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 TargetLibraryInfo *TLI,
248 const TargetTransformInfo *TTI, unsigned VecWidth,
249 unsigned UnrollFactor)
250 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), TLI(TLI), TTI(TTI),
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.
406 /// Target Library Info.
407 const TargetLibraryInfo *TLI;
408 /// Target Transform Info.
409 const TargetTransformInfo *TTI;
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 TargetLibraryInfo *TLI,
460 const TargetTransformInfo *TTI, unsigned UnrollFactor)
461 : InnerLoopVectorizer(OrigLoop, SE, LI, DT, TLI, TTI, 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, DominatorTree *DT,
564 TargetLibraryInfo *TLI, AliasAnalysis *AA,
565 Function *F, const TargetTransformInfo *TTI,
566 LoopAccessAnalysis *LAA)
567 : NumPredStores(0), TheLoop(L), SE(SE), TLI(TLI), TheFunction(F),
568 TTI(TTI), DT(DT), LAA(LAA), LAI(nullptr), Induction(nullptr),
569 WidestIndTy(nullptr), HasFunNoNaNAttr(false) {}
571 /// This enum represents the kinds of reductions that we support.
573 RK_NoReduction, ///< Not a reduction.
574 RK_IntegerAdd, ///< Sum of integers.
575 RK_IntegerMult, ///< Product of integers.
576 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
577 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
578 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
579 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
580 RK_FloatAdd, ///< Sum of floats.
581 RK_FloatMult, ///< Product of floats.
582 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
585 /// This enum represents the kinds of inductions that we support.
587 IK_NoInduction, ///< Not an induction variable.
588 IK_IntInduction, ///< Integer induction variable. Step = C.
589 IK_PtrInduction ///< Pointer induction var. Step = C / sizeof(elem).
592 // This enum represents the kind of minmax reduction.
593 enum MinMaxReductionKind {
603 /// This struct holds information about reduction variables.
604 struct ReductionDescriptor {
605 ReductionDescriptor() : StartValue(nullptr), LoopExitInstr(nullptr),
606 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
608 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
609 MinMaxReductionKind MK)
610 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
612 // The starting value of the reduction.
613 // It does not have to be zero!
614 TrackingVH<Value> StartValue;
615 // The instruction who's value is used outside the loop.
616 Instruction *LoopExitInstr;
617 // The kind of the reduction.
619 // If this a min/max reduction the kind of reduction.
620 MinMaxReductionKind MinMaxKind;
623 /// This POD struct holds information about a potential reduction operation.
624 struct ReductionInstDesc {
625 ReductionInstDesc(bool IsRedux, Instruction *I) :
626 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
628 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
629 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
631 // Is this instruction a reduction candidate.
633 // The last instruction in a min/max pattern (select of the select(icmp())
634 // pattern), or the current reduction instruction otherwise.
635 Instruction *PatternLastInst;
636 // If this is a min/max pattern the comparison predicate.
637 MinMaxReductionKind MinMaxKind;
640 /// A struct for saving information about induction variables.
641 struct InductionInfo {
642 InductionInfo(Value *Start, InductionKind K, ConstantInt *Step)
643 : StartValue(Start), IK(K), StepValue(Step) {
644 assert(IK != IK_NoInduction && "Not an induction");
645 assert(StartValue && "StartValue is null");
646 assert(StepValue && !StepValue->isZero() && "StepValue is zero");
647 assert((IK != IK_PtrInduction || StartValue->getType()->isPointerTy()) &&
648 "StartValue is not a pointer for pointer induction");
649 assert((IK != IK_IntInduction || StartValue->getType()->isIntegerTy()) &&
650 "StartValue is not an integer for integer induction");
651 assert(StepValue->getType()->isIntegerTy() &&
652 "StepValue is not an integer");
655 : StartValue(nullptr), IK(IK_NoInduction), StepValue(nullptr) {}
657 /// Get the consecutive direction. Returns:
658 /// 0 - unknown or non-consecutive.
659 /// 1 - consecutive and increasing.
660 /// -1 - consecutive and decreasing.
661 int getConsecutiveDirection() const {
662 if (StepValue && (StepValue->isOne() || StepValue->isMinusOne()))
663 return StepValue->getSExtValue();
667 /// Compute the transformed value of Index at offset StartValue using step
669 /// For integer induction, returns StartValue + Index * StepValue.
670 /// For pointer induction, returns StartValue[Index * StepValue].
671 /// FIXME: The newly created binary instructions should contain nsw/nuw
672 /// flags, which can be found from the original scalar operations.
673 Value *transform(IRBuilder<> &B, Value *Index) const {
675 case IK_IntInduction:
676 assert(Index->getType() == StartValue->getType() &&
677 "Index type does not match StartValue type");
678 if (StepValue->isMinusOne())
679 return B.CreateSub(StartValue, Index);
680 if (!StepValue->isOne())
681 Index = B.CreateMul(Index, StepValue);
682 return B.CreateAdd(StartValue, Index);
684 case IK_PtrInduction:
685 if (StepValue->isMinusOne())
686 Index = B.CreateNeg(Index);
687 else if (!StepValue->isOne())
688 Index = B.CreateMul(Index, StepValue);
689 return B.CreateGEP(StartValue, Index);
694 llvm_unreachable("invalid enum");
698 TrackingVH<Value> StartValue;
702 ConstantInt *StepValue;
705 /// ReductionList contains the reduction descriptors for all
706 /// of the reductions that were found in the loop.
707 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
709 /// InductionList saves induction variables and maps them to the
710 /// induction descriptor.
711 typedef MapVector<PHINode*, InductionInfo> InductionList;
713 /// Returns true if it is legal to vectorize this loop.
714 /// This does not mean that it is profitable to vectorize this
715 /// loop, only that it is legal to do so.
718 /// Returns the Induction variable.
719 PHINode *getInduction() { return Induction; }
721 /// Returns the reduction variables found in the loop.
722 ReductionList *getReductionVars() { return &Reductions; }
724 /// Returns the induction variables found in the loop.
725 InductionList *getInductionVars() { return &Inductions; }
727 /// Returns the widest induction type.
728 Type *getWidestInductionType() { return WidestIndTy; }
730 /// Returns True if V is an induction variable in this loop.
731 bool isInductionVariable(const Value *V);
733 /// Return true if the block BB needs to be predicated in order for the loop
734 /// to be vectorized.
735 bool blockNeedsPredication(BasicBlock *BB);
737 /// Check if this pointer is consecutive when vectorizing. This happens
738 /// when the last index of the GEP is the induction variable, or that the
739 /// pointer itself is an induction variable.
740 /// This check allows us to vectorize A[idx] into a wide load/store.
742 /// 0 - Stride is unknown or non-consecutive.
743 /// 1 - Address is consecutive.
744 /// -1 - Address is consecutive, and decreasing.
745 int isConsecutivePtr(Value *Ptr);
747 /// Returns true if the value V is uniform within the loop.
748 bool isUniform(Value *V);
750 /// Returns true if this instruction will remain scalar after vectorization.
751 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
753 /// Returns the information that we collected about runtime memory check.
754 const LoopAccessInfo::RuntimePointerCheck *getRuntimePointerCheck() const {
755 return LAI->getRuntimePointerCheck();
758 const LoopAccessInfo *getLAI() const {
762 /// This function returns the identity element (or neutral element) for
764 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
766 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
768 bool hasStride(Value *V) { return StrideSet.count(V); }
769 bool mustCheckStrides() { return !StrideSet.empty(); }
770 SmallPtrSet<Value *, 8>::iterator strides_begin() {
771 return StrideSet.begin();
773 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
775 /// Returns true if the target machine supports masked store operation
776 /// for the given \p DataType and kind of access to \p Ptr.
777 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
778 return TTI->isLegalMaskedStore(DataType, isConsecutivePtr(Ptr));
780 /// Returns true if the target machine supports masked load operation
781 /// for the given \p DataType and kind of access to \p Ptr.
782 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
783 return TTI->isLegalMaskedLoad(DataType, isConsecutivePtr(Ptr));
785 /// Returns true if vector representation of the instruction \p I
787 bool isMaskRequired(const Instruction* I) {
788 return (MaskedOp.count(I) != 0);
790 unsigned getNumStores() const {
791 return LAI->getNumStores();
793 unsigned getNumLoads() const {
794 return LAI->getNumLoads();
796 unsigned getNumPredStores() const {
797 return NumPredStores;
800 /// Check if a single basic block loop is vectorizable.
801 /// At this point we know that this is a loop with a constant trip count
802 /// and we only need to check individual instructions.
803 bool canVectorizeInstrs();
805 /// When we vectorize loops we may change the order in which
806 /// we read and write from memory. This method checks if it is
807 /// legal to vectorize the code, considering only memory constrains.
808 /// Returns true if the loop is vectorizable
809 bool canVectorizeMemory();
811 /// Return true if we can vectorize this loop using the IF-conversion
813 bool canVectorizeWithIfConvert();
815 /// Collect the variables that need to stay uniform after vectorization.
816 void collectLoopUniforms();
818 /// Return true if all of the instructions in the block can be speculatively
819 /// executed. \p SafePtrs is a list of addresses that are known to be legal
820 /// and we know that we can read from them without segfault.
821 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
823 /// Returns True, if 'Phi' is the kind of reduction variable for type
824 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
825 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
826 /// Returns a struct describing if the instruction 'I' can be a reduction
827 /// variable of type 'Kind'. If the reduction is a min/max pattern of
828 /// select(icmp()) this function advances the instruction pointer 'I' from the
829 /// compare instruction to the select instruction and stores this pointer in
830 /// 'PatternLastInst' member of the returned struct.
831 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
832 ReductionInstDesc &Desc);
833 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
834 /// pattern corresponding to a min(X, Y) or max(X, Y).
835 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
836 ReductionInstDesc &Prev);
837 /// Returns the induction kind of Phi and record the step. This function may
838 /// return NoInduction if the PHI is not an induction variable.
839 InductionKind isInductionVariable(PHINode *Phi, ConstantInt *&StepValue);
841 /// \brief Collect memory access with loop invariant strides.
843 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
845 void collectStridedAccess(Value *LoadOrStoreInst);
847 /// Report an analysis message to assist the user in diagnosing loops that are
848 /// not vectorized. These are handled as LoopAccessReport rather than
849 /// VectorizationReport because the << operator of VectorizationReport returns
850 /// LoopAccessReport.
851 void emitAnalysis(const LoopAccessReport &Message) {
852 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, LV_NAME);
855 unsigned NumPredStores;
857 /// The loop that we evaluate.
861 /// Target Library Info.
862 TargetLibraryInfo *TLI;
864 Function *TheFunction;
865 /// Target Transform Info
866 const TargetTransformInfo *TTI;
869 // LoopAccess analysis.
870 LoopAccessAnalysis *LAA;
871 // And the loop-accesses info corresponding to this loop. This pointer is
872 // null until canVectorizeMemory sets it up.
873 const LoopAccessInfo *LAI;
875 // --- vectorization state --- //
877 /// Holds the integer induction variable. This is the counter of the
880 /// Holds the reduction variables.
881 ReductionList Reductions;
882 /// Holds all of the induction variables that we found in the loop.
883 /// Notice that inductions don't need to start at zero and that induction
884 /// variables can be pointers.
885 InductionList Inductions;
886 /// Holds the widest induction type encountered.
889 /// Allowed outside users. This holds the reduction
890 /// vars which can be accessed from outside the loop.
891 SmallPtrSet<Value*, 4> AllowedExit;
892 /// This set holds the variables which are known to be uniform after
894 SmallPtrSet<Instruction*, 4> Uniforms;
896 /// Can we assume the absence of NaNs.
897 bool HasFunNoNaNAttr;
899 ValueToValueMap Strides;
900 SmallPtrSet<Value *, 8> StrideSet;
902 /// While vectorizing these instructions we have to generate a
903 /// call to the appropriate masked intrinsic
904 SmallPtrSet<const Instruction*, 8> MaskedOp;
907 /// LoopVectorizationCostModel - estimates the expected speedups due to
909 /// In many cases vectorization is not profitable. This can happen because of
910 /// a number of reasons. In this class we mainly attempt to predict the
911 /// expected speedup/slowdowns due to the supported instruction set. We use the
912 /// TargetTransformInfo to query the different backends for the cost of
913 /// different operations.
914 class LoopVectorizationCostModel {
916 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
917 LoopVectorizationLegality *Legal,
918 const TargetTransformInfo &TTI,
919 const TargetLibraryInfo *TLI, AssumptionCache *AC,
920 const Function *F, const LoopVectorizeHints *Hints)
921 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), TLI(TLI),
922 TheFunction(F), Hints(Hints) {
923 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
926 /// Information about vectorization costs
927 struct VectorizationFactor {
928 unsigned Width; // Vector width with best cost
929 unsigned Cost; // Cost of the loop with that width
931 /// \return The most profitable vectorization factor and the cost of that VF.
932 /// This method checks every power of two up to VF. If UserVF is not ZERO
933 /// then this vectorization factor will be selected if vectorization is
935 VectorizationFactor selectVectorizationFactor(bool OptForSize);
937 /// \return The size (in bits) of the widest type in the code that
938 /// needs to be vectorized. We ignore values that remain scalar such as
939 /// 64 bit loop indices.
940 unsigned getWidestType();
942 /// \return The most profitable unroll factor.
943 /// If UserUF is non-zero then this method finds the best unroll-factor
944 /// based on register pressure and other parameters.
945 /// VF and LoopCost are the selected vectorization factor and the cost of the
947 unsigned selectUnrollFactor(bool OptForSize, unsigned VF, unsigned LoopCost);
949 /// \brief A struct that represents some properties of the register usage
951 struct RegisterUsage {
952 /// Holds the number of loop invariant values that are used in the loop.
953 unsigned LoopInvariantRegs;
954 /// Holds the maximum number of concurrent live intervals in the loop.
955 unsigned MaxLocalUsers;
956 /// Holds the number of instructions in the loop.
957 unsigned NumInstructions;
960 /// \return information about the register usage of the loop.
961 RegisterUsage calculateRegisterUsage();
964 /// Returns the expected execution cost. The unit of the cost does
965 /// not matter because we use the 'cost' units to compare different
966 /// vector widths. The cost that is returned is *not* normalized by
967 /// the factor width.
968 unsigned expectedCost(unsigned VF);
970 /// Returns the execution time cost of an instruction for a given vector
971 /// width. Vector width of one means scalar.
972 unsigned getInstructionCost(Instruction *I, unsigned VF);
974 /// Returns whether the instruction is a load or store and will be a emitted
975 /// as a vector operation.
976 bool isConsecutiveLoadOrStore(Instruction *I);
978 /// Report an analysis message to assist the user in diagnosing loops that are
979 /// not vectorized. These are handled as LoopAccessReport rather than
980 /// VectorizationReport because the << operator of VectorizationReport returns
981 /// LoopAccessReport.
982 void emitAnalysis(const LoopAccessReport &Message) {
983 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, LV_NAME);
986 /// Values used only by @llvm.assume calls.
987 SmallPtrSet<const Value *, 32> EphValues;
989 /// The loop that we evaluate.
993 /// Loop Info analysis.
995 /// Vectorization legality.
996 LoopVectorizationLegality *Legal;
997 /// Vector target information.
998 const TargetTransformInfo &TTI;
999 /// Target Library Info.
1000 const TargetLibraryInfo *TLI;
1001 const Function *TheFunction;
1002 // Loop Vectorize Hint.
1003 const LoopVectorizeHints *Hints;
1006 /// Utility class for getting and setting loop vectorizer hints in the form
1007 /// of loop metadata.
1008 /// This class keeps a number of loop annotations locally (as member variables)
1009 /// and can, upon request, write them back as metadata on the loop. It will
1010 /// initially scan the loop for existing metadata, and will update the local
1011 /// values based on information in the loop.
1012 /// We cannot write all values to metadata, as the mere presence of some info,
1013 /// for example 'force', means a decision has been made. So, we need to be
1014 /// careful NOT to add them if the user hasn't specifically asked so.
1015 class LoopVectorizeHints {
1022 /// Hint - associates name and validation with the hint value.
1025 unsigned Value; // This may have to change for non-numeric values.
1028 Hint(const char * Name, unsigned Value, HintKind Kind)
1029 : Name(Name), Value(Value), Kind(Kind) { }
1031 bool validate(unsigned Val) {
1034 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
1036 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1044 /// Vectorization width.
1046 /// Vectorization interleave factor.
1048 /// Vectorization forced
1051 /// Return the loop metadata prefix.
1052 static StringRef Prefix() { return "llvm.loop."; }
1056 FK_Undefined = -1, ///< Not selected.
1057 FK_Disabled = 0, ///< Forcing disabled.
1058 FK_Enabled = 1, ///< Forcing enabled.
1061 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
1062 : Width("vectorize.width", VectorizerParams::VectorizationFactor,
1064 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1065 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1067 // Populate values with existing loop metadata.
1068 getHintsFromMetadata();
1070 // force-vector-interleave overrides DisableInterleaving.
1071 if (VectorizerParams::isInterleaveForced())
1072 Interleave.Value = VectorizerParams::VectorizationInterleave;
1074 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
1075 << "LV: Interleaving disabled by the pass manager\n");
1078 /// Mark the loop L as already vectorized by setting the width to 1.
