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"
96 #include "llvm/Transforms/Utils/LoopUtils.h"
101 using namespace llvm;
102 using namespace llvm::PatternMatch;
104 #define LV_NAME "loop-vectorize"
105 #define DEBUG_TYPE LV_NAME
107 STATISTIC(LoopsVectorized, "Number of loops vectorized");
108 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
111 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
112 cl::desc("Enable if-conversion during vectorization."));
114 /// We don't vectorize loops with a known constant trip count below this number.
115 static cl::opt<unsigned>
116 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
118 cl::desc("Don't vectorize loops with a constant "
119 "trip count that is smaller than this "
122 /// This enables versioning on the strides of symbolically striding memory
123 /// accesses in code like the following.
124 /// for (i = 0; i < N; ++i)
125 /// A[i * Stride1] += B[i * Stride2] ...
127 /// Will be roughly translated to
128 /// if (Stride1 == 1 && Stride2 == 1) {
129 /// for (i = 0; i < N; i+=4)
133 static cl::opt<bool> EnableMemAccessVersioning(
134 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
135 cl::desc("Enable symblic stride memory access versioning"));
137 /// We don't unroll loops with a known constant trip count below this number.
138 static const unsigned TinyTripCountUnrollThreshold = 128;
140 static cl::opt<unsigned> ForceTargetNumScalarRegs(
141 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
142 cl::desc("A flag that overrides the target's number of scalar registers."));
144 static cl::opt<unsigned> ForceTargetNumVectorRegs(
145 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
146 cl::desc("A flag that overrides the target's number of vector registers."));
148 /// Maximum vectorization interleave count.
149 static const unsigned MaxInterleaveFactor = 16;
151 static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
152 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
153 cl::desc("A flag that overrides the target's max interleave factor for "
156 static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
157 "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
158 cl::desc("A flag that overrides the target's max interleave factor for "
159 "vectorized loops."));
161 static cl::opt<unsigned> ForceTargetInstructionCost(
162 "force-target-instruction-cost", cl::init(0), cl::Hidden,
163 cl::desc("A flag that overrides the target's expected cost for "
164 "an instruction to a single constant value. Mostly "
165 "useful for getting consistent testing."));
167 static cl::opt<unsigned> SmallLoopCost(
168 "small-loop-cost", cl::init(20), cl::Hidden,
169 cl::desc("The cost of a loop that is considered 'small' by the unroller."));
171 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
172 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
173 cl::desc("Enable the use of the block frequency analysis to access PGO "
174 "heuristics minimizing code growth in cold regions and being more "
175 "aggressive in hot regions."));
177 // Runtime unroll loops for load/store throughput.
178 static cl::opt<bool> EnableLoadStoreRuntimeUnroll(
179 "enable-loadstore-runtime-unroll", cl::init(true), cl::Hidden,
180 cl::desc("Enable runtime unrolling until load/store ports are saturated"));
182 /// The number of stores in a loop that are allowed to need predication.
183 static cl::opt<unsigned> NumberOfStoresToPredicate(
184 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
185 cl::desc("Max number of stores to be predicated behind an if."));
187 static cl::opt<bool> EnableIndVarRegisterHeur(
188 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
189 cl::desc("Count the induction variable only once when unrolling"));
191 static cl::opt<bool> EnableCondStoresVectorization(
192 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
193 cl::desc("Enable if predication of stores during vectorization."));
195 static cl::opt<unsigned> MaxNestedScalarReductionUF(
196 "max-nested-scalar-reduction-unroll", cl::init(2), cl::Hidden,
197 cl::desc("The maximum unroll factor to use when unrolling a scalar "
198 "reduction in a nested loop."));
202 // Forward declarations.
203 class LoopVectorizationLegality;
204 class LoopVectorizationCostModel;
205 class LoopVectorizeHints;
207 /// \brief This modifies LoopAccessReport to initialize message with
208 /// loop-vectorizer-specific part.
209 class VectorizationReport : public LoopAccessReport {
211 VectorizationReport(Instruction *I = nullptr)
212 : LoopAccessReport("loop not vectorized: ", I) {}
214 /// \brief This allows promotion of the loop-access analysis report into the
215 /// loop-vectorizer report. It modifies the message to add the
216 /// loop-vectorizer-specific part of the message.
217 explicit VectorizationReport(const LoopAccessReport &R)
218 : LoopAccessReport(Twine("loop not vectorized: ") + R.str(),
222 /// A helper function for converting Scalar types to vector types.
223 /// If the incoming type is void, we return void. If the VF is 1, we return
225 static Type* ToVectorTy(Type *Scalar, unsigned VF) {
226 if (Scalar->isVoidTy() || VF == 1)
228 return VectorType::get(Scalar, VF);
231 /// InnerLoopVectorizer vectorizes loops which contain only one basic
232 /// block to a specified vectorization factor (VF).
233 /// This class performs the widening of scalars into vectors, or multiple
234 /// scalars. This class also implements the following features:
235 /// * It inserts an epilogue loop for handling loops that don't have iteration
236 /// counts that are known to be a multiple of the vectorization factor.
237 /// * It handles the code generation for reduction variables.
238 /// * Scalarization (implementation using scalars) of un-vectorizable
240 /// InnerLoopVectorizer does not perform any vectorization-legality
241 /// checks, and relies on the caller to check for the different legality
242 /// aspects. The InnerLoopVectorizer relies on the
243 /// LoopVectorizationLegality class to provide information about the induction
244 /// and reduction variables that were found to a given vectorization factor.
245 class InnerLoopVectorizer {
247 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
248 DominatorTree *DT, const TargetLibraryInfo *TLI,
249 const TargetTransformInfo *TTI, unsigned VecWidth,
250 unsigned UnrollFactor)
251 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), TLI(TLI), TTI(TTI),
252 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
253 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
254 Legal(nullptr), AddedSafetyChecks(false) {}
256 // Perform the actual loop widening (vectorization).
257 void vectorize(LoopVectorizationLegality *L) {
259 // Create a new empty loop. Unlink the old loop and connect the new one.
261 // Widen each instruction in the old loop to a new one in the new loop.
262 // Use the Legality module to find the induction and reduction variables.
264 // Register the new loop and update the analysis passes.
268 // Return true if any runtime check is added.
269 bool IsSafetyChecksAdded() {
270 return AddedSafetyChecks;
273 virtual ~InnerLoopVectorizer() {}
276 /// A small list of PHINodes.
277 typedef SmallVector<PHINode*, 4> PhiVector;
278 /// When we unroll loops we have multiple vector values for each scalar.
279 /// This data structure holds the unrolled and vectorized values that
280 /// originated from one scalar instruction.
281 typedef SmallVector<Value*, 2> VectorParts;
283 // When we if-convert we need create edge masks. We have to cache values so
284 // that we don't end up with exponential recursion/IR.
285 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
286 VectorParts> EdgeMaskCache;
288 /// \brief Add checks for strides that where assumed to be 1.
290 /// Returns the last check instruction and the first check instruction in the
291 /// pair as (first, last).
292 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
294 /// Create an empty loop, based on the loop ranges of the old loop.
295 void createEmptyLoop();
296 /// Copy and widen the instructions from the old loop.
297 virtual void vectorizeLoop();
299 /// \brief The Loop exit block may have single value PHI nodes where the
300 /// incoming value is 'Undef'. While vectorizing we only handled real values
301 /// that were defined inside the loop. Here we fix the 'undef case'.
305 /// A helper function that computes the predicate of the block BB, assuming
306 /// that the header block of the loop is set to True. It returns the *entry*
307 /// mask for the block BB.
308 VectorParts createBlockInMask(BasicBlock *BB);
309 /// A helper function that computes the predicate of the edge between SRC
311 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
313 /// A helper function to vectorize a single BB within the innermost loop.
314 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
316 /// Vectorize a single PHINode in a block. This method handles the induction
317 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
318 /// arbitrary length vectors.
319 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
320 unsigned UF, unsigned VF, PhiVector *PV);
322 /// Insert the new loop to the loop hierarchy and pass manager
323 /// and update the analysis passes.
324 void updateAnalysis();
326 /// This instruction is un-vectorizable. Implement it as a sequence
327 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
328 /// scalarized instruction behind an if block predicated on the control
329 /// dependence of the instruction.
330 virtual void scalarizeInstruction(Instruction *Instr,
331 bool IfPredicateStore=false);
333 /// Vectorize Load and Store instructions,
334 virtual void vectorizeMemoryInstruction(Instruction *Instr);
336 /// Create a broadcast instruction. This method generates a broadcast
337 /// instruction (shuffle) for loop invariant values and for the induction
338 /// value. If this is the induction variable then we extend it to N, N+1, ...
339 /// this is needed because each iteration in the loop corresponds to a SIMD
341 virtual Value *getBroadcastInstrs(Value *V);
343 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
344 /// to each vector element of Val. The sequence starts at StartIndex.
345 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step);
347 /// When we go over instructions in the basic block we rely on previous
348 /// values within the current basic block or on loop invariant values.
349 /// When we widen (vectorize) values we place them in the map. If the values
350 /// are not within the map, they have to be loop invariant, so we simply
351 /// broadcast them into a vector.
352 VectorParts &getVectorValue(Value *V);
354 /// Generate a shuffle sequence that will reverse the vector Vec.
355 virtual Value *reverseVector(Value *Vec);
357 /// This is a helper class that holds the vectorizer state. It maps scalar
358 /// instructions to vector instructions. When the code is 'unrolled' then
359 /// then a single scalar value is mapped to multiple vector parts. The parts
360 /// are stored in the VectorPart type.
362 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
364 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
366 /// \return True if 'Key' is saved in the Value Map.
367 bool has(Value *Key) const { return MapStorage.count(Key); }
369 /// Initializes a new entry in the map. Sets all of the vector parts to the
370 /// save value in 'Val'.
371 /// \return A reference to a vector with splat values.
372 VectorParts &splat(Value *Key, Value *Val) {
373 VectorParts &Entry = MapStorage[Key];
374 Entry.assign(UF, Val);
378 ///\return A reference to the value that is stored at 'Key'.
379 VectorParts &get(Value *Key) {
380 VectorParts &Entry = MapStorage[Key];
383 assert(Entry.size() == UF);
388 /// The unroll factor. Each entry in the map stores this number of vector
392 /// Map storage. We use std::map and not DenseMap because insertions to a
393 /// dense map invalidates its iterators.
394 std::map<Value *, VectorParts> MapStorage;
397 /// The original loop.
399 /// Scev analysis to use.
407 /// Target Library Info.
408 const TargetLibraryInfo *TLI;
409 /// Target Transform Info.
410 const TargetTransformInfo *TTI;
412 /// The vectorization SIMD factor to use. Each vector will have this many
417 /// The vectorization unroll factor to use. Each scalar is vectorized to this
418 /// many different vector instructions.
421 /// The builder that we use
424 // --- Vectorization state ---
426 /// The vector-loop preheader.
427 BasicBlock *LoopVectorPreHeader;
428 /// The scalar-loop preheader.
429 BasicBlock *LoopScalarPreHeader;
430 /// Middle Block between the vector and the scalar.
431 BasicBlock *LoopMiddleBlock;
432 ///The ExitBlock of the scalar loop.
433 BasicBlock *LoopExitBlock;
434 ///The vector loop body.
435 SmallVector<BasicBlock *, 4> LoopVectorBody;
436 ///The scalar loop body.
437 BasicBlock *LoopScalarBody;
438 /// A list of all bypass blocks. The first block is the entry of the loop.
439 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
441 /// The new Induction variable which was added to the new block.
443 /// The induction variable of the old basic block.
444 PHINode *OldInduction;
445 /// Holds the extended (to the widest induction type) start index.
447 /// Maps scalars to widened vectors.
449 EdgeMaskCache MaskCache;
451 LoopVectorizationLegality *Legal;
453 // Record whether runtime check is added.
454 bool AddedSafetyChecks;
457 class InnerLoopUnroller : public InnerLoopVectorizer {
459 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
460 DominatorTree *DT, const TargetLibraryInfo *TLI,
461 const TargetTransformInfo *TTI, unsigned UnrollFactor)
462 : InnerLoopVectorizer(OrigLoop, SE, LI, DT, TLI, TTI, 1, UnrollFactor) {}
465 void scalarizeInstruction(Instruction *Instr,
466 bool IfPredicateStore = false) override;
467 void vectorizeMemoryInstruction(Instruction *Instr) override;
468 Value *getBroadcastInstrs(Value *V) override;
469 Value *getStepVector(Value *Val, int StartIdx, Value *Step) override;
470 Value *reverseVector(Value *Vec) override;
473 /// \brief Look for a meaningful debug location on the instruction or it's
475 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
480 if (I->getDebugLoc() != Empty)
483 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
484 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
485 if (OpInst->getDebugLoc() != Empty)
492 /// \brief Set the debug location in the builder using the debug location in the
494 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
495 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
496 B.SetCurrentDebugLocation(Inst->getDebugLoc());
498 B.SetCurrentDebugLocation(DebugLoc());
502 /// \return string containing a file name and a line # for the given loop.
503 static std::string getDebugLocString(const Loop *L) {
506 raw_string_ostream OS(Result);
507 const DebugLoc LoopDbgLoc = L->getStartLoc();
508 if (!LoopDbgLoc.isUnknown())
509 LoopDbgLoc.print(OS);
511 // Just print the module name.
512 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
519 /// \brief Propagate known metadata from one instruction to another.
520 static void propagateMetadata(Instruction *To, const Instruction *From) {
521 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
522 From->getAllMetadataOtherThanDebugLoc(Metadata);
524 for (auto M : Metadata) {
525 unsigned Kind = M.first;
527 // These are safe to transfer (this is safe for TBAA, even when we
528 // if-convert, because should that metadata have had a control dependency
529 // on the condition, and thus actually aliased with some other
530 // non-speculated memory access when the condition was false, this would be
531 // caught by the runtime overlap checks).
532 if (Kind != LLVMContext::MD_tbaa &&
533 Kind != LLVMContext::MD_alias_scope &&
534 Kind != LLVMContext::MD_noalias &&
535 Kind != LLVMContext::MD_fpmath)
538 To->setMetadata(Kind, M.second);
542 /// \brief Propagate known metadata from one instruction to a vector of others.
543 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
545 if (Instruction *I = dyn_cast<Instruction>(V))
546 propagateMetadata(I, From);
549 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
550 /// to what vectorization factor.
551 /// This class does not look at the profitability of vectorization, only the
552 /// legality. This class has two main kinds of checks:
553 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
554 /// will change the order of memory accesses in a way that will change the
555 /// correctness of the program.
556 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
557 /// checks for a number of different conditions, such as the availability of a
558 /// single induction variable, that all types are supported and vectorize-able,
559 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
560 /// This class is also used by InnerLoopVectorizer for identifying
561 /// induction variable and the different reduction variables.
562 class LoopVectorizationLegality {
564 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DominatorTree *DT,
565 TargetLibraryInfo *TLI, AliasAnalysis *AA,
566 Function *F, const TargetTransformInfo *TTI,
567 LoopAccessAnalysis *LAA)
568 : NumPredStores(0), TheLoop(L), SE(SE), TLI(TLI), TheFunction(F),
569 TTI(TTI), DT(DT), LAA(LAA), LAI(nullptr), Induction(nullptr),
570 WidestIndTy(nullptr), HasFunNoNaNAttr(false) {}
572 /// This enum represents the kinds of reductions that we support.
574 RK_NoReduction, ///< Not a reduction.
575 RK_IntegerAdd, ///< Sum of integers.
576 RK_IntegerMult, ///< Product of integers.
577 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
578 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
579 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
580 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
581 RK_FloatAdd, ///< Sum of floats.
582 RK_FloatMult, ///< Product of floats.
583 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
586 /// This enum represents the kinds of inductions that we support.
588 IK_NoInduction, ///< Not an induction variable.
589 IK_IntInduction, ///< Integer induction variable. Step = C.
590 IK_PtrInduction ///< Pointer induction var. Step = C / sizeof(elem).
593 // This enum represents the kind of minmax reduction.
594 enum MinMaxReductionKind {
604 /// This struct holds information about reduction variables.
605 struct ReductionDescriptor {
606 ReductionDescriptor() : StartValue(nullptr), LoopExitInstr(nullptr),
607 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
609 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
610 MinMaxReductionKind MK)
611 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
613 // The starting value of the reduction.
614 // It does not have to be zero!
615 TrackingVH<Value> StartValue;
616 // The instruction who's value is used outside the loop.
617 Instruction *LoopExitInstr;
618 // The kind of the reduction.
620 // If this a min/max reduction the kind of reduction.
621 MinMaxReductionKind MinMaxKind;
624 /// This POD struct holds information about a potential reduction operation.
625 struct ReductionInstDesc {
626 ReductionInstDesc(bool IsRedux, Instruction *I) :
627 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
629 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
630 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
632 // Is this instruction a reduction candidate.
634 // The last instruction in a min/max pattern (select of the select(icmp())
635 // pattern), or the current reduction instruction otherwise.
636 Instruction *PatternLastInst;
637 // If this is a min/max pattern the comparison predicate.
638 MinMaxReductionKind MinMaxKind;
641 /// A struct for saving information about induction variables.
642 struct InductionInfo {
643 InductionInfo(Value *Start, InductionKind K, ConstantInt *Step)
644 : StartValue(Start), IK(K), StepValue(Step) {
645 assert(IK != IK_NoInduction && "Not an induction");
646 assert(StartValue && "StartValue is null");
647 assert(StepValue && !StepValue->isZero() && "StepValue is zero");
648 assert((IK != IK_PtrInduction || StartValue->getType()->isPointerTy()) &&
649 "StartValue is not a pointer for pointer induction");
650 assert((IK != IK_IntInduction || StartValue->getType()->isIntegerTy()) &&
651 "StartValue is not an integer for integer induction");
652 assert(StepValue->getType()->isIntegerTy() &&
653 "StepValue is not an integer");
656 : StartValue(nullptr), IK(IK_NoInduction), StepValue(nullptr) {}
658 /// Get the consecutive direction. Returns:
659 /// 0 - unknown or non-consecutive.
660 /// 1 - consecutive and increasing.
661 /// -1 - consecutive and decreasing.
662 int getConsecutiveDirection() const {
663 if (StepValue && (StepValue->isOne() || StepValue->isMinusOne()))
664 return StepValue->getSExtValue();
668 /// Compute the transformed value of Index at offset StartValue using step
670 /// For integer induction, returns StartValue + Index * StepValue.
671 /// For pointer induction, returns StartValue[Index * StepValue].
672 /// FIXME: The newly created binary instructions should contain nsw/nuw
673 /// flags, which can be found from the original scalar operations.
674 Value *transform(IRBuilder<> &B, Value *Index) const {
676 case IK_IntInduction:
677 assert(Index->getType() == StartValue->getType() &&
678 "Index type does not match StartValue type");
679 if (StepValue->isMinusOne())
680 return B.CreateSub(StartValue, Index);
681 if (!StepValue->isOne())
682 Index = B.CreateMul(Index, StepValue);
683 return B.CreateAdd(StartValue, Index);
685 case IK_PtrInduction:
686 if (StepValue->isMinusOne())
687 Index = B.CreateNeg(Index);
688 else if (!StepValue->isOne())
689 Index = B.CreateMul(Index, StepValue);
690 return B.CreateGEP(StartValue, Index);
695 llvm_unreachable("invalid enum");
699 TrackingVH<Value> StartValue;
703 ConstantInt *StepValue;
706 /// ReductionList contains the reduction descriptors for all
707 /// of the reductions that were found in the loop.
