if it is consecutive.
4201 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4203 return IK_NoInduction;
4205 ConstantInt *CV = C->getValue();
4206 if (PhiTy->isIntegerTy()) {
4208 return IK_IntInduction;
4211 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4212 Type *PointerElementType = PhiTy->getPointerElementType();
4213 // The pointer stride cannot be determined if the pointer element type is not
4215 if (!PointerElementType->isSized())
4216 return IK_NoInduction;
4218 int64_t Size = static_cast<int64_t>(DL->getTypeAllocSize(PointerElementType));
4219 int64_t CVSize = CV->getSExtValue();
4221 return IK_NoInduction;
4222 StepValue = ConstantInt::getSigned(CV->getType(), CVSize / Size);
4223 return IK_PtrInduction;
4226 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4227 Value *In0 = const_cast<Value*>(V);
4228 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4232 return Inductions.count(PN);
4235 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4236 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4239 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4240 SmallPtrSetImpl<Value *> &SafePtrs) {
4242 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4243 // Check that we don't have a constant expression that can trap as operand.
4244 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4246 if (Constant *C = dyn_cast<Constant>(*OI))
4250 // We might be able to hoist the load.
4251 if (it->mayReadFromMemory()) {
4252 LoadInst *LI = dyn_cast<LoadInst>(it);
4255 if (!SafePtrs.count(LI->getPointerOperand())) {
4256 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
4257 MaskedOp.insert(LI);
4264 // We don't predicate stores at the moment.
4265 if (it->mayWriteToMemory()) {
4266 StoreInst *SI = dyn_cast<StoreInst>(it);
4267 // We only support predication of stores in basic blocks with one
4272 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
4273 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
4275 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
4276 !isSinglePredecessor) {
4277 // Build a masked store if it is legal for the target, otherwise scalarize
4279 bool isLegalMaskedOp =
4280 isLegalMaskedStore(SI->getValueOperand()->getType(),
4281 SI->getPointerOperand());
4282 if (isLegalMaskedOp) {
4284 MaskedOp.insert(SI);
4293 // The instructions below can trap.
4294 switch (it->getOpcode()) {
4296 case Instruction::UDiv:
4297 case Instruction::SDiv:
4298 case Instruction::URem:
4299 case Instruction::SRem:
4307 LoopVectorizationCostModel::VectorizationFactor
4308 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
4309 // Width 1 means no vectorize
4310 VectorizationFactor Factor = { 1U, 0U };
4311 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4312 emitAnalysis(VectorizationReport() <<
4313 "runtime pointer checks needed. Enable vectorization of this "
4314 "loop with '#pragma clang loop vectorize(enable)' when "
4315 "compiling with -Os");
4316 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4320 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
4321 emitAnalysis(VectorizationReport() <<
4322 "store that is conditionally executed prevents vectorization");
4323 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
4327 // Find the trip count.
4328 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4329 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4331 unsigned WidestType = getWidestType();
4332 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4333 unsigned MaxSafeDepDist = -1U;
4334 if (Legal->getMaxSafeDepDistBytes() != -1U)
4335 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4336 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4337 WidestRegister : MaxSafeDepDist);
4338 unsigned MaxVectorSize = WidestRegister / WidestType;
4339 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4340 DEBUG(dbgs() << "LV: The Widest register is: "
4341 << WidestRegister << " bits.\n");
4343 if (MaxVectorSize == 0) {
4344 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4348 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
4349 " into one vector!");
4351 unsigned VF = MaxVectorSize;
4353 // If we optimize the program for size, avoid creating the tail loop.
4355 // If we are unable to calculate the trip count then don't try to vectorize.
4358 (VectorizationReport() <<
4359 "unable to calculate the loop count due to complex control flow");
4360 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4364 // Find the maximum SIMD width that can fit within the trip count.
4365 VF = TC % MaxVectorSize;
4370 // If the trip count that we found modulo the vectorization factor is not
4371 // zero then we require a tail.