1079 void setAlreadyVectorized() {
1080 Width.Value = Interleave.Value = 1;
1081 Hint Hints[] = {Width, Interleave};
1082 writeHintsToMetadata(Hints);
1085 /// Dumps all the hint information.
1086 std::string emitRemark() const {
1087 VectorizationReport R;
1088 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1089 R << "vectorization is explicitly disabled";
1091 R << "use -Rpass-analysis=loop-vectorize for more info";
1092 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1093 R << " (Force=true";
1094 if (Width.Value != 0)
1095 R << ", Vector Width=" << Width.Value;
1096 if (Interleave.Value != 0)
1097 R << ", Interleave Count=" << Interleave.Value;
1105 unsigned getWidth() const { return Width.Value; }
1106 unsigned getInterleave() const { return Interleave.Value; }
1107 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1110 /// Find hints specified in the loop metadata and update local values.
1111 void getHintsFromMetadata() {
1112 MDNode *LoopID = TheLoop->getLoopID();
1116 // First operand should refer to the loop id itself.
1117 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1118 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1120 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1121 const MDString *S = nullptr;
1122 SmallVector<Metadata *, 4> Args;
1124 // The expected hint is either a MDString or a MDNode with the first
1125 // operand a MDString.
1126 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1127 if (!MD || MD->getNumOperands() == 0)
1129 S = dyn_cast<MDString>(MD->getOperand(0));
1130 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1131 Args.push_back(MD->getOperand(i));
1133 S = dyn_cast<MDString>(LoopID->getOperand(i));
1134 assert(Args.size() == 0 && "too many arguments for MDString");
1140 // Check if the hint starts with the loop metadata prefix.
1141 StringRef Name = S->getString();
1142 if (Args.size() == 1)
1143 setHint(Name, Args[0]);
1147 /// Checks string hint with one operand and set value if valid.
1148 void setHint(StringRef Name, Metadata *Arg) {
1149 if (!Name.startswith(Prefix()))
1151 Name = Name.substr(Prefix().size(), StringRef::npos);
1153 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1155 unsigned Val = C->getZExtValue();
1157 Hint *Hints[] = {&Width, &Interleave, &Force};
1158 for (auto H : Hints) {
1159 if (Name == H->Name) {
1160 if (H->validate(Val))
1163 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1169 /// Create a new hint from name / value pair.
1170 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1171 LLVMContext &Context = TheLoop->getHeader()->getContext();
1172 Metadata *MDs[] = {MDString::get(Context, Name),
1173 ConstantAsMetadata::get(
1174 ConstantInt::get(Type::getInt32Ty(Context), V))};
1175 return MDNode::get(Context, MDs);
1178 /// Matches metadata with hint name.
1179 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1180 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1184 for (auto H : HintTypes)
1185 if (Name->getString().endswith(H.Name))
1190 /// Sets current hints into loop metadata, keeping other values intact.
1191 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1192 if (HintTypes.size() == 0)
1195 // Reserve the first element to LoopID (see below).
1196 SmallVector<Metadata *, 4> MDs(1);
1197 // If the loop already has metadata, then ignore the existing operands.
1198 MDNode *LoopID = TheLoop->getLoopID();
1200 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1201 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1202 // If node in update list, ignore old value.
1203 if (!matchesHintMetadataName(Node, HintTypes))
1204 MDs.push_back(Node);
1208 // Now, add the missing hints.
1209 for (auto H : HintTypes)
1210 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1212 // Replace current metadata node with new one.
1213 LLVMContext &Context = TheLoop->getHeader()->getContext();
1214 MDNode *NewLoopID = MDNode::get(Context, MDs);
1215 // Set operand 0 to refer to the loop id itself.
1216 NewLoopID->replaceOperandWith(0, NewLoopID);
1218 TheLoop->setLoopID(NewLoopID);
1221 /// The loop these hints belong to.
1222 const Loop *TheLoop;
1225 static void emitMissedWarning(Function *F, Loop *L,
1226 const LoopVectorizeHints &LH) {
1227 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1228 L->getStartLoc(), LH.emitRemark());
1230 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1231 if (LH.getWidth() != 1)
1232 emitLoopVectorizeWarning(
1233 F->getContext(), *F, L->getStartLoc(),
1234 "failed explicitly specified loop vectorization");
1235 else if (LH.getInterleave() != 1)
1236 emitLoopInterleaveWarning(
1237 F->getContext(), *F, L->getStartLoc(),
1238 "failed explicitly specified loop interleaving");
1242 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1244 return V.push_back(&L);
1246 for (Loop *InnerL : L)
1247 addInnerLoop(*InnerL, V);
1250 /// The LoopVectorize Pass.
1251 struct LoopVectorize : public FunctionPass {
1252 /// Pass identification, replacement for typeid
1255 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1257 DisableUnrolling(NoUnrolling),
1258 AlwaysVectorize(AlwaysVectorize) {
1259 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1262 ScalarEvolution *SE;
1264 TargetTransformInfo *TTI;
1266 BlockFrequencyInfo *BFI;
1267 TargetLibraryInfo *TLI;
1269 AssumptionCache *AC;
1270 LoopAccessAnalysis *LAA;
1271 bool DisableUnrolling;
1272 bool AlwaysVectorize;
1274 BlockFrequency ColdEntryFreq;
1276 bool runOnFunction(Function &F) override {
1277 SE = &getAnalysis<ScalarEvolution>();
1278 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1279 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1280 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1281 BFI = &getAnalysis<BlockFrequencyInfo>();
1282 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1283 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1284 AA = &getAnalysis<AliasAnalysis>();
1285 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1286 LAA = &getAnalysis<LoopAccessAnalysis>();
1288 // Compute some weights outside of the loop over the loops. Compute this
1289 // using a BranchProbability to re-use its scaling math.
1290 const BranchProbability ColdProb(1, 5); // 20%
1291 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1293 // If the target claims to have no vector registers don't attempt
1295 if (!TTI->getNumberOfRegisters(true))
1298 // Build up a worklist of inner-loops to vectorize. This is necessary as
1299 // the act of vectorizing or partially unrolling a loop creates new loops
1300 // and can invalidate iterators across the loops.
1301 SmallVector<Loop *, 8> Worklist;
1304 addInnerLoop(*L, Worklist);
1306 LoopsAnalyzed += Worklist.size();
1308 // Now walk the identified inner loops.
1309 bool Changed = false;
1310 while (!Worklist.empty())
1311 Changed |= processLoop(Worklist.pop_back_val());
1313 // Process each loop nest in the function.
1317 static void AddRuntimeUnrollDisableMetaData(Loop *L) {
1318 SmallVector<Metadata *, 4> MDs;
1319 // Reserve first location for self reference to the LoopID metadata node.
1320 MDs.push_back(nullptr);
1321 bool IsUnrollMetadata = false;
1322 MDNode *LoopID = L->getLoopID();
1324 // First find existing loop unrolling disable metadata.
1325 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1326 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
1328 const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
1330 S && S->getString().startswith("llvm.loop.unroll.disable");
1332 MDs.push_back(LoopID->getOperand(i));
1336 if (!IsUnrollMetadata) {
1337 // Add runtime unroll disable metadata.
1338 LLVMContext &Context = L->getHeader()->getContext();
1339 SmallVector<Metadata *, 1> DisableOperands;
1340 DisableOperands.push_back(
1341 MDString::get(Context, "llvm.loop.unroll.runtime.disable"));
1342 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
1343 MDs.push_back(DisableNode);
1344 MDNode *NewLoopID = MDNode::get(Context, MDs);
1345 // Set operand 0 to refer to the loop id itself.
1346 NewLoopID->replaceOperandWith(0, NewLoopID);
1347 L->setLoopID(NewLoopID);
1351 bool processLoop(Loop *L) {
1352 assert(L->empty() && "Only process inner loops.");
1355 const std::string DebugLocStr = getDebugLocString(L);
1358 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1359 << L->getHeader()->getParent()->getName() << "\" from "
1360 << DebugLocStr << "\n");
1362 LoopVectorizeHints Hints(L, DisableUnrolling);
1364 DEBUG(dbgs() << "LV: Loop hints:"
1366 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1368 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1370 : "?")) << " width=" << Hints.getWidth()
1371 << " unroll=" << Hints.getInterleave() << "\n");
1373 // Function containing loop
1374 Function *F = L->getHeader()->getParent();
1376 // Looking at the diagnostic output is the only way to determine if a loop
1377 // was vectorized (other than looking at the IR or machine code), so it
1378 // is important to generate an optimization remark for each loop. Most of
1379 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1380 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1381 // less verbose reporting vectorized loops and unvectorized loops that may
1382 // benefit from vectorization, respectively.
1384 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1385 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1386 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1387 L->getStartLoc(), Hints.emitRemark());
1391 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1392 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1393 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1394 L->getStartLoc(), Hints.emitRemark());
1398 if (Hints.getWidth() == 1 && Hints.getInterleave() == 1) {
1399 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1400 emitOptimizationRemarkAnalysis(
1401 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1402 "loop not vectorized: vector width and interleave count are "
1403 "explicitly set to 1");
1407 // Check the loop for a trip count threshold:
1408 // do not vectorize loops with a tiny trip count.
1409 const unsigned TC = SE->getSmallConstantTripCount(L);
1410 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1411 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1412 << "This loop is not worth vectorizing.");
1413 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1414 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1416 DEBUG(dbgs() << "\n");
1417 emitOptimizationRemarkAnalysis(
1418 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1419 "vectorization is not beneficial and is not explicitly forced");
1424 // Check if it is legal to vectorize the loop.
1425 LoopVectorizationLegality LVL(L, SE, DT, TLI, AA, F, TTI, LAA);
1426 if (!LVL.canVectorize()) {
1427 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1428 emitMissedWarning(F, L, Hints);
1432 // Use the cost model.
1433 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, TLI, AC, F, &Hints);
1435 // Check the function attributes to find out if this function should be
1436 // optimized for size.
1437 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1438 F->hasFnAttribute(Attribute::OptimizeForSize);
1440 // Compute the weighted frequency of this loop being executed and see if it
1441 // is less than 20% of the function entry baseline frequency. Note that we
1442 // always have a canonical loop here because we think we *can* vectoriez.
1443 // FIXME: This is hidden behind a flag due to pervasive problems with
1444 // exactly what block frequency models.
1445 if (LoopVectorizeWithBlockFrequency) {
1446 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1447 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1448 LoopEntryFreq < ColdEntryFreq)
1452 // Check the function attributes to see if implicit floats are allowed.a
1453 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1454 // an integer loop and the vector instructions selected are purely integer
1455 // vector instructions?
1456 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1457 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1458 "attribute is used.\n");
1459 emitOptimizationRemarkAnalysis(
1460 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1461 "loop not vectorized due to NoImplicitFloat attribute");
1462 emitMissedWarning(F, L, Hints);
1466 // Select the optimal vectorization factor.
1467 const LoopVectorizationCostModel::VectorizationFactor VF =
1468 CM.selectVectorizationFactor(OptForSize);
1470 // Select the unroll factor.
1472 CM.selectUnrollFactor(OptForSize, VF.Width, VF.Cost);
1474 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1475 << DebugLocStr << '\n');
1476 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1478 if (VF.Width == 1) {
1479 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
1482 emitOptimizationRemarkAnalysis(
1483 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1484 "not beneficial to vectorize and user disabled interleaving");
1487 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1489 // Report the unrolling decision.
1490 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1491 Twine("unrolled with interleaving factor " +
1493 " (vectorization not beneficial)"));
1495 // We decided not to vectorize, but we may want to unroll.
1497 InnerLoopUnroller Unroller(L, SE, LI, DT, TLI, TTI, UF);
1498 Unroller.vectorize(&LVL);
1500 // If we decided that it is *legal* to vectorize the loop then do it.
1501 InnerLoopVectorizer LB(L, SE, LI, DT, TLI, TTI, VF.Width, UF);
1505 // Add metadata to disable runtime unrolling scalar loop when there's no
1506 // runtime check about strides and memory. Because at this situation,
1507 // scalar loop is rarely used not worthy to be unrolled.
1508 if (!LB.IsSafetyChecksAdded())
1509 AddRuntimeUnrollDisableMetaData(L);
1511 // Report the vectorization decision.
1512 emitOptimizationRemark(
1513 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1514 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1515 ", unrolling interleave factor: " + Twine(UF) + ")");
1518 // Mark the loop as already vectorized to avoid vectorizing again.
1519 Hints.setAlreadyVectorized();
1521 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1525 void getAnalysisUsage(AnalysisUsage &AU) const override {
1526 AU.addRequired<AssumptionCacheTracker>();
1527 AU.addRequiredID(LoopSimplifyID);
1528 AU.addRequiredID(LCSSAID);
1529 AU.addRequired<BlockFrequencyInfo>();
1530 AU.addRequired<DominatorTreeWrapperPass>();
1531 AU.addRequired<LoopInfoWrapperPass>();
1532 AU.addRequired<ScalarEvolution>();
1533 AU.addRequired<TargetTransformInfoWrapperPass>();
1534 AU.addRequired<AliasAnalysis>();
1535 AU.addRequired<LoopAccessAnalysis>();
1536 AU.addPreserved<LoopInfoWrapperPass>();
1537 AU.addPreserved<DominatorTreeWrapperPass>();
1538 AU.addPreserved<AliasAnalysis>();
1543 } // end anonymous namespace
1545 //===----------------------------------------------------------------------===//
1546 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1547 // LoopVectorizationCostModel.
1548 //===----------------------------------------------------------------------===//
1550 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1551 // We need to place the broadcast of invariant variables outside the loop.
1552 Instruction *Instr = dyn_cast<Instruction>(V);
1554 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1555 Instr->getParent()) != LoopVectorBody.end());
1556 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1558 // Place the code for broadcasting invariant variables in the new preheader.
1559 IRBuilder<>::InsertPointGuard Guard(Builder);
1561 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1563 // Broadcast the scalar into all locations in the vector.
1564 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1569 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
1571 assert(Val->getType()->isVectorTy() && "Must be a vector");
1572 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1573 "Elem must be an integer");
1574 assert(Step->getType() == Val->getType()->getScalarType() &&
1575 "Step has wrong type");
1576 // Create the types.
1577 Type *ITy = Val->getType()->getScalarType();
1578 VectorType *Ty = cast<VectorType>(Val->getType());
1579 int VLen = Ty->getNumElements();
1580 SmallVector<Constant*, 8> Indices;
1582 // Create a vector of consecutive numbers from zero to VF.
1583 for (int i = 0; i < VLen; ++i)
1584 Indices.push_back(ConstantInt::get(ITy, StartIdx + i));
1586 // Add the consecutive indices to the vector value.
1587 Constant *Cv = ConstantVector::get(Indices);
1588 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1589 Step = Builder.CreateVectorSplat(VLen, Step);
1590 assert(Step->getType() == Val->getType() && "Invalid step vec");
1591 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
1592 // which can be found from the original scalar operations.
1593 Step = Builder.CreateMul(Cv, Step);
1594 return Builder.CreateAdd(Val, Step, "induction");
1597 /// \brief Find the operand of the GEP that should be checked for consecutive
1598 /// stores. This ignores trailing indices that have no effect on the final
1600 static unsigned getGEPInductionOperand(const GetElementPtrInst *Gep) {
1601 const DataLayout &DL = Gep->getModule()->getDataLayout();
1602 unsigned LastOperand = Gep->getNumOperands() - 1;
1603 unsigned GEPAllocSize = DL.getTypeAllocSize(
1604 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1606 // Walk backwards and try to peel off zeros.
1607 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1608 // Find the type we're currently indexing into.
1609 gep_type_iterator GEPTI = gep_type_begin(Gep);
1610 std::advance(GEPTI, LastOperand - 1);
1612 // If it's a type with the same allocation size as the result of the GEP we
1613 // can peel off the zero index.
1614 if (DL.getTypeAllocSize(*GEPTI) != GEPAllocSize)
1622 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1623 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1624 // Make sure that the pointer does not point to structs.
1625 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1628 // If this value is a pointer induction variable we know it is consecutive.
1629 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1630 if (Phi && Inductions.count(Phi)) {
1631 InductionInfo II = Inductions[Phi];
1632 return II.getConsecutiveDirection();
1635 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1639 unsigned NumOperands = Gep->getNumOperands();
1640 Value *GpPtr = Gep->getPointerOperand();
1641 // If this GEP value is a consecutive pointer induction variable and all of
1642 // the indices are constant then we know it is consecutive. We can
1643 Phi = dyn_cast<PHINode>(GpPtr);
1644 if (Phi && Inductions.count(Phi)) {
1646 // Make sure that the pointer does not point to structs.
1647 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1648 if (GepPtrType->getElementType()->isAggregateType())
1651 // Make sure that all of the index operands are loop invariant.
1652 for (unsigned i = 1; i < NumOperands; ++i)
1653 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1656 InductionInfo II = Inductions[Phi];
1657 return II.getConsecutiveDirection();
1660 unsigned InductionOperand = getGEPInductionOperand(Gep);
1662 // Check that all of the gep indices are uniform except for our induction
1664 for (unsigned i = 0; i != NumOperands; ++i)
1665 if (i != InductionOperand &&
1666 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1669 // We can emit wide load/stores only if the last non-zero index is the
1670 // induction variable.
1671 const SCEV *Last = nullptr;
1672 if (!Strides.count(Gep))
1673 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1675 // Because of the multiplication by a stride we can have a s/zext cast.