708 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
710 /// InductionList saves induction variables and maps them to the
711 /// induction descriptor.
712 typedef MapVector<PHINode*, InductionInfo> InductionList;
714 /// Returns true if it is legal to vectorize this loop.
715 /// This does not mean that it is profitable to vectorize this
716 /// loop, only that it is legal to do so.
719 /// Returns the Induction variable.
720 PHINode *getInduction() { return Induction; }
722 /// Returns the reduction variables found in the loop.
723 ReductionList *getReductionVars() { return &Reductions; }
725 /// Returns the induction variables found in the loop.
726 InductionList *getInductionVars() { return &Inductions; }
728 /// Returns the widest induction type.
729 Type *getWidestInductionType() { return WidestIndTy; }
731 /// Returns True if V is an induction variable in this loop.
732 bool isInductionVariable(const Value *V);
734 /// Return true if the block BB needs to be predicated in order for the loop
735 /// to be vectorized.
736 bool blockNeedsPredication(BasicBlock *BB);
738 /// Check if this pointer is consecutive when vectorizing. This happens
739 /// when the last index of the GEP is the induction variable, or that the
740 /// pointer itself is an induction variable.
741 /// This check allows us to vectorize A[idx] into a wide load/store.
743 /// 0 - Stride is unknown or non-consecutive.
744 /// 1 - Address is consecutive.
745 /// -1 - Address is consecutive, and decreasing.
746 int isConsecutivePtr(Value *Ptr);
748 /// Returns true if the value V is uniform within the loop.
749 bool isUniform(Value *V);
751 /// Returns true if this instruction will remain scalar after vectorization.
752 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
754 /// Returns the information that we collected about runtime memory check.
755 const LoopAccessInfo::RuntimePointerCheck *getRuntimePointerCheck() const {
756 return LAI->getRuntimePointerCheck();
759 const LoopAccessInfo *getLAI() const {
763 /// This function returns the identity element (or neutral element) for
765 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
767 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
769 bool hasStride(Value *V) { return StrideSet.count(V); }
770 bool mustCheckStrides() { return !StrideSet.empty(); }
771 SmallPtrSet<Value *, 8>::iterator strides_begin() {
772 return StrideSet.begin();
774 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
776 /// Returns true if the target machine supports masked store operation
777 /// for the given \p DataType and kind of access to \p Ptr.
778 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
779 return TTI->isLegalMaskedStore(DataType, isConsecutivePtr(Ptr));
781 /// Returns true if the target machine supports masked load operation
782 /// for the given \p DataType and kind of access to \p Ptr.
783 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
784 return TTI->isLegalMaskedLoad(DataType, isConsecutivePtr(Ptr));
786 /// Returns true if vector representation of the instruction \p I
788 bool isMaskRequired(const Instruction* I) {
789 return (MaskedOp.count(I) != 0);
791 unsigned getNumStores() const {
792 return LAI->getNumStores();
794 unsigned getNumLoads() const {
795 return LAI->getNumLoads();
797 unsigned getNumPredStores() const {
798 return NumPredStores;
801 /// Check if a single basic block loop is vectorizable.
802 /// At this point we know that this is a loop with a constant trip count
803 /// and we only need to check individual instructions.
804 bool canVectorizeInstrs();
806 /// When we vectorize loops we may change the order in which
807 /// we read and write from memory. This method checks if it is
808 /// legal to vectorize the code, considering only memory constrains.
809 /// Returns true if the loop is vectorizable
810 bool canVectorizeMemory();
812 /// Return true if we can vectorize this loop using the IF-conversion
814 bool canVectorizeWithIfConvert();
816 /// Collect the variables that need to stay uniform after vectorization.
817 void collectLoopUniforms();
819 /// Return true if all of the instructions in the block can be speculatively
820 /// executed. \p SafePtrs is a list of addresses that are known to be legal
821 /// and we know that we can read from them without segfault.
822 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
824 /// Returns True, if 'Phi' is the kind of reduction variable for type
825 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
826 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
827 /// Returns a struct describing if the instruction 'I' can be a reduction
828 /// variable of type 'Kind'. If the reduction is a min/max pattern of
829 /// select(icmp()) this function advances the instruction pointer 'I' from the
830 /// compare instruction to the select instruction and stores this pointer in
831 /// 'PatternLastInst' member of the returned struct.
832 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
833 ReductionInstDesc &Desc);
834 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
835 /// pattern corresponding to a min(X, Y) or max(X, Y).
836 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
837 ReductionInstDesc &Prev);
838 /// Returns the induction kind of Phi and record the step. This function may
839 /// return NoInduction if the PHI is not an induction variable.
840 InductionKind isInductionVariable(PHINode *Phi, ConstantInt *&StepValue);
842 /// \brief Collect memory access with loop invariant strides.
844 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
846 void collectStridedAccess(Value *LoadOrStoreInst);
848 /// Report an analysis message to assist the user in diagnosing loops that are
849 /// not vectorized. These are handled as LoopAccessReport rather than
850 /// VectorizationReport because the << operator of VectorizationReport returns
851 /// LoopAccessReport.
852 void emitAnalysis(const LoopAccessReport &Message) {
853 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, LV_NAME);
856 unsigned NumPredStores;
858 /// The loop that we evaluate.
862 /// Target Library Info.
863 TargetLibraryInfo *TLI;
865 Function *TheFunction;
866 /// Target Transform Info
867 const TargetTransformInfo *TTI;
870 // LoopAccess analysis.
871 LoopAccessAnalysis *LAA;
872 // And the loop-accesses info corresponding to this loop. This pointer is
873 // null until canVectorizeMemory sets it up.
874 const LoopAccessInfo *LAI;
876 // --- vectorization state --- //
878 /// Holds the integer induction variable. This is the counter of the
881 /// Holds the reduction variables.
882 ReductionList Reductions;
883 /// Holds all of the induction variables that we found in the loop.
884 /// Notice that inductions don't need to start at zero and that induction
885 /// variables can be pointers.
886 InductionList Inductions;
887 /// Holds the widest induction type encountered.
890 /// Allowed outside users. This holds the reduction
891 /// vars which can be accessed from outside the loop.
892 SmallPtrSet<Value*, 4> AllowedExit;
893 /// This set holds the variables which are known to be uniform after
895 SmallPtrSet<Instruction*, 4> Uniforms;
897 /// Can we assume the absence of NaNs.
898 bool HasFunNoNaNAttr;
900 ValueToValueMap Strides;
901 SmallPtrSet<Value *, 8> StrideSet;
903 /// While vectorizing these instructions we have to generate a
904 /// call to the appropriate masked intrinsic
905 SmallPtrSet<const Instruction*, 8> MaskedOp;
908 /// LoopVectorizationCostModel - estimates the expected speedups due to
910 /// In many cases vectorization is not profitable. This can happen because of
911 /// a number of reasons. In this class we mainly attempt to predict the
912 /// expected speedup/slowdowns due to the supported instruction set. We use the
913 /// TargetTransformInfo to query the different backends for the cost of
914 /// different operations.
915 class LoopVectorizationCostModel {
917 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
918 LoopVectorizationLegality *Legal,
919 const TargetTransformInfo &TTI,
920 const TargetLibraryInfo *TLI, AssumptionCache *AC,
921 const Function *F, const LoopVectorizeHints *Hints)
922 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), TLI(TLI),
923 TheFunction(F), Hints(Hints) {
924 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
927 /// Information about vectorization costs
928 struct VectorizationFactor {
929 unsigned Width; // Vector width with best cost
930 unsigned Cost; // Cost of the loop with that width
932 /// \return The most profitable vectorization factor and the cost of that VF.
933 /// This method checks every power of two up to VF. If UserVF is not ZERO
934 /// then this vectorization factor will be selected if vectorization is
936 VectorizationFactor selectVectorizationFactor(bool OptForSize);
938 /// \return The size (in bits) of the widest type in the code that
939 /// needs to be vectorized. We ignore values that remain scalar such as
940 /// 64 bit loop indices.
941 unsigned getWidestType();
943 /// \return The most profitable unroll factor.
944 /// If UserUF is non-zero then this method finds the best unroll-factor
945 /// based on register pressure and other parameters.
946 /// VF and LoopCost are the selected vectorization factor and the cost of the
948 unsigned selectUnrollFactor(bool OptForSize, unsigned VF, unsigned LoopCost);
950 /// \brief A struct that represents some properties of the register usage
952 struct RegisterUsage {
953 /// Holds the number of loop invariant values that are used in the loop.
954 unsigned LoopInvariantRegs;
955 /// Holds the maximum number of concurrent live intervals in the loop.
956 unsigned MaxLocalUsers;
957 /// Holds the number of instructions in the loop.
958 unsigned NumInstructions;
961 /// \return information about the register usage of the loop.
962 RegisterUsage calculateRegisterUsage();
965 /// Returns the expected execution cost. The unit of the cost does
966 /// not matter because we use the 'cost' units to compare different
967 /// vector widths. The cost that is returned is *not* normalized by
968 /// the factor width.
969 unsigned expectedCost(unsigned VF);
971 /// Returns the execution time cost of an instruction for a given vector
972 /// width. Vector width of one means scalar.
973 unsigned getInstructionCost(Instruction *I, unsigned VF);
975 /// Returns whether the instruction is a load or store and will be a emitted
976 /// as a vector operation.
977 bool isConsecutiveLoadOrStore(Instruction *I);
979 /// Report an analysis message to assist the user in diagnosing loops that are
980 /// not vectorized. These are handled as LoopAccessReport rather than
981 /// VectorizationReport because the << operator of VectorizationReport returns
982 /// LoopAccessReport.
983 void emitAnalysis(const LoopAccessReport &Message) {
984 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, LV_NAME);
987 /// Values used only by @llvm.assume calls.
988 SmallPtrSet<const Value *, 32> EphValues;
990 /// The loop that we evaluate.
994 /// Loop Info analysis.
996 /// Vectorization legality.
997 LoopVectorizationLegality *Legal;
998 /// Vector target information.
999 const TargetTransformInfo &TTI;
1000 /// Target Library Info.
1001 const TargetLibraryInfo *TLI;
1002 const Function *TheFunction;
1003 // Loop Vectorize Hint.
1004 const LoopVectorizeHints *Hints;
1007 /// Utility class for getting and setting loop vectorizer hints in the form
1008 /// of loop metadata.
1009 /// This class keeps a number of loop annotations locally (as member variables)
1010 /// and can, upon request, write them back as metadata on the loop. It will
1011 /// initially scan the loop for existing metadata, and will update the local
1012 /// values based on information in the loop.
1013 /// We cannot write all values to metadata, as the mere presence of some info,
1014 /// for example 'force', means a decision has been made. So, we need to be
1015 /// careful NOT to add them if the user hasn't specifically asked so.
1016 class LoopVectorizeHints {
1023 /// Hint - associates name and validation with the hint value.
1026 unsigned Value; // This may have to change for non-numeric values.
1029 Hint(const char * Name, unsigned Value, HintKind Kind)
1030 : Name(Name), Value(Value), Kind(Kind) { }
1032 bool validate(unsigned Val) {
1035 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
1037 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1045 /// Vectorization width.
1047 /// Vectorization interleave factor.
1049 /// Vectorization forced
1052 /// Return the loop metadata prefix.
1053 static StringRef Prefix() { return "llvm.loop."; }
1057 FK_Undefined = -1, ///< Not selected.
1058 FK_Disabled = 0, ///< Forcing disabled.
1059 FK_Enabled = 1, ///< Forcing enabled.
1062 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
1063 : Width("vectorize.width", VectorizerParams::VectorizationFactor,
1065 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1066 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1068 // Populate values with existing loop metadata.
1069 getHintsFromMetadata();
1071 // force-vector-interleave overrides DisableInterleaving.
1072 if (VectorizerParams::isInterleaveForced())
1073 Interleave.Value = VectorizerParams::VectorizationInterleave;
1075 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
1076 << "LV: Interleaving disabled by the pass manager\n");
1079 /// Mark the loop L as already vectorized by setting the width to 1.
1080 void setAlreadyVectorized() {
1081 Width.Value = Interleave.Value = 1;
1082 Hint Hints[] = {Width, Interleave};
1083 writeHintsToMetadata(Hints);
1086 /// Dumps all the hint information.
1087 std::string emitRemark() const {
1088 VectorizationReport R;
1089 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1090 R << "vectorization is explicitly disabled";
1092 R << "use -Rpass-analysis=loop-vectorize for more info";
1093 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1094 R << " (Force=true";
1095 if (Width.Value != 0)
1096 R << ", Vector Width=" << Width.Value;
1097 if (Interleave.Value != 0)
1098 R << ", Interleave Count=" << Interleave.Value;
1106 unsigned getWidth() const { return Width.Value; }
1107 unsigned getInterleave() const { return Interleave.Value; }
1108 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1111 /// Find hints specified in the loop metadata and update local values.
1112 void getHintsFromMetadata() {
1113 MDNode *LoopID = TheLoop->getLoopID();
1117 // First operand should refer to the loop id itself.
1118 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1119 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1121 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1122 const MDString *S = nullptr;
1123 SmallVector<Metadata *, 4> Args;
1125 // The expected hint is either a MDString or a MDNode with the first
1126 // operand a MDString.
1127 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1128 if (!MD || MD->getNumOperands() == 0)
1130 S = dyn_cast<MDString>(MD->getOperand(0));
1131 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1132 Args.push_back(MD->getOperand(i));
1134 S = dyn_cast<MDString>(LoopID->getOperand(i));
1135 assert(Args.size() == 0 && "too many arguments for MDString");
1141 // Check if the hint starts with the loop metadata prefix.
1142 StringRef Name = S->getString();
1143 if (Args.size() == 1)
1144 setHint(Name, Args[0]);
1148 /// Checks string hint with one operand and set value if valid.
1149 void setHint(StringRef Name, Metadata *Arg) {
1150 if (!Name.startswith(Prefix()))
1152 Name = Name.substr(Prefix().size(), StringRef::npos);
1154 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1156 unsigned Val = C->getZExtValue();
1158 Hint *Hints[] = {&Width, &Interleave, &Force};
1159 for (auto H : Hints) {
1160 if (Name == H->Name) {
1161 if (H->validate(Val))
1164 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1170 /// Create a new hint from name / value pair.
1171 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1172 LLVMContext &Context = TheLoop->getHeader()->getContext();
1173 Metadata *MDs[] = {MDString::get(Context, Name),
1174 ConstantAsMetadata::get(
1175 ConstantInt::get(Type::getInt32Ty(Context), V))};
1176 return MDNode::get(Context, MDs);
1179 /// Matches metadata with hint name.
1180 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1181 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1185 for (auto H : HintTypes)
1186 if (Name->getString().endswith(H.Name))
1191 /// Sets current hints into loop metadata, keeping other values intact.
1192 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1193 if (HintTypes.size() == 0)
1196 // Reserve the first element to LoopID (see below).
1197 SmallVector<Metadata *, 4> MDs(1);
1198 // If the loop already has metadata, then ignore the existing operands.
1199 MDNode *LoopID = TheLoop->getLoopID();
1201 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1202 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1203 // If node in update list, ignore old value.
1204 if (!matchesHintMetadataName(Node, HintTypes))
1205 MDs.push_back(Node);
1209 // Now, add the missing hints.
1210 for (auto H : HintTypes)
1211 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1213 // Replace current metadata node with new one.
1214 LLVMContext &Context = TheLoop->getHeader()->getContext();
1215 MDNode *NewLoopID = MDNode::get(Context, MDs);
1216 // Set operand 0 to refer to the loop id itself.
1217 NewLoopID->replaceOperandWith(0, NewLoopID);
1219 TheLoop->setLoopID(NewLoopID);
1222 /// The loop these hints belong to.
1223 const Loop *TheLoop;
1226 static void emitMissedWarning(Function *F, Loop *L,
1227 const LoopVectorizeHints &LH) {
1228 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1229 L->getStartLoc(), LH.emitRemark());
1231 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1232 if (LH.getWidth() != 1)
1233 emitLoopVectorizeWarning(
1234 F->getContext(), *F, L->getStartLoc(),
1235 "failed explicitly specified loop vectorization");
1236 else if (LH.getInterleave() != 1)
1237 emitLoopInterleaveWarning(
1238 F->getContext(), *F, L->getStartLoc(),
1239 "failed explicitly specified loop interleaving");
1243 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1245 return V.push_back(&L);
1247 for (Loop *InnerL : L)
1248 addInnerLoop(*InnerL, V);
1251 /// The LoopVectorize Pass.
1252 struct LoopVectorize : public FunctionPass {
1253 /// Pass identification, replacement for typeid
1256 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1258 DisableUnrolling(NoUnrolling),
1259 AlwaysVectorize(AlwaysVectorize) {
1260 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1263 ScalarEvolution *SE;
1265 TargetTransformInfo *TTI;
1267 BlockFrequencyInfo *BFI;
1268 TargetLibraryInfo *TLI;
1270 AssumptionCache *AC;
1271 LoopAccessAnalysis *LAA;
1272 bool DisableUnrolling;
1273 bool AlwaysVectorize;
1275 BlockFrequency ColdEntryFreq;
1277 bool runOnFunction(Function &F) override {
1278 SE = &getAnalysis<ScalarEvolution>();
1279 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1280 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1281 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1282 BFI = &getAnalysis<BlockFrequencyInfo>();
1283 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1284 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1285 AA = &getAnalysis<AliasAnalysis>();
1286 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1287 LAA = &getAnalysis<LoopAccessAnalysis>();
1289 // Compute some weights outside of the loop over the loops. Compute this
1290 // using a BranchProbability to re-use its scaling math.
1291 const BranchProbability ColdProb(1, 5); // 20%
1292 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1294 // If the target claims to have no vector registers don't attempt
1296 if (!TTI->getNumberOfRegisters(true))
1299 // Build up a worklist of inner-loops to vectorize. This is necessary as
1300 // the act of vectorizing or partially unrolling a loop creates new loops
1301 // and can invalidate iterators across the loops.
1302 SmallVector<Loop *, 8> Worklist;
1305 addInnerLoop(*L, Worklist);
1307 LoopsAnalyzed += Worklist.size();
1309 // Now walk the identified inner loops.
1310 bool Changed = false;
1311 while (!Worklist.empty())
1312 Changed |= processLoop(Worklist.pop_back_val());
1314 // Process each loop nest in the function.
1318 static void AddRuntimeUnrollDisableMetaData(Loop *L) {
1319 SmallVector<Metadata *, 4> MDs;
1320 // Reserve first location for self reference to the LoopID metadata node.
1321 MDs.push_back(nullptr);
1322 bool IsUnrollMetadata = false;
1323 MDNode *LoopID = L->getLoopID();
1325 // First find existing loop unrolling disable metadata.