4373 emitAnalysis(VectorizationReport() <<
4374 "cannot optimize for size and vectorize at the "
4375 "same time. Enable vectorization of this loop "
4376 "with '#pragma clang loop vectorize(enable)' "
4377 "when compiling with -Os");
4378 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4383 int UserVF = Hints->getWidth();
4385 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4386 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4388 Factor.Width = UserVF;
4392 float Cost = expectedCost(1);
4394 const float ScalarCost = Cost;
4397 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
4399 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
4400 // Ignore scalar width, because the user explicitly wants vectorization.
4401 if (ForceVectorization && VF > 1) {
4403 Cost = expectedCost(Width) / (float)Width;
4406 for (unsigned i=2; i <= VF; i*=2) {
4407 // Notice that the vector loop needs to be executed less times, so
4408 // we need to divide the cost of the vector loops by the width of
4409 // the vector elements.
4410 float VectorCost = expectedCost(i) / (float)i;
4411 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
4412 (int)VectorCost << ".\n");
4413 if (VectorCost < Cost) {
4419 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
4420 << "LV: Vectorization seems to be not beneficial, "
4421 << "but was forced by a user.\n");
4422 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
4423 Factor.Width = Width;
4424 Factor.Cost = Width * Cost;
4428 unsigned LoopVectorizationCostModel::getWidestType() {
4429 unsigned MaxWidth = 8;
4432 for (Loop::block_iterator bb = TheLoop->block_begin(),
4433 be = TheLoop->block_end(); bb != be; ++bb) {
4434 BasicBlock *BB = *bb;
4436 // For each instruction in the loop.
4437 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4438 Type *T = it->getType();
4440 // Ignore ephemeral values.
4441 if (EphValues.count(it))
4444 // Only examine Loads, Stores and PHINodes.
4445 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4448 // Examine PHI nodes that are reduction variables.
4449 if (PHINode *PN = dyn_cast<PHINode>(it))
4450 if (!Legal->getReductionVars()->count(PN))
4453 // Examine the stored values.
4454 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4455 T = ST->getValueOperand()->getType();
4457 // Ignore loaded pointer types and stored pointer types that are not
4458 // consecutive. However, we do want to take consecutive stores/loads of
4459 // pointer vectors into account.
4460 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4463 MaxWidth = std::max(MaxWidth,
4464 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
4472 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
4474 unsigned LoopCost) {
4476 // -- The unroll heuristics --
4477 // We unroll the loop in order to expose ILP and reduce the loop overhead.
4478 // There are many micro-architectural considerations that we can't predict
4479 // at this level. For example, frontend pressure (on decode or fetch) due to
4480 // code size, or the number and capabilities of the execution ports.
4482 // We use the following heuristics to select the unroll factor:
4483 // 1. If the code has reductions, then we unroll in order to break the cross
4484 // iteration dependency.
4485 // 2. If the loop is really small, then we unroll in order to reduce the loop
4487 // 3. We don't unroll if we think that we will spill registers to memory due
4488 // to the increased register pressure.
4490 // Use the user preference, unless 'auto' is selected.
4491 int UserUF = Hints->getInterleave();
4495 // When we optimize for size, we don't unroll.
4499 // We used the distance for the unroll factor.
4500 if (Legal->getMaxSafeDepDistBytes() != -1U)
4503 // Do not unroll loops with a relatively small trip count.
4504 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4505 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
4508 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
4509 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
4513 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
4514 TargetNumRegisters = ForceTargetNumScalarRegs;
4516 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
4517 TargetNumRegisters = ForceTargetNumVectorRegs;
4520 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4521 // We divide by these constants so assume that we have at least one
4522 // instruction that uses at least one register.
4523 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4524 R.NumInstructions = std::max(R.NumInstructions, 1U);
4526 // We calculate the unroll factor using the following formula.
4527 // Subtract the number of loop invariants from the number of available
4528 // registers. These registers are used by all of the unrolled instances.
4529 // Next, divide the remaining registers by the number of registers that is
4530 // required by the loop, in order to estimate how many parallel instances
4531 // fit without causing spills. All of this is rounded down if necessary to be
4532 // a power of two. We want power of two unroll factors to simplify any
4533 // addressing operations or alignment considerations.