1676 // We are going to replace this stride by 1 so the cast is safe to ignore.
1678 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1679 // %0 = trunc i64 %indvars.iv to i32
1680 // %mul = mul i32 %0, %Stride1
1681 // %idxprom = zext i32 %mul to i64 << Safe cast.
1682 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1684 Last = replaceSymbolicStrideSCEV(SE, Strides,
1685 Gep->getOperand(InductionOperand), Gep);
1686 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1688 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1692 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1693 const SCEV *Step = AR->getStepRecurrence(*SE);
1695 // The memory is consecutive because the last index is consecutive
1696 // and all other indices are loop invariant.
1699 if (Step->isAllOnesValue())
1706 bool LoopVectorizationLegality::isUniform(Value *V) {
1707 return LAI->isUniform(V);
1710 InnerLoopVectorizer::VectorParts&
1711 InnerLoopVectorizer::getVectorValue(Value *V) {
1712 assert(V != Induction && "The new induction variable should not be used.");
1713 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1715 // If we have a stride that is replaced by one, do it here.
1716 if (Legal->hasStride(V))
1717 V = ConstantInt::get(V->getType(), 1);
1719 // If we have this scalar in the map, return it.
1720 if (WidenMap.has(V))
1721 return WidenMap.get(V);
1723 // If this scalar is unknown, assume that it is a constant or that it is
1724 // loop invariant. Broadcast V and save the value for future uses.
1725 Value *B = getBroadcastInstrs(V);
1726 return WidenMap.splat(V, B);
1729 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1730 assert(Vec->getType()->isVectorTy() && "Invalid type");
1731 SmallVector<Constant*, 8> ShuffleMask;
1732 for (unsigned i = 0; i < VF; ++i)
1733 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1735 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1736 ConstantVector::get(ShuffleMask),
1740 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1741 // Attempt to issue a wide load.
1742 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1743 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1745 assert((LI || SI) && "Invalid Load/Store instruction");
1747 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1748 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1749 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1750 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1751 // An alignment of 0 means target abi alignment. We need to use the scalar's
1752 // target abi alignment in such a case.
1753 const DataLayout &DL = Instr->getModule()->getDataLayout();
1755 Alignment = DL.getABITypeAlignment(ScalarDataTy);
1756 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1757 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ScalarDataTy);
1758 unsigned VectorElementSize = DL.getTypeStoreSize(DataTy) / VF;
1760 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
1761 !Legal->isMaskRequired(SI))
1762 return scalarizeInstruction(Instr, true);
1764 if (ScalarAllocatedSize != VectorElementSize)
1765 return scalarizeInstruction(Instr);
1767 // If the pointer is loop invariant or if it is non-consecutive,
1768 // scalarize the load.
1769 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1770 bool Reverse = ConsecutiveStride < 0;
1771 bool UniformLoad = LI && Legal->isUniform(Ptr);
1772 if (!ConsecutiveStride || UniformLoad)
1773 return scalarizeInstruction(Instr);
1775 Constant *Zero = Builder.getInt32(0);
1776 VectorParts &Entry = WidenMap.get(Instr);
1778 // Handle consecutive loads/stores.
1779 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1780 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1781 setDebugLocFromInst(Builder, Gep);
1782 Value *PtrOperand = Gep->getPointerOperand();
1783 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1784 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1786 // Create the new GEP with the new induction variable.
1787 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1788 Gep2->setOperand(0, FirstBasePtr);
1789 Gep2->setName("gep.indvar.base");
1790 Ptr = Builder.Insert(Gep2);
1792 setDebugLocFromInst(Builder, Gep);
1793 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1794 OrigLoop) && "Base ptr must be invariant");
1796 // The last index does not have to be the induction. It can be
1797 // consecutive and be a function of the index. For example A[I+1];
1798 unsigned NumOperands = Gep->getNumOperands();
1799 unsigned InductionOperand = getGEPInductionOperand(Gep);
1800 // Create the new GEP with the new induction variable.
1801 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1803 for (unsigned i = 0; i < NumOperands; ++i) {
1804 Value *GepOperand = Gep->getOperand(i);
1805 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1807 // Update last index or loop invariant instruction anchored in loop.
1808 if (i == InductionOperand ||
1809 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1810 assert((i == InductionOperand ||
1811 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1812 "Must be last index or loop invariant");
1814 VectorParts &GEPParts = getVectorValue(GepOperand);
1815 Value *Index = GEPParts[0];
1816 Index = Builder.CreateExtractElement(Index, Zero);
1817 Gep2->setOperand(i, Index);
1818 Gep2->setName("gep.indvar.idx");
1821 Ptr = Builder.Insert(Gep2);
1823 // Use the induction element ptr.
1824 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1825 setDebugLocFromInst(Builder, Ptr);
1826 VectorParts &PtrVal = getVectorValue(Ptr);
1827 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1830 VectorParts Mask = createBlockInMask(Instr->getParent());
1833 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1834 "We do not allow storing to uniform addresses");
1835 setDebugLocFromInst(Builder, SI);
1836 // We don't want to update the value in the map as it might be used in
1837 // another expression. So don't use a reference type for "StoredVal".
1838 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1840 for (unsigned Part = 0; Part < UF; ++Part) {
1841 // Calculate the pointer for the specific unroll-part.
1842 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1845 // If we store to reverse consecutive memory locations then we need
1846 // to reverse the order of elements in the stored value.
1847 StoredVal[Part] = reverseVector(StoredVal[Part]);
1848 // If the address is consecutive but reversed, then the
1849 // wide store needs to start at the last vector element.
1850 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1851 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1852 Mask[Part] = reverseVector(Mask[Part]);
1855 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1856 DataTy->getPointerTo(AddressSpace));
1859 if (Legal->isMaskRequired(SI))
1860 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
1863 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
1864 propagateMetadata(NewSI, SI);
1870 assert(LI && "Must have a load instruction");
1871 setDebugLocFromInst(Builder, LI);
1872 for (unsigned Part = 0; Part < UF; ++Part) {
1873 // Calculate the pointer for the specific unroll-part.
1874 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1877 // If the address is consecutive but reversed, then the
1878 // wide load needs to start at the last vector element.
1879 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1880 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1881 Mask[Part] = reverseVector(Mask[Part]);
1885 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1886 DataTy->getPointerTo(AddressSpace));
1887 if (Legal->isMaskRequired(LI))
1888 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
1889 UndefValue::get(DataTy),
1890 "wide.masked.load");
1892 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
1893 propagateMetadata(NewLI, LI);
1894 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
1898 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1899 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1900 // Holds vector parameters or scalars, in case of uniform vals.
1901 SmallVector<VectorParts, 4> Params;
1903 setDebugLocFromInst(Builder, Instr);
1905 // Find all of the vectorized parameters.
1906 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1907 Value *SrcOp = Instr->getOperand(op);
1909 // If we are accessing the old induction variable, use the new one.
1910 if (SrcOp == OldInduction) {
1911 Params.push_back(getVectorValue(SrcOp));
1915 // Try using previously calculated values.
1916 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1918 // If the src is an instruction that appeared earlier in the basic block
1919 // then it should already be vectorized.
1920 if (SrcInst && OrigLoop->contains(SrcInst)) {
1921 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1922 // The parameter is a vector value from earlier.
1923 Params.push_back(WidenMap.get(SrcInst));
1925 // The parameter is a scalar from outside the loop. Maybe even a constant.
1926 VectorParts Scalars;
1927 Scalars.append(UF, SrcOp);
1928 Params.push_back(Scalars);
1932 assert(Params.size() == Instr->getNumOperands() &&
1933 "Invalid number of operands");
1935 // Does this instruction return a value ?
1936 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1938 Value *UndefVec = IsVoidRetTy ? nullptr :
1939 UndefValue::get(VectorType::get(Instr->getType(), VF));
1940 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1941 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1943 Instruction *InsertPt = Builder.GetInsertPoint();
1944 BasicBlock *IfBlock = Builder.GetInsertBlock();
1945 BasicBlock *CondBlock = nullptr;
1948 Loop *VectorLp = nullptr;
1949 if (IfPredicateStore) {
1950 assert(Instr->getParent()->getSinglePredecessor() &&
1951 "Only support single predecessor blocks");
1952 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1953 Instr->getParent());
1954 VectorLp = LI->getLoopFor(IfBlock);
1955 assert(VectorLp && "Must have a loop for this block");
1958 // For each vector unroll 'part':
1959 for (unsigned Part = 0; Part < UF; ++Part) {
1960 // For each scalar that we create:
1961 for (unsigned Width = 0; Width < VF; ++Width) {
1964 Value *Cmp = nullptr;
1965 if (IfPredicateStore) {
1966 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1967 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1968 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1969 LoopVectorBody.push_back(CondBlock);
1970 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
1971 // Update Builder with newly created basic block.
1972 Builder.SetInsertPoint(InsertPt);
1975 Instruction *Cloned = Instr->clone();
1977 Cloned->setName(Instr->getName() + ".cloned");
1978 // Replace the operands of the cloned instructions with extracted scalars.
1979 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1980 Value *Op = Params[op][Part];
1981 // Param is a vector. Need to extract the right lane.
1982 if (Op->getType()->isVectorTy())
1983 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1984 Cloned->setOperand(op, Op);
1987 // Place the cloned scalar in the new loop.
1988 Builder.Insert(Cloned);
1990 // If the original scalar returns a value we need to place it in a vector
1991 // so that future users will be able to use it.
1993 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1994 Builder.getInt32(Width));
1996 if (IfPredicateStore) {
1997 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1998 LoopVectorBody.push_back(NewIfBlock);
1999 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
2000 Builder.SetInsertPoint(InsertPt);
2001 Instruction *OldBr = IfBlock->getTerminator();
2002 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2003 OldBr->eraseFromParent();
2004 IfBlock = NewIfBlock;
2010 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
2014 if (Instruction *I = dyn_cast<Instruction>(V))
2015 return I->getParent() == Loc->getParent() ? I : nullptr;
2019 std::pair<Instruction *, Instruction *>
2020 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
2021 Instruction *tnullptr = nullptr;
2022 if (!Legal->mustCheckStrides())
2023 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2025 IRBuilder<> ChkBuilder(Loc);
2028 Value *Check = nullptr;
2029 Instruction *FirstInst = nullptr;
2030 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
2031 SE = Legal->strides_end();
2033 Value *Ptr = stripIntegerCast(*SI);
2034 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
2036 // Store the first instruction we create.
2037 FirstInst = getFirstInst(FirstInst, C, Loc);
2039 Check = ChkBuilder.CreateOr(Check, C);
2044 // We have to do this trickery because the IRBuilder might fold the check to a
2045 // constant expression in which case there is no Instruction anchored in a
2047 LLVMContext &Ctx = Loc->getContext();
2048 Instruction *TheCheck =
2049 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2050 ChkBuilder.Insert(TheCheck, "stride.not.one");
2051 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2053 return std::make_pair(FirstInst, TheCheck);
2056 void InnerLoopVectorizer::createEmptyLoop() {
2058 In this function we generate a new loop. The new loop will contain
2059 the vectorized instructions while the old loop will continue to run the
2062 [ ] <-- Back-edge taken count overflow check.
2065 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2068 || [ ] <-- vector pre header.
2072 || [ ]_| <-- vector loop.
2075 | >[ ] <--- middle-block.
2078 -|- >[ ] <--- new preheader.
2082 | [ ]_| <-- old scalar loop to handle remainder.
2085 >[ ] <-- exit block.
2089 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2090 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2091 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2092 assert(BypassBlock && "Invalid loop structure");
2093 assert(ExitBlock && "Must have an exit block");
2095 // Some loops have a single integer induction variable, while other loops
2096 // don't. One example is c++ iterators that often have multiple pointer
2097 // induction variables. In the code below we also support a case where we
2098 // don't have a single induction variable.
2099 OldInduction = Legal->getInduction();
2100 Type *IdxTy = Legal->getWidestInductionType();
2102 // Find the loop boundaries.
2103 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2104 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2106 // The exit count might have the type of i64 while the phi is i32. This can
2107 // happen if we have an induction variable that is sign extended before the
2108 // compare. The only way that we get a backedge taken count is that the
2109 // induction variable was signed and as such will not overflow. In such a case
2110 // truncation is legal.
2111 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2112 IdxTy->getPrimitiveSizeInBits())
2113 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2115 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2116 // Get the total trip count from the count by adding 1.
2117 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2118 SE->getConstant(BackedgeTakeCount->getType(), 1));
2120 const DataLayout &DL = OldBasicBlock->getModule()->getDataLayout();
2122 // Expand the trip count and place the new instructions in the preheader.
2123 // Notice that the pre-header does not change, only the loop body.
2124 SCEVExpander Exp(*SE, DL, "induction");
2126 // We need to test whether the backedge-taken count is uint##_max. Adding one
2127 // to it will cause overflow and an incorrect loop trip count in the vector
2128 // body. In case of overflow we want to directly jump to the scalar remainder
2130 Value *BackedgeCount =
2131 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2132 BypassBlock->getTerminator());
2133 if (BackedgeCount->getType()->isPointerTy())
2134 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2135 "backedge.ptrcnt.to.int",
2136 BypassBlock->getTerminator());
2137 Instruction *CheckBCOverflow =
2138 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2139 Constant::getAllOnesValue(BackedgeCount->getType()),
2140 "backedge.overflow", BypassBlock->getTerminator());
2142 // The loop index does not have to start at Zero. Find the original start
2143 // value from the induction PHI node. If we don't have an induction variable
2144 // then we know that it starts at zero.
2145 Builder.SetInsertPoint(BypassBlock->getTerminator());
2146 Value *StartIdx = ExtendedIdx = OldInduction ?
2147 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2149 ConstantInt::get(IdxTy, 0);
2151 // We need an instruction to anchor the overflow check on. StartIdx needs to
2152 // be defined before the overflow check branch. Because the scalar preheader
2153 // is going to merge the start index and so the overflow branch block needs to
2154 // contain a definition of the start index.
2155 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2156 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2157 BypassBlock->getTerminator());
2159 // Count holds the overall loop count (N).
2160 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2161 BypassBlock->getTerminator());
2163 LoopBypassBlocks.push_back(BypassBlock);
2165 // Split the single block loop into the two loop structure described above.
2166 BasicBlock *VectorPH =
2167 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2168 BasicBlock *VecBody =
2169 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2170 BasicBlock *MiddleBlock =
2171 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2172 BasicBlock *ScalarPH =
2173 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2175 // Create and register the new vector loop.
2176 Loop* Lp = new Loop();
2177 Loop *ParentLoop = OrigLoop->getParentLoop();
2179 // Insert the new loop into the loop nest and register the new basic blocks
2180 // before calling any utilities such as SCEV that require valid LoopInfo.
2182 ParentLoop->addChildLoop(Lp);
2183 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
2184 ParentLoop->addBasicBlockToLoop(VectorPH, *LI);
2185 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
2187 LI->addTopLevelLoop(Lp);
2189 Lp->addBasicBlockToLoop(VecBody, *LI);
2191 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2193 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2195 // Generate the induction variable.
2196 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2197 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2198 // The loop step is equal to the vectorization factor (num of SIMD elements)
2199 // times the unroll factor (num of SIMD instructions).
2200 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2202 // This is the IR builder that we use to add all of the logic for bypassing
2203 // the new vector loop.
2204 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2205 setDebugLocFromInst(BypassBuilder,
2206 getDebugLocFromInstOrOperands(OldInduction));
2208 // We may need to extend the index in case there is a type mismatch.
2209 // We know that the count starts at zero and does not overflow.
2210 if (Count->getType() != IdxTy) {
2211 // The exit count can be of pointer type. Convert it to the correct
2213 if (ExitCount->getType()->isPointerTy())
2214 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2216 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2219 // Add the start index to the loop count to get the new end index.
2220 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2222 // Now we need to generate the expression for N - (N % VF), which is
2223 // the part that the vectorized body will execute.
2224 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2225 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2226 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2227 "end.idx.rnd.down");
2229 // Now, compare the new count to zero. If it is zero skip the vector loop and
2230 // jump to the scalar loop.
2232 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2234 BasicBlock *LastBypassBlock = BypassBlock;
2236 // Generate code to check that the loops trip count that we computed by adding
2237 // one to the backedge-taken count will not overflow.
2239 auto PastOverflowCheck =
2240 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2241 BasicBlock *CheckBlock =
2242 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2244 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2245 LoopBypassBlocks.push_back(CheckBlock);
2246 Instruction *OldTerm = LastBypassBlock->getTerminator();
2247 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2248 OldTerm->eraseFromParent();
2249 LastBypassBlock = CheckBlock;
2252 // Generate the code to check that the strides we assumed to be one are really
2253 // one. We want the new basic block to start at the first instruction in a
2254 // sequence of instructions that form a check.
2255 Instruction *StrideCheck;
2256 Instruction *FirstCheckInst;
2257 std::tie(FirstCheckInst, StrideCheck) =
2258 addStrideCheck(LastBypassBlock->getTerminator());
2260 AddedSafetyChecks = true;
2261 // Create a new block containing the stride check.
2262 BasicBlock *CheckBlock =
2263 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2265 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2266 LoopBypassBlocks.push_back(CheckBlock);
2268 // Replace the branch into the memory check block with a conditional branch
2269 // for the "few elements case".