1326 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1327 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
1329 const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
1331 S && S->getString().startswith("llvm.loop.unroll.disable");
1333 MDs.push_back(LoopID->getOperand(i));
1337 if (!IsUnrollMetadata) {
1338 // Add runtime unroll disable metadata.
1339 LLVMContext &Context = L->getHeader()->getContext();
1340 SmallVector<Metadata *, 1> DisableOperands;
1341 DisableOperands.push_back(
1342 MDString::get(Context, "llvm.loop.unroll.runtime.disable"));
1343 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
1344 MDs.push_back(DisableNode);
1345 MDNode *NewLoopID = MDNode::get(Context, MDs);
1346 // Set operand 0 to refer to the loop id itself.
1347 NewLoopID->replaceOperandWith(0, NewLoopID);
1348 L->setLoopID(NewLoopID);
1352 bool processLoop(Loop *L) {
1353 assert(L->empty() && "Only process inner loops.");
1356 const std::string DebugLocStr = getDebugLocString(L);
1359 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1360 << L->getHeader()->getParent()->getName() << "\" from "
1361 << DebugLocStr << "\n");
1363 LoopVectorizeHints Hints(L, DisableUnrolling);
1365 DEBUG(dbgs() << "LV: Loop hints:"
1367 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1369 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1371 : "?")) << " width=" << Hints.getWidth()
1372 << " unroll=" << Hints.getInterleave() << "\n");
1374 // Function containing loop
1375 Function *F = L->getHeader()->getParent();
1377 // Looking at the diagnostic output is the only way to determine if a loop
1378 // was vectorized (other than looking at the IR or machine code), so it
1379 // is important to generate an optimization remark for each loop. Most of
1380 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1381 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1382 // less verbose reporting vectorized loops and unvectorized loops that may
1383 // benefit from vectorization, respectively.
1385 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1386 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1387 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1388 L->getStartLoc(), Hints.emitRemark());
1392 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1393 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1394 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1395 L->getStartLoc(), Hints.emitRemark());
1399 if (Hints.getWidth() == 1 && Hints.getInterleave() == 1) {
1400 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1401 emitOptimizationRemarkAnalysis(
1402 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1403 "loop not vectorized: vector width and interleave count are "
1404 "explicitly set to 1");
1408 // Check the loop for a trip count threshold:
1409 // do not vectorize loops with a tiny trip count.
1410 const unsigned TC = SE->getSmallConstantTripCount(L);
1411 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1412 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1413 << "This loop is not worth vectorizing.");
1414 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1415 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1417 DEBUG(dbgs() << "\n");
1418 emitOptimizationRemarkAnalysis(
1419 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1420 "vectorization is not beneficial and is not explicitly forced");
1425 // Check if it is legal to vectorize the loop.
1426 LoopVectorizationLegality LVL(L, SE, DT, TLI, AA, F, TTI, LAA);
1427 if (!LVL.canVectorize()) {
1428 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1429 emitMissedWarning(F, L, Hints);
1433 // Use the cost model.
1434 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, TLI, AC, F, &Hints);
1436 // Check the function attributes to find out if this function should be
1437 // optimized for size.
1438 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1439 F->hasFnAttribute(Attribute::OptimizeForSize);
1441 // Compute the weighted frequency of this loop being executed and see if it
1442 // is less than 20% of the function entry baseline frequency. Note that we
1443 // always have a canonical loop here because we think we *can* vectoriez.
1444 // FIXME: This is hidden behind a flag due to pervasive problems with
1445 // exactly what block frequency models.
1446 if (LoopVectorizeWithBlockFrequency) {
1447 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1448 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1449 LoopEntryFreq < ColdEntryFreq)
1453 // Check the function attributes to see if implicit floats are allowed.a
1454 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1455 // an integer loop and the vector instructions selected are purely integer
1456 // vector instructions?
1457 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1458 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1459 "attribute is used.\n");
1460 emitOptimizationRemarkAnalysis(
1461 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1462 "loop not vectorized due to NoImplicitFloat attribute");
1463 emitMissedWarning(F, L, Hints);
1467 // Select the optimal vectorization factor.
1468 const LoopVectorizationCostModel::VectorizationFactor VF =
1469 CM.selectVectorizationFactor(OptForSize);
1471 // Select the unroll factor.
1473 CM.selectUnrollFactor(OptForSize, VF.Width, VF.Cost);
1475 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1476 << DebugLocStr << '\n');
1477 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1479 if (VF.Width == 1) {
1480 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
1483 emitOptimizationRemarkAnalysis(
1484 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1485 "not beneficial to vectorize and user disabled interleaving");
1488 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1490 // Report the unrolling decision.
1491 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1492 Twine("unrolled with interleaving factor " +
1494 " (vectorization not beneficial)"));
1496 // We decided not to vectorize, but we may want to unroll.
1498 InnerLoopUnroller Unroller(L, SE, LI, DT, TLI, TTI, UF);
1499 Unroller.vectorize(&LVL);
1501 // If we decided that it is *legal* to vectorize the loop then do it.
1502 InnerLoopVectorizer LB(L, SE, LI, DT, TLI, TTI, VF.Width, UF);
1506 // Add metadata to disable runtime unrolling scalar loop when there's no
1507 // runtime check about strides and memory. Because at this situation,
1508 // scalar loop is rarely used not worthy to be unrolled.
1509 if (!LB.IsSafetyChecksAdded())
1510 AddRuntimeUnrollDisableMetaData(L);
1512 // Report the vectorization decision.
1513 emitOptimizationRemark(
1514 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1515 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1516 ", unrolling interleave factor: " + Twine(UF) + ")");
1519 // Mark the loop as already vectorized to avoid vectorizing again.
1520 Hints.setAlreadyVectorized();
1522 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1526 void getAnalysisUsage(AnalysisUsage &AU) const override {
1527 AU.addRequired<AssumptionCacheTracker>();
1528 AU.addRequiredID(LoopSimplifyID);
1529 AU.addRequiredID(LCSSAID);
1530 AU.addRequired<BlockFrequencyInfo>();
1531 AU.addRequired<DominatorTreeWrapperPass>();
1532 AU.addRequired<LoopInfoWrapperPass>();
1533 AU.addRequired<ScalarEvolution>();
1534 AU.addRequired<TargetTransformInfoWrapperPass>();
1535 AU.addRequired<AliasAnalysis>();
1536 AU.addRequired<LoopAccessAnalysis>();
1537 AU.addPreserved<LoopInfoWrapperPass>();
1538 AU.addPreserved<DominatorTreeWrapperPass>();
1539 AU.addPreserved<AliasAnalysis>();
1544 } // end anonymous namespace
1546 //===----------------------------------------------------------------------===//
1547 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1548 // LoopVectorizationCostModel.
1549 //===----------------------------------------------------------------------===//
1551 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1552 // We need to place the broadcast of invariant variables outside the loop.
1553 Instruction *Instr = dyn_cast<Instruction>(V);
1555 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1556 Instr->getParent()) != LoopVectorBody.end());
1557 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1559 // Place the code for broadcasting invariant variables in the new preheader.
1560 IRBuilder<>::InsertPointGuard Guard(Builder);
1562 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1564 // Broadcast the scalar into all locations in the vector.
1565 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1570 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
1572 assert(Val->getType()->isVectorTy() && "Must be a vector");
1573 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1574 "Elem must be an integer");
1575 assert(Step->getType() == Val->getType()->getScalarType() &&
1576 "Step has wrong type");
1577 // Create the types.
1578 Type *ITy = Val->getType()->getScalarType();
1579 VectorType *Ty = cast<VectorType>(Val->getType());
1580 int VLen = Ty->getNumElements();
1581 SmallVector<Constant*, 8> Indices;
1583 // Create a vector of consecutive numbers from zero to VF.
1584 for (int i = 0; i < VLen; ++i)
1585 Indices.push_back(ConstantInt::get(ITy, StartIdx + i));
1587 // Add the consecutive indices to the vector value.
1588 Constant *Cv = ConstantVector::get(Indices);
1589 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1590 Step = Builder.CreateVectorSplat(VLen, Step);
1591 assert(Step->getType() == Val->getType() && "Invalid step vec");
1592 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
1593 // which can be found from the original scalar operations.
1594 Step = Builder.CreateMul(Cv, Step);
1595 return Builder.CreateAdd(Val, Step, "induction");
1598 /// \brief Find the operand of the GEP that should be checked for consecutive
1599 /// stores. This ignores trailing indices that have no effect on the final
1601 static unsigned getGEPInductionOperand(const GetElementPtrInst *Gep) {
1602 const DataLayout &DL = Gep->getModule()->getDataLayout();
1603 unsigned LastOperand = Gep->getNumOperands() - 1;
1604 unsigned GEPAllocSize = DL.getTypeAllocSize(
1605 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1607 // Walk backwards and try to peel off zeros.
1608 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1609 // Find the type we're currently indexing into.
1610 gep_type_iterator GEPTI = gep_type_begin(Gep);
1611 std::advance(GEPTI, LastOperand - 1);
1613 // If it's a type with the same allocation size as the result of the GEP we
1614 // can peel off the zero index.
1615 if (DL.getTypeAllocSize(*GEPTI) != GEPAllocSize)
1623 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1624 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1625 // Make sure that the pointer does not point to structs.
1626 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1629 // If this value is a pointer induction variable we know it is consecutive.
1630 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1631 if (Phi && Inductions.count(Phi)) {
1632 InductionInfo II = Inductions[Phi];
1633 return II.getConsecutiveDirection();
1636 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1640 unsigned NumOperands = Gep->getNumOperands();
1641 Value *GpPtr = Gep->getPointerOperand();
1642 // If this GEP value is a consecutive pointer induction variable and all of
1643 // the indices are constant then we know it is consecutive. We can
1644 Phi = dyn_cast<PHINode>(GpPtr);
1645 if (Phi && Inductions.count(Phi)) {
1647 // Make sure that the pointer does not point to structs.
1648 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1649 if (GepPtrType->getElementType()->isAggregateType())
1652 // Make sure that all of the index operands are loop invariant.
1653 for (unsigned i = 1; i < NumOperands; ++i)
1654 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1657 InductionInfo II = Inductions[Phi];
1658 return II.getConsecutiveDirection();
1661 unsigned InductionOperand = getGEPInductionOperand(Gep);
1663 // Check that all of the gep indices are uniform except for our induction
1665 for (unsigned i = 0; i != NumOperands; ++i)
1666 if (i != InductionOperand &&
1667 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1670 // We can emit wide load/stores only if the last non-zero index is the
1671 // induction variable.
1672 const SCEV *Last = nullptr;
1673 if (!Strides.count(Gep))
1674 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1676 // Because of the multiplication by a stride we can have a s/zext cast.
1677 // We are going to replace this stride by 1 so the cast is safe to ignore.
1679 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1680 // %0 = trunc i64 %indvars.iv to i32
1681 // %mul = mul i32 %0, %Stride1
1682 // %idxprom = zext i32 %mul to i64 << Safe cast.
1683 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1685 Last = replaceSymbolicStrideSCEV(SE, Strides,
1686 Gep->getOperand(InductionOperand), Gep);
1687 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1689 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1693 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1694 const SCEV *Step = AR->getStepRecurrence(*SE);
1696 // The memory is consecutive because the last index is consecutive
1697 // and all other indices are loop invariant.
1700 if (Step->isAllOnesValue())
1707 bool LoopVectorizationLegality::isUniform(Value *V) {
1708 return LAI->isUniform(V);
1711 InnerLoopVectorizer::VectorParts&
1712 InnerLoopVectorizer::getVectorValue(Value *V) {
1713 assert(V != Induction && "The new induction variable should not be used.");
1714 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1716 // If we have a stride that is replaced by one, do it here.
1717 if (Legal->hasStride(V))
1718 V = ConstantInt::get(V->getType(), 1);
1720 // If we have this scalar in the map, return it.
1721 if (WidenMap.has(V))
1722 return WidenMap.get(V);
1724 // If this scalar is unknown, assume that it is a constant or that it is
1725 // loop invariant. Broadcast V and save the value for future uses.
1726 Value *B = getBroadcastInstrs(V);
1727 return WidenMap.splat(V, B);
1730 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1731 assert(Vec->getType()->isVectorTy() && "Invalid type");
1732 SmallVector<Constant*, 8> ShuffleMask;
1733 for (unsigned i = 0; i < VF; ++i)
1734 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1736 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1737 ConstantVector::get(ShuffleMask),
1741 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1742 // Attempt to issue a wide load.
1743 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1744 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1746 assert((LI || SI) && "Invalid Load/Store instruction");
1748 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1749 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1750 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1751 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1752 // An alignment of 0 means target abi alignment. We need to use the scalar's
1753 // target abi alignment in such a case.
1754 const DataLayout &DL = Instr->getModule()->getDataLayout();
1756 Alignment = DL.getABITypeAlignment(ScalarDataTy);
1757 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1758 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ScalarDataTy);
1759 unsigned VectorElementSize = DL.getTypeStoreSize(DataTy) / VF;
1761 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
1762 !Legal->isMaskRequired(SI))
1763 return scalarizeInstruction(Instr, true);
1765 if (ScalarAllocatedSize != VectorElementSize)
1766 return scalarizeInstruction(Instr);
1768 // If the pointer is loop invariant or if it is non-consecutive,
1769 // scalarize the load.
1770 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1771 bool Reverse = ConsecutiveStride < 0;
1772 bool UniformLoad = LI && Legal->isUniform(Ptr);
1773 if (!ConsecutiveStride || UniformLoad)
1774 return scalarizeInstruction(Instr);
1776 Constant *Zero = Builder.getInt32(0);
1777 VectorParts &Entry = WidenMap.get(Instr);
1779 // Handle consecutive loads/stores.
1780 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1781 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1782 setDebugLocFromInst(Builder, Gep);
1783 Value *PtrOperand = Gep->getPointerOperand();
1784 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1785 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1787 // Create the new GEP with the new induction variable.
1788 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1789 Gep2->setOperand(0, FirstBasePtr);
1790 Gep2->setName("gep.indvar.base");
1791 Ptr = Builder.Insert(Gep2);
1793 setDebugLocFromInst(Builder, Gep);
1794 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1795 OrigLoop) && "Base ptr must be invariant");
1797 // The last index does not have to be the induction. It can be
1798 // consecutive and be a function of the index. For example A[I+1];
1799 unsigned NumOperands = Gep->getNumOperands();
1800 unsigned InductionOperand = getGEPInductionOperand(Gep);
1801 // Create the new GEP with the new induction variable.
1802 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1804 for (unsigned i = 0; i < NumOperands; ++i) {
1805 Value *GepOperand = Gep->getOperand(i);
1806 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1808 // Update last index or loop invariant instruction anchored in loop.
1809 if (i == InductionOperand ||
1810 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1811 assert((i == InductionOperand ||
1812 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1813 "Must be last index or loop invariant");
1815 VectorParts &GEPParts = getVectorValue(GepOperand);
1816 Value *Index = GEPParts[0];
1817 Index = Builder.CreateExtractElement(Index, Zero);
1818 Gep2->setOperand(i, Index);
1819 Gep2->setName("gep.indvar.idx");
1822 Ptr = Builder.Insert(Gep2);
1824 // Use the induction element ptr.
1825 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1826 setDebugLocFromInst(Builder, Ptr);
1827 VectorParts &PtrVal = getVectorValue(Ptr);
1828 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1831 VectorParts Mask = createBlockInMask(Instr->getParent());
1834 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1835 "We do not allow storing to uniform addresses");
1836 setDebugLocFromInst(Builder, SI);
1837 // We don't want to update the value in the map as it might be used in
1838 // another expression. So don't use a reference type for "StoredVal".
1839 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1841 for (unsigned Part = 0; Part < UF; ++Part) {
1842 // Calculate the pointer for the specific unroll-part.
1843 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1846 // If we store to reverse consecutive memory locations then we need
1847 // to reverse the order of elements in the stored value.
1848 StoredVal[Part] = reverseVector(StoredVal[Part]);
1849 // If the address is consecutive but reversed, then the
1850 // wide store needs to start at the last vector element.
1851 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1852 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1853 Mask[Part] = reverseVector(Mask[Part]);
1856 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1857 DataTy->getPointerTo(AddressSpace));
1860 if (Legal->isMaskRequired(SI))
1861 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
1864 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
1865 propagateMetadata(NewSI, SI);
1871 assert(LI && "Must have a load instruction");
1872 setDebugLocFromInst(Builder, LI);
1873 for (unsigned Part = 0; Part < UF; ++Part) {
1874 // Calculate the pointer for the specific unroll-part.
1875 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1878 // If the address is consecutive but reversed, then the
1879 // wide load needs to start at the last vector element.
1880 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1881 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1882 Mask[Part] = reverseVector(Mask[Part]);
1886 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1887 DataTy->getPointerTo(AddressSpace));
1888 if (Legal->isMaskRequired(LI))
1889 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
1890 UndefValue::get(DataTy),
1891 "wide.masked.load");
1893 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
1894 propagateMetadata(NewLI, LI);
1895 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
1899 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1900 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1901 // Holds vector parameters or scalars, in case of uniform vals.
1902 SmallVector<VectorParts, 4> Params;
1904 setDebugLocFromInst(Builder, Instr);
1906 // Find all of the vectorized parameters.
1907 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1908 Value *SrcOp = Instr->getOperand(op);
1910 // If we are accessing the old induction variable, use the new one.
1911 if (SrcOp == OldInduction) {
1912 Params.push_back(getVectorValue(SrcOp));
1916 // Try using previously calculated values.
1917 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1919 // If the src is an instruction that appeared earlier in the basic block
1920 // then it should already be vectorized.
1921 if (SrcInst && OrigLoop->contains(SrcInst)) {
1922 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1923 // The parameter is a vector value from earlier.
1924 Params.push_back(WidenMap.get(SrcInst));
1926 // The parameter is a scalar from outside the loop. Maybe even a constant.
1927 VectorParts Scalars;
1928 Scalars.append(UF, SrcOp);
1929 Params.push_back(Scalars);
1933 assert(Params.size() == Instr->getNumOperands() &&
1934 "Invalid number of operands");
1936 // Does this instruction return a value ?
1937 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1939 Value *UndefVec = IsVoidRetTy ? nullptr :
1940 UndefValue::get(VectorType::get(Instr->getType(), VF));
1941 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1942 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1944 Instruction *InsertPt = Builder.GetInsertPoint();
1945 BasicBlock *IfBlock = Builder.GetInsertBlock();
1946 BasicBlock *CondBlock = nullptr;
1949 Loop *VectorLp = nullptr;
1950 if (IfPredicateStore) {
1951 assert(Instr->getParent()->getSinglePredecessor() &&
1952 "Only support single predecessor blocks");
1953 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1954 Instr->getParent());
1955 VectorLp = LI->getLoopFor(IfBlock);
1956 assert(VectorLp && "Must have a loop for this block");
1959 // For each vector unroll 'part':
1960 for (unsigned Part = 0; Part < UF; ++Part) {
1961 // For each scalar that we create:
1962 for (unsigned Width = 0; Width < VF; ++Width) {
1965 Value *Cmp = nullptr;
1966 if (IfPredicateStore) {
1967 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1968 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1969 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1970 LoopVectorBody.push_back(CondBlock);
1971 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
1972 // Update Builder with newly created basic block.