4534 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
4537 // Don't count the induction variable as unrolled.
4538 if (EnableIndVarRegisterHeur)
4539 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
4540 std::max(1U, (R.MaxLocalUsers - 1)));
4542 // Clamp the unroll factor ranges to reasonable factors.
4543 unsigned MaxInterleaveSize = TTI.getMaxInterleaveFactor();
4545 // Check if the user has overridden the unroll max.
4547 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
4548 MaxInterleaveSize = ForceTargetMaxScalarInterleaveFactor;
4550 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
4551 MaxInterleaveSize = ForceTargetMaxVectorInterleaveFactor;
4554 // If we did not calculate the cost for VF (because the user selected the VF)
4555 // then we calculate the cost of VF here.
4557 LoopCost = expectedCost(VF);
4559 // Clamp the calculated UF to be between the 1 and the max unroll factor
4560 // that the target allows.
4561 if (UF > MaxInterleaveSize)
4562 UF = MaxInterleaveSize;
4566 // Unroll if we vectorized this loop and there is a reduction that could
4567 // benefit from unrolling.
4568 if (VF > 1 && Legal->getReductionVars()->size()) {
4569 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
4573 // Note that if we've already vectorized the loop we will have done the
4574 // runtime check and so unrolling won't require further checks.
4575 bool UnrollingRequiresRuntimePointerCheck =
4576 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
4578 // We want to unroll small loops in order to reduce the loop overhead and
4579 // potentially expose ILP opportunities.
4580 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
4581 if (!UnrollingRequiresRuntimePointerCheck &&
4582 LoopCost < SmallLoopCost) {
4583 // We assume that the cost overhead is 1 and we use the cost model
4584 // to estimate the cost of the loop and unroll until the cost of the
4585 // loop overhead is about 5% of the cost of the loop.
4586 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
4588 // Unroll until store/load ports (estimated by max unroll factor) are
4590 unsigned NumStores = Legal->getNumStores();
4591 unsigned NumLoads = Legal->getNumLoads();
4592 unsigned StoresUF = UF / (NumStores ? NumStores : 1);
4593 unsigned LoadsUF = UF / (NumLoads ? NumLoads : 1);
4595 // If we have a scalar reduction (vector reductions are already dealt with
4596 // by this point), we can increase the critical path length if the loop
4597 // we're unrolling is inside another loop. Limit, by default to 2, so the
4598 // critical path only gets increased by one reduction operation.
4599 if (Legal->getReductionVars()->size() &&
4600 TheLoop->getLoopDepth() > 1) {
4601 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionUF);
4602 SmallUF = std::min(SmallUF, F);
4603 StoresUF = std::min(StoresUF, F);
4604 LoadsUF = std::min(LoadsUF, F);
4607 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
4608 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
4609 return std::max(StoresUF, LoadsUF);
4612 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
4616 // Unroll if this is a large loop (small loops are already dealt with by this
4617 // point) that could benefit from interleaved unrolling.
4618 bool HasReductions = (Legal->getReductionVars()->size() > 0);
4619 if (TTI.enableAggressiveInterleaving(HasReductions)) {
4620 DEBUG(dbgs() << "LV: Unrolling to expose ILP.\n");
4624 DEBUG(dbgs() << "LV: Not Unrolling.\n");
4628 LoopVectorizationCostModel::RegisterUsage
4629 LoopVectorizationCostModel::calculateRegisterUsage() {
4630 // This function calculates the register usage by measuring the highest number
4631 // of values that are alive at a single location. Obviously, this is a very
4632 // rough estimation. We scan the loop in a topological order in order and
4633 // assign a number to each instruction. We use RPO to ensure that defs are
4634 // met before their users. We assume that each instruction that has in-loop
4635 // users starts an interval. We record every time that an in-loop value is
4636 // used, so we have a list of the first and last occurrences of each
4637 // instruction. Next, we transpose this data structure into a multi map that
4638 // holds the list of intervals that *end* at a specific location. This multi
4639 // map allows us to perform a linear search. We scan the instructions linearly
4640 // and record each time that a new interval starts, by placing it in a set.