2270 Instruction *OldTerm = LastBypassBlock->getTerminator();
2271 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2272 OldTerm->eraseFromParent();
2275 LastBypassBlock = CheckBlock;
2278 // Generate the code that checks in runtime if arrays overlap. We put the
2279 // checks into a separate block to make the more common case of few elements
2281 Instruction *MemRuntimeCheck;
2282 std::tie(FirstCheckInst, MemRuntimeCheck) =
2283 Legal->getLAI()->addRuntimeCheck(LastBypassBlock->getTerminator());
2284 if (MemRuntimeCheck) {
2285 AddedSafetyChecks = true;
2286 // Create a new block containing the memory check.
2287 BasicBlock *CheckBlock =
2288 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.memcheck");
2290 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2291 LoopBypassBlocks.push_back(CheckBlock);
2293 // Replace the branch into the memory check block with a conditional branch
2294 // for the "few elements case".
2295 Instruction *OldTerm = LastBypassBlock->getTerminator();
2296 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2297 OldTerm->eraseFromParent();
2299 Cmp = MemRuntimeCheck;
2300 LastBypassBlock = CheckBlock;
2303 LastBypassBlock->getTerminator()->eraseFromParent();
2304 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2307 // We are going to resume the execution of the scalar loop.
2308 // Go over all of the induction variables that we found and fix the
2309 // PHIs that are left in the scalar version of the loop.
2310 // The starting values of PHI nodes depend on the counter of the last
2311 // iteration in the vectorized loop.
2312 // If we come from a bypass edge then we need to start from the original
2315 // This variable saves the new starting index for the scalar loop.
2316 PHINode *ResumeIndex = nullptr;
2317 LoopVectorizationLegality::InductionList::iterator I, E;
2318 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2319 // Set builder to point to last bypass block.
2320 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2321 for (I = List->begin(), E = List->end(); I != E; ++I) {
2322 PHINode *OrigPhi = I->first;
2323 LoopVectorizationLegality::InductionInfo II = I->second;
2325 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2326 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2327 MiddleBlock->getTerminator());
2328 // We might have extended the type of the induction variable but we need a
2329 // truncated version for the scalar loop.
2330 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2331 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2332 MiddleBlock->getTerminator()) : nullptr;
2334 // Create phi nodes to merge from the backedge-taken check block.
2335 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2336 ScalarPH->getTerminator());
2337 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2339 PHINode *BCTruncResumeVal = nullptr;
2340 if (OrigPhi == OldInduction) {
2342 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2343 ScalarPH->getTerminator());
2344 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2347 Value *EndValue = nullptr;
2349 case LoopVectorizationLegality::IK_NoInduction:
2350 llvm_unreachable("Unknown induction");
2351 case LoopVectorizationLegality::IK_IntInduction: {
2352 // Handle the integer induction counter.
2353 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2355 // We have the canonical induction variable.
2356 if (OrigPhi == OldInduction) {
2357 // Create a truncated version of the resume value for the scalar loop,
2358 // we might have promoted the type to a larger width.
2360 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2361 // The new PHI merges the original incoming value, in case of a bypass,
2362 // or the value at the end of the vectorized loop.
2363 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2364 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2365 TruncResumeVal->addIncoming(EndValue, VecBody);
2367 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2369 // We know what the end value is.
2370 EndValue = IdxEndRoundDown;
2371 // We also know which PHI node holds it.
2372 ResumeIndex = ResumeVal;
2376 // Not the canonical induction variable - add the vector loop count to the
2378 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2379 II.StartValue->getType(),
2381 EndValue = II.transform(BypassBuilder, CRD);
2382 EndValue->setName("ind.end");
2385 case LoopVectorizationLegality::IK_PtrInduction: {
2386 EndValue = II.transform(BypassBuilder, CountRoundDown);
2387 EndValue->setName("ptr.ind.end");
2392 // The new PHI merges the original incoming value, in case of a bypass,
2393 // or the value at the end of the vectorized loop.
2394 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2395 if (OrigPhi == OldInduction)
2396 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2398 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2400 ResumeVal->addIncoming(EndValue, VecBody);
2402 // Fix the scalar body counter (PHI node).
2403 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2405 // The old induction's phi node in the scalar body needs the truncated
2407 if (OrigPhi == OldInduction) {
2408 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2409 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2411 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2412 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2416 // If we are generating a new induction variable then we also need to
2417 // generate the code that calculates the exit value. This value is not
2418 // simply the end of the counter because we may skip the vectorized body
2419 // in case of a runtime check.
2421 assert(!ResumeIndex && "Unexpected resume value found");
2422 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2423 MiddleBlock->getTerminator());
2424 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2425 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2426 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2429 // Make sure that we found the index where scalar loop needs to continue.
2430 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2431 "Invalid resume Index");
2433 // Add a check in the middle block to see if we have completed
2434 // all of the iterations in the first vector loop.
2435 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2436 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2437 ResumeIndex, "cmp.n",
2438 MiddleBlock->getTerminator());
2440 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2441 // Remove the old terminator.
2442 MiddleBlock->getTerminator()->eraseFromParent();
2444 // Create i+1 and fill the PHINode.
2445 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2446 Induction->addIncoming(StartIdx, VectorPH);
2447 Induction->addIncoming(NextIdx, VecBody);
2448 // Create the compare.
2449 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2450 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2452 // Now we have two terminators. Remove the old one from the block.
2453 VecBody->getTerminator()->eraseFromParent();
2455 // Get ready to start creating new instructions into the vectorized body.
2456 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2459 LoopVectorPreHeader = VectorPH;
2460 LoopScalarPreHeader = ScalarPH;
2461 LoopMiddleBlock = MiddleBlock;
2462 LoopExitBlock = ExitBlock;
2463 LoopVectorBody.push_back(VecBody);
2464 LoopScalarBody = OldBasicBlock;
2466 LoopVectorizeHints Hints(Lp, true);
2467 Hints.setAlreadyVectorized();
2470 /// This function returns the identity element (or neutral element) for
2471 /// the operation K.
2473 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2478 // Adding, Xoring, Oring zero to a number does not change it.
2479 return ConstantInt::get(Tp, 0);
2480 case RK_IntegerMult:
2481 // Multiplying a number by 1 does not change it.
2482 return ConstantInt::get(Tp, 1);
2484 // AND-ing a number with an all-1 value does not change it.
2485 return ConstantInt::get(Tp, -1, true);
2487 // Multiplying a number by 1 does not change it.
2488 return ConstantFP::get(Tp, 1.0L);
2490 // Adding zero to a number does not change it.
2491 return ConstantFP::get(Tp, 0.0L);
2493 llvm_unreachable("Unknown reduction kind");
2497 /// This function translates the reduction kind to an LLVM binary operator.
2499 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2501 case LoopVectorizationLegality::RK_IntegerAdd:
2502 return Instruction::Add;
2503 case LoopVectorizationLegality::RK_IntegerMult:
2504 return Instruction::Mul;
2505 case LoopVectorizationLegality::RK_IntegerOr:
2506 return Instruction::Or;
2507 case LoopVectorizationLegality::RK_IntegerAnd:
2508 return Instruction::And;
2509 case LoopVectorizationLegality::RK_IntegerXor:
2510 return Instruction::Xor;
2511 case LoopVectorizationLegality::RK_FloatMult:
2512 return Instruction::FMul;
2513 case LoopVectorizationLegality::RK_FloatAdd:
2514 return Instruction::FAdd;
2515 case LoopVectorizationLegality::RK_IntegerMinMax:
2516 return Instruction::ICmp;
2517 case LoopVectorizationLegality::RK_FloatMinMax:
2518 return Instruction::FCmp;
2520 llvm_unreachable("Unknown reduction operation");
2524 static Value *createMinMaxOp(IRBuilder<> &Builder,
2525 LoopVectorizationLegality::MinMaxReductionKind RK,
2526 Value *Left, Value *Right) {
2527 CmpInst::Predicate P = CmpInst::ICMP_NE;
2530 llvm_unreachable("Unknown min/max reduction kind");
2531 case LoopVectorizationLegality::MRK_UIntMin:
2532 P = CmpInst::ICMP_ULT;
2534 case LoopVectorizationLegality::MRK_UIntMax:
2535 P = CmpInst::ICMP_UGT;
2537 case LoopVectorizationLegality::MRK_SIntMin:
2538 P = CmpInst::ICMP_SLT;
2540 case LoopVectorizationLegality::MRK_SIntMax:
2541 P = CmpInst::ICMP_SGT;
2543 case LoopVectorizationLegality::MRK_FloatMin:
2544 P = CmpInst::FCMP_OLT;
2546 case LoopVectorizationLegality::MRK_FloatMax:
2547 P = CmpInst::FCMP_OGT;
2552 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2553 RK == LoopVectorizationLegality::MRK_FloatMax)
2554 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2556 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2558 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2563 struct CSEDenseMapInfo {
2564 static bool canHandle(Instruction *I) {
2565 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2566 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2568 static inline Instruction *getEmptyKey() {
2569 return DenseMapInfo<Instruction *>::getEmptyKey();
2571 static inline Instruction *getTombstoneKey() {
2572 return DenseMapInfo<Instruction *>::getTombstoneKey();
2574 static unsigned getHashValue(Instruction *I) {
2575 assert(canHandle(I) && "Unknown instruction!");
2576 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2577 I->value_op_end()));
2579 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2580 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2581 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2583 return LHS->isIdenticalTo(RHS);
2588 /// \brief Check whether this block is a predicated block.
2589 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2590 /// = ...; " blocks. We start with one vectorized basic block. For every
2591 /// conditional block we split this vectorized block. Therefore, every second
2592 /// block will be a predicated one.
2593 static bool isPredicatedBlock(unsigned BlockNum) {
2594 return BlockNum % 2;
2597 ///\brief Perform cse of induction variable instructions.
2598 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2599 // Perform simple cse.
2600 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2601 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2602 BasicBlock *BB = BBs[i];
2603 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2604 Instruction *In = I++;
2606 if (!CSEDenseMapInfo::canHandle(In))
2609 // Check if we can replace this instruction with any of the
2610 // visited instructions.
2611 if (Instruction *V = CSEMap.lookup(In)) {
2612 In->replaceAllUsesWith(V);
2613 In->eraseFromParent();
2616 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2617 // ...;" blocks for predicated stores. Every second block is a predicated
2619 if (isPredicatedBlock(i))
2627 /// \brief Adds a 'fast' flag to floating point operations.
2628 static Value *addFastMathFlag(Value *V) {
2629 if (isa<FPMathOperator>(V)){
2630 FastMathFlags Flags;
2631 Flags.setUnsafeAlgebra();
2632 cast<Instruction>(V)->setFastMathFlags(Flags);
2637 /// Estimate the overhead of scalarizing a value. Insert and Extract are set if
2638 /// the result needs to be inserted and/or extracted from vectors.
2639 static unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract,
2640 const TargetTransformInfo &TTI) {
2644 assert(Ty->isVectorTy() && "Can only scalarize vectors");
2647 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
2649 Cost += TTI.getVectorInstrCost(Instruction::InsertElement, Ty, i);
2651 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, Ty, i);
2657 // Estimate cost of a call instruction CI if it were vectorized with factor VF.
2658 // Return the cost of the instruction, including scalarization overhead if it's
2659 // needed. The flag NeedToScalarize shows if the call needs to be scalarized -
2660 // i.e. either vector version isn't available, or is too expensive.
2661 static unsigned getVectorCallCost(CallInst *CI, unsigned VF,
2662 const TargetTransformInfo &TTI,
2663 const TargetLibraryInfo *TLI,
2664 bool &NeedToScalarize) {
2665 Function *F = CI->getCalledFunction();
2666 StringRef FnName = CI->getCalledFunction()->getName();
2667 Type *ScalarRetTy = CI->getType();
2668 SmallVector<Type *, 4> Tys, ScalarTys;
2669 for (auto &ArgOp : CI->arg_operands())
2670 ScalarTys.push_back(ArgOp->getType());
2672 // Estimate cost of scalarized vector call. The source operands are assumed
2673 // to be vectors, so we need to extract individual elements from there,
2674 // execute VF scalar calls, and then gather the result into the vector return
2676 unsigned ScalarCallCost = TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys);
2678 return ScalarCallCost;
2680 // Compute corresponding vector type for return value and arguments.
2681 Type *RetTy = ToVectorTy(ScalarRetTy, VF);
2682 for (unsigned i = 0, ie = ScalarTys.size(); i != ie; ++i)
2683 Tys.push_back(ToVectorTy(ScalarTys[i], VF));
2685 // Compute costs of unpacking argument values for the scalar calls and
2686 // packing the return values to a vector.
2687 unsigned ScalarizationCost =
2688 getScalarizationOverhead(RetTy, true, false, TTI);
2689 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i)
2690 ScalarizationCost += getScalarizationOverhead(Tys[i], false, true, TTI);
2692 unsigned Cost = ScalarCallCost * VF + ScalarizationCost;
2694 // If we can't emit a vector call for this function, then the currently found
2695 // cost is the cost we need to return.
2696 NeedToScalarize = true;
2697 if (!TLI || !TLI->isFunctionVectorizable(FnName, VF) || CI->isNoBuiltin())
2700 // If the corresponding vector cost is cheaper, return its cost.
2701 unsigned VectorCallCost = TTI.getCallInstrCost(nullptr, RetTy, Tys);
2702 if (VectorCallCost < Cost) {
2703 NeedToScalarize = false;
2704 return VectorCallCost;
2709 // Estimate cost of an intrinsic call instruction CI if it were vectorized with
2710 // factor VF. Return the cost of the instruction, including scalarization
2711 // overhead if it's needed.
2712 static unsigned getVectorIntrinsicCost(CallInst *CI, unsigned VF,
2713 const TargetTransformInfo &TTI,
2714 const TargetLibraryInfo *TLI) {
2715 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
2716 assert(ID && "Expected intrinsic call!");
2718 Type *RetTy = ToVectorTy(CI->getType(), VF);
2719 SmallVector<Type *, 4> Tys;
2720 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
2721 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
2723 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
2726 void InnerLoopVectorizer::vectorizeLoop() {
2727 //===------------------------------------------------===//
2729 // Notice: any optimization or new instruction that go
2730 // into the code below should be also be implemented in
2733 //===------------------------------------------------===//
2734 Constant *Zero = Builder.getInt32(0);
2736 // In order to support reduction variables we need to be able to vectorize
2737 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2738 // stages. First, we create a new vector PHI node with no incoming edges.
2739 // We use this value when we vectorize all of the instructions that use the
2740 // PHI. Next, after all of the instructions in the block are complete we
2741 // add the new incoming edges to the PHI. At this point all of the
2742 // instructions in the basic block are vectorized, so we can use them to
2743 // construct the PHI.
2744 PhiVector RdxPHIsToFix;
2746 // Scan the loop in a topological order to ensure that defs are vectorized
2748 LoopBlocksDFS DFS(OrigLoop);
2751 // Vectorize all of the blocks in the original loop.
2752 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2753 be = DFS.endRPO(); bb != be; ++bb)
2754 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2756 // At this point every instruction in the original loop is widened to
2757 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2758 // that we vectorized. The PHI nodes are currently empty because we did
2759 // not want to introduce cycles. Notice that the remaining PHI nodes
2760 // that we need to fix are reduction variables.
2762 // Create the 'reduced' values for each of the induction vars.
2763 // The reduced values are the vector values that we scalarize and combine
2764 // after the loop is finished.
2765 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2767 PHINode *RdxPhi = *it;
2768 assert(RdxPhi && "Unable to recover vectorized PHI");
2770 // Find the reduction variable descriptor.
2771 assert(Legal->getReductionVars()->count(RdxPhi) &&
2772 "Unable to find the reduction variable");
2773 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2774 (*Legal->getReductionVars())[RdxPhi];
2776 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2778 // We need to generate a reduction vector from the incoming scalar.
2779 // To do so, we need to generate the 'identity' vector and override
2780 // one of the elements with the incoming scalar reduction. We need
2781 // to do it in the vector-loop preheader.
2782 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
2784 // This is the vector-clone of the value that leaves the loop.
2785 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2786 Type *VecTy = VectorExit[0]->getType();
2788 // Find the reduction identity variable. Zero for addition, or, xor,
2789 // one for multiplication, -1 for And.
2792 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2793 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2794 // MinMax reduction have the start value as their identify.
2796 VectorStart = Identity = RdxDesc.StartValue;
2798 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2803 // Handle other reduction kinds:
2805 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2806 VecTy->getScalarType());
2809 // This vector is the Identity vector where the first element is the
2810 // incoming scalar reduction.
2811 VectorStart = RdxDesc.StartValue;
2813 Identity = ConstantVector::getSplat(VF, Iden);
2815 // This vector is the Identity vector where the first element is the
2816 // incoming scalar reduction.
2817 VectorStart = Builder.CreateInsertElement(Identity,
2818 RdxDesc.StartValue, Zero);
2822 // Fix the vector-loop phi.
2824 // Reductions do not have to start at zero. They can start with
2825 // any loop invariant values.
2826 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2827 BasicBlock *Latch = OrigLoop->getLoopLatch();
2828 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2829 VectorParts &Val = getVectorValue(LoopVal);
2830 for (unsigned part = 0; part < UF; ++part) {
2831 // Make sure to add the reduction stat value only to the
2832 // first unroll part.