1973 Builder.SetInsertPoint(InsertPt);
1976 Instruction *Cloned = Instr->clone();
1978 Cloned->setName(Instr->getName() + ".cloned");
1979 // Replace the operands of the cloned instructions with extracted scalars.
1980 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1981 Value *Op = Params[op][Part];
1982 // Param is a vector. Need to extract the right lane.
1983 if (Op->getType()->isVectorTy())
1984 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1985 Cloned->setOperand(op, Op);
1988 // Place the cloned scalar in the new loop.
1989 Builder.Insert(Cloned);
1991 // If the original scalar returns a value we need to place it in a vector
1992 // so that future users will be able to use it.
1994 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1995 Builder.getInt32(Width));
1997 if (IfPredicateStore) {
1998 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1999 LoopVectorBody.push_back(NewIfBlock);
2000 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
2001 Builder.SetInsertPoint(InsertPt);
2002 Instruction *OldBr = IfBlock->getTerminator();
2003 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2004 OldBr->eraseFromParent();
2005 IfBlock = NewIfBlock;
2011 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
2015 if (Instruction *I = dyn_cast<Instruction>(V))
2016 return I->getParent() == Loc->getParent() ? I : nullptr;
2020 std::pair<Instruction *, Instruction *>
2021 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
2022 Instruction *tnullptr = nullptr;
2023 if (!Legal->mustCheckStrides())
2024 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2026 IRBuilder<> ChkBuilder(Loc);
2029 Value *Check = nullptr;
2030 Instruction *FirstInst = nullptr;
2031 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
2032 SE = Legal->strides_end();
2034 Value *Ptr = stripIntegerCast(*SI);
2035 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
2037 // Store the first instruction we create.
2038 FirstInst = getFirstInst(FirstInst, C, Loc);
2040 Check = ChkBuilder.CreateOr(Check, C);
2045 // We have to do this trickery because the IRBuilder might fold the check to a
2046 // constant expression in which case there is no Instruction anchored in a
2048 LLVMContext &Ctx = Loc->getContext();
2049 Instruction *TheCheck =
2050 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2051 ChkBuilder.Insert(TheCheck, "stride.not.one");
2052 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2054 return std::make_pair(FirstInst, TheCheck);
2057 void InnerLoopVectorizer::createEmptyLoop() {
2059 In this function we generate a new loop. The new loop will contain
2060 the vectorized instructions while the old loop will continue to run the
2063 [ ] <-- Back-edge taken count overflow check.
2066 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2069 || [ ] <-- vector pre header.
2073 || [ ]_| <-- vector loop.
2076 | >[ ] <--- middle-block.
2079 -|- >[ ] <--- new preheader.
2083 | [ ]_| <-- old scalar loop to handle remainder.
2086 >[ ] <-- exit block.
2090 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2091 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2092 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2093 assert(BypassBlock && "Invalid loop structure");
2094 assert(ExitBlock && "Must have an exit block");
2096 // Some loops have a single integer induction variable, while other loops
2097 // don't. One example is c++ iterators that often have multiple pointer
2098 // induction variables. In the code below we also support a case where we
2099 // don't have a single induction variable.
2100 OldInduction = Legal->getInduction();
2101 Type *IdxTy = Legal->getWidestInductionType();
2103 // Find the loop boundaries.
2104 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2105 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2107 // The exit count might have the type of i64 while the phi is i32. This can
2108 // happen if we have an induction variable that is sign extended before the
2109 // compare. The only way that we get a backedge taken count is that the
2110 // induction variable was signed and as such will not overflow. In such a case
2111 // truncation is legal.
2112 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2113 IdxTy->getPrimitiveSizeInBits())
2114 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2116 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2117 // Get the total trip count from the count by adding 1.
2118 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2119 SE->getConstant(BackedgeTakeCount->getType(), 1));
2121 const DataLayout &DL = OldBasicBlock->getModule()->getDataLayout();
2123 // Expand the trip count and place the new instructions in the preheader.
2124 // Notice that the pre-header does not change, only the loop body.
2125 SCEVExpander Exp(*SE, DL, "induction");
2127 // We need to test whether the backedge-taken count is uint##_max. Adding one
2128 // to it will cause overflow and an incorrect loop trip count in the vector
2129 // body. In case of overflow we want to directly jump to the scalar remainder
2131 Value *BackedgeCount =
2132 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2133 BypassBlock->getTerminator());
2134 if (BackedgeCount->getType()->isPointerTy())
2135 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2136 "backedge.ptrcnt.to.int",
2137 BypassBlock->getTerminator());
2138 Instruction *CheckBCOverflow =
2139 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2140 Constant::getAllOnesValue(BackedgeCount->getType()),
2141 "backedge.overflow", BypassBlock->getTerminator());
2143 // The loop index does not have to start at Zero. Find the original start
2144 // value from the induction PHI node. If we don't have an induction variable
2145 // then we know that it starts at zero.
2146 Builder.SetInsertPoint(BypassBlock->getTerminator());
2147 Value *StartIdx = ExtendedIdx = OldInduction ?
2148 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2150 ConstantInt::get(IdxTy, 0);
2152 // We need an instruction to anchor the overflow check on. StartIdx needs to
2153 // be defined before the overflow check branch. Because the scalar preheader
2154 // is going to merge the start index and so the overflow branch block needs to
2155 // contain a definition of the start index.
2156 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2157 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2158 BypassBlock->getTerminator());
2160 // Count holds the overall loop count (N).
2161 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2162 BypassBlock->getTerminator());
2164 LoopBypassBlocks.push_back(BypassBlock);
2166 // Split the single block loop into the two loop structure described above.
2167 BasicBlock *VectorPH =
2168 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2169 BasicBlock *VecBody =
2170 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2171 BasicBlock *MiddleBlock =
2172 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2173 BasicBlock *ScalarPH =
2174 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2176 // Create and register the new vector loop.
2177 Loop* Lp = new Loop();
2178 Loop *ParentLoop = OrigLoop->getParentLoop();
2180 // Insert the new loop into the loop nest and register the new basic blocks
2181 // before calling any utilities such as SCEV that require valid LoopInfo.
2183 ParentLoop->addChildLoop(Lp);
2184 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
2185 ParentLoop->addBasicBlockToLoop(VectorPH, *LI);
2186 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
2188 LI->addTopLevelLoop(Lp);
2190 Lp->addBasicBlockToLoop(VecBody, *LI);
2192 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2194 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2196 // Generate the induction variable.
2197 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2198 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2199 // The loop step is equal to the vectorization factor (num of SIMD elements)
2200 // times the unroll factor (num of SIMD instructions).
2201 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2203 // This is the IR builder that we use to add all of the logic for bypassing
2204 // the new vector loop.
2205 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2206 setDebugLocFromInst(BypassBuilder,
2207 getDebugLocFromInstOrOperands(OldInduction));
2209 // We may need to extend the index in case there is a type mismatch.
2210 // We know that the count starts at zero and does not overflow.
2211 if (Count->getType() != IdxTy) {
2212 // The exit count can be of pointer type. Convert it to the correct
2214 if (ExitCount->getType()->isPointerTy())
2215 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2217 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2220 // Add the start index to the loop count to get the new end index.
2221 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2223 // Now we need to generate the expression for N - (N % VF), which is
2224 // the part that the vectorized body will execute.
2225 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2226 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2227 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2228 "end.idx.rnd.down");
2230 // Now, compare the new count to zero. If it is zero skip the vector loop and
2231 // jump to the scalar loop.
2233 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2235 BasicBlock *LastBypassBlock = BypassBlock;
2237 // Generate code to check that the loops trip count that we computed by adding
2238 // one to the backedge-taken count will not overflow.
2240 auto PastOverflowCheck =
2241 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2242 BasicBlock *CheckBlock =
2243 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2245 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2246 LoopBypassBlocks.push_back(CheckBlock);
2247 Instruction *OldTerm = LastBypassBlock->getTerminator();
2248 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2249 OldTerm->eraseFromParent();
2250 LastBypassBlock = CheckBlock;
2253 // Generate the code to check that the strides we assumed to be one are really
2254 // one. We want the new basic block to start at the first instruction in a
2255 // sequence of instructions that form a check.
2256 Instruction *StrideCheck;
2257 Instruction *FirstCheckInst;
2258 std::tie(FirstCheckInst, StrideCheck) =
2259 addStrideCheck(LastBypassBlock->getTerminator());
2261 AddedSafetyChecks = true;
2262 // Create a new block containing the stride check.
2263 BasicBlock *CheckBlock =
2264 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2266 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2267 LoopBypassBlocks.push_back(CheckBlock);
2269 // Replace the branch into the memory check block with a conditional branch
2270 // for the "few elements case".
2271 Instruction *OldTerm = LastBypassBlock->getTerminator();
2272 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2273 OldTerm->eraseFromParent();
2276 LastBypassBlock = CheckBlock;
2279 // Generate the code that checks in runtime if arrays overlap. We put the
2280 // checks into a separate block to make the more common case of few elements
2282 Instruction *MemRuntimeCheck;
2283 std::tie(FirstCheckInst, MemRuntimeCheck) =
2284 Legal->getLAI()->addRuntimeCheck(LastBypassBlock->getTerminator());
2285 if (MemRuntimeCheck) {
2286 AddedSafetyChecks = true;
2287 // Create a new block containing the memory check.
2288 BasicBlock *CheckBlock =
2289 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.memcheck");
2291 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2292 LoopBypassBlocks.push_back(CheckBlock);
2294 // Replace the branch into the memory check block with a conditional branch
2295 // for the "few elements case".
2296 Instruction *OldTerm = LastBypassBlock->getTerminator();
2297 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2298 OldTerm->eraseFromParent();
2300 Cmp = MemRuntimeCheck;
2301 LastBypassBlock = CheckBlock;
2304 LastBypassBlock->getTerminator()->eraseFromParent();
2305 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2308 // We are going to resume the execution of the scalar loop.
2309 // Go over all of the induction variables that we found and fix the
2310 // PHIs that are left in the scalar version of the loop.
2311 // The starting values of PHI nodes depend on the counter of the last
2312 // iteration in the vectorized loop.
2313 // If we come from a bypass edge then we need to start from the original
2316 // This variable saves the new starting index for the scalar loop.
2317 PHINode *ResumeIndex = nullptr;
2318 LoopVectorizationLegality::InductionList::iterator I, E;
2319 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2320 // Set builder to point to last bypass block.
2321 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2322 for (I = List->begin(), E = List->end(); I != E; ++I) {
2323 PHINode *OrigPhi = I->first;
2324 LoopVectorizationLegality::InductionInfo II = I->second;
2326 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2327 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2328 MiddleBlock->getTerminator());
2329 // We might have extended the type of the induction variable but we need a
2330 // truncated version for the scalar loop.
2331 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2332 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2333 MiddleBlock->getTerminator()) : nullptr;
2335 // Create phi nodes to merge from the backedge-taken check block.
2336 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2337 ScalarPH->getTerminator());
2338 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2340 PHINode *BCTruncResumeVal = nullptr;
2341 if (OrigPhi == OldInduction) {
2343 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2344 ScalarPH->getTerminator());
2345 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2348 Value *EndValue = nullptr;
2350 case LoopVectorizationLegality::IK_NoInduction:
2351 llvm_unreachable("Unknown induction");
2352 case LoopVectorizationLegality::IK_IntInduction: {
2353 // Handle the integer induction counter.
2354 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2356 // We have the canonical induction variable.
2357 if (OrigPhi == OldInduction) {
2358 // Create a truncated version of the resume value for the scalar loop,
2359 // we might have promoted the type to a larger width.
2361 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2362 // The new PHI merges the original incoming value, in case of a bypass,
2363 // or the value at the end of the vectorized loop.
2364 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2365 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2366 TruncResumeVal->addIncoming(EndValue, VecBody);
2368 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2370 // We know what the end value is.
2371 EndValue = IdxEndRoundDown;
2372 // We also know which PHI node holds it.
2373 ResumeIndex = ResumeVal;
2377 // Not the canonical induction variable - add the vector loop count to the
2379 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2380 II.StartValue->getType(),
2382 EndValue = II.transform(BypassBuilder, CRD);
2383 EndValue->setName("ind.end");
2386 case LoopVectorizationLegality::IK_PtrInduction: {
2387 EndValue = II.transform(BypassBuilder, CountRoundDown);
2388 EndValue->setName("ptr.ind.end");
2393 // The new PHI merges the original incoming value, in case of a bypass,
2394 // or the value at the end of the vectorized loop.
2395 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2396 if (OrigPhi == OldInduction)
2397 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2399 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2401 ResumeVal->addIncoming(EndValue, VecBody);
2403 // Fix the scalar body counter (PHI node).
2404 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2406 // The old induction's phi node in the scalar body needs the truncated
2408 if (OrigPhi == OldInduction) {
2409 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2410 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2412 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2413 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2417 // If we are generating a new induction variable then we also need to
2418 // generate the code that calculates the exit value. This value is not
2419 // simply the end of the counter because we may skip the vectorized body
2420 // in case of a runtime check.
2422 assert(!ResumeIndex && "Unexpected resume value found");
2423 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2424 MiddleBlock->getTerminator());
2425 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2426 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2427 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2430 // Make sure that we found the index where scalar loop needs to continue.
2431 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2432 "Invalid resume Index");
2434 // Add a check in the middle block to see if we have completed
2435 // all of the iterations in the first vector loop.
2436 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2437 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2438 ResumeIndex, "cmp.n",
2439 MiddleBlock->getTerminator());
2441 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2442 // Remove the old terminator.
2443 MiddleBlock->getTerminator()->eraseFromParent();
2445 // Create i+1 and fill the PHINode.
2446 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2447 Induction->addIncoming(StartIdx, VectorPH);
2448 Induction->addIncoming(NextIdx, VecBody);
2449 // Create the compare.
2450 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2451 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2453 // Now we have two terminators. Remove the old one from the block.
2454 VecBody->getTerminator()->eraseFromParent();
2456 // Get ready to start creating new instructions into the vectorized body.
2457 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2460 LoopVectorPreHeader = VectorPH;
2461 LoopScalarPreHeader = ScalarPH;
2462 LoopMiddleBlock = MiddleBlock;
2463 LoopExitBlock = ExitBlock;
2464 LoopVectorBody.push_back(VecBody);
2465 LoopScalarBody = OldBasicBlock;
2467 LoopVectorizeHints Hints(Lp, true);
2468 Hints.setAlreadyVectorized();
2471 /// This function returns the identity element (or neutral element) for
2472 /// the operation K.
2474 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2479 // Adding, Xoring, Oring zero to a number does not change it.
2480 return ConstantInt::get(Tp, 0);
2481 case RK_IntegerMult:
2482 // Multiplying a number by 1 does not change it.
2483 return ConstantInt::get(Tp, 1);
2485 // AND-ing a number with an all-1 value does not change it.
2486 return ConstantInt::get(Tp, -1, true);
2488 // Multiplying a number by 1 does not change it.
2489 return ConstantFP::get(Tp, 1.0L);
2491 // Adding zero to a number does not change it.
2492 return ConstantFP::get(Tp, 0.0L);
2494 llvm_unreachable("Unknown reduction kind");
2498 /// This function translates the reduction kind to an LLVM binary operator.
2500 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2502 case LoopVectorizationLegality::RK_IntegerAdd:
2503 return Instruction::Add;
2504 case LoopVectorizationLegality::RK_IntegerMult:
2505 return Instruction::Mul;
2506 case LoopVectorizationLegality::RK_IntegerOr:
2507 return Instruction::Or;
2508 case LoopVectorizationLegality::RK_IntegerAnd:
2509 return Instruction::And;
2510 case LoopVectorizationLegality::RK_IntegerXor:
2511 return Instruction::Xor;
2512 case LoopVectorizationLegality::RK_FloatMult:
2513 return Instruction::FMul;
2514 case LoopVectorizationLegality::RK_FloatAdd:
2515 return Instruction::FAdd;
2516 case LoopVectorizationLegality::RK_IntegerMinMax:
2517 return Instruction::ICmp;
2518 case LoopVectorizationLegality::RK_FloatMinMax:
2519 return Instruction::FCmp;
2521 llvm_unreachable("Unknown reduction operation");
2525 static Value *createMinMaxOp(IRBuilder<> &Builder,
2526 LoopVectorizationLegality::MinMaxReductionKind RK,
2527 Value *Left, Value *Right) {
2528 CmpInst::Predicate P = CmpInst::ICMP_NE;
2531 llvm_unreachable("Unknown min/max reduction kind");
2532 case LoopVectorizationLegality::MRK_UIntMin:
2533 P = CmpInst::ICMP_ULT;
2535 case LoopVectorizationLegality::MRK_UIntMax:
2536 P = CmpInst::ICMP_UGT;
2538 case LoopVectorizationLegality::MRK_SIntMin:
2539 P = CmpInst::ICMP_SLT;
2541 case LoopVectorizationLegality::MRK_SIntMax:
2542 P = CmpInst::ICMP_SGT;
2544 case LoopVectorizationLegality::MRK_FloatMin:
2545 P = CmpInst::FCMP_OLT;
2547 case LoopVectorizationLegality::MRK_FloatMax:
2548 P = CmpInst::FCMP_OGT;
2553 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2554 RK == LoopVectorizationLegality::MRK_FloatMax)
2555 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2557 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2559 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2564 struct CSEDenseMapInfo {
2565 static bool canHandle(Instruction *I) {
2566 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2567 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2569 static inline Instruction *getEmptyKey() {
2570 return DenseMapInfo<Instruction *>::getEmptyKey();
2572 static inline Instruction *getTombstoneKey() {
2573 return DenseMapInfo<Instruction *>::getTombstoneKey();
2575 static unsigned getHashValue(Instruction *I) {
2576 assert(canHandle(I) && "Unknown instruction!");
2577 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2578 I->value_op_end()));
2580 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2581 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2582 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2584 return LHS->isIdenticalTo(RHS);
2589 /// \brief Check whether this block is a predicated block.
2590 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2591 /// = ...; " blocks. We start with one vectorized basic block. For every
2592 /// conditional block we split this vectorized block. Therefore, every second
2593 /// block will be a predicated one.
2594 static bool isPredicatedBlock(unsigned BlockNum) {
2595 return BlockNum % 2;
2598 ///\brief Perform cse of induction variable instructions.
2599 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2600 // Perform simple cse.
2601 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2602 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2603 BasicBlock *BB = BBs[i];
2604 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2605 Instruction *In = I++;
2607 if (!CSEDenseMapInfo::canHandle(In))
2610 // Check if we can replace this instruction with any of the
2611 // visited instructions.
2612 if (Instruction *V = CSEMap.lookup(In)) {
2613 In->replaceAllUsesWith(V);
2614 In->eraseFromParent();
2617 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2618 // ...;" blocks for predicated stores. Every second block is a predicated
2620 if (isPredicatedBlock(i))
2628 /// \brief Adds a 'fast' flag to floating point operations.
