1 //===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops
11 // and generates target-independent LLVM-IR.
12 // The vectorizer uses the TargetTransformInfo analysis to estimate the costs
13 // of instructions in order to estimate the profitability of vectorization.
15 // The loop vectorizer combines consecutive loop iterations into a single
16 // 'wide' iteration. After this transformation the index is incremented
17 // by the SIMD vector width, and not by one.
19 // This pass has three parts:
20 // 1. The main loop pass that drives the different parts.
21 // 2. LoopVectorizationLegality - A unit that checks for the legality
22 // of the vectorization.
23 // 3. InnerLoopVectorizer - A unit that performs the actual
24 // widening of instructions.
25 // 4. LoopVectorizationCostModel - A unit that checks for the profitability
26 // of vectorization. It decides on the optimal vector width, which
27 // can be