2833 Value *StartVal = (part == 0) ? VectorStart : Identity;
2834 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
2835 LoopVectorPreHeader);
2836 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2837 LoopVectorBody.back());
2840 // Before each round, move the insertion point right between
2841 // the PHIs and the values we are going to write.
2842 // This allows us to write both PHINodes and the extractelement
2844 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2846 VectorParts RdxParts;
2847 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2848 for (unsigned part = 0; part < UF; ++part) {
2849 // This PHINode contains the vectorized reduction variable, or
2850 // the initial value vector, if we bypass the vector loop.
2851 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2852 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2853 Value *StartVal = (part == 0) ? VectorStart : Identity;
2854 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2855 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2856 NewPhi->addIncoming(RdxExitVal[part],
2857 LoopVectorBody.back());
2858 RdxParts.push_back(NewPhi);
2861 // Reduce all of the unrolled parts into a single vector.
2862 Value *ReducedPartRdx = RdxParts[0];
2863 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2864 setDebugLocFromInst(Builder, ReducedPartRdx);
2865 for (unsigned part = 1; part < UF; ++part) {
2866 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2867 // Floating point operations had to be 'fast' to enable the reduction.
2868 ReducedPartRdx = addFastMathFlag(
2869 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2870 ReducedPartRdx, "bin.rdx"));
2872 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2873 ReducedPartRdx, RdxParts[part]);
2877 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2878 // and vector ops, reducing the set of values being computed by half each
2880 assert(isPowerOf2_32(VF) &&
2881 "Reduction emission only supported for pow2 vectors!");
2882 Value *TmpVec = ReducedPartRdx;
2883 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2884 for (unsigned i = VF; i != 1; i >>= 1) {
2885 // Move the upper half of the vector to the lower half.
2886 for (unsigned j = 0; j != i/2; ++j)
2887 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2889 // Fill the rest of the mask with undef.
2890 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2891 UndefValue::get(Builder.getInt32Ty()));
2894 Builder.CreateShuffleVector(TmpVec,
2895 UndefValue::get(TmpVec->getType()),
2896 ConstantVector::get(ShuffleMask),
2899 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2900 // Floating point operations had to be 'fast' to enable the reduction.
2901 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2902 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2904 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2907 // The result is in the first element of the vector.
2908 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2909 Builder.getInt32(0));
2912 // Create a phi node that merges control-flow from the backedge-taken check
2913 // block and the middle block.
2914 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
2915 LoopScalarPreHeader->getTerminator());
2916 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
2917 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2919 // Now, we need to fix the users of the reduction variable
2920 // inside and outside of the scalar remainder loop.
2921 // We know that the loop is in LCSSA form. We need to update the
2922 // PHI nodes in the exit blocks.
2923 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2924 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2925 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2926 if (!LCSSAPhi) break;
2928 // All PHINodes need to have a single entry edge, or two if
2929 // we already fixed them.
2930 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2932 // We found our reduction value exit-PHI. Update it with the
2933 // incoming bypass edge.
2934 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2935 // Add an edge coming from the bypass.
2936 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2939 }// end of the LCSSA phi scan.
2941 // Fix the scalar loop reduction variable with the incoming reduction sum
2942 // from the vector body and from the backedge value.
2943 int IncomingEdgeBlockIdx =
2944 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2945 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2946 // Pick the other block.
2947 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2948 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
2949 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2950 }// end of for each redux variable.
2954 // Remove redundant induction instructions.
2955 cse(LoopVectorBody);
2958 void InnerLoopVectorizer::fixLCSSAPHIs() {
2959 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2960 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2961 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2962 if (!LCSSAPhi) break;
2963 if (LCSSAPhi->getNumIncomingValues() == 1)
2964 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2969 InnerLoopVectorizer::VectorParts
2970 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2971 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2974 // Look for cached value.
2975 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2976 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2977 if (ECEntryIt != MaskCache.end())
2978 return ECEntryIt->second;
2980 VectorParts SrcMask = createBlockInMask(Src);
2982 // The terminator has to be a branch inst!
2983 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2984 assert(BI && "Unexpected terminator found");
2986 if (BI->isConditional()) {
2987 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2989 if (BI->getSuccessor(0) != Dst)
2990 for (unsigned part = 0; part < UF; ++part)
2991 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2993 for (unsigned part = 0; part < UF; ++part)
2994 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2996 MaskCache[Edge] = EdgeMask;
3000 MaskCache[Edge] = SrcMask;
3004 InnerLoopVectorizer::VectorParts
3005 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
3006 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
3008 // Loop incoming mask is all-one.
3009 if (OrigLoop->getHeader() == BB) {
3010 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
3011 return getVectorValue(C);
3014 // This is the block mask. We OR all incoming edges, and with zero.
3015 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
3016 VectorParts BlockMask = getVectorValue(Zero);
3019 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
3020 VectorParts EM = createEdgeMask(*it, BB);
3021 for (unsigned part = 0; part < UF; ++part)
3022 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
3028 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
3029 InnerLoopVectorizer::VectorParts &Entry,
3030 unsigned UF, unsigned VF, PhiVector *PV) {
3031 PHINode* P = cast<PHINode>(PN);
3032 // Handle reduction variables:
3033 if (Legal->getReductionVars()->count(P)) {
3034 for (unsigned part = 0; part < UF; ++part) {
3035 // This is phase one of vectorizing PHIs.
3036 Type *VecTy = (VF == 1) ? PN->getType() :
3037 VectorType::get(PN->getType(), VF);
3038 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
3039 LoopVectorBody.back()-> getFirstInsertionPt());
3045 setDebugLocFromInst(Builder, P);
3046 // Check for PHI nodes that are lowered to vector selects.
3047 if (P->getParent() != OrigLoop->getHeader()) {
3048 // We know that all PHIs in non-header blocks are converted into
3049 // selects, so we don't have to worry about the insertion order and we
3050 // can just use the builder.
3051 // At this point we generate the predication tree. There may be
3052 // duplications since this is a simple recursive scan, but future
3053 // optimizations will clean it up.
3055 unsigned NumIncoming = P->getNumIncomingValues();
3057 // Generate a sequence of selects of the form:
3058 // SELECT(Mask3, In3,
3059 // SELECT(Mask2, In2,
3061 for (unsigned In = 0; In < NumIncoming; In++) {
3062 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
3064 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3066 for (unsigned part = 0; part < UF; ++part) {
3067 // We might have single edge PHIs (blocks) - use an identity
3068 // 'select' for the first PHI operand.
3070 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3073 // Select between the current value and the previous incoming edge
3074 // based on the incoming mask.
3075 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3076 Entry[part], "predphi");
3082 // This PHINode must be an induction variable.
3083 // Make sure that we know about it.
3084 assert(Legal->getInductionVars()->count(P) &&
3085 "Not an induction variable");
3087 LoopVectorizationLegality::InductionInfo II =
3088 Legal->getInductionVars()->lookup(P);
3090 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
3091 // which can be found from the original scalar operations.
3093 case LoopVectorizationLegality::IK_NoInduction:
3094 llvm_unreachable("Unknown induction");
3095 case LoopVectorizationLegality::IK_IntInduction: {
3096 assert(P->getType() == II.StartValue->getType() && "Types must match");
3097 Type *PhiTy = P->getType();
3099 if (P == OldInduction) {
3100 // Handle the canonical induction variable. We might have had to
3102 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
3104 // Handle other induction variables that are now based on the
3106 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
3108 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
3109 Broadcasted = II.transform(Builder, NormalizedIdx);
3110 Broadcasted->setName("offset.idx");
3112 Broadcasted = getBroadcastInstrs(Broadcasted);
3113 // After broadcasting the induction variable we need to make the vector
3114 // consecutive by adding 0, 1, 2, etc.
3115 for (unsigned part = 0; part < UF; ++part)
3116 Entry[part] = getStepVector(Broadcasted, VF * part, II.StepValue);
3119 case LoopVectorizationLegality::IK_PtrInduction:
3120 // Handle the pointer induction variable case.
3121 assert(P->getType()->isPointerTy() && "Unexpected type.");
3122 // This is the normalized GEP that starts counting at zero.
3123 Value *NormalizedIdx =
3124 Builder.CreateSub(Induction, ExtendedIdx, "normalized.idx");
3125 // This is the vector of results. Notice that we don't generate
3126 // vector geps because scalar geps result in better code.
3127 for (unsigned part = 0; part < UF; ++part) {
3129 int EltIndex = part;
3130 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3131 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3132 Value *SclrGep = II.transform(Builder, GlobalIdx);
3133 SclrGep->setName("next.gep");
3134 Entry[part] = SclrGep;
3138 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3139 for (unsigned int i = 0; i < VF; ++i) {
3140 int EltIndex = i + part * VF;
3141 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3142 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3143 Value *SclrGep = II.transform(Builder, GlobalIdx);
3144 SclrGep->setName("next.gep");
3145 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3146 Builder.getInt32(i),
3149 Entry[part] = VecVal;
3155 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3156 // For each instruction in the old loop.
3157 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3158 VectorParts &Entry = WidenMap.get(it);
3159 switch (it->getOpcode()) {
3160 case Instruction::Br:
3161 // Nothing to do for PHIs and BR, since we already took care of the
3162 // loop control flow instructions.
3164 case Instruction::PHI: {
3165 // Vectorize PHINodes.
3166 widenPHIInstruction(it, Entry, UF, VF, PV);
3170 case Instruction::Add:
3171 case Instruction::FAdd:
3172 case Instruction::Sub:
3173 case Instruction::FSub:
3174 case Instruction::Mul:
3175 case Instruction::FMul:
3176 case Instruction::UDiv:
3177 case Instruction::SDiv:
3178 case Instruction::FDiv:
3179 case Instruction::URem:
3180 case Instruction::SRem:
3181 case Instruction::FRem:
3182 case Instruction::Shl:
3183 case Instruction::LShr:
3184 case Instruction::AShr:
3185 case Instruction::And:
3186 case Instruction::Or:
3187 case Instruction::Xor: {
3188 // Just widen binops.
3189 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3190 setDebugLocFromInst(Builder, BinOp);
3191 VectorParts &A = getVectorValue(it->getOperand(0));
3192 VectorParts &B = getVectorValue(it->getOperand(1));
3194 // Use this vector value for all users of the original instruction.
3195 for (unsigned Part = 0; Part < UF; ++Part) {
3196 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3198 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3199 VecOp->copyIRFlags(BinOp);
3204 propagateMetadata(Entry, it);
3207 case Instruction::Select: {
3209 // If the selector is loop invariant we can create a select
3210 // instruction with a scalar condition. Otherwise, use vector-select.
3211 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3213 setDebugLocFromInst(Builder, it);
3215 // The condition can be loop invariant but still defined inside the
3216 // loop. This means that we can't just use the original 'cond' value.
3217 // We have to take the 'vectorized' value and pick the first lane.
3218 // Instcombine will make this a no-op.
3219 VectorParts &Cond = getVectorValue(it->getOperand(0));
3220 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3221 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3223 Value *ScalarCond = (VF == 1) ? Cond[0] :
3224 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3226 for (unsigned Part = 0; Part < UF; ++Part) {
3227 Entry[Part] = Builder.CreateSelect(
3228 InvariantCond ? ScalarCond : Cond[Part],
3233 propagateMetadata(Entry, it);
3237 case Instruction::ICmp:
3238 case Instruction::FCmp: {
3239 // Widen compares. Generate vector compares.
3240 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3241 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3242 setDebugLocFromInst(Builder, it);
3243 VectorParts &A = getVectorValue(it->getOperand(0));
3244 VectorParts &B = getVectorValue(it->getOperand(1));
3245 for (unsigned Part = 0; Part < UF; ++Part) {
3248 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3250 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3254 propagateMetadata(Entry, it);
3258 case Instruction::Store:
3259 case Instruction::Load:
3260 vectorizeMemoryInstruction(it);
3262 case Instruction::ZExt:
3263 case Instruction::SExt:
3264 case Instruction::FPToUI:
3265 case Instruction::FPToSI:
3266 case Instruction::FPExt:
3267 case Instruction::PtrToInt:
3268 case Instruction::IntToPtr:
3269 case Instruction::SIToFP:
3270 case Instruction::UIToFP:
3271 case Instruction::Trunc:
3272 case Instruction::FPTrunc:
3273 case Instruction::BitCast: {
3274 CastInst *CI = dyn_cast<CastInst>(it);
3275 setDebugLocFromInst(Builder, it);
3276 /// Optimize the special case where the source is the induction
3277 /// variable. Notice that we can only optimize the 'trunc' case
3278 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3279 /// c. other casts depend on pointer size.
3280 if (CI->getOperand(0) == OldInduction &&
3281 it->getOpcode() == Instruction::Trunc) {
3282 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3284 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3285 LoopVectorizationLegality::InductionInfo II =
3286 Legal->getInductionVars()->lookup(OldInduction);
3288 ConstantInt::getSigned(CI->getType(), II.StepValue->getSExtValue());
3289 for (unsigned Part = 0; Part < UF; ++Part)
3290 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
3291 propagateMetadata(Entry, it);
3294 /// Vectorize casts.
3295 Type *DestTy = (VF == 1) ? CI->getType() :
3296 VectorType::get(CI->getType(), VF);
3298 VectorParts &A = getVectorValue(it->getOperand(0));
3299 for (unsigned Part = 0; Part < UF; ++Part)
3300 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3301 propagateMetadata(Entry, it);
3305 case Instruction::Call: {
3306 // Ignore dbg intrinsics.
3307 if (isa<DbgInfoIntrinsic>(it))
3309 setDebugLocFromInst(Builder, it);
3311 Module *M = BB->getParent()->getParent();
3312 CallInst *CI = cast<CallInst>(it);
3314 StringRef FnName = CI->getCalledFunction()->getName();
3315 Function *F = CI->getCalledFunction();
3316 Type *RetTy = ToVectorTy(CI->getType(), VF);
3317 SmallVector<Type *, 4> Tys;
3318 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3319 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3321 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3323 (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end ||
3324 ID == Intrinsic::lifetime_start)) {
3325 scalarizeInstruction(it);
3328 // The flag shows whether we use Intrinsic or a usual Call for vectorized
3329 // version of the instruction.
3330 // Is it beneficial to perform intrinsic call compared to lib call?
3331 bool NeedToScalarize;
3332 unsigned CallCost = getVectorCallCost(CI, VF, *TTI, TLI, NeedToScalarize);
3333 bool UseVectorIntrinsic =
3334 ID && getVectorIntrinsicCost(CI, VF, *TTI, TLI) <= CallCost;
3335 if (!UseVectorIntrinsic && NeedToScalarize) {
3336 scalarizeInstruction(it);
3340 for (unsigned Part = 0; Part < UF; ++Part) {
3341 SmallVector<Value *, 4> Args;
3342 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3343 Value *Arg = CI->getArgOperand(i);
3344 // Some intrinsics have a scalar argument - don't replace it with a
3346 if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, i)) {
3347 VectorParts &VectorArg = getVectorValue(CI->getArgOperand(i));
3348 Arg = VectorArg[Part];
3350 Args.push_back(Arg);
3354 if (UseVectorIntrinsic) {
3355 // Use vector version of the intrinsic.
3356 Type *TysForDecl[] = {CI->getType()};
3358 TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3359 VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl);
3361 // Use vector version of the library call.
3362 StringRef VFnName = TLI->getVectorizedFunction(FnName, VF);
3363 assert(!VFnName.empty() && "Vector function name is empty.");
3364 VectorF = M->getFunction(VFnName);
3366 // Generate a declaration
3367 FunctionType *FTy = FunctionType::get(RetTy, Tys, false);
3369 Function::Create(FTy, Function::ExternalLinkage, VFnName, M);
3370 VectorF->copyAttributesFrom(F);
3373 assert(VectorF && "Can't create vector function.");
3374 Entry[Part] = Builder.CreateCall(VectorF, Args);
3377 propagateMetadata(Entry, it);
3382 // All other instructions are unsupported. Scalarize them.
3383 scalarizeInstruction(it);
3386 }// end of for_each instr.
3389 void InnerLoopVectorizer::updateAnalysis() {
3390 // Forget the original basic block.
3391 SE->forgetLoop(OrigLoop);
3393 // Update the dominator tree information.
3394 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3395 "Entry does not dominate exit.");
3397 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3398 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3399 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3401 // Due to if predication of stores we might create a sequence of "if(pred)
3402 // a[i] = ...; " blocks.
3403 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3405 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3406 else if (isPredicatedBlock(i)) {
3407 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3409 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3413 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3414 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3415 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3416 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3418 DEBUG(DT->verifyDomTree());
3421 /// \brief Check whether it is safe to if-convert this phi node.
3423 /// Phi nodes with constant expressions that can trap are not safe to if
3425 static bool canIfConvertPHINodes(BasicBlock *BB) {
3426 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3427 PHINode *Phi = dyn_cast<PHINode>(I);
3430 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3431 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3438 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3439 if (!EnableIfConversion) {
3440 emitAnalysis(VectorizationReport() << "if-conversion is disabled");
3444 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3446 // A list of pointers that we can safely read and write to.
3447 SmallPtrSet<Value *, 8> SafePointes;
3449 // Collect safe addresses.
3450 for (Loop::block_iterator BI = TheLoop->block_begin(),
3451 BE = TheLoop->block_end(); BI != BE; ++BI) {
3452 BasicBlock *BB = *BI;
3454 if (blockNeedsPredication(BB))
3457 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3458 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3459 SafePointes.insert(LI->getPointerOperand());
3460 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3461 SafePointes.insert(SI->getPointerOperand());
3465 // Collect the blocks that need predication.