2629 static Value *addFastMathFlag(Value *V) {
2630 if (isa<FPMathOperator>(V)){
2631 FastMathFlags Flags;
2632 Flags.setUnsafeAlgebra();
2633 cast<Instruction>(V)->setFastMathFlags(Flags);
2638 /// Estimate the overhead of scalarizing a value. Insert and Extract are set if
2639 /// the result needs to be inserted and/or extracted from vectors.
2640 static unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract,
2641 const TargetTransformInfo &TTI) {
2645 assert(Ty->isVectorTy() && "Can only scalarize vectors");
2648 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
2650 Cost += TTI.getVectorInstrCost(Instruction::InsertElement, Ty, i);
2652 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, Ty, i);
2658 // Estimate cost of a call instruction CI if it were vectorized with factor VF.
2659 // Return the cost of the instruction, including scalarization overhead if it's
2660 // needed. The flag NeedToScalarize shows if the call needs to be scalarized -
2661 // i.e. either vector version isn't available, or is too expensive.
2662 static unsigned getVectorCallCost(CallInst *CI, unsigned VF,
2663 const TargetTransformInfo &TTI,
2664 const TargetLibraryInfo *TLI,
2665 bool &NeedToScalarize) {
2666 Function *F = CI->getCalledFunction();
2667 StringRef FnName = CI->getCalledFunction()->getName();
2668 Type *ScalarRetTy = CI->getType();
2669 SmallVector<Type *, 4> Tys, ScalarTys;
2670 for (auto &ArgOp : CI->arg_operands())
2671 ScalarTys.push_back(ArgOp->getType());
2673 // Estimate cost of scalarized vector call. The source operands are assumed
2674 // to be vectors, so we need to extract individual elements from there,
2675 // execute VF scalar calls, and then gather the result into the vector return
2677 unsigned ScalarCallCost = TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys);
2679 return ScalarCallCost;
2681 // Compute corresponding vector type for return value and arguments.
2682 Type *RetTy = ToVectorTy(ScalarRetTy, VF);
2683 for (unsigned i = 0, ie = ScalarTys.size(); i != ie; ++i)
2684 Tys.push_back(ToVectorTy(ScalarTys[i], VF));
2686 // Compute costs of unpacking argument values for the scalar calls and
2687 // packing the return values to a vector.
2688 unsigned ScalarizationCost =
2689 getScalarizationOverhead(RetTy, true, false, TTI);
2690 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i)
2691 ScalarizationCost += getScalarizationOverhead(Tys[i], false, true, TTI);
2693 unsigned Cost = ScalarCallCost * VF + ScalarizationCost;
2695 // If we can't emit a vector call for this function, then the currently found
2696 // cost is the cost we need to return.
2697 NeedToScalarize = true;
2698 if (!TLI || !TLI->isFunctionVectorizable(FnName, VF) || CI->isNoBuiltin())
2701 // If the corresponding vector cost is cheaper, return its cost.
2702 unsigned VectorCallCost = TTI.getCallInstrCost(nullptr, RetTy, Tys);
2703 if (VectorCallCost < Cost) {
2704 NeedToScalarize = false;
2705 return VectorCallCost;
2710 // Estimate cost of an intrinsic call instruction CI if it were vectorized with
2711 // factor VF. Return the cost of the instruction, including scalarization
2712 // overhead if it's needed.
2713 static unsigned getVectorIntrinsicCost(CallInst *CI, unsigned VF,
2714 const TargetTransformInfo &TTI,
2715 const TargetLibraryInfo *TLI) {
2716 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
2717 assert(ID && "Expected intrinsic call!");
2719 Type *RetTy = ToVectorTy(CI->getType(), VF);
2720 SmallVector<Type *, 4> Tys;
2721 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
2722 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
2724 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
2727 void InnerLoopVectorizer::vectorizeLoop() {
2728 //===------------------------------------------------===//
2730 // Notice: any optimization or new instruction that go
2731 // into the code below should be also be implemented in
2734 //===------------------------------------------------===//
2735 Constant *Zero = Builder.getInt32(0);
2737 // In order to support reduction variables we need to be able to vectorize
2738 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2739 // stages. First, we create a new vector PHI node with no incoming edges.
2740 // We use this value when we vectorize all of the instructions that use the
2741 // PHI. Next, after all of the instructions in the block are complete we
2742 // add the new incoming edges to the PHI. At this point all of the
2743 // instructions in the basic block are vectorized, so we can use them to
2744 // construct the PHI.
2745 PhiVector RdxPHIsToFix;
2747 // Scan the loop in a topological order to ensure that defs are vectorized
2749 LoopBlocksDFS DFS(OrigLoop);
2752 // Vectorize all of the blocks in the original loop.
2753 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2754 be = DFS.endRPO(); bb != be; ++bb)
2755 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2757 // At this point every instruction in the original loop is widened to
2758 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2759 // that we vectorized. The PHI nodes are currently empty because we did
2760 // not want to introduce cycles. Notice that the remaining PHI nodes
2761 // that we need to fix are reduction variables.
2763 // Create the 'reduced' values for each of the induction vars.
2764 // The reduced values are the vector values that we scalarize and combine
2765 // after the loop is finished.
2766 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2768 PHINode *RdxPhi = *it;
2769 assert(RdxPhi && "Unable to recover vectorized PHI");
2771 // Find the reduction variable descriptor.
2772 assert(Legal->getReductionVars()->count(RdxPhi) &&
2773 "Unable to find the reduction variable");
2774 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2775 (*Legal->getReductionVars())[RdxPhi];
2777 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2779 // We need to generate a reduction vector from the incoming scalar.
2780 // To do so, we need to generate the 'identity' vector and override
2781 // one of the elements with the incoming scalar reduction. We need
2782 // to do it in the vector-loop preheader.
2783 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
2785 // This is the vector-clone of the value that leaves the loop.
2786 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2787 Type *VecTy = VectorExit[0]->getType();
2789 // Find the reduction identity variable. Zero for addition, or, xor,
2790 // one for multiplication, -1 for And.
2793 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2794 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2795 // MinMax reduction have the start value as their identify.
2797 VectorStart = Identity = RdxDesc.StartValue;
2799 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2804 // Handle other reduction kinds:
2806 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2807 VecTy->getScalarType());
2810 // This vector is the Identity vector where the first element is the
2811 // incoming scalar reduction.
2812 VectorStart = RdxDesc.StartValue;
2814 Identity = ConstantVector::getSplat(VF, Iden);
2816 // This vector is the Identity vector where the first element is the
2817 // incoming scalar reduction.
2818 VectorStart = Builder.CreateInsertElement(Identity,
2819 RdxDesc.StartValue, Zero);
2823 // Fix the vector-loop phi.
2825 // Reductions do not have to start at zero. They can start with
2826 // any loop invariant values.
2827 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2828 BasicBlock *Latch = OrigLoop->getLoopLatch();
2829 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2830 VectorParts &Val = getVectorValue(LoopVal);
2831 for (unsigned part = 0; part < UF; ++part) {
2832 // Make sure to add the reduction stat value only to the
2833 // first unroll part.
2834 Value *StartVal = (part == 0) ? VectorStart : Identity;
2835 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
2836 LoopVectorPreHeader);
2837 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2838 LoopVectorBody.back());
2841 // Before each round, move the insertion point right between
2842 // the PHIs and the values we are going to write.
2843 // This allows us to write both PHINodes and the extractelement
2845 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2847 VectorParts RdxParts;
2848 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2849 for (unsigned part = 0; part < UF; ++part) {
2850 // This PHINode contains the vectorized reduction variable, or
2851 // the initial value vector, if we bypass the vector loop.
2852 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2853 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2854 Value *StartVal = (part == 0) ? VectorStart : Identity;
2855 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2856 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2857 NewPhi->addIncoming(RdxExitVal[part],
2858 LoopVectorBody.back());
2859 RdxParts.push_back(NewPhi);
2862 // Reduce all of the unrolled parts into a single vector.
2863 Value *ReducedPartRdx = RdxParts[0];
2864 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2865 setDebugLocFromInst(Builder, ReducedPartRdx);
2866 for (unsigned part = 1; part < UF; ++part) {
2867 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2868 // Floating point operations had to be 'fast' to enable the reduction.
2869 ReducedPartRdx = addFastMathFlag(
2870 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2871 ReducedPartRdx, "bin.rdx"));
2873 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2874 ReducedPartRdx, RdxParts[part]);
2878 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2879 // and vector ops, reducing the set of values being computed by half each
2881 assert(isPowerOf2_32(VF) &&
2882 "Reduction emission only supported for pow2 vectors!");
2883 Value *TmpVec = ReducedPartRdx;
2884 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2885 for (unsigned i = VF; i != 1; i >>= 1) {
2886 // Move the upper half of the vector to the lower half.
2887 for (unsigned j = 0; j != i/2; ++j)
2888 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2890 // Fill the rest of the mask with undef.
2891 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2892 UndefValue::get(Builder.getInt32Ty()));
2895 Builder.CreateShuffleVector(TmpVec,
2896 UndefValue::get(TmpVec->getType()),
2897 ConstantVector::get(ShuffleMask),
2900 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2901 // Floating point operations had to be 'fast' to enable the reduction.
2902 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2903 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2905 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2908 // The result is in the first element of the vector.
2909 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2910 Builder.getInt32(0));
2913 // Create a phi node that merges control-flow from the backedge-taken check
2914 // block and the middle block.
2915 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
2916 LoopScalarPreHeader->getTerminator());
2917 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
2918 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2920 // Now, we need to fix the users of the reduction variable
2921 // inside and outside of the scalar remainder loop.
2922 // We know that the loop is in LCSSA form. We need to update the
2923 // PHI nodes in the exit blocks.
2924 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2925 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2926 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2927 if (!LCSSAPhi) break;
2929 // All PHINodes need to have a single entry edge, or two if
2930 // we already fixed them.
2931 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2933 // We found our reduction value exit-PHI. Update it with the
2934 // incoming bypass edge.
2935 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2936 // Add an edge coming from the bypass.
2937 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2940 }// end of the LCSSA phi scan.
2942 // Fix the scalar loop reduction variable with the incoming reduction sum
2943 // from the vector body and from the backedge value.
2944 int IncomingEdgeBlockIdx =
2945 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2946 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2947 // Pick the other block.
2948 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2949 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
2950 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2951 }// end of for each redux variable.
2955 // Remove redundant induction instructions.
2956 cse(LoopVectorBody);
2959 void InnerLoopVectorizer::fixLCSSAPHIs() {
2960 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2961 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2962 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2963 if (!LCSSAPhi) break;
2964 if (LCSSAPhi->getNumIncomingValues() == 1)
2965 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2970 InnerLoopVectorizer::VectorParts
2971 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2972 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2975 // Look for cached value.
2976 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2977 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2978 if (ECEntryIt != MaskCache.end())
2979 return ECEntryIt->second;
2981 VectorParts SrcMask = createBlockInMask(Src);
2983 // The terminator has to be a branch inst!
2984 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2985 assert(BI && "Unexpected terminator found");
2987 if (BI->isConditional()) {
2988 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2990 if (BI->getSuccessor(0) != Dst)
2991 for (unsigned part = 0; part < UF; ++part)
2992 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2994 for (unsigned part = 0; part < UF; ++part)
2995 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2997 MaskCache[Edge] = EdgeMask;
3001 MaskCache[Edge] = SrcMask;
3005 InnerLoopVectorizer::VectorParts
3006 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
3007 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
3009 // Loop incoming mask is all-one.
3010 if (OrigLoop->getHeader() == BB) {
3011 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
3012 return getVectorValue(C);
3015 // This is the block mask. We OR all incoming edges, and with zero.
3016 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
3017 VectorParts BlockMask = getVectorValue(Zero);
3020 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
3021 VectorParts EM = createEdgeMask(*it, BB);
3022 for (unsigned part = 0; part < UF; ++part)
3023 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
3029 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
3030 InnerLoopVectorizer::VectorParts &Entry,
3031 unsigned UF, unsigned VF, PhiVector *PV) {
3032 PHINode* P = cast<PHINode>(PN);
3033 // Handle reduction variables:
3034 if (Legal->getReductionVars()->count(P)) {
3035 for (unsigned part = 0; part < UF; ++part) {
3036 // This is phase one of vectorizing PHIs.
3037 Type *VecTy = (VF == 1) ? PN->getType() :
3038 VectorType::get(PN->getType(), VF);
3039 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
3040 LoopVectorBody.back()-> getFirstInsertionPt());
3046 setDebugLocFromInst(Builder, P);
3047 // Check for PHI nodes that are lowered to vector selects.
3048 if (P->getParent() != OrigLoop->getHeader()) {
3049 // We know that all PHIs in non-header blocks are converted into
3050 // selects, so we don't have to worry about the insertion order and we
3051 // can just use the builder.
3052 // At this point we generate the predication tree. There may be
3053 // duplications since this is a simple recursive scan, but future
3054 // optimizations will clean it up.
3056 unsigned NumIncoming = P->getNumIncomingValues();
3058 // Generate a sequence of selects of the form:
3059 // SELECT(Mask3, In3,
3060 // SELECT(Mask2, In2,
3062 for (unsigned In = 0; In < NumIncoming; In++) {
3063 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
3065 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3067 for (unsigned part = 0; part < UF; ++part) {
3068 // We might have single edge PHIs (blocks) - use an identity
3069 // 'select' for the first PHI operand.
3071 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3074 // Select between the current value and the previous incoming edge
3075 // based on the incoming mask.
3076 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3077 Entry[part], "predphi");
3083 // This PHINode must be an induction variable.
3084 // Make sure that we know about it.
3085 assert(Legal->getInductionVars()->count(P) &&
3086 "Not an induction variable");
3088 LoopVectorizationLegality::InductionInfo II =
3089 Legal->getInductionVars()->lookup(P);
3091 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
3092 // which can be found from the original scalar operations.
3094 case LoopVectorizationLegality::IK_NoInduction:
3095 llvm_unreachable("Unknown induction");
3096 case LoopVectorizationLegality::IK_IntInduction: {
3097 assert(P->getType() == II.StartValue->getType() && "Types must match");
3098 Type *PhiTy = P->getType();
3100 if (P == OldInduction) {
3101 // Handle the canonical induction variable. We might have had to
3103 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
3105 // Handle other induction variables that are now based on the
3107 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
3109 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
3110 Broadcasted = II.transform(Builder, NormalizedIdx);
3111 Broadcasted->setName("offset.idx");
3113 Broadcasted = getBroadcastInstrs(Broadcasted);
3114 // After broadcasting the induction variable we need to make the vector
3115 // consecutive by adding 0, 1, 2, etc.
3116 for (unsigned part = 0; part < UF; ++part)
3117 Entry[part] = getStepVector(Broadcasted, VF * part, II.StepValue);
3120 case LoopVectorizationLegality::IK_PtrInduction:
3121 // Handle the pointer induction variable case.
3122 assert(P->getType()->isPointerTy() && "Unexpected type.");
3123 // This is the normalized GEP that starts counting at zero.
3124 Value *NormalizedIdx =
3125 Builder.CreateSub(Induction, ExtendedIdx, "normalized.idx");
3126 // This is the vector of results. Notice that we don't generate
3127 // vector geps because scalar geps result in better code.
3128 for (unsigned part = 0; part < UF; ++part) {
3130 int EltIndex = part;
3131 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3132 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3133 Value *SclrGep = II.transform(Builder, GlobalIdx);
3134 SclrGep->setName("next.gep");
3135 Entry[part] = SclrGep;
3139 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3140 for (unsigned int i = 0; i < VF; ++i) {
3141 int EltIndex = i + part * VF;
3142 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3143 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3144 Value *SclrGep = II.transform(Builder, GlobalIdx);
3145 SclrGep->setName("next.gep");
3146 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3147 Builder.getInt32(i),
3150 Entry[part] = VecVal;
3156 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3157 // For each instruction in the old loop.
3158 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3159 VectorParts &Entry = WidenMap.get(it);
3160 switch (it->getOpcode()) {
3161 case Instruction::Br:
3162 // Nothing to do for PHIs and BR, since we already took care of the
3163 // loop control flow instructions.
3165 case Instruction::PHI: {
3166 // Vectorize PHINodes.
3167 widenPHIInstruction(it, Entry, UF, VF, PV);
3171 case Instruction::Add:
3172 case Instruction::FAdd:
3173 case Instruction::Sub:
3174 case Instruction::FSub:
3175 case Instruction::Mul:
3176 case Instruction::FMul:
3177 case Instruction::UDiv:
3178 case Instruction::SDiv:
3179 case Instruction::FDiv:
3180 case Instruction::URem:
3181 case Instruction::SRem:
3182 case Instruction::FRem:
3183 case Instruction::Shl:
3184 case Instruction::LShr:
3185 case Instruction::AShr:
3186 case Instruction::And:
3187 case Instruction::Or:
3188 case Instruction::Xor: {
3189 // Just widen binops.
3190 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3191 setDebugLocFromInst(Builder, BinOp);
3192 VectorParts &A = getVectorValue(it->getOperand(0));
3193 VectorParts &B = getVectorValue(it->getOperand(1));
3195 // Use this vector value for all users of the original instruction.
3196 for (unsigned Part = 0; Part < UF; ++Part) {
3197 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3199 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3200 VecOp->copyIRFlags(BinOp);
3205 propagateMetadata(Entry, it);
3208 case Instruction::Select: {
3210 // If the selector is loop invariant we can create a select
3211 // instruction with a scalar condition. Otherwise, use vector-select.
3212 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3214 setDebugLocFromInst(Builder, it);
3216 // The condition can be loop invariant but still defined inside the
3217 // loop. This means that we can't just use the original 'cond' value.
3218 // We have to take the 'vectorized' value and pick the first lane.
3219 // Instcombine will make this a no-op.
3220 VectorParts &Cond = getVectorValue(it->getOperand(0));
3221 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3222 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3224 Value *ScalarCond = (VF == 1) ? Cond[0] :
3225 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3227 for (unsigned Part = 0; Part < UF; ++Part) {
3228 Entry[Part] = Builder.CreateSelect(
3229 InvariantCond ? ScalarCond : Cond[Part],
3234 propagateMetadata(Entry, it);
3238 case Instruction::ICmp:
3239 case Instruction::FCmp: {
3240 // Widen compares. Generate vector compares.
3241 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3242 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3243 setDebugLocFromInst(Builder, it);
3244 VectorParts &A = getVectorValue(it->getOperand(0));
3245 VectorParts &B = getVectorValue(it->getOperand(1));
3246 for (unsigned Part = 0; Part < UF; ++Part) {
3249 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3251 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3255 propagateMetadata(Entry, it);
3259 case Instruction::Store:
3260 case Instruction::Load:
3261 vectorizeMemoryInstruction(it);
3263 case Instruction::ZExt:
3264 case Instruction::SExt:
3265 case Instruction::FPToUI:
3266 case Instruction::FPToSI:
3267 case Instruction::FPExt:
3268 case Instruction::PtrToInt:
3269 case Instruction::IntToPtr:
3270 case Instruction::SIToFP:
3271 case Instruction::UIToFP:
3272 case Instruction::Trunc:
3273 case Instruction::FPTrunc:
3274 case Instruction::BitCast: {
3275 CastInst *CI = dyn_cast<CastInst>(it);
3276 setDebugLocFromInst(Builder, it);
3277 /// Optimize the special case where the source is the induction
3278 /// variable. Notice that we can only optimize the 'trunc' case
3279 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3280 /// c. other casts depend on pointer size.