4641 // If we find this value in the multi-map then we remove it from the set.
4642 // The max register usage is the maximum size of the set.
4643 // We also search for instructions that are defined outside the loop, but are
4644 // used inside the loop. We need this number separately from the max-interval
4645 // usage number because when we unroll, loop-invariant values do not take
4647 LoopBlocksDFS DFS(TheLoop);
4651 R.NumInstructions = 0;
4653 // Each 'key' in the map opens a new interval. The values
4654 // of the map are the index of the 'last seen' usage of the
4655 // instruction that is the key.
4656 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4657 // Maps instruction to its index.
4658 DenseMap<unsigned, Instruction*> IdxToInstr;
4659 // Marks the end of each interval.
4660 IntervalMap EndPoint;
4661 // Saves the list of instruction indices that are used in the loop.
4662 SmallSet<Instruction*, 8> Ends;
4663 // Saves the list of values that are used in the loop but are
4664 // defined outside the loop, such as arguments and constants.
4665 SmallPtrSet<Value*, 8> LoopInvariants;
4668 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4669 be = DFS.endRPO(); bb != be; ++bb) {
4670 R.NumInstructions += (*bb)->size();
4671 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4673 Instruction *I = it;
4674 IdxToInstr[Index++] = I;
4676 // Save the end location of each USE.
4677 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4678 Value *U = I->getOperand(i);
4679 Instruction *Instr = dyn_cast<Instruction>(U);
4681 // Ignore non-instruction values such as arguments, constants, etc.
4682 if (!Instr) continue;
4684 // If this instruction is outside the loop then record it and continue.
4685 if (!TheLoop->contains(Instr)) {
4686 LoopInvariants.insert(Instr);
4690 // Overwrite previous end points.
4691 EndPoint[Instr] = Index;
4697 // Saves the list of intervals that end with the index in 'key'.
4698 typedef SmallVector<Instruction*, 2> InstrList;
4699 DenseMap<unsigned, InstrList> TransposeEnds;
4701 // Transpose the EndPoints to a list of values that end at each index.
4702 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4704 TransposeEnds[it->second].push_back(it->first);
4706 SmallSet<Instruction*, 8> OpenIntervals;
4707 unsigned MaxUsage = 0;
4710 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4711 for (unsigned int i = 0; i < Index; ++i) {
4712 Instruction *I = IdxToInstr[i];
4713 // Ignore instructions that are never used within the loop.
4714 if (!Ends.count(I)) continue;
4716 // Ignore ephemeral values.
4717 if (EphValues.count(I))
4720 // Remove all of the instructions that end at this location.
4721 InstrList &List = TransposeEnds[i];
4722 for (unsigned int j=0, e = List.size(); j < e; ++j)
4723 OpenIntervals.erase(List[j]);
4725 // Count the number of live interals.
4726 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4728 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4729 OpenIntervals.size() << '\n');
4731 // Add the current instruction to the list of open intervals.
4732 OpenIntervals.insert(I);
4735 unsigned Invariant = LoopInvariants.size();
4736 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
4737 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
4738 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
4740 R.LoopInvariantRegs = Invariant;
4741 R.MaxLocalUsers = MaxUsage;
4745 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4749 for (Loop::block_iterator bb = TheLoop->block_begin(),
4750 be = TheLoop->block_end(); bb != be; ++bb) {
4751 unsigned BlockCost = 0;
4752 BasicBlock *BB = *bb;
4754 // For each instruction in the old loop.
4755 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4756 // Skip dbg intrinsics.
4757 if (isa<DbgInfoIntrinsic>(it))
4760 // Ignore ephemeral values.
4761 if (EphValues.count(it))
4764 unsigned C = getInstructionCost(it, VF);
4766 // Check if we should override the cost.
4767 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
4768 C = ForceTargetInstructionCost;
4771 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
4772 VF << " For instruction: " << *it << '\n');
4775 // We assume that if-converted blocks have a 50% chance of being executed.
4776 // When the code is scalar then some of the blocks are avoided due to CF.
4777 // When the code is vectorized we execute all code paths.