3466 BasicBlock *Header = TheLoop->getHeader();
3467 for (Loop::block_iterator BI = TheLoop->block_begin(),
3468 BE = TheLoop->block_end(); BI != BE; ++BI) {
3469 BasicBlock *BB = *BI;
3471 // We don't support switch statements inside loops.
3472 if (!isa<BranchInst>(BB->getTerminator())) {
3473 emitAnalysis(VectorizationReport(BB->getTerminator())
3474 << "loop contains a switch statement");
3478 // We must be able to predicate all blocks that need to be predicated.
3479 if (blockNeedsPredication(BB)) {
3480 if (!blockCanBePredicated(BB, SafePointes)) {
3481 emitAnalysis(VectorizationReport(BB->getTerminator())
3482 << "control flow cannot be substituted for a select");
3485 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3486 emitAnalysis(VectorizationReport(BB->getTerminator())
3487 << "control flow cannot be substituted for a select");
3492 // We can if-convert this loop.
3496 bool LoopVectorizationLegality::canVectorize() {
3497 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3498 // be canonicalized.
3499 if (!TheLoop->getLoopPreheader()) {
3501 VectorizationReport() <<
3502 "loop control flow is not understood by vectorizer");
3506 // We can only vectorize innermost loops.
3507 if (!TheLoop->getSubLoopsVector().empty()) {
3508 emitAnalysis(VectorizationReport() << "loop is not the innermost loop");
3512 // We must have a single backedge.
3513 if (TheLoop->getNumBackEdges() != 1) {
3515 VectorizationReport() <<
3516 "loop control flow is not understood by vectorizer");
3520 // We must have a single exiting block.
3521 if (!TheLoop->getExitingBlock()) {
3523 VectorizationReport() <<
3524 "loop control flow is not understood by vectorizer");
3528 // We only handle bottom-tested loops, i.e. loop in which the condition is
3529 // checked at the end of each iteration. With that we can assume that all
3530 // instructions in the loop are executed the same number of times.
3531 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
3533 VectorizationReport() <<
3534 "loop control flow is not understood by vectorizer");
3538 // We need to have a loop header.
3539 DEBUG(dbgs() << "LV: Found a loop: " <<
3540 TheLoop->getHeader()->getName() << '\n');
3542 // Check if we can if-convert non-single-bb loops.
3543 unsigned NumBlocks = TheLoop->getNumBlocks();
3544 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3545 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3549 // ScalarEvolution needs to be able to find the exit count.
3550 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3551 if (ExitCount == SE->getCouldNotCompute()) {
3552 emitAnalysis(VectorizationReport() <<
3553 "could not determine number of loop iterations");
3554 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3558 // Check if we can vectorize the instructions and CFG in this loop.
3559 if (!canVectorizeInstrs()) {
3560 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3564 // Go over each instruction and look at memory deps.
3565 if (!canVectorizeMemory()) {
3566 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3570 // Collect all of the variables that remain uniform after vectorization.
3571 collectLoopUniforms();
3573 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3574 (LAI->getRuntimePointerCheck()->Need ? " (with a runtime bound check)" :
3578 // Okay! We can vectorize. At this point we don't have any other mem analysis
3579 // which may limit our maximum vectorization factor, so just return true with
3584 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3585 if (Ty->isPointerTy())
3586 return DL.getIntPtrType(Ty);
3588 // It is possible that char's or short's overflow when we ask for the loop's
3589 // trip count, work around this by changing the type size.
3590 if (Ty->getScalarSizeInBits() < 32)
3591 return Type::getInt32Ty(Ty->getContext());
3596 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3597 Ty0 = convertPointerToIntegerType(DL, Ty0);
3598 Ty1 = convertPointerToIntegerType(DL, Ty1);
3599 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3604 /// \brief Check that the instruction has outside loop users and is not an
3605 /// identified reduction variable.
3606 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3607 SmallPtrSetImpl<Value *> &Reductions) {
3608 // Reduction instructions are allowed to have exit users. All other
3609 // instructions must not have external users.
3610 if (!Reductions.count(Inst))
3611 //Check that all of the users of the loop are inside the BB.
3612 for (User *U : Inst->users()) {
3613 Instruction *UI = cast<Instruction>(U);
3614 // This user may be a reduction exit value.
3615 if (!TheLoop->contains(UI)) {
3616 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3623 bool LoopVectorizationLegality::canVectorizeInstrs() {
3624 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3625 BasicBlock *Header = TheLoop->getHeader();
3627 // Look for the attribute signaling the absence of NaNs.
3628 Function &F = *Header->getParent();
3629 const DataLayout &DL = F.getParent()->getDataLayout();
3630 if (F.hasFnAttribute("no-nans-fp-math"))
3632 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
3634 // For each block in the loop.
3635 for (Loop::block_iterator bb = TheLoop->block_begin(),
3636 be = TheLoop->block_end(); bb != be; ++bb) {
3638 // Scan the instructions in the block and look for hazards.
3639 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3642 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3643 Type *PhiTy = Phi->getType();
3644 // Check that this PHI type is allowed.
3645 if (!PhiTy->isIntegerTy() &&
3646 !PhiTy->isFloatingPointTy() &&
3647 !PhiTy->isPointerTy()) {
3648 emitAnalysis(VectorizationReport(it)
3649 << "loop control flow is not understood by vectorizer");
3650 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3654 // If this PHINode is not in the header block, then we know that we
3655 // can convert it to select during if-conversion. No need to check if
3656 // the PHIs in this block are induction or reduction variables.
3657 if (*bb != Header) {
3658 // Check that this instruction has no outside users or is an
3659 // identified reduction value with an outside user.
3660 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3662 emitAnalysis(VectorizationReport(it) <<
3663 "value could not be identified as "
3664 "an induction or reduction variable");
3668 // We only allow if-converted PHIs with exactly two incoming values.
3669 if (Phi->getNumIncomingValues() != 2) {
3670 emitAnalysis(VectorizationReport(it)
3671 << "control flow not understood by vectorizer");
3672 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3676 // This is the value coming from the preheader.
3677 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3678 ConstantInt *StepValue = nullptr;
3679 // Check if this is an induction variable.
3680 InductionKind IK = isInductionVariable(Phi, StepValue);
3682 if (IK_NoInduction != IK) {
3683 // Get the widest type.
3685 WidestIndTy = convertPointerToIntegerType(DL, PhiTy);
3687 WidestIndTy = getWiderType(DL, PhiTy, WidestIndTy);
3689 // Int inductions are special because we only allow one IV.
3690 if (IK == IK_IntInduction && StepValue->isOne()) {
3691 // Use the phi node with the widest type as induction. Use the last
3692 // one if there are multiple (no good reason for doing this other
3693 // than it is expedient).
3694 if (!Induction || PhiTy == WidestIndTy)
3698 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3699 Inductions[Phi] = InductionInfo(StartValue, IK, StepValue);
3701 // Until we explicitly handle the case of an induction variable with
3702 // an outside loop user we have to give up vectorizing this loop.
3703 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3704 emitAnalysis(VectorizationReport(it) <<
3705 "use of induction value outside of the "
3706 "loop is not handled by vectorizer");
3713 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3714 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3717 if (AddReductionVar(Phi, RK_IntegerMult)) {
3718 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3721 if (AddReductionVar(Phi, RK_IntegerOr)) {
3722 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3725 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3726 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3729 if (AddReductionVar(Phi, RK_IntegerXor)) {
3730 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3733 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3734 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3737 if (AddReductionVar(Phi, RK_FloatMult)) {
3738 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3741 if (AddReductionVar(Phi, RK_FloatAdd)) {
3742 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3745 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3746 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3751 emitAnalysis(VectorizationReport(it) <<
3752 "value that could not be identified as "
3753 "reduction is used outside the loop");
3754 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3756 }// end of PHI handling
3758 // We handle calls that:
3759 // * Are debug info intrinsics.
3760 // * Have a mapping to an IR intrinsic.
3761 // * Have a vector version available.
3762 CallInst *CI = dyn_cast<CallInst>(it);
3763 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI) &&
3764 !(CI->getCalledFunction() && TLI &&
3765 TLI->isFunctionVectorizable(CI->getCalledFunction()->getName()))) {
3766 emitAnalysis(VectorizationReport(it) <<
3767 "call instruction cannot be vectorized");
3768 DEBUG(dbgs() << "LV: Found a non-intrinsic, non-libfunc callsite.\n");
3772 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
3773 // second argument is the same (i.e. loop invariant)
3775 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
3776 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
3777 emitAnalysis(VectorizationReport(it)
3778 << "intrinsic instruction cannot be vectorized");
3779 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
3784 // Check that the instruction return type is vectorizable.
3785 // Also, we can't vectorize extractelement instructions.
3786 if ((!VectorType::isValidElementType(it->getType()) &&
3787 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3788 emitAnalysis(VectorizationReport(it)
3789 << "instruction return type cannot be vectorized");
3790 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3794 // Check that the stored type is vectorizable.
3795 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3796 Type *T = ST->getValueOperand()->getType();
3797 if (!VectorType::isValidElementType(T)) {
3798 emitAnalysis(VectorizationReport(ST) <<
3799 "store instruction cannot be vectorized");
3802 if (EnableMemAccessVersioning)
3803 collectStridedAccess(ST);
3806 if (EnableMemAccessVersioning)
3807 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3808 collectStridedAccess(LI);
3810 // Reduction instructions are allowed to have exit users.
3811 // All other instructions must not have external users.
3812 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3813 emitAnalysis(VectorizationReport(it) <<
3814 "value cannot be used outside the loop");
3823 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3824 if (Inductions.empty()) {
3825 emitAnalysis(VectorizationReport()
3826 << "loop induction variable could not be identified");
3834 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3835 /// return the induction operand of the gep pointer.
3836 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE, Loop *Lp) {
3837 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3841 unsigned InductionOperand = getGEPInductionOperand(GEP);
3843 // Check that all of the gep indices are uniform except for our induction
3845 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3846 if (i != InductionOperand &&
3847 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3849 return GEP->getOperand(InductionOperand);
3852 ///\brief Look for a cast use of the passed value.
3853 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3854 Value *UniqueCast = nullptr;
3855 for (User *U : Ptr->users()) {
3856 CastInst *CI = dyn_cast<CastInst>(U);
3857 if (CI && CI->getType() == Ty) {
3867 ///\brief Get the stride of a pointer access in a loop.
3868 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3869 /// pointer to the Value, or null otherwise.
3870 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE, Loop *Lp) {
3871 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3872 if (!PtrTy || PtrTy->isAggregateType())
3875 // Try to remove a gep instruction to make the pointer (actually index at this
3876 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3877 // pointer, otherwise, we are analyzing the index.
3878 Value *OrigPtr = Ptr;
3880 // The size of the pointer access.
3881 int64_t PtrAccessSize = 1;
3883 Ptr = stripGetElementPtr(Ptr, SE, Lp);
3884 const SCEV *V = SE->getSCEV(Ptr);
3888 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3889 V = C->getOperand();
3891 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3895 V = S->getStepRecurrence(*SE);
3899 // Strip off the size of access multiplication if we are still analyzing the
3901 if (OrigPtr == Ptr) {
3902 const DataLayout &DL = Lp->getHeader()->getModule()->getDataLayout();
3903 DL.getTypeAllocSize(PtrTy->getElementType());
3904 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3905 if (M->getOperand(0)->getSCEVType() != scConstant)
3908 const APInt &APStepVal =
3909 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3911 // Huge step value - give up.
3912 if (APStepVal.getBitWidth() > 64)
3915 int64_t StepVal = APStepVal.getSExtValue();
3916 if (PtrAccessSize != StepVal)
3918 V = M->getOperand(1);
3923 Type *StripedOffRecurrenceCast = nullptr;
3924 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3925 StripedOffRecurrenceCast = C->getType();
3926 V = C->getOperand();
3929 // Look for the loop invariant symbolic value.
3930 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3934 Value *Stride = U->getValue();
3935 if (!Lp->isLoopInvariant(Stride))
3938 // If we have stripped off the recurrence cast we have to make sure that we
3939 // return the value that is used in this loop so that we can replace it later.
3940 if (StripedOffRecurrenceCast)
3941 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3946 void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) {
3947 Value *Ptr = nullptr;
3948 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3949 Ptr = LI->getPointerOperand();
3950 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3951 Ptr = SI->getPointerOperand();
3955 Value *Stride = getStrideFromPointer(Ptr, SE, TheLoop);
3959 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3960 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3961 Strides[Ptr] = Stride;
3962 StrideSet.insert(Stride);
3965 void LoopVectorizationLegality::collectLoopUniforms() {
3966 // We now know that the loop is vectorizable!
3967 // Collect variables that will remain uniform after vectorization.
3968 std::vector<Value*> Worklist;
3969 BasicBlock *Latch = TheLoop->getLoopLatch();
3971 // Start with the conditional branch and walk up the block.
3972 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3974 // Also add all consecutive pointer values; these values will be uniform
3975 // after vectorization (and subsequent cleanup) and, until revectorization is
3976 // supported, all dependencies must also be uniform.
3977 for (Loop::block_iterator B = TheLoop->block_begin(),
3978 BE = TheLoop->block_end(); B != BE; ++B)
3979 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
3981 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
3982 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3984 while (!Worklist.empty()) {
3985 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3986 Worklist.pop_back();
3988 // Look at instructions inside this loop.
3989 // Stop when reaching PHI nodes.
3990 // TODO: we need to follow values all over the loop, not only in this block.
3991 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3994 // This is a known uniform.
3997 // Insert all operands.
3998 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4002 bool LoopVectorizationLegality::canVectorizeMemory() {
4003 LAI = &LAA->getInfo(TheLoop, Strides);
4004 auto &OptionalReport = LAI->getReport();
4006 emitAnalysis(VectorizationReport(*OptionalReport));
4007 if (!LAI->canVectorizeMemory())
4010 if (LAI->getNumRuntimePointerChecks() >
4011 VectorizerParams::RuntimeMemoryCheckThreshold) {
4012 emitAnalysis(VectorizationReport()
4013 << LAI->getNumRuntimePointerChecks() << " exceeds limit of "
4014 << VectorizerParams::RuntimeMemoryCheckThreshold
4015 << " dependent memory operations checked at runtime");
4016 DEBUG(dbgs() << "LV: Too many memory checks needed.\n");
4022 static bool hasMultipleUsesOf(Instruction *I,
4023 SmallPtrSetImpl<Instruction *> &Insts) {
4024 unsigned NumUses = 0;
4025 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
4026 if (Insts.count(dyn_cast<Instruction>(*Use)))
4035 static bool areAllUsesIn(Instruction *I, SmallPtrSetImpl<Instruction *> &Set) {
4036 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
4037 if (!Set.count(dyn_cast<Instruction>(*Use)))
4042 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
4043 ReductionKind Kind) {
4044 if (Phi->getNumIncomingValues() != 2)
4047 // Reduction variables are only found in the loop header block.
4048 if (Phi->getParent() != TheLoop->getHeader())
4051 // Obtain the reduction start value from the value that comes from the loop
4053 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
4055 // ExitInstruction is the single value which is used outside the loop.
4056 // We only allow for a single reduction value to be used outside the loop.
4057 // This includes users of the reduction, variables (which form a cycle
4058 // which ends in the phi node).
4059 Instruction *ExitInstruction = nullptr;
4060 // Indicates that we found a reduction operation in our scan.
4061 bool FoundReduxOp = false;
4063 // We start with the PHI node and scan for all of the users of this
4064 // instruction. All users must be instructions that can be used as reduction
4065 // variables (such as ADD). We must have a single out-of-block user. The cycle
4066 // must include the original PHI.
4067 bool FoundStartPHI = false;
4069 // To recognize min/max patterns formed by a icmp select sequence, we store
4070 // the number of instruction we saw from the recognized min/max pattern,
4071 // to make sure we only see exactly the two instructions.
4072 unsigned NumCmpSelectPatternInst = 0;
4073 ReductionInstDesc ReduxDesc(false, nullptr);
4075 SmallPtrSet<Instruction *, 8> VisitedInsts;
4076 SmallVector<Instruction *, 8> Worklist;
4077 Worklist.push_back(Phi);
4078 VisitedInsts.insert(Phi);
4080 // A value in the reduction can be used:
4081 // - By the reduction:
4082 // - Reduction operation:
4083 // - One use of reduction value (safe).
4084 // - Multiple use of reduction value (not safe).
4086 // - All uses of the PHI must be the reduction (safe).
4087 // - Otherwise, not safe.
4088 // - By one instruction outside of the loop (safe).
4089 // - By further instructions outside of the loop (not safe).
4090 // - By an instruction that is not part of the reduction (not safe).
4092 // * An instruction type other than PHI or the reduction operation.
4093 // * A PHI in the header other than the initial PHI.
4094 while (!Worklist.empty()) {
4095 Instruction *Cur = Worklist.back();
4096 Worklist.pop_back();
4099 // If the instruction has no users then this is a broken chain and can't be
4100 // a reduction variable.
4101 if (Cur->use_empty())
4104 bool IsAPhi = isa<PHINode>(Cur);
4106 // A header PHI use other than the original PHI.
4107 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
4110 // Reductions of instructions such as Div, and Sub is only possible if the
4111 // LHS is the reduction variable.