3281 if (CI->getOperand(0) == OldInduction &&
3282 it->getOpcode() == Instruction::Trunc) {
3283 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3285 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3286 LoopVectorizationLegality::InductionInfo II =
3287 Legal->getInductionVars()->lookup(OldInduction);
3289 ConstantInt::getSigned(CI->getType(), II.StepValue->getSExtValue());
3290 for (unsigned Part = 0; Part < UF; ++Part)
3291 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
3292 propagateMetadata(Entry, it);
3295 /// Vectorize casts.
3296 Type *DestTy = (VF == 1) ? CI->getType() :
3297 VectorType::get(CI->getType(), VF);
3299 VectorParts &A = getVectorValue(it->getOperand(0));
3300 for (unsigned Part = 0; Part < UF; ++Part)
3301 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3302 propagateMetadata(Entry, it);
3306 case Instruction::Call: {
3307 // Ignore dbg intrinsics.
3308 if (isa<DbgInfoIntrinsic>(it))
3310 setDebugLocFromInst(Builder, it);
3312 Module *M = BB->getParent()->getParent();
3313 CallInst *CI = cast<CallInst>(it);
3315 StringRef FnName = CI->getCalledFunction()->getName();
3316 Function *F = CI->getCalledFunction();
3317 Type *RetTy = ToVectorTy(CI->getType(), VF);
3318 SmallVector<Type *, 4> Tys;
3319 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
3320 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
3322 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3324 (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end ||
3325 ID == Intrinsic::lifetime_start)) {
3326 scalarizeInstruction(it);
3329 // The flag shows whether we use Intrinsic or a usual Call for vectorized
3330 // version of the instruction.
3331 // Is it beneficial to perform intrinsic call compared to lib call?
3332 bool NeedToScalarize;
3333 unsigned CallCost = getVectorCallCost(CI, VF, *TTI, TLI, NeedToScalarize);
3334 bool UseVectorIntrinsic =
3335 ID && getVectorIntrinsicCost(CI, VF, *TTI, TLI) <= CallCost;
3336 if (!UseVectorIntrinsic && NeedToScalarize) {
3337 scalarizeInstruction(it);
3341 for (unsigned Part = 0; Part < UF; ++Part) {
3342 SmallVector<Value *, 4> Args;
3343 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3344 Value *Arg = CI->getArgOperand(i);
3345 // Some intrinsics have a scalar argument - don't replace it with a
3347 if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, i)) {
3348 VectorParts &VectorArg = getVectorValue(CI->getArgOperand(i));
3349 Arg = VectorArg[Part];
3351 Args.push_back(Arg);
3355 if (UseVectorIntrinsic) {
3356 // Use vector version of the intrinsic.
3357 Type *TysForDecl[] = {CI->getType()};
3359 TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3360 VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl);
3362 // Use vector version of the library call.
3363 StringRef VFnName = TLI->getVectorizedFunction(FnName, VF);
3364 assert(!VFnName.empty() && "Vector function name is empty.");
3365 VectorF = M->getFunction(VFnName);
3367 // Generate a declaration
3368 FunctionType *FTy = FunctionType::get(RetTy, Tys, false);
3370 Function::Create(FTy, Function::ExternalLinkage, VFnName, M);
3371 VectorF->copyAttributesFrom(F);
3374 assert(VectorF && "Can't create vector function.");
3375 Entry[Part] = Builder.CreateCall(VectorF, Args);
3378 propagateMetadata(Entry, it);
3383 // All other instructions are unsupported. Scalarize them.
3384 scalarizeInstruction(it);
3387 }// end of for_each instr.
3390 void InnerLoopVectorizer::updateAnalysis() {
3391 // Forget the original basic block.
3392 SE->forgetLoop(OrigLoop);
3394 // Update the dominator tree information.
3395 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3396 "Entry does not dominate exit.");
3398 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3399 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3400 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3402 // Due to if predication of stores we might create a sequence of "if(pred)
3403 // a[i] = ...; " blocks.
3404 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3406 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3407 else if (isPredicatedBlock(i)) {
3408 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3410 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3414 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3415 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3416 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3417 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3419 DEBUG(DT->verifyDomTree());
3422 /// \brief Check whether it is safe to if-convert this phi node.
3424 /// Phi nodes with constant expressions that can trap are not safe to if
3426 static bool canIfConvertPHINodes(BasicBlock *BB) {
3427 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3428 PHINode *Phi = dyn_cast<PHINode>(I);
3431 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3432 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3439 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3440 if (!EnableIfConversion) {
3441 emitAnalysis(VectorizationReport() << "if-conversion is disabled");
3445 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3447 // A list of pointers that we can safely read and write to.
3448 SmallPtrSet<Value *, 8> SafePointes;
3450 // Collect safe addresses.
3451 for (Loop::block_iterator BI = TheLoop->block_begin(),
3452 BE = TheLoop->block_end(); BI != BE; ++BI) {
3453 BasicBlock *BB = *BI;
3455 if (blockNeedsPredication(BB))
3458 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3459 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3460 SafePointes.insert(LI->getPointerOperand());
3461 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3462 SafePointes.insert(SI->getPointerOperand());
3466 // Collect the blocks that need predication.
3467 BasicBlock *Header = TheLoop->getHeader();
3468 for (Loop::block_iterator BI = TheLoop->block_begin(),
3469 BE = TheLoop->block_end(); BI != BE; ++BI) {
3470 BasicBlock *BB = *BI;
3472 // We don't support switch statements inside loops.
3473 if (!isa<BranchInst>(BB->getTerminator())) {
3474 emitAnalysis(VectorizationReport(BB->getTerminator())
3475 << "loop contains a switch statement");
3479 // We must be able to predicate all blocks that need to be predicated.
3480 if (blockNeedsPredication(BB)) {
3481 if (!blockCanBePredicated(BB, SafePointes)) {
3482 emitAnalysis(VectorizationReport(BB->getTerminator())
3483 << "control flow cannot be substituted for a select");
3486 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3487 emitAnalysis(VectorizationReport(BB->getTerminator())
3488 << "control flow cannot be substituted for a select");
3493 // We can if-convert this loop.
3497 bool LoopVectorizationLegality::canVectorize() {
3498 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3499 // be canonicalized.
3500 if (!TheLoop->getLoopPreheader()) {
3502 VectorizationReport() <<
3503 "loop control flow is not understood by vectorizer");
3507 // We can only vectorize innermost loops.
3508 if (!TheLoop->getSubLoopsVector().empty()) {
3509 emitAnalysis(VectorizationReport() << "loop is not the innermost loop");
3513 // We must have a single backedge.
3514 if (TheLoop->getNumBackEdges() != 1) {
3516 VectorizationReport() <<
3517 "loop control flow is not understood by vectorizer");
3521 // We must have a single exiting block.
3522 if (!TheLoop->getExitingBlock()) {
3524 VectorizationReport() <<
3525 "loop control flow is not understood by vectorizer");
3529 // We only handle bottom-tested loops, i.e. loop in which the condition is
3530 // checked at the end of each iteration. With that we can assume that all
3531 // instructions in the loop are executed the same number of times.
3532 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
3534 VectorizationReport() <<
3535 "loop control flow is not understood by vectorizer");
3539 // We need to have a loop header.
3540 DEBUG(dbgs() << "LV: Found a loop: " <<
3541 TheLoop->getHeader()->getName() << '\n');
3543 // Check if we can if-convert non-single-bb loops.
3544 unsigned NumBlocks = TheLoop->getNumBlocks();
3545 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3546 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3550 // ScalarEvolution needs to be able to find the exit count.
3551 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3552 if (ExitCount == SE->getCouldNotCompute()) {
3553 emitAnalysis(VectorizationReport() <<
3554 "could not determine number of loop iterations");
3555 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3559 // Check if we can vectorize the instructions and CFG in this loop.
3560 if (!canVectorizeInstrs()) {
3561 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3565 // Go over each instruction and look at memory deps.
3566 if (!canVectorizeMemory()) {
3567 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3571 // Collect all of the variables that remain uniform after vectorization.
3572 collectLoopUniforms();
3574 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3575 (LAI->getRuntimePointerCheck()->Need ? " (with a runtime bound check)" :
3579 // Okay! We can vectorize. At this point we don't have any other mem analysis
3580 // which may limit our maximum vectorization factor, so just return true with
3585 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3586 if (Ty->isPointerTy())
3587 return DL.getIntPtrType(Ty);
3589 // It is possible that char's or short's overflow when we ask for the loop's
3590 // trip count, work around this by changing the type size.
3591 if (Ty->getScalarSizeInBits() < 32)
3592 return Type::getInt32Ty(Ty->getContext());
3597 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3598 Ty0 = convertPointerToIntegerType(DL, Ty0);
3599 Ty1 = convertPointerToIntegerType(DL, Ty1);
3600 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3605 /// \brief Check that the instruction has outside loop users and is not an
3606 /// identified reduction variable.
3607 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3608 SmallPtrSetImpl<Value *> &Reductions) {
3609 // Reduction instructions are allowed to have exit users. All other
3610 // instructions must not have external users.
3611 if (!Reductions.count(Inst))
3612 //Check that all of the users of the loop are inside the BB.
3613 for (User *U : Inst->users()) {
3614 Instruction *UI = cast<Instruction>(U);
3615 // This user may be a reduction exit value.
3616 if (!TheLoop->contains(UI)) {
3617 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3624 bool LoopVectorizationLegality::canVectorizeInstrs() {
3625 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3626 BasicBlock *Header = TheLoop->getHeader();
3628 // Look for the attribute signaling the absence of NaNs.
3629 Function &F = *Header->getParent();
3630 const DataLayout &DL = F.getParent()->getDataLayout();
3631 if (F.hasFnAttribute("no-nans-fp-math"))
3633 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
3635 // For each block in the loop.
3636 for (Loop::block_iterator bb = TheLoop->block_begin(),
3637 be = TheLoop->block_end(); bb != be; ++bb) {
3639 // Scan the instructions in the block and look for hazards.
3640 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3643 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3644 Type *PhiTy = Phi->getType();
3645 // Check that this PHI type is allowed.
3646 if (!PhiTy->isIntegerTy() &&
3647 !PhiTy->isFloatingPointTy() &&
3648 !PhiTy->isPointerTy()) {
3649 emitAnalysis(VectorizationReport(it)
3650 << "loop control flow is not understood by vectorizer");
3651 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3655 // If this PHINode is not in the header block, then we know that we
3656 // can convert it to select during if-conversion. No need to check if
3657 // the PHIs in this block are induction or reduction variables.
3658 if (*bb != Header) {
3659 // Check that this instruction has no outside users or is an
3660 // identified reduction value with an outside user.
3661 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3663 emitAnalysis(VectorizationReport(it) <<
3664 "value could not be identified as "
3665 "an induction or reduction variable");
3669 // We only allow if-converted PHIs with exactly two incoming values.
3670 if (Phi->getNumIncomingValues() != 2) {
3671 emitAnalysis(VectorizationReport(it)
3672 << "control flow not understood by vectorizer");
3673 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3677 // This is the value coming from the preheader.
3678 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3679 ConstantInt *StepValue = nullptr;
3680 // Check if this is an induction variable.
3681 InductionKind IK = isInductionVariable(Phi, StepValue);
3683 if (IK_NoInduction != IK) {
3684 // Get the widest type.
3686 WidestIndTy = convertPointerToIntegerType(DL, PhiTy);
3688 WidestIndTy = getWiderType(DL, PhiTy, WidestIndTy);
3690 // Int inductions are special because we only allow one IV.
3691 if (IK == IK_IntInduction && StepValue->isOne()) {
3692 // Use the phi node with the widest type as induction. Use the last
3693 // one if there are multiple (no good reason for doing this other
3694 // than it is expedient).
3695 if (!Induction || PhiTy == WidestIndTy)
3699 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3700 Inductions[Phi] = InductionInfo(StartValue, IK, StepValue);
3702 // Until we explicitly handle the case of an induction variable with
3703 // an outside loop user we have to give up vectorizing this loop.
3704 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3705 emitAnalysis(VectorizationReport(it) <<
3706 "use of induction value outside of the "
3707 "loop is not handled by vectorizer");
3714 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3715 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3718 if (AddReductionVar(Phi, RK_IntegerMult)) {
3719 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3722 if (AddReductionVar(Phi, RK_IntegerOr)) {
3723 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3726 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3727 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3730 if (AddReductionVar(Phi, RK_IntegerXor)) {
3731 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3734 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3735 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3738 if (AddReductionVar(Phi, RK_FloatMult)) {
3739 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3742 if (AddReductionVar(Phi, RK_FloatAdd)) {
3743 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3746 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3747 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3752 emitAnalysis(VectorizationReport(it) <<
3753 "value that could not be identified as "
3754 "reduction is used outside the loop");
3755 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3757 }// end of PHI handling
3759 // We handle calls that:
3760 // * Are debug info intrinsics.
3761 // * Have a mapping to an IR intrinsic.
3762 // * Have a vector version available.
3763 CallInst *CI = dyn_cast<CallInst>(it);
3764 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI) &&
3765 !(CI->getCalledFunction() && TLI &&
3766 TLI->isFunctionVectorizable(CI->getCalledFunction()->getName()))) {
3767 emitAnalysis(VectorizationReport(it) <<
3768 "call instruction cannot be vectorized");
3769 DEBUG(dbgs() << "LV: Found a non-intrinsic, non-libfunc callsite.\n");
3773 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
3774 // second argument is the same (i.e. loop invariant)
3776 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
3777 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
3778 emitAnalysis(VectorizationReport(it)
3779 << "intrinsic instruction cannot be vectorized");
3780 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
3785 // Check that the instruction return type is vectorizable.
3786 // Also, we can't vectorize extractelement instructions.
3787 if ((!VectorType::isValidElementType(it->getType()) &&
3788 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3789 emitAnalysis(VectorizationReport(it)
3790 << "instruction return type cannot be vectorized");
3791 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3795 // Check that the stored type is vectorizable.
3796 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3797 Type *T = ST->getValueOperand()->getType();
3798 if (!VectorType::isValidElementType(T)) {
3799 emitAnalysis(VectorizationReport(ST) <<
3800 "store instruction cannot be vectorized");
3803 if (EnableMemAccessVersioning)
3804 collectStridedAccess(ST);
3807 if (EnableMemAccessVersioning)
3808 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3809 collectStridedAccess(LI);
3811 // Reduction instructions are allowed to have exit users.
3812 // All other instructions must not have external users.
3813 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3814 emitAnalysis(VectorizationReport(it) <<
3815 "value cannot be used outside the loop");
3824 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3825 if (Inductions.empty()) {
3826 emitAnalysis(VectorizationReport()
3827 << "loop induction variable could not be identified");
3835 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3836 /// return the induction operand of the gep pointer.
3837 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE, Loop *Lp) {
3838 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3842 unsigned InductionOperand = getGEPInductionOperand(GEP);
3844 // Check that all of the gep indices are uniform except for our induction
3846 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3847 if (i != InductionOperand &&
3848 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3850 return GEP->getOperand(InductionOperand);
3853 ///\brief Look for a cast use of the passed value.
3854 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3855 Value *UniqueCast = nullptr;
3856 for (User *U : Ptr->users()) {
3857 CastInst *CI = dyn_cast<CastInst>(U);
3858 if (CI && CI->getType() == Ty) {
3868 ///\brief Get the stride of a pointer access in a loop.
3869 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3870 /// pointer to the Value, or null otherwise.
3871 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE, Loop *Lp) {
3872 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3873 if (!PtrTy || PtrTy->isAggregateType())
3876 // Try to remove a gep instruction to make the pointer (actually index at this
3877 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3878 // pointer, otherwise, we are analyzing the index.
3879 Value *OrigPtr = Ptr;
3881 // The size of the pointer access.
3882 int64_t PtrAccessSize = 1;
3884 Ptr = stripGetElementPtr(Ptr, SE, Lp);
3885 const SCEV *V = SE->getSCEV(Ptr);
3889 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3890 V = C->getOperand();
3892 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3896 V = S->getStepRecurrence(*SE);
3900 // Strip off the size of access multiplication if we are still analyzing the
3902 if (OrigPtr == Ptr) {
3903 const DataLayout &DL = Lp->getHeader()->getModule()->getDataLayout();
3904 DL.getTypeAllocSize(PtrTy->getElementType());
3905 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3906 if (M->getOperand(0)->getSCEVType() != scConstant)
3909 const APInt &APStepVal =
3910 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3912 // Huge step value - give up.
3913 if (APStepVal.getBitWidth() > 64)
3916 int64_t StepVal = APStepVal.getSExtValue();
3917 if (PtrAccessSize != StepVal)
3919 V = M->getOperand(1);
3924 Type *StripedOffRecurrenceCast = nullptr;
3925 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3926 StripedOffRecurrenceCast = C->getType();
3927 V = C->getOperand();
3930 // Look for the loop invariant symbolic value.
3931 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3935 Value *Stride = U->getValue();
3936 if (!Lp->isLoopInvariant(Stride))
3939 // If we have stripped off the recurrence cast we have to make sure that we
3940 // return the value that is used in this loop so that we can replace it later.
3941 if (StripedOffRecurrenceCast)
3942 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3947 void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) {
3948 Value *Ptr = nullptr;
3949 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3950 Ptr = LI->getPointerOperand();
3951 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3952 Ptr = SI->getPointerOperand();
3956 Value *Stride = getStrideFromPointer(Ptr, SE, TheLoop);
3960 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3961 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3962 Strides[Ptr] = Stride;
3963 StrideSet.insert(Stride);
3966 void LoopVectorizationLegality::collectLoopUniforms() {
3967 // We now know that the loop is vectorizable!
3968 // Collect variables that will remain uniform after vectorization.
3969 std::vector<Value*> Worklist;
3970 BasicBlock *Latch = TheLoop->getLoopLatch();
3972 // Start with the conditional branch and walk up the block.
3973 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3975 // Also add all consecutive pointer values; these values will be uniform
3976 // after vectorization (and subsequent cleanup) and, until revectorization is
3977 // supported, all dependencies must also be uniform.
3978 for (Loop::block_iterator B = TheLoop->block_begin(),
3979 BE = TheLoop->block_end(); B != BE; ++B)
3980 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
3982 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
3983 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3985 while (!Worklist.empty()) {
3986 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3987 Worklist.pop_back();
3989 // Look at instructions inside this loop.
3990 // Stop when reaching PHI nodes.
3991 // TODO: we need to follow values all over the loop, not only in this block.
3992 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3995 // This is a known uniform.
3998 // Insert all operands.