4778 if (VF == 1 && Legal->blockNeedsPredication(*bb))
4787 /// \brief Check whether the address computation for a non-consecutive memory
4788 /// access looks like an unlikely candidate for being merged into the indexing
4791 /// We look for a GEP which has one index that is an induction variable and all
4792 /// other indices are loop invariant. If the stride of this access is also
4793 /// within a small bound we decide that this address computation can likely be
4794 /// merged into the addressing mode.
4795 /// In all other cases, we identify the address computation as complex.
4796 static bool isLikelyComplexAddressComputation(Value *Ptr,
4797 LoopVectorizationLegality *Legal,
4798 ScalarEvolution *SE,
4799 const Loop *TheLoop) {
4800 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
4804 // We are looking for a gep with all loop invariant indices except for one
4805 // which should be an induction variable.
4806 unsigned NumOperands = Gep->getNumOperands();
4807 for (unsigned i = 1; i < NumOperands; ++i) {
4808 Value *Opd = Gep->getOperand(i);
4809 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
4810 !Legal->isInductionVariable(Opd))
4814 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
4815 // can likely be merged into the address computation.
4816 unsigned MaxMergeDistance = 64;
4818 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
4822 // Check the step is constant.
4823 const SCEV *Step = AddRec->getStepRecurrence(*SE);
4824 // Calculate the pointer stride and check if it is consecutive.
4825 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4829 const APInt &APStepVal = C->getValue()->getValue();
4831 // Huge step value - give up.
4832 if (APStepVal.getBitWidth() > 64)
4835 int64_t StepVal = APStepVal.getSExtValue();
4837 return StepVal > MaxMergeDistance;
4840 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
4841 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
4847 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
4848 // If we know that this instruction will remain uniform, check the cost of
4849 // the scalar version.
4850 if (Legal->isUniformAfterVectorization(I))
4853 Type *RetTy = I->getType();
4854 Type *VectorTy = ToVectorTy(RetTy, VF);
4856 // TODO: We need to estimate the cost of intrinsic calls.
4857 switch (I->getOpcode()) {
4858 case Instruction::GetElementPtr:
4859 // We mark this instruction as zero-cost because the cost of GEPs in
4860 // vectorized code depends on whether the corresponding memory instruction
4861 // is scalarized or not. Therefore, we handle GEPs with the memory
4862 // instruction cost.
4864 case Instruction::Br: {
4865 return TTI.getCFInstrCost(I->getOpcode());
4867 case Instruction::PHI:
4868 //TODO: IF-converted IFs become selects.
4870 case Instruction::Add:
4871 case Instruction::FAdd:
4872 case Instruction::Sub:
4873 case Instruction::FSub:
4874 case Instruction::Mul:
4875 case Instruction::FMul:
4876 case Instruction::UDiv:
4877 case Instruction::SDiv:
4878 case Instruction::FDiv:
4879 case Instruction::URem:
4880 case Instruction::SRem:
4881 case Instruction::FRem:
4882 case Instruction::Shl:
4883 case Instruction::LShr:
4884 case Instruction::AShr:
4885 case Instruction::And:
4886 case Instruction::Or:
4887 case Instruction::Xor: {
4888 // Since we will replace the stride by 1 the multiplication should go away.
4889 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
4891 // Certain instructions can be cheaper to vectorize if they have a constant
4892 // second vector operand. One example of this are shifts on x86.
4893 TargetTransformInfo::OperandValueKind Op1VK =
4894 TargetTransformInfo::OK_AnyValue;
4895 TargetTransformInfo::OperandValueKind Op2VK =
4896 TargetTransformInfo::OK_AnyValue;
4897 TargetTransformInfo::OperandValueProperties Op1VP =
4898 TargetTransformInfo::OP_None;
4899 TargetTransformInfo::OperandValueProperties Op2VP =
4900 TargetTransformInfo::OP_None;
4901 Value *Op2 = I->getOperand(1);
4903 // Check for a splat of a constant or for a non uniform vector of constants.