4112 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
4113 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
4114 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
4117 // Any reduction instruction must be of one of the allowed kinds.
4118 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
4119 if (!ReduxDesc.IsReduction)
4122 // A reduction operation must only have one use of the reduction value.
4123 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
4124 hasMultipleUsesOf(Cur, VisitedInsts))
4127 // All inputs to a PHI node must be a reduction value.
4128 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
4131 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
4132 isa<SelectInst>(Cur)))
4133 ++NumCmpSelectPatternInst;
4134 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
4135 isa<SelectInst>(Cur)))
4136 ++NumCmpSelectPatternInst;
4138 // Check whether we found a reduction operator.
4139 FoundReduxOp |= !IsAPhi;
4141 // Process users of current instruction. Push non-PHI nodes after PHI nodes
4142 // onto the stack. This way we are going to have seen all inputs to PHI
4143 // nodes once we get to them.
4144 SmallVector<Instruction *, 8> NonPHIs;
4145 SmallVector<Instruction *, 8> PHIs;
4146 for (User *U : Cur->users()) {
4147 Instruction *UI = cast<Instruction>(U);
4149 // Check if we found the exit user.
4150 BasicBlock *Parent = UI->getParent();
4151 if (!TheLoop->contains(Parent)) {
4152 // Exit if you find multiple outside users or if the header phi node is
4153 // being used. In this case the user uses the value of the previous
4154 // iteration, in which case we would loose "VF-1" iterations of the
4155 // reduction operation if we vectorize.
4156 if (ExitInstruction != nullptr || Cur == Phi)
4159 // The instruction used by an outside user must be the last instruction
4160 // before we feed back to the reduction phi. Otherwise, we loose VF-1
4161 // operations on the value.
4162 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
4165 ExitInstruction = Cur;
4169 // Process instructions only once (termination). Each reduction cycle
4170 // value must only be used once, except by phi nodes and min/max
4171 // reductions which are represented as a cmp followed by a select.
4172 ReductionInstDesc IgnoredVal(false, nullptr);
4173 if (VisitedInsts.insert(UI).second) {
4174 if (isa<PHINode>(UI))
4177 NonPHIs.push_back(UI);
4178 } else if (!isa<PHINode>(UI) &&
4179 ((!isa<FCmpInst>(UI) &&
4180 !isa<ICmpInst>(UI) &&
4181 !isa<SelectInst>(UI)) ||
4182 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
4185 // Remember that we completed the cycle.
4187 FoundStartPHI = true;
4189 Worklist.append(PHIs.begin(), PHIs.end());
4190 Worklist.append(NonPHIs.begin(), NonPHIs.end());
4193 // This means we have seen one but not the other instruction of the
4194 // pattern or more than just a select and cmp.
4195 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
4196 NumCmpSelectPatternInst != 2)
4199 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
4202 // We found a reduction var if we have reached the original phi node and we
4203 // only have a single instruction with out-of-loop users.
4205 // This instruction is allowed to have out-of-loop users.
4206 AllowedExit.insert(ExitInstruction);
4208 // Save the description of this reduction variable.
4209 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
4210 ReduxDesc.MinMaxKind);
4211 Reductions[Phi] = RD;
4212 // We've ended the cycle. This is a reduction variable if we have an
4213 // outside user and it has a binary op.
4218 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
4219 /// pattern corresponding to a min(X, Y) or max(X, Y).
4220 LoopVectorizationLegality::ReductionInstDesc
4221 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
4222 ReductionInstDesc &Prev) {
4224 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
4225 "Expect a select instruction");
4226 Instruction *Cmp = nullptr;
4227 SelectInst *Select = nullptr;
4229 // We must handle the select(cmp()) as a single instruction. Advance to the
4231 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
4232 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
4233 return ReductionInstDesc(false, I);
4234 return ReductionInstDesc(Select, Prev.MinMaxKind);
4237 // Only handle single use cases for now.
4238 if (!(Select = dyn_cast<SelectInst>(I)))
4239 return ReductionInstDesc(false, I);
4240 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
4241 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
4242 return ReductionInstDesc(false, I);
4243 if (!Cmp->hasOneUse())
4244 return ReductionInstDesc(false, I);
4249 // Look for a min/max pattern.
4250 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4251 return ReductionInstDesc(Select, MRK_UIntMin);
4252 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4253 return ReductionInstDesc(Select, MRK_UIntMax);
4254 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4255 return ReductionInstDesc(Select, MRK_SIntMax);
4256 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4257 return ReductionInstDesc(Select, MRK_SIntMin);
4258 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4259 return ReductionInstDesc(Select, MRK_FloatMin);
4260 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4261 return ReductionInstDesc(Select, MRK_FloatMax);
4262 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4263 return ReductionInstDesc(Select, MRK_FloatMin);
4264 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4265 return ReductionInstDesc(Select, MRK_FloatMax);
4267 return ReductionInstDesc(false, I);
4270 LoopVectorizationLegality::ReductionInstDesc
4271 LoopVectorizationLegality::isReductionInstr(Instruction *I,
4273 ReductionInstDesc &Prev) {
4274 bool FP = I->getType()->isFloatingPointTy();
4275 bool FastMath = FP && I->hasUnsafeAlgebra();
4276 switch (I->getOpcode()) {
4278 return ReductionInstDesc(false, I);
4279 case Instruction::PHI:
4280 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
4281 Kind != RK_FloatMinMax))
4282 return ReductionInstDesc(false, I);
4283 return ReductionInstDesc(I, Prev.MinMaxKind);
4284 case Instruction::Sub:
4285 case Instruction::Add:
4286 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4287 case Instruction::Mul:
4288 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4289 case Instruction::And:
4290 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4291 case Instruction::Or:
4292 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4293 case Instruction::Xor:
4294 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4295 case Instruction::FMul:
4296 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4297 case Instruction::FSub:
4298 case Instruction::FAdd:
4299 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4300 case Instruction::FCmp:
4301 case Instruction::ICmp:
4302 case Instruction::Select:
4303 if (Kind != RK_IntegerMinMax &&
4304 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4305 return ReductionInstDesc(false, I);
4306 return isMinMaxSelectCmpPattern(I, Prev);
4310 LoopVectorizationLegality::InductionKind
4311 LoopVectorizationLegality::isInductionVariable(PHINode *Phi,
4312 ConstantInt *&StepValue) {
4313 Type *PhiTy = Phi->getType();
4314 // We only handle integer and pointer inductions variables.
4315 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4316 return IK_NoInduction;
4318 // Check that the PHI is consecutive.
4319 const SCEV *PhiScev = SE->getSCEV(Phi);
4320 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4322 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4323 return IK_NoInduction;
4326 const SCEV *Step = AR->getStepRecurrence(*SE);
4327 // Calculate the pointer stride and check if it is consecutive.
4328 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4330 return IK_NoInduction;
4332 ConstantInt *CV = C->getValue();
4333 if (PhiTy->isIntegerTy()) {
4335 return IK_IntInduction;
4338 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4339 Type *PointerElementType = PhiTy->getPointerElementType();
4340 // The pointer stride cannot be determined if the pointer element type is not
4342 if (!PointerElementType->isSized())
4343 return IK_NoInduction;
4345 const DataLayout &DL = Phi->getModule()->getDataLayout();
4346 int64_t Size = static_cast<int64_t>(DL.getTypeAllocSize(PointerElementType));
4347 int64_t CVSize = CV->getSExtValue();
4349 return IK_NoInduction;
4350 StepValue = ConstantInt::getSigned(CV->getType(), CVSize / Size);
4351 return IK_PtrInduction;
4354 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4355 Value *In0 = const_cast<Value*>(V);
4356 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4360 return Inductions.count(PN);
4363 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4364 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4367 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4368 SmallPtrSetImpl<Value *> &SafePtrs) {
4370 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4371 // Check that we don't have a constant expression that can trap as operand.
4372 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4374 if (Constant *C = dyn_cast<Constant>(*OI))
4378 // We might be able to hoist the load.
4379 if (it->mayReadFromMemory()) {
4380 LoadInst *LI = dyn_cast<LoadInst>(it);
4383 if (!SafePtrs.count(LI->getPointerOperand())) {
4384 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
4385 MaskedOp.insert(LI);
4392 // We don't predicate stores at the moment.
4393 if (it->mayWriteToMemory()) {
4394 StoreInst *SI = dyn_cast<StoreInst>(it);
4395 // We only support predication of stores in basic blocks with one
4400 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
4401 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
4403 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
4404 !isSinglePredecessor) {
4405 // Build a masked store if it is legal for the target, otherwise scalarize
4407 bool isLegalMaskedOp =
4408 isLegalMaskedStore(SI->getValueOperand()->getType(),
4409 SI->getPointerOperand());
4410 if (isLegalMaskedOp) {
4412 MaskedOp.insert(SI);
4421 // The instructions below can trap.
4422 switch (it->getOpcode()) {
4424 case Instruction::UDiv:
4425 case Instruction::SDiv:
4426 case Instruction::URem:
4427 case Instruction::SRem:
4435 LoopVectorizationCostModel::VectorizationFactor
4436 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
4437 // Width 1 means no vectorize
4438 VectorizationFactor Factor = { 1U, 0U };
4439 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4440 emitAnalysis(VectorizationReport() <<
4441 "runtime pointer checks needed. Enable vectorization of this "
4442 "loop with '#pragma clang loop vectorize(enable)' when "
4443 "compiling with -Os");
4444 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4448 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
4449 emitAnalysis(VectorizationReport() <<
4450 "store that is conditionally executed prevents vectorization");
4451 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
4455 // Find the trip count.
4456 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4457 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4459 unsigned WidestType = getWidestType();
4460 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4461 unsigned MaxSafeDepDist = -1U;
4462 if (Legal->getMaxSafeDepDistBytes() != -1U)
4463 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4464 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4465 WidestRegister : MaxSafeDepDist);
4466 unsigned MaxVectorSize = WidestRegister / WidestType;
4467 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4468 DEBUG(dbgs() << "LV: The Widest register is: "
4469 << WidestRegister << " bits.\n");
4471 if (MaxVectorSize == 0) {
4472 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4476 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
4477 " into one vector!");
4479 unsigned VF = MaxVectorSize;
4481 // If we optimize the program for size, avoid creating the tail loop.
4483 // If we are unable to calculate the trip count then don't try to vectorize.
4486 (VectorizationReport() <<
4487 "unable to calculate the loop count due to complex control flow");
4488 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4492 // Find the maximum SIMD width that can fit within the trip count.
4493 VF = TC % MaxVectorSize;
4498 // If the trip count that we found modulo the vectorization factor is not
4499 // zero then we require a tail.
4501 emitAnalysis(VectorizationReport() <<
4502 "cannot optimize for size and vectorize at the "
4503 "same time. Enable vectorization of this loop "
4504 "with '#pragma clang loop vectorize(enable)' "
4505 "when compiling with -Os");
4506 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4511 int UserVF = Hints->getWidth();
4513 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4514 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4516 Factor.Width = UserVF;
4520 float Cost = expectedCost(1);
4522 const float ScalarCost = Cost;
4525 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
4527 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
4528 // Ignore scalar width, because the user explicitly wants vectorization.
4529 if (ForceVectorization && VF > 1) {
4531 Cost = expectedCost(Width) / (float)Width;
4534 for (unsigned i=2; i <= VF; i*=2) {
4535 // Notice that the vector loop needs to be executed less times, so
4536 // we need to divide the cost of the vector loops by the width of
4537 // the vector elements.
4538 float VectorCost = expectedCost(i) / (float)i;
4539 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
4540 (int)VectorCost << ".\n");
4541 if (VectorCost < Cost) {
4547 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
4548 << "LV: Vectorization seems to be not beneficial, "
4549 << "but was forced by a user.\n");
4550 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
4551 Factor.Width = Width;
4552 Factor.Cost = Width * Cost;
4556 unsigned LoopVectorizationCostModel::getWidestType() {
4557 unsigned MaxWidth = 8;
4558 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
4561 for (Loop::block_iterator bb = TheLoop->block_begin(),
4562 be = TheLoop->block_end(); bb != be; ++bb) {
4563 BasicBlock *BB = *bb;
4565 // For each instruction in the loop.
4566 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4567 Type *T = it->getType();
4569 // Ignore ephemeral values.
4570 if (EphValues.count(it))
4573 // Only examine Loads, Stores and PHINodes.
4574 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4577 // Examine PHI nodes that are reduction variables.
4578 if (PHINode *PN = dyn_cast<PHINode>(it))
4579 if (!Legal->getReductionVars()->count(PN))
4582 // Examine the stored values.
4583 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4584 T = ST->getValueOperand()->getType();
4586 // Ignore loaded pointer types and stored pointer types that are not
4587 // consecutive. However, we do want to take consecutive stores/loads of
4588 // pointer vectors into account.
4589 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4592 MaxWidth = std::max(MaxWidth,
4593 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
4601 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
4603 unsigned LoopCost) {
4605 // -- The unroll heuristics --
4606 // We unroll the loop in order to expose ILP and reduce the loop overhead.
4607 // There are many micro-architectural considerations that we can't predict
4608 // at this level. For example, frontend pressure (on decode or fetch) due to
4609 // code size, or the number and capabilities of the execution ports.
4611 // We use the following heuristics to select the unroll factor:
4612 // 1. If the code has reductions, then we unroll in order to break the cross
4613 // iteration dependency.
4614 // 2. If the loop is really small, then we unroll in order to reduce the loop
4616 // 3. We don't unroll if we think that we will spill registers to memory due
4617 // to the increased register pressure.
4619 // Use the user preference, unless 'auto' is selected.
4620 int UserUF = Hints->getInterleave();
4624 // When we optimize for size, we don't unroll.
4628 // We used the distance for the unroll factor.
4629 if (Legal->getMaxSafeDepDistBytes() != -1U)
4632 // Do not unroll loops with a relatively small trip count.
4633 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4634 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
4637 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
4638 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
4642 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
4643 TargetNumRegisters = ForceTargetNumScalarRegs;
4645 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
4646 TargetNumRegisters = ForceTargetNumVectorRegs;
4649 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4650 // We divide by these constants so assume that we have at least one
4651 // instruction that uses at least one register.
4652 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4653 R.NumInstructions = std::max(R.NumInstructions, 1U);
4655 // We calculate the unroll factor using the following formula.
4656 // Subtract the number of loop invariants from the number of available
4657 // registers. These registers are used by all of the unrolled instances.
4658 // Next, divide the remaining registers by the number of registers that is
4659 // required by the loop, in order to estimate how many parallel instances
4660 // fit without causing spills. All of this is rounded down if necessary to be
4661 // a power of two. We want power of two unroll factors to simplify any
4662 // addressing operations or alignment considerations.
4663 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
4666 // Don't count the induction variable as unrolled.
4667 if (EnableIndVarRegisterHeur)
4668 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
4669 std::max(1U, (R.MaxLocalUsers - 1)));
4671 // Clamp the unroll factor ranges to reasonable factors.
4672 unsigned MaxInterleaveSize = TTI.getMaxInterleaveFactor();
4674 // Check if the user has overridden the unroll max.
4676 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
4677 MaxInterleaveSize = ForceTargetMaxScalarInterleaveFactor;
4679 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
4680 MaxInterleaveSize = ForceTargetMaxVectorInterleaveFactor;
4683 // If we did not calculate the cost for VF (because the user selected the VF)
4684 // then we calculate the cost of VF here.
4686 LoopCost = expectedCost(VF);
4688 // Clamp the calculated UF to be between the 1 and the max unroll factor
4689 // that the target allows.
4690 if (UF > MaxInterleaveSize)
4691 UF = MaxInterleaveSize;
4695 // Unroll if we vectorized this loop and there is a reduction that could
4696 // benefit from unrolling.
4697 if (VF > 1 && Legal->getReductionVars()->size()) {
4698 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
4702 // Note that if we've already vectorized the loop we will have done the
4703 // runtime check and so unrolling won't require further checks.
4704 bool UnrollingRequiresRuntimePointerCheck =
4705 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
4707 // We want to unroll small loops in order to reduce the loop overhead and
4708 // potentially expose ILP opportunities.
4709 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
4710 if (!UnrollingRequiresRuntimePointerCheck &&
4711 LoopCost < SmallLoopCost) {
4712 // We assume that the cost overhead is 1 and we use the cost model
4713 // to estimate the cost of the loop and unroll until the cost of the
4714 // loop overhead is about 5% of the cost of the loop.
4715 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
4717 // Unroll until store/load ports (estimated by max unroll factor) are
4719 unsigned NumStores = Legal->getNumStores();
4720 unsigned NumLoads = Legal->getNumLoads();
4721 unsigned StoresUF = UF / (NumStores ? NumStores : 1);
4722 unsigned LoadsUF = UF / (NumLoads ? NumLoads : 1);
4724 // If we have a scalar reduction (vector reductions are already dealt with
4725 // by this point), we can increase the critical path length if the loop
4726 // we're unrolling is inside another loop. Limit, by default to 2, so the
4727 // critical path only gets increased by one reduction operation.
4728 if (Legal->getReductionVars()->size() &&
4729 TheLoop->getLoopDepth() > 1) {
4730 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionUF);
4731 SmallUF = std::min(SmallUF, F);
4732 StoresUF = std::min(StoresUF, F);
4733 LoadsUF = std::min(LoadsUF, F);
4736 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
4737 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
4738 return std::max(StoresUF, LoadsUF);
4741 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
4745 // Unroll if this is a large loop (small loops are already dealt with by this
4746 // point) that could benefit from interleaved unrolling.