3999 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4003 bool LoopVectorizationLegality::canVectorizeMemory() {
4004 LAI = &LAA->getInfo(TheLoop, Strides);
4005 auto &OptionalReport = LAI->getReport();
4007 emitAnalysis(VectorizationReport(*OptionalReport));
4008 if (!LAI->canVectorizeMemory())
4011 if (LAI->getNumRuntimePointerChecks() >
4012 VectorizerParams::RuntimeMemoryCheckThreshold) {
4013 emitAnalysis(VectorizationReport()
4014 << LAI->getNumRuntimePointerChecks() << " exceeds limit of "
4015 << VectorizerParams::RuntimeMemoryCheckThreshold
4016 << " dependent memory operations checked at runtime");
4017 DEBUG(dbgs() << "LV: Too many memory checks needed.\n");
4023 static bool hasMultipleUsesOf(Instruction *I,
4024 SmallPtrSetImpl<Instruction *> &Insts) {
4025 unsigned NumUses = 0;
4026 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
4027 if (Insts.count(dyn_cast<Instruction>(*Use)))
4036 static bool areAllUsesIn(Instruction *I, SmallPtrSetImpl<Instruction *> &Set) {
4037 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
4038 if (!Set.count(dyn_cast<Instruction>(*Use)))
4043 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
4044 ReductionKind Kind) {
4045 if (Phi->getNumIncomingValues() != 2)
4048 // Reduction variables are only found in the loop header block.
4049 if (Phi->getParent() != TheLoop->getHeader())
4052 // Obtain the reduction start value from the value that comes from the loop
4054 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
4056 // ExitInstruction is the single value which is used outside the loop.
4057 // We only allow for a single reduction value to be used outside the loop.
4058 // This includes users of the reduction, variables (which form a cycle
4059 // which ends in the phi node).
4060 Instruction *ExitInstruction = nullptr;
4061 // Indicates that we found a reduction operation in our scan.
4062 bool FoundReduxOp = false;
4064 // We start with the PHI node and scan for all of the users of this
4065 // instruction. All users must be instructions that can be used as reduction
4066 // variables (such as ADD). We must have a single out-of-block user. The cycle
4067 // must include the original PHI.
4068 bool FoundStartPHI = false;
4070 // To recognize min/max patterns formed by a icmp select sequence, we store
4071 // the number of instruction we saw from the recognized min/max pattern,
4072 // to make sure we only see exactly the two instructions.
4073 unsigned NumCmpSelectPatternInst = 0;
4074 ReductionInstDesc ReduxDesc(false, nullptr);
4076 SmallPtrSet<Instruction *, 8> VisitedInsts;
4077 SmallVector<Instruction *, 8> Worklist;
4078 Worklist.push_back(Phi);
4079 VisitedInsts.insert(Phi);
4081 // A value in the reduction can be used:
4082 // - By the reduction:
4083 // - Reduction operation:
4084 // - One use of reduction value (safe).
4085 // - Multiple use of reduction value (not safe).
4087 // - All uses of the PHI must be the reduction (safe).
4088 // - Otherwise, not safe.
4089 // - By one instruction outside of the loop (safe).
4090 // - By further instructions outside of the loop (not safe).
4091 // - By an instruction that is not part of the reduction (not safe).
4093 // * An instruction type other than PHI or the reduction operation.
4094 // * A PHI in the header other than the initial PHI.
4095 while (!Worklist.empty()) {
4096 Instruction *Cur = Worklist.back();
4097 Worklist.pop_back();
4100 // If the instruction has no users then this is a broken chain and can't be
4101 // a reduction variable.
4102 if (Cur->use_empty())
4105 bool IsAPhi = isa<PHINode>(Cur);
4107 // A header PHI use other than the original PHI.
4108 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
4111 // Reductions of instructions such as Div, and Sub is only possible if the
4112 // LHS is the reduction variable.
4113 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
4114 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
4115 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
4118 // Any reduction instruction must be of one of the allowed kinds.
4119 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
4120 if (!ReduxDesc.IsReduction)
4123 // A reduction operation must only have one use of the reduction value.
4124 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
4125 hasMultipleUsesOf(Cur, VisitedInsts))
4128 // All inputs to a PHI node must be a reduction value.
4129 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
4132 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
4133 isa<SelectInst>(Cur)))
4134 ++NumCmpSelectPatternInst;
4135 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
4136 isa<SelectInst>(Cur)))
4137 ++NumCmpSelectPatternInst;
4139 // Check whether we found a reduction operator.
4140 FoundReduxOp |= !IsAPhi;
4142 // Process users of current instruction. Push non-PHI nodes after PHI nodes
4143 // onto the stack. This way we are going to have seen all inputs to PHI
4144 // nodes once we get to them.
4145 SmallVector<Instruction *, 8> NonPHIs;
4146 SmallVector<Instruction *, 8> PHIs;
4147 for (User *U : Cur->users()) {
4148 Instruction *UI = cast<Instruction>(U);
4150 // Check if we found the exit user.
4151 BasicBlock *Parent = UI->getParent();
4152 if (!TheLoop->contains(Parent)) {
4153 // Exit if you find multiple outside users or if the header phi node is
4154 // being used. In this case the user uses the value of the previous
4155 // iteration, in which case we would loose "VF-1" iterations of the
4156 // reduction operation if we vectorize.
4157 if (ExitInstruction != nullptr || Cur == Phi)
4160 // The instruction used by an outside user must be the last instruction
4161 // before we feed back to the reduction phi. Otherwise, we loose VF-1
4162 // operations on the value.
4163 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
4166 ExitInstruction = Cur;
4170 // Process instructions only once (termination). Each reduction cycle
4171 // value must only be used once, except by phi nodes and min/max
4172 // reductions which are represented as a cmp followed by a select.
4173 ReductionInstDesc IgnoredVal(false, nullptr);
4174 if (VisitedInsts.insert(UI).second) {
4175 if (isa<PHINode>(UI))
4178 NonPHIs.push_back(UI);
4179 } else if (!isa<PHINode>(UI) &&
4180 ((!isa<FCmpInst>(UI) &&
4181 !isa<ICmpInst>(UI) &&
4182 !isa<SelectInst>(UI)) ||
4183 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
4186 // Remember that we completed the cycle.
4188 FoundStartPHI = true;
4190 Worklist.append(PHIs.begin(), PHIs.end());
4191 Worklist.append(NonPHIs.begin(), NonPHIs.end());
4194 // This means we have seen one but not the other instruction of the
4195 // pattern or more than just a select and cmp.
4196 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
4197 NumCmpSelectPatternInst != 2)
4200 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
4203 // We found a reduction var if we have reached the original phi node and we
4204 // only have a single instruction with out-of-loop users.
4206 // This instruction is allowed to have out-of-loop users.
4207 AllowedExit.insert(ExitInstruction);
4209 // Save the description of this reduction variable.
4210 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
4211 ReduxDesc.MinMaxKind);
4212 Reductions[Phi] = RD;
4213 // We've ended the cycle. This is a reduction variable if we have an
4214 // outside user and it has a binary op.
4219 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
4220 /// pattern corresponding to a min(X, Y) or max(X, Y).
4221 LoopVectorizationLegality::ReductionInstDesc
4222 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
4223 ReductionInstDesc &Prev) {
4225 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
4226 "Expect a select instruction");
4227 Instruction *Cmp = nullptr;
4228 SelectInst *Select = nullptr;
4230 // We must handle the select(cmp()) as a single instruction. Advance to the
4232 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
4233 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
4234 return ReductionInstDesc(false, I);
4235 return ReductionInstDesc(Select, Prev.MinMaxKind);
4238 // Only handle single use cases for now.
4239 if (!(Select = dyn_cast<SelectInst>(I)))
4240 return ReductionInstDesc(false, I);
4241 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
4242 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
4243 return ReductionInstDesc(false, I);
4244 if (!Cmp->hasOneUse())
4245 return ReductionInstDesc(false, I);
4250 // Look for a min/max pattern.
4251 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4252 return ReductionInstDesc(Select, MRK_UIntMin);
4253 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4254 return ReductionInstDesc(Select, MRK_UIntMax);
4255 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4256 return ReductionInstDesc(Select, MRK_SIntMax);
4257 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4258 return ReductionInstDesc(Select, MRK_SIntMin);
4259 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4260 return ReductionInstDesc(Select, MRK_FloatMin);
4261 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4262 return ReductionInstDesc(Select, MRK_FloatMax);
4263 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4264 return ReductionInstDesc(Select, MRK_FloatMin);
4265 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4266 return ReductionInstDesc(Select, MRK_FloatMax);
4268 return ReductionInstDesc(false, I);
4271 LoopVectorizationLegality::ReductionInstDesc
4272 LoopVectorizationLegality::isReductionInstr(Instruction *I,
4274 ReductionInstDesc &Prev) {
4275 bool FP = I->getType()->isFloatingPointTy();
4276 bool FastMath = FP && I->hasUnsafeAlgebra();
4277 switch (I->getOpcode()) {
4279 return ReductionInstDesc(false, I);
4280 case Instruction::PHI:
4281 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
4282 Kind != RK_FloatMinMax))
4283 return ReductionInstDesc(false, I);
4284 return ReductionInstDesc(I, Prev.MinMaxKind);
4285 case Instruction::Sub:
4286 case Instruction::Add:
4287 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4288 case Instruction::Mul:
4289 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4290 case Instruction::And:
4291 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4292 case Instruction::Or:
4293 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4294 case Instruction::Xor:
4295 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4296 case Instruction::FMul:
4297 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4298 case Instruction::FSub:
4299 case Instruction::FAdd:
4300 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4301 case Instruction::FCmp:
4302 case Instruction::ICmp:
4303 case Instruction::Select:
4304 if (Kind != RK_IntegerMinMax &&
4305 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4306 return ReductionInstDesc(false, I);
4307 return isMinMaxSelectCmpPattern(I, Prev);
4311 bool llvm::isInductionPHI(PHINode *Phi, ScalarEvolution *SE,
4312 ConstantInt *&StepValue) {
4313 Type *PhiTy = Phi->getType();
4314 // We only handle integer and pointer inductions variables.
4315 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
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");
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);
4332 ConstantInt *CV = C->getValue();
4333 if (PhiTy->isIntegerTy()) {
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())
4345 const DataLayout &DL = Phi->getModule()->getDataLayout();
4346 int64_t Size = static_cast<int64_t>(DL.getTypeAllocSize(PointerElementType));
4347 int64_t CVSize = CV->getSExtValue();
4350 StepValue = ConstantInt::getSigned(CV->getType(), CVSize / Size);
4354 LoopVectorizationLegality::InductionKind
4355 LoopVectorizationLegality::isInductionVariable(PHINode *Phi,
4356 ConstantInt *&StepValue) {
4357 if (!isInductionPHI(Phi, SE, StepValue))
4358 return IK_NoInduction;
4360 Type *PhiTy = Phi->getType();
4361 // Found an Integer induction variable.
4362 if (PhiTy->isIntegerTy())
4363 return IK_IntInduction;
4364 // Found an Pointer induction variable.
4365 return IK_PtrInduction;
4368 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4369 Value *In0 = const_cast<Value*>(V);
4370 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4374 return Inductions.count(PN);
4377 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4378 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4381 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4382 SmallPtrSetImpl<Value *> &SafePtrs) {
4384 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4385 // Check that we don't have a constant expression that can trap as operand.
4386 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4388 if (Constant *C = dyn_cast<Constant>(*OI))
4392 // We might be able to hoist the load.
4393 if (it->mayReadFromMemory()) {
4394 LoadInst *LI = dyn_cast<LoadInst>(it);
4397 if (!SafePtrs.count(LI->getPointerOperand())) {
4398 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
4399 MaskedOp.insert(LI);
4406 // We don't predicate stores at the moment.
4407 if (it->mayWriteToMemory()) {
4408 StoreInst *SI = dyn_cast<StoreInst>(it);
4409 // We only support predication of stores in basic blocks with one
4414 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
4415 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
4417 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
4418 !isSinglePredecessor) {
4419 // Build a masked store if it is legal for the target, otherwise scalarize
4421 bool isLegalMaskedOp =
4422 isLegalMaskedStore(SI->getValueOperand()->getType(),
4423 SI->getPointerOperand());
4424 if (isLegalMaskedOp) {
4426 MaskedOp.insert(SI);
4435 // The instructions below can trap.
4436 switch (it->getOpcode()) {
4438 case Instruction::UDiv:
4439 case Instruction::SDiv:
4440 case Instruction::URem:
4441 case Instruction::SRem:
4449 LoopVectorizationCostModel::VectorizationFactor
4450 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
4451 // Width 1 means no vectorize
4452 VectorizationFactor Factor = { 1U, 0U };
4453 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4454 emitAnalysis(VectorizationReport() <<
4455 "runtime pointer checks needed. Enable vectorization of this "
4456 "loop with '#pragma clang loop vectorize(enable)' when "
4457 "compiling with -Os");
4458 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4462 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
4463 emitAnalysis(VectorizationReport() <<
4464 "store that is conditionally executed prevents vectorization");
4465 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
4469 // Find the trip count.
4470 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4471 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4473 unsigned WidestType = getWidestType();
4474 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4475 unsigned MaxSafeDepDist = -1U;
4476 if (Legal->getMaxSafeDepDistBytes() != -1U)
4477 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4478 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4479 WidestRegister : MaxSafeDepDist);
4480 unsigned MaxVectorSize = WidestRegister / WidestType;
4481 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4482 DEBUG(dbgs() << "LV: The Widest register is: "
4483 << WidestRegister << " bits.\n");
4485 if (MaxVectorSize == 0) {
4486 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4490 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
4491 " into one vector!");
4493 unsigned VF = MaxVectorSize;
4495 // If we optimize the program for size, avoid creating the tail loop.
4497 // If we are unable to calculate the trip count then don't try to vectorize.
4500 (VectorizationReport() <<
4501 "unable to calculate the loop count due to complex control flow");
4502 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4506 // Find the maximum SIMD width that can fit within the trip count.
4507 VF = TC % MaxVectorSize;
4512 // If the trip count that we found modulo the vectorization factor is not
4513 // zero then we require a tail.
4515 emitAnalysis(VectorizationReport() <<
4516 "cannot optimize for size and vectorize at the "
4517 "same time. Enable vectorization of this loop "
4518 "with '#pragma clang loop vectorize(enable)' "
4519 "when compiling with -Os");
4520 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4525 int UserVF = Hints->getWidth();
4527 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4528 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4530 Factor.Width = UserVF;
4534 float Cost = expectedCost(1);
4536 const float ScalarCost = Cost;
4539 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
4541 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
4542 // Ignore scalar width, because the user explicitly wants vectorization.
4543 if (ForceVectorization && VF > 1) {
4545 Cost = expectedCost(Width) / (float)Width;
4548 for (unsigned i=2; i <= VF; i*=2) {
4549 // Notice that the vector loop needs to be executed less times, so
4550 // we need to divide the cost of the vector loops by the width of
4551 // the vector elements.
4552 float VectorCost = expectedCost(i) / (float)i;
4553 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
4554 (int)VectorCost << ".\n");
4555 if (VectorCost < Cost) {
4561 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
4562 << "LV: Vectorization seems to be not beneficial, "
4563 << "but was forced by a user.\n");
4564 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
4565 Factor.Width = Width;
4566 Factor.Cost = Width * Cost;
4570 unsigned LoopVectorizationCostModel::getWidestType() {
4571 unsigned MaxWidth = 8;
4572 const DataLayout &DL = TheFunction->getParent()->getDataLayout();
4575 for (Loop::block_iterator bb = TheLoop->block_begin(),
4576 be = TheLoop->block_end(); bb != be; ++bb) {
4577 BasicBlock *BB = *bb;
4579 // For each instruction in the loop.
4580 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4581 Type *T = it->getType();
4583 // Ignore ephemeral values.
4584 if (EphValues.count(it))
4587 // Only examine Loads, Stores and PHINodes.
4588 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4591 // Examine PHI nodes that are reduction variables.
4592 if (PHINode *PN = dyn_cast<PHINode>(it))
4593 if (!Legal->getReductionVars()->count(PN))
4596 // Examine the stored values.
4597 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4598 T = ST->getValueOperand()->getType();
4600 // Ignore loaded pointer types and stored pointer types that are not
4601 // consecutive. However, we do want to take consecutive stores/loads of
4602 // pointer vectors into account.
4603 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4606 MaxWidth = std::max(MaxWidth,
4607 (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
4615 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
4617 unsigned LoopCost) {
4619 // -- The unroll heuristics --
4620 // We unroll the loop in order to expose ILP and reduce the loop overhead.
4621 // There are many micro-architectural considerations that we can't predict
4622 // at this level. For example, frontend pressure (on decode or fetch) due to
4623 // code size, or the number and capabilities of the execution ports.
4625 // We use the following heuristics to select the unroll factor:
4626 // 1. If the code has reductions, then we unroll in order to break the cross
4627 // iteration dependency.
4628 // 2. If the loop is really small, then we unroll in order to reduce the loop
4630 // 3. We don't unroll if we think that we will spill registers to memory due
4631 // to the increased register pressure.
4633 // Use the user preference, unless 'auto' is selected.
4634 int UserUF = Hints->getInterleave();
4638 // When we optimize for size, we don't unroll.
4642 // We used the distance for the unroll factor.
4643 if (Legal->getMaxSafeDepDistBytes() != -1U)
4646 // Do not unroll loops with a relatively small trip count.
4647 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4648 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
4651 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
4652 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
4656 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
4657 TargetNumRegisters = ForceTargetNumScalarRegs;
4659 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
4660 TargetNumRegisters = ForceTargetNumVectorRegs;
4663 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4664 // We divide by these constants so assume that we have at least one
4665 // instruction that uses at least one register.
4666 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4667 R.NumInstructions = std::max(R.NumInstructions, 1U);
4669 // We calculate the unroll factor using the following formula.
4670 // Subtract the number of loop invariants from the number of available
4671 // registers. These registers are used by all of the unrolled instances.
4672 // Next, divide the remaining registers by the number of registers that is
4673 // required by the loop, in order to estimate how many parallel instances
4674 // fit without causing spills. All of this is rounded down if necessary to be
4675 // a power of two. We want power of two unroll factors to simplify any
4676 // addressing operations or alignment considerations.
4677 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
4680 // Don't count the induction variable as unrolled.
4681 if (EnableIndVarRegisterHeur)
4682 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
4683 std::max(1U, (R.MaxLocalUsers - 1)));
4685 // Clamp the unroll factor ranges to reasonable factors.
4686 unsigned MaxInterleaveSize = TTI.getMaxInterleaveFactor();
4688 // Check if the user has overridden the unroll max.
4690 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
4691 MaxInterleaveSize = ForceTargetMaxScalarInterleaveFactor;
4693 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
4694 MaxInterleaveSize = ForceTargetMaxVectorInterleaveFactor;
4697 // If we did not calculate the cost for VF (because the user selected the VF)
4698 // then we calculate the cost of VF here.
4700 LoopCost = expectedCost(VF);
4702 // Clamp the calculated UF to be between the 1 and the max unroll factor
4703 // that the target allows.
4704 if (UF > MaxInterleaveSize)
4705 UF = MaxInterleaveSize;
4709 // Unroll if we vectorized this loop and there is a reduction that could
4710 // benefit from unrolling.