4904 if (isa<ConstantInt>(Op2)) {
4905 ConstantInt *CInt = cast<ConstantInt>(Op2);
4906 if (CInt && CInt->getValue().isPowerOf2())
4907 Op2VP = TargetTransformInfo::OP_PowerOf2;
4908 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4909 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
4910 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
4911 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
4913 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
4914 if (CInt && CInt->getValue().isPowerOf2())
4915 Op2VP = TargetTransformInfo::OP_PowerOf2;
4916 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4920 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
4923 case Instruction::Select: {
4924 SelectInst *SI = cast<SelectInst>(I);
4925 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
4926 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
4927 Type *CondTy = SI->getCondition()->getType();
4929 CondTy = VectorType::get(CondTy, VF);
4931 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
4933 case Instruction::ICmp:
4934 case Instruction::FCmp: {
4935 Type *ValTy = I->getOperand(0)->getType();
4936 VectorTy = ToVectorTy(ValTy, VF);
4937 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
4939 case Instruction::Store:
4940 case Instruction::Load: {
4941 StoreInst *SI = dyn_cast<StoreInst>(I);
4942 LoadInst *LI = dyn_cast<LoadInst>(I);
4943 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
4945 VectorTy = ToVectorTy(ValTy, VF);
4947 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
4948 unsigned AS = SI ? SI->getPointerAddressSpace() :
4949 LI->getPointerAddressSpace();
4950 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
4951 // We add the cost of address computation here instead of with the gep
4952 // instruction because only here we know whether the operation is
4955 return TTI.getAddressComputationCost(VectorTy) +
4956 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4958 // Scalarized loads/stores.
4959 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
4960 bool Reverse = ConsecutiveStride < 0;
4961 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
4962 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
4963 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
4964 bool IsComplexComputation =
4965 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
4967 // The cost of extracting from the value vector and pointer vector.
4968 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
4969 for (unsigned i = 0; i < VF; ++i) {
4970 // The cost of extracting the pointer operand.
4971 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
4972 // In case of STORE, the cost of ExtractElement from the vector.
4973 // In case of LOAD, the cost of InsertElement into the returned
4975 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
4976 Instruction::InsertElement,
4980 // The cost of the scalar loads/stores.
4981 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
4982 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
4987 // Wide load/stores.
4988 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
4989 if (Legal->isMaskRequired(I))
4990 Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment,
4993 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4996 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5000 case Instruction::ZExt:
5001 case Instruction::SExt:
5002 case Instruction::FPToUI:
5003 case Instruction::FPToSI:
5004 case Instruction::FPExt:
5005 case Instruction::PtrToInt:
5006 case Instruction::IntToPtr:
5007 case Instruction::SIToFP:
5008 case Instruction::UIToFP:
5009 case Instruction::Trunc:
5010 case Instruction::FPTrunc:
5011 case Instruction::BitCast: {
5012 // We optimize the truncation of induction variable.
5013 // The cost of these is the same as the scalar operation.
5014 if (I->getOpcode() == Instruction::Trunc &&
5015 Legal->isInductionVariable(I->getOperand(0)))
5016 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5017 I->getOperand(0)->getType());
5019 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5020 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5022 case Instruction::Call: {
5023 CallInst *CI = cast<CallInst>(I);
5024 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
5025 assert(ID && "Not an intrinsic call!");
5026 Type *RetTy = ToVectorTy(CI->getType(), VF);
5027 SmallVector<Type*, 4> Tys;
5028 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
5029 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
5030 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
5033 // We are scalarizing the instruction. Return the cost of the scalar
5034 // instruction, plus the cost of insert and extract into vector
5035 // elements, times the vector width.
5038 if (!RetTy->isVoidTy() && VF != 1) {
5039 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5041 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5044 // The cost of inserting the results plus extracting each one of the
5046 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5049 // The cost of executing VF copies of the scalar instruction. This opcode
5050 // is unknown. Assume that it is the same as 'mul'.
5051 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5057 char LoopVectorize::ID = 0;
5058 static const char lv_name[] = "Loop Vectorization";
5059 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5060 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
5061 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
5062 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
5063 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
5064 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5065 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5066 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5067 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
5068 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5069 INITIALIZE_PASS_DEPENDENCY(LoopAccessAnalysis)
5070 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5073 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5074 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5078 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5079 // Check for a store.