4747 bool HasReductions = (Legal->getReductionVars()->size() > 0);
4748 if (TTI.enableAggressiveInterleaving(HasReductions)) {
4749 DEBUG(dbgs() << "LV: Unrolling to expose ILP.\n");
4753 DEBUG(dbgs() << "LV: Not Unrolling.\n");
4757 LoopVectorizationCostModel::RegisterUsage
4758 LoopVectorizationCostModel::calculateRegisterUsage() {
4759 // This function calculates the register usage by measuring the highest number
4760 // of values that are alive at a single location. Obviously, this is a very
4761 // rough estimation. We scan the loop in a topological order in order and
4762 // assign a number to each instruction. We use RPO to ensure that defs are
4763 // met before their users. We assume that each instruction that has in-loop
4764 // users starts an interval. We record every time that an in-loop value is
4765 // used, so we have a list of the first and last occurrences of each
4766 // instruction. Next, we transpose this data structure into a multi map that
4767 // holds the list of intervals that *end* at a specific location. This multi
4768 // map allows us to perform a linear search. We scan the instructions linearly
4769 // and record each time that a new interval starts, by placing it in a set.
4770 // If we find this value in the multi-map then we remove it from the set.
4771 // The max register usage is the maximum size of the set.
4772 // We also search for instructions that are defined outside the loop, but are
4773 // used inside the loop. We need this number separately from the max-interval
4774 // usage number because when we unroll, loop-invariant values do not take
4776 LoopBlocksDFS DFS(TheLoop);
4780 R.NumInstructions = 0;
4782 // Each 'key' in the map opens a new interval. The values
4783 // of the map are the index of the 'last seen' usage of the
4784 // instruction that is the key.
4785 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4786 // Maps instruction to its index.
4787 DenseMap<unsigned, Instruction*> IdxToInstr;
4788 // Marks the end of each interval.
4789 IntervalMap EndPoint;
4790 // Saves the list of instruction indices that are used in the loop.
4791 SmallSet<Instruction*, 8> Ends;
4792 // Saves the list of values that are used in the loop but are
4793 // defined outside the loop, such as arguments and constants.
4794 SmallPtrSet<Value*, 8> LoopInvariants;
4797 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4798 be = DFS.endRPO(); bb != be; ++bb) {
4799 R.NumInstructions += (*bb)->size();
4800 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4802 Instruction *I = it;
4803 IdxToInstr[Index++] = I;
4805 // Save the end location of each USE.
4806 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4807 Value *U = I->getOperand(i);
4808 Instruction *Instr = dyn_cast<Instruction>(U);
4810 // Ignore non-instruction values such as arguments, constants, etc.
4811 if (!Instr) continue;
4813 // If this instruction is outside the loop then record it and continue.
4814 if (!TheLoop->contains(Instr)) {
4815 LoopInvariants.insert(Instr);
4819 // Overwrite previous end points.
4820 EndPoint[Instr] = Index;
4826 // Saves the list of intervals that end with the index in 'key'.
4827 typedef SmallVector<Instruction*, 2> InstrList;
4828 DenseMap<unsigned, InstrList> TransposeEnds;
4830 // Transpose the EndPoints to a list of values that end at each index.
4831 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4833 TransposeEnds[it->second].push_back(it->first);
4835 SmallSet<Instruction*, 8> OpenIntervals;
4836 unsigned MaxUsage = 0;
4839 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4840 for (unsigned int i = 0; i < Index; ++i) {
4841 Instruction *I = IdxToInstr[i];
4842 // Ignore instructions that are never used within the loop.
4843 if (!Ends.count(I)) continue;
4845 // Ignore ephemeral values.
4846 if (EphValues.count(I))
4849 // Remove all of the instructions that end at this location.
4850 InstrList &List = TransposeEnds[i];
4851 for (unsigned int j=0, e = List.size(); j < e; ++j)
4852 OpenIntervals.erase(List[j]);
4854 // Count the number of live interals.
4855 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4857 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4858 OpenIntervals.size() << '\n');
4860 // Add the current instruction to the list of open intervals.
4861 OpenIntervals.insert(I);
4864 unsigned Invariant = LoopInvariants.size();
4865 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
4866 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
4867 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
4869 R.LoopInvariantRegs = Invariant;
4870 R.MaxLocalUsers = MaxUsage;
4874 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4878 for (Loop::block_iterator bb = TheLoop->block_begin(),
4879 be = TheLoop->block_end(); bb != be; ++bb) {
4880 unsigned BlockCost = 0;
4881 BasicBlock *BB = *bb;
4883 // For each instruction in the old loop.
4884 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4885 // Skip dbg intrinsics.
4886 if (isa<DbgInfoIntrinsic>(it))
4889 // Ignore ephemeral values.
4890 if (EphValues.count(it))
4893 unsigned C = getInstructionCost(it, VF);
4895 // Check if we should override the cost.
4896 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
4897 C = ForceTargetInstructionCost;
4900 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
4901 VF << " For instruction: " << *it << '\n');
4904 // We assume that if-converted blocks have a 50% chance of being executed.
4905 // When the code is scalar then some of the blocks are avoided due to CF.
4906 // When the code is vectorized we execute all code paths.
4907 if (VF == 1 && Legal->blockNeedsPredication(*bb))
4916 /// \brief Check whether the address computation for a non-consecutive memory
4917 /// access looks like an unlikely candidate for being merged into the indexing
4920 /// We look for a GEP which has one index that is an induction variable and all
4921 /// other indices are loop invariant. If the stride of this access is also
4922 /// within a small bound we decide that this address computation can likely be
4923 /// merged into the addressing mode.
4924 /// In all other cases, we identify the address computation as complex.
4925 static bool isLikelyComplexAddressComputation(Value *Ptr,
4926 LoopVectorizationLegality *Legal,
4927 ScalarEvolution *SE,
4928 const Loop *TheLoop) {
4929 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
4933 // We are looking for a gep with all loop invariant indices except for one
4934 // which should be an induction variable.
4935 unsigned NumOperands = Gep->getNumOperands();
4936 for (unsigned i = 1; i < NumOperands; ++i) {
4937 Value *Opd = Gep->getOperand(i);
4938 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
4939 !Legal->isInductionVariable(Opd))
4943 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
4944 // can likely be merged into the address computation.
4945 unsigned MaxMergeDistance = 64;
4947 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
4951 // Check the step is constant.
4952 const SCEV *Step = AddRec->getStepRecurrence(*SE);
4953 // Calculate the pointer stride and check if it is consecutive.
4954 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4958 const APInt &APStepVal = C->getValue()->getValue();
4960 // Huge step value - give up.
4961 if (APStepVal.getBitWidth() > 64)
4964 int64_t StepVal = APStepVal.getSExtValue();
4966 return StepVal > MaxMergeDistance;
4969 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
4970 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
4976 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
4977 // If we know that this instruction will remain uniform, check the cost of
4978 // the scalar version.
4979 if (Legal->isUniformAfterVectorization(I))
4982 Type *RetTy = I->getType();
4983 Type *VectorTy = ToVectorTy(RetTy, VF);
4985 // TODO: We need to estimate the cost of intrinsic calls.
4986 switch (I->getOpcode()) {
4987 case Instruction::GetElementPtr:
4988 // We mark this instruction as zero-cost because the cost of GEPs in
4989 // vectorized code depends on whether the corresponding memory instruction
4990 // is scalarized or not. Therefore, we handle GEPs with the memory
4991 // instruction cost.
4993 case Instruction::Br: {
4994 return TTI.getCFInstrCost(I->getOpcode());
4996 case Instruction::PHI:
4997 //TODO: IF-converted IFs become selects.
4999 case Instruction::Add:
5000 case Instruction::FAdd:
5001 case Instruction::Sub:
5002 case Instruction::FSub:
5003 case Instruction::Mul:
5004 case Instruction::FMul:
5005 case Instruction::UDiv:
5006 case Instruction::SDiv:
5007 case Instruction::FDiv:
5008 case Instruction::URem:
5009 case Instruction::SRem:
5010 case Instruction::FRem:
5011 case Instruction::Shl:
5012 case Instruction::LShr:
5013 case Instruction::AShr:
5014 case Instruction::And:
5015 case Instruction::Or:
5016 case Instruction::Xor: {
5017 // Since we will replace the stride by 1 the multiplication should go away.
5018 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5020 // Certain instructions can be cheaper to vectorize if they have a constant
5021 // second vector operand. One example of this are shifts on x86.
5022 TargetTransformInfo::OperandValueKind Op1VK =
5023 TargetTransformInfo::OK_AnyValue;
5024 TargetTransformInfo::OperandValueKind Op2VK =
5025 TargetTransformInfo::OK_AnyValue;
5026 TargetTransformInfo::OperandValueProperties Op1VP =
5027 TargetTransformInfo::OP_None;
5028 TargetTransformInfo::OperandValueProperties Op2VP =
5029 TargetTransformInfo::OP_None;
5030 Value *Op2 = I->getOperand(1);
5032 // Check for a splat of a constant or for a non uniform vector of constants.
5033 if (isa<ConstantInt>(Op2)) {
5034 ConstantInt *CInt = cast<ConstantInt>(Op2);
5035 if (CInt && CInt->getValue().isPowerOf2())
5036 Op2VP = TargetTransformInfo::OP_PowerOf2;
5037 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5038 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5039 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5040 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
5042 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
5043 if (CInt && CInt->getValue().isPowerOf2())
5044 Op2VP = TargetTransformInfo::OP_PowerOf2;
5045 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5049 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
5052 case Instruction::Select: {
5053 SelectInst *SI = cast<SelectInst>(I);
5054 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5055 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5056 Type *CondTy = SI->getCondition()->getType();
5058 CondTy = VectorType::get(CondTy, VF);
5060 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5062 case Instruction::ICmp:
5063 case Instruction::FCmp: {
5064 Type *ValTy = I->getOperand(0)->getType();
5065 VectorTy = ToVectorTy(ValTy, VF);
5066 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5068 case Instruction::Store:
5069 case Instruction::Load: {
5070 StoreInst *SI = dyn_cast<StoreInst>(I);
5071 LoadInst *LI = dyn_cast<LoadInst>(I);
5072 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5074 VectorTy = ToVectorTy(ValTy, VF);
5076 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5077 unsigned AS = SI ? SI->getPointerAddressSpace() :
5078 LI->getPointerAddressSpace();
5079 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5080 // We add the cost of address computation here instead of with the gep
5081 // instruction because only here we know whether the operation is
5084 return TTI.getAddressComputationCost(VectorTy) +
5085 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5087 // Scalarized loads/stores.
5088 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5089 bool Reverse = ConsecutiveStride < 0;
5090 const DataLayout &DL = I->getModule()->getDataLayout();
5091 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ValTy);
5092 unsigned VectorElementSize = DL.getTypeStoreSize(VectorTy) / VF;
5093 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5094 bool IsComplexComputation =
5095 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5097 // The cost of extracting from the value vector and pointer vector.
5098 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5099 for (unsigned i = 0; i < VF; ++i) {
5100 // The cost of extracting the pointer operand.
5101 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5102 // In case of STORE, the cost of ExtractElement from the vector.
5103 // In case of LOAD, the cost of InsertElement into the returned
5105 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5106 Instruction::InsertElement,
5110 // The cost of the scalar loads/stores.
5111 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5112 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5117 // Wide load/stores.
5118 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5119 if (Legal->isMaskRequired(I))
5120 Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment,
5123 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5126 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5130 case Instruction::ZExt:
5131 case Instruction::SExt:
5132 case Instruction::FPToUI:
5133 case Instruction::FPToSI:
5134 case Instruction::FPExt:
5135 case Instruction::PtrToInt:
5136 case Instruction::IntToPtr:
5137 case Instruction::SIToFP:
5138 case Instruction::UIToFP:
5139 case Instruction::Trunc:
5140 case Instruction::FPTrunc:
5141 case Instruction::BitCast: {
5142 // We optimize the truncation of induction variable.
5143 // The cost of these is the same as the scalar operation.
5144 if (I->getOpcode() == Instruction::Trunc &&
5145 Legal->isInductionVariable(I->getOperand(0)))
5146 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5147 I->getOperand(0)->getType());
5149 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5150 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5152 case Instruction::Call: {
5153 bool NeedToScalarize;
5154 CallInst *CI = cast<CallInst>(I);
5155 unsigned CallCost = getVectorCallCost(CI, VF, TTI, TLI, NeedToScalarize);
5156 if (getIntrinsicIDForCall(CI, TLI))
5157 return std::min(CallCost, getVectorIntrinsicCost(CI, VF, TTI, TLI));
5161 // We are scalarizing the instruction. Return the cost of the scalar
5162 // instruction, plus the cost of insert and extract into vector
5163 // elements, times the vector width.
5166 if (!RetTy->isVoidTy() && VF != 1) {
5167 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5169 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5172 // The cost of inserting the results plus extracting each one of the
5174 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5177 // The cost of executing VF copies of the scalar instruction. This opcode
5178 // is unknown. Assume that it is the same as 'mul'.
5179 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5185 char LoopVectorize::ID = 0;
5186 static const char lv_name[] = "Loop Vectorization";
5187 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5188 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
5189 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
5190 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
5191 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
5192 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5193 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5194 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5195 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
5196 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5197 INITIALIZE_PASS_DEPENDENCY(LoopAccessAnalysis)
5198 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5201 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5202 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5206 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5207 // Check for a store.
5208 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5209 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5211 // Check for a load.
5212 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5213 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5219 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5220 bool IfPredicateStore) {
5221 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5222 // Holds vector parameters or scalars, in case of uniform vals.
5223 SmallVector<VectorParts, 4> Params;
5225 setDebugLocFromInst(Builder, Instr);
5227 // Find all of the vectorized parameters.
5228 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5229 Value *SrcOp = Instr->getOperand(op);
5231 // If we are accessing the old induction variable, use the new one.
5232 if (SrcOp == OldInduction) {
5233 Params.push_back(getVectorValue(SrcOp));
5237 // Try using previously calculated values.
5238 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5240 // If the src is an instruction that appeared earlier in the basic block
5241 // then it should already be vectorized.
5242 if (SrcInst && OrigLoop->contains(SrcInst)) {
5243 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5244 // The parameter is a vector value from earlier.
5245 Params.push_back(WidenMap.get(SrcInst));
5247 // The parameter is a scalar from outside the loop. Maybe even a constant.
5248 VectorParts Scalars;
5249 Scalars.append(UF, SrcOp);
5250 Params.push_back(Scalars);
5254 assert(Params.size() == Instr->getNumOperands() &&
5255 "Invalid number of operands");
5257 // Does this instruction return a value ?
5258 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5260 Value *UndefVec = IsVoidRetTy ? nullptr :
5261 UndefValue::get(Instr->getType());
5262 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5263 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5265 Instruction *InsertPt = Builder.GetInsertPoint();
5266 BasicBlock *IfBlock = Builder.GetInsertBlock();
5267 BasicBlock *CondBlock = nullptr;
5270 Loop *VectorLp = nullptr;
5271 if (IfPredicateStore) {
5272 assert(Instr->getParent()->getSinglePredecessor() &&
5273 "Only support single predecessor blocks");
5274 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5275 Instr->getParent());
5276 VectorLp = LI->getLoopFor(IfBlock);
5277 assert(VectorLp && "Must have a loop for this block");
5280 // For each vector unroll 'part':
5281 for (unsigned Part = 0; Part < UF; ++Part) {
5282 // For each scalar that we create:
5284 // Start an "if (pred) a[i] = ..." block.
5285 Value *Cmp = nullptr;
5286 if (IfPredicateStore) {
5287 if (Cond[Part]->getType()->isVectorTy())
5289 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5290 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5291 ConstantInt::get(Cond[Part]->getType(), 1));
5292 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5293 LoopVectorBody.push_back(CondBlock);
5294 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
5295 // Update Builder with newly created basic block.
5296 Builder.SetInsertPoint(InsertPt);
5299 Instruction *Cloned = Instr->clone();
5301 Cloned->setName(Instr->getName() + ".cloned");
5302 // Replace the operands of the cloned instructions with extracted scalars.
5303 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5304 Value *Op = Params[op][Part];
5305 Cloned->setOperand(op, Op);
5308 // Place the cloned scalar in the new loop.
5309 Builder.Insert(Cloned);
5311 // If the original scalar returns a value we need to place it in a vector
5312 // so that future users will be able to use it.
5314 VecResults[Part] = Cloned;
5317 if (IfPredicateStore) {
5318 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
5319 LoopVectorBody.push_back(NewIfBlock);
5320 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
5321 Builder.SetInsertPoint(InsertPt);
5322 Instruction *OldBr = IfBlock->getTerminator();
5323 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
5324 OldBr->eraseFromParent();
5325 IfBlock = NewIfBlock;
5330 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5331 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5332 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5334 return scalarizeInstruction(Instr, IfPredicateStore);
5337 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5341 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5345 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) {
5346 // When unrolling and the VF is 1, we only need to add a simple scalar.
5347 Type *ITy = Val->getType();
5348 assert(!ITy->isVectorTy() && "Val must be a scalar");
5349 Constant *C = ConstantInt::get(ITy, StartIdx);
5350 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");