4711 if (VF > 1 && Legal->getReductionVars()->size()) {
4712 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
4716 // Note that if we've already vectorized the loop we will have done the
4717 // runtime check and so unrolling won't require further checks.
4718 bool UnrollingRequiresRuntimePointerCheck =
4719 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
4721 // We want to unroll small loops in order to reduce the loop overhead and
4722 // potentially expose ILP opportunities.
4723 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
4724 if (!UnrollingRequiresRuntimePointerCheck &&
4725 LoopCost < SmallLoopCost) {
4726 // We assume that the cost overhead is 1 and we use the cost model
4727 // to estimate the cost of the loop and unroll until the cost of the
4728 // loop overhead is about 5% of the cost of the loop.
4729 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
4731 // Unroll until store/load ports (estimated by max unroll factor) are
4733 unsigned NumStores = Legal->getNumStores();
4734 unsigned NumLoads = Legal->getNumLoads();
4735 unsigned StoresUF = UF / (NumStores ? NumStores : 1);
4736 unsigned LoadsUF = UF / (NumLoads ? NumLoads : 1);
4738 // If we have a scalar reduction (vector reductions are already dealt with
4739 // by this point), we can increase the critical path length if the loop
4740 // we're unrolling is inside another loop. Limit, by default to 2, so the
4741 // critical path only gets increased by one reduction operation.
4742 if (Legal->getReductionVars()->size() &&
4743 TheLoop->getLoopDepth() > 1) {
4744 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionUF);
4745 SmallUF = std::min(SmallUF, F);
4746 StoresUF = std::min(StoresUF, F);
4747 LoadsUF = std::min(LoadsUF, F);
4750 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
4751 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
4752 return std::max(StoresUF, LoadsUF);
4755 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
4759 // Unroll if this is a large loop (small loops are already dealt with by this
4760 // point) that could benefit from interleaved unrolling.
4761 bool HasReductions = (Legal->getReductionVars()->size() > 0);
4762 if (TTI.enableAggressiveInterleaving(HasReductions)) {
4763 DEBUG(dbgs() << "LV: Unrolling to expose ILP.\n");
4767 DEBUG(dbgs() << "LV: Not Unrolling.\n");
4771 LoopVectorizationCostModel::RegisterUsage
4772 LoopVectorizationCostModel::calculateRegisterUsage() {
4773 // This function calculates the register usage by measuring the highest number
4774 // of values that are alive at a single location. Obviously, this is a very
4775 // rough estimation. We scan the loop in a topological order in order and
4776 // assign a number to each instruction. We use RPO to ensure that defs are
4777 // met before their users. We assume that each instruction that has in-loop
4778 // users starts an interval. We record every time that an in-loop value is
4779 // used, so we have a list of the first and last occurrences of each
4780 // instruction. Next, we transpose this data structure into a multi map that
4781 // holds the list of intervals that *end* at a specific location. This multi
4782 // map allows us to perform a linear search. We scan the instructions linearly
4783 // and record each time that a new interval starts, by placing it in a set.
4784 // If we find this value in the multi-map then we remove it from the set.
4785 // The max register usage is the maximum size of the set.
4786 // We also search for instructions that are defined outside the loop, but are
4787 // used inside the loop. We need this number separately from the max-interval
4788 // usage number because when we unroll, loop-invariant values do not take
4790 LoopBlocksDFS DFS(TheLoop);
4794 R.NumInstructions = 0;
4796 // Each 'key' in the map opens a new interval. The values
4797 // of the map are the index of the 'last seen' usage of the
4798 // instruction that is the key.
4799 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4800 // Maps instruction to its index.
4801 DenseMap<unsigned, Instruction*> IdxToInstr;
4802 // Marks the end of each interval.
4803 IntervalMap EndPoint;
4804 // Saves the list of instruction indices that are used in the loop.
4805 SmallSet<Instruction*, 8> Ends;
4806 // Saves the list of values that are used in the loop but are
4807 // defined outside the loop, such as arguments and constants.
4808 SmallPtrSet<Value*, 8> LoopInvariants;
4811 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4812 be = DFS.endRPO(); bb != be; ++bb) {
4813 R.NumInstructions += (*bb)->size();
4814 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4816 Instruction *I = it;
4817 IdxToInstr[Index++] = I;
4819 // Save the end location of each USE.
4820 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4821 Value *U = I->getOperand(i);
4822 Instruction *Instr = dyn_cast<Instruction>(U);
4824 // Ignore non-instruction values such as arguments, constants, etc.
4825 if (!Instr) continue;
4827 // If this instruction is outside the loop then record it and continue.
4828 if (!TheLoop->contains(Instr)) {
4829 LoopInvariants.insert(Instr);
4833 // Overwrite previous end points.
4834 EndPoint[Instr] = Index;
4840 // Saves the list of intervals that end with the index in 'key'.
4841 typedef SmallVector<Instruction*, 2> InstrList;
4842 DenseMap<unsigned, InstrList> TransposeEnds;
4844 // Transpose the EndPoints to a list of values that end at each index.
4845 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4847 TransposeEnds[it->second].push_back(it->first);
4849 SmallSet<Instruction*, 8> OpenIntervals;
4850 unsigned MaxUsage = 0;
4853 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4854 for (unsigned int i = 0; i < Index; ++i) {
4855 Instruction *I = IdxToInstr[i];
4856 // Ignore instructions that are never used within the loop.
4857 if (!Ends.count(I)) continue;
4859 // Ignore ephemeral values.
4860 if (EphValues.count(I))
4863 // Remove all of the instructions that end at this location.
4864 InstrList &List = TransposeEnds[i];
4865 for (unsigned int j=0, e = List.size(); j < e; ++j)
4866 OpenIntervals.erase(List[j]);
4868 // Count the number of live interals.
4869 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4871 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4872 OpenIntervals.size() << '\n');
4874 // Add the current instruction to the list of open intervals.
4875 OpenIntervals.insert(I);
4878 unsigned Invariant = LoopInvariants.size();
4879 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
4880 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
4881 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
4883 R.LoopInvariantRegs = Invariant;
4884 R.MaxLocalUsers = MaxUsage;
4888 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4892 for (Loop::block_iterator bb = TheLoop->block_begin(),
4893 be = TheLoop->block_end(); bb != be; ++bb) {
4894 unsigned BlockCost = 0;
4895 BasicBlock *BB = *bb;
4897 // For each instruction in the old loop.
4898 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4899 // Skip dbg intrinsics.
4900 if (isa<DbgInfoIntrinsic>(it))
4903 // Ignore ephemeral values.
4904 if (EphValues.count(it))
4907 unsigned C = getInstructionCost(it, VF);
4909 // Check if we should override the cost.
4910 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
4911 C = ForceTargetInstructionCost;
4914 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
4915 VF << " For instruction: " << *it << '\n');
4918 // We assume that if-converted blocks have a 50% chance of being executed.
4919 // When the code is scalar then some of the blocks are avoided due to CF.
4920 // When the code is vectorized we execute all code paths.
4921 if (VF == 1 && Legal->blockNeedsPredication(*bb))
4930 /// \brief Check whether the address computation for a non-consecutive memory
4931 /// access looks like an unlikely candidate for being merged into the indexing
4934 /// We look for a GEP which has one index that is an induction variable and all
4935 /// other indices are loop invariant. If the stride of this access is also
4936 /// within a small bound we decide that this address computation can likely be
4937 /// merged into the addressing mode.
4938 /// In all other cases, we identify the address computation as complex.
4939 static bool isLikelyComplexAddressComputation(Value *Ptr,
4940 LoopVectorizationLegality *Legal,
4941 ScalarEvolution *SE,
4942 const Loop *TheLoop) {
4943 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
4947 // We are looking for a gep with all loop invariant indices except for one
4948 // which should be an induction variable.
4949 unsigned NumOperands = Gep->getNumOperands();
4950 for (unsigned i = 1; i < NumOperands; ++i) {
4951 Value *Opd = Gep->getOperand(i);
4952 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
4953 !Legal->isInductionVariable(Opd))
4957 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
4958 // can likely be merged into the address computation.
4959 unsigned MaxMergeDistance = 64;
4961 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
4965 // Check the step is constant.
4966 const SCEV *Step = AddRec->getStepRecurrence(*SE);
4967 // Calculate the pointer stride and check if it is consecutive.
4968 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4972 const APInt &APStepVal = C->getValue()->getValue();
4974 // Huge step value - give up.
4975 if (APStepVal.getBitWidth() > 64)
4978 int64_t StepVal = APStepVal.getSExtValue();
4980 return StepVal > MaxMergeDistance;
4983 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
4984 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
4990 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
4991 // If we know that this instruction will remain uniform, check the cost of
4992 // the scalar version.
4993 if (Legal->isUniformAfterVectorization(I))
4996 Type *RetTy = I->getType();
4997 Type *VectorTy = ToVectorTy(RetTy, VF);
4999 // TODO: We need to estimate the cost of intrinsic calls.
5000 switch (I->getOpcode()) {
5001 case Instruction::GetElementPtr:
5002 // We mark this instruction as zero-cost because the cost of GEPs in
5003 // vectorized code depends on whether the corresponding memory instruction
5004 // is scalarized or not. Therefore, we handle GEPs with the memory
5005 // instruction cost.
5007 case Instruction::Br: {
5008 return TTI.getCFInstrCost(I->getOpcode());
5010 case Instruction::PHI:
5011 //TODO: IF-converted IFs become selects.
5013 case Instruction::Add:
5014 case Instruction::FAdd:
5015 case Instruction::Sub:
5016 case Instruction::FSub:
5017 case Instruction::Mul:
5018 case Instruction::FMul:
5019 case Instruction::UDiv:
5020 case Instruction::SDiv:
5021 case Instruction::FDiv:
5022 case Instruction::URem:
5023 case Instruction::SRem:
5024 case Instruction::FRem:
5025 case Instruction::Shl:
5026 case Instruction::LShr:
5027 case Instruction::AShr:
5028 case Instruction::And:
5029 case Instruction::Or:
5030 case Instruction::Xor: {
5031 // Since we will replace the stride by 1 the multiplication should go away.
5032 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5034 // Certain instructions can be cheaper to vectorize if they have a constant
5035 // second vector operand. One example of this are shifts on x86.
5036 TargetTransformInfo::OperandValueKind Op1VK =
5037 TargetTransformInfo::OK_AnyValue;
5038 TargetTransformInfo::OperandValueKind Op2VK =
5039 TargetTransformInfo::OK_AnyValue;
5040 TargetTransformInfo::OperandValueProperties Op1VP =
5041 TargetTransformInfo::OP_None;
5042 TargetTransformInfo::OperandValueProperties Op2VP =
5043 TargetTransformInfo::OP_None;
5044 Value *Op2 = I->getOperand(1);
5046 // Check for a splat of a constant or for a non uniform vector of constants.
5047 if (isa<ConstantInt>(Op2)) {
5048 ConstantInt *CInt = cast<ConstantInt>(Op2);
5049 if (CInt && CInt->getValue().isPowerOf2())
5050 Op2VP = TargetTransformInfo::OP_PowerOf2;
5051 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5052 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5053 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5054 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
5056 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
5057 if (CInt && CInt->getValue().isPowerOf2())
5058 Op2VP = TargetTransformInfo::OP_PowerOf2;
5059 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5063 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
5066 case Instruction::Select: {
5067 SelectInst *SI = cast<SelectInst>(I);
5068 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5069 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5070 Type *CondTy = SI->getCondition()->getType();
5072 CondTy = VectorType::get(CondTy, VF);
5074 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5076 case Instruction::ICmp:
5077 case Instruction::FCmp: {
5078 Type *ValTy = I->getOperand(0)->getType();
5079 VectorTy = ToVectorTy(ValTy, VF);
5080 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5082 case Instruction::Store:
5083 case Instruction::Load: {
5084 StoreInst *SI = dyn_cast<StoreInst>(I);
5085 LoadInst *LI = dyn_cast<LoadInst>(I);
5086 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5088 VectorTy = ToVectorTy(ValTy, VF);
5090 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5091 unsigned AS = SI ? SI->getPointerAddressSpace() :
5092 LI->getPointerAddressSpace();
5093 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5094 // We add the cost of address computation here instead of with the gep
5095 // instruction because only here we know whether the operation is
5098 return TTI.getAddressComputationCost(VectorTy) +
5099 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5101 // Scalarized loads/stores.
5102 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5103 bool Reverse = ConsecutiveStride < 0;
5104 const DataLayout &DL = I->getModule()->getDataLayout();
5105 unsigned ScalarAllocatedSize = DL.getTypeAllocSize(ValTy);
5106 unsigned VectorElementSize = DL.getTypeStoreSize(VectorTy) / VF;
5107 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5108 bool IsComplexComputation =
5109 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5111 // The cost of extracting from the value vector and pointer vector.
5112 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5113 for (unsigned i = 0; i < VF; ++i) {
5114 // The cost of extracting the pointer operand.
5115 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5116 // In case of STORE, the cost of ExtractElement from the vector.
5117 // In case of LOAD, the cost of InsertElement into the returned
5119 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5120 Instruction::InsertElement,
5124 // The cost of the scalar loads/stores.
5125 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5126 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5131 // Wide load/stores.
5132 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5133 if (Legal->isMaskRequired(I))
5134 Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment,
5137 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5140 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5144 case Instruction::ZExt:
5145 case Instruction::SExt:
5146 case Instruction::FPToUI:
5147 case Instruction::FPToSI:
5148 case Instruction::FPExt:
5149 case Instruction::PtrToInt:
5150 case Instruction::IntToPtr:
5151 case Instruction::SIToFP:
5152 case Instruction::UIToFP:
5153 case Instruction::Trunc:
5154 case Instruction::FPTrunc:
5155 case Instruction::BitCast: {
5156 // We optimize the truncation of induction variable.
5157 // The cost of these is the same as the scalar operation.
5158 if (I->getOpcode() == Instruction::Trunc &&
5159 Legal->isInductionVariable(I->getOperand(0)))
5160 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5161 I->getOperand(0)->getType());
5163 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5164 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5166 case Instruction::Call: {
5167 bool NeedToScalarize;
5168 CallInst *CI = cast<CallInst>(I);
5169 unsigned CallCost = getVectorCallCost(CI, VF, TTI, TLI, NeedToScalarize);
5170 if (getIntrinsicIDForCall(CI, TLI))
5171 return std::min(CallCost, getVectorIntrinsicCost(CI, VF, TTI, TLI));
5175 // We are scalarizing the instruction. Return the cost of the scalar
5176 // instruction, plus the cost of insert and extract into vector
5177 // elements, times the vector width.
5180 if (!RetTy->isVoidTy() && VF != 1) {
5181 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5183 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5186 // The cost of inserting the results plus extracting each one of the
5188 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5191 // The cost of executing VF copies of the scalar instruction. This opcode
5192 // is unknown. Assume that it is the same as 'mul'.
5193 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5199 char LoopVectorize::ID = 0;
5200 static const char lv_name[] = "Loop Vectorization";
5201 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5202 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
5203 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
5204 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
5205 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
5206 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5207 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5208 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5209 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
5210 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5211 INITIALIZE_PASS_DEPENDENCY(LoopAccessAnalysis)
5212 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5215 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5216 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5220 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5221 // Check for a store.
5222 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5223 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5225 // Check for a load.
5226 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5227 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5233 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5234 bool IfPredicateStore) {
5235 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5236 // Holds vector parameters or scalars, in case of uniform vals.
5237 SmallVector<VectorParts, 4> Params;
5239 setDebugLocFromInst(Builder, Instr);
5241 // Find all of the vectorized parameters.
5242 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5243 Value *SrcOp = Instr->getOperand(op);
5245 // If we are accessing the old induction variable, use the new one.
5246 if (SrcOp == OldInduction) {
5247 Params.push_back(getVectorValue(SrcOp));
5251 // Try using previously calculated values.
5252 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5254 // If the src is an instruction that appeared earlier in the basic block
5255 // then it should already be vectorized.
5256 if (SrcInst && OrigLoop->contains(SrcInst)) {
5257 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5258 // The parameter is a vector value from earlier.
5259 Params.push_back(WidenMap.get(SrcInst));
5261 // The parameter is a scalar from outside the loop. Maybe even a constant.
5262 VectorParts Scalars;
5263 Scalars.append(UF, SrcOp);
5264 Params.push_back(Scalars);
5268 assert(Params.size() == Instr->getNumOperands() &&
5269 "Invalid number of operands");
5271 // Does this instruction return a value ?
5272 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5274 Value *UndefVec = IsVoidRetTy ? nullptr :
5275 UndefValue::get(Instr->getType());
5276 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5277 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5279 Instruction *InsertPt = Builder.GetInsertPoint();
5280 BasicBlock *IfBlock = Builder.GetInsertBlock();
5281 BasicBlock *CondBlock = nullptr;
5284 Loop *VectorLp = nullptr;
5285 if (IfPredicateStore) {
5286 assert(Instr->getParent()->getSinglePredecessor() &&
5287 "Only support single predecessor blocks");
5288 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5289 Instr->getParent());
5290 VectorLp = LI->getLoopFor(IfBlock);
5291 assert(VectorLp && "Must have a loop for this block");
5294 // For each vector unroll 'part':
5295 for (unsigned Part = 0; Part < UF; ++Part) {
5296 // For each scalar that we create:
5298 // Start an "if (pred) a[i] = ..." block.
5299 Value *Cmp = nullptr;
5300 if (IfPredicateStore) {
5301 if (Cond[Part]->getType()->isVectorTy())
5303 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5304 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5305 ConstantInt::get(Cond[Part]->getType(), 1));
5306 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5307 LoopVectorBody.push_back(CondBlock);
5308 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
5309 // Update Builder with newly created basic block.
5310 Builder.SetInsertPoint(InsertPt);
5313 Instruction *Cloned = Instr->clone();
5315 Cloned->setName(Instr->getName() + ".cloned");
5316 // Replace the operands of the cloned instructions with extracted scalars.
5317 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5318 Value *Op = Params[op][Part];
5319 Cloned->setOperand(op, Op);
5322 // Place the cloned scalar in the new loop.
5323 Builder.Insert(Cloned);
5325 // If the original scalar returns a value we need to place it in a vector
5326 // so that future users will be able to use it.
5328 VecResults[Part] = Cloned;
5331 if (IfPredicateStore) {
5332 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
5333 LoopVectorBody.push_back(NewIfBlock);
5334 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
5335 Builder.SetInsertPoint(InsertPt);
5336 Instruction *OldBr = IfBlock->getTerminator();
5337 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
5338 OldBr->eraseFromParent();
5339 IfBlock = NewIfBlock;
5344 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5345 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5346 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5348 return scalarizeInstruction(Instr, IfPredicateStore);
5351 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5355 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5359 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) {
5360 // When unrolling and the VF is 1, we only need to add a simple scalar.
5361 Type *ITy = Val->getType();
5362 assert(!ITy->isVectorTy() && "Val must be a scalar");
5363 Constant *C = ConstantInt::get(ITy, StartIdx);
5364 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");