5080 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5081 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5083 // Check for a load.
5084 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5085 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5091 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5092 bool IfPredicateStore) {
5093 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5094 // Holds vector parameters or scalars, in case of uniform vals.
5095 SmallVector<VectorParts, 4> Params;
5097 setDebugLocFromInst(Builder, Instr);
5099 // Find all of the vectorized parameters.
5100 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5101 Value *SrcOp = Instr->getOperand(op);
5103 // If we are accessing the old induction variable, use the new one.
5104 if (SrcOp == OldInduction) {
5105 Params.push_back(getVectorValue(SrcOp));
5109 // Try using previously calculated values.
5110 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5112 // If the src is an instruction that appeared earlier in the basic block
5113 // then it should already be vectorized.
5114 if (SrcInst && OrigLoop->contains(SrcInst)) {
5115 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5116 // The parameter is a vector value from earlier.
5117 Params.push_back(WidenMap.get(SrcInst));
5119 // The parameter is a scalar from outside the loop. Maybe even a constant.
5120 VectorParts Scalars;
5121 Scalars.append(UF, SrcOp);
5122 Params.push_back(Scalars);
5126 assert(Params.size() == Instr->getNumOperands() &&
5127 "Invalid number of operands");
5129 // Does this instruction return a value ?
5130 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5132 Value *UndefVec = IsVoidRetTy ? nullptr :
5133 UndefValue::get(Instr->getType());
5134 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5135 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5137 Instruction *InsertPt = Builder.GetInsertPoint();
5138 BasicBlock *IfBlock = Builder.GetInsertBlock();
5139 BasicBlock *CondBlock = nullptr;
5142 Loop *VectorLp = nullptr;
5143 if (IfPredicateStore) {
5144 assert(Instr->getParent()->getSinglePredecessor() &&
5145 "Only support single predecessor blocks");
5146 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5147 Instr->getParent());
5148 VectorLp = LI->getLoopFor(IfBlock);
5149 assert(VectorLp && "Must have a loop for this block");
5152 // For each vector unroll 'part':
5153 for (unsigned Part = 0; Part < UF; ++Part) {
5154 // For each scalar that we create:
5156 // Start an "if (pred) a[i] = ..." block.
5157 Value *Cmp = nullptr;
5158 if (IfPredicateStore) {
5159 if (Cond[Part]->getType()->isVectorTy())
5161 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5162 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5163 ConstantInt::get(Cond[Part]->getType(), 1));
5164 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5165 LoopVectorBody.push_back(CondBlock);
5166 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
5167 // Update Builder with newly created basic block.
5168 Builder.SetInsertPoint(InsertPt);
5171 Instruction *Cloned = Instr->clone();
5173 Cloned->setName(Instr->getName() + ".cloned");
5174 // Replace the operands of the cloned instructions with extracted scalars.
5175 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5176 Value *Op = Params[op][Part];
5177 Cloned->setOperand(op, Op);
5180 // Place the cloned scalar in the new loop.
5181 Builder.Insert(Cloned);
5183 // If the original scalar returns a value we need to place it in a vector
5184 // so that future users will be able to use it.
5186 VecResults[Part] = Cloned;
5189 if (IfPredicateStore) {
5190 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
5191 LoopVectorBody.push_back(NewIfBlock);
5192 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
5193 Builder.SetInsertPoint(InsertPt);
5194 Instruction *OldBr = IfBlock->getTerminator();
5195 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
5196 OldBr->eraseFromParent();
5197 IfBlock = NewIfBlock;
5202 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5203 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5204 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5206 return scalarizeInstruction(Instr, IfPredicateStore);
5209 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5213 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5217 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) {
5218 // When unrolling and the VF is 1, we only need to add a simple scalar.
5219 Type *ITy = Val->getType();
5220 assert(!ITy->isVectorTy() && "Val must be a scalar");
5221 Constant *C = ConstantInt::get(ITy, StartIdx);
5222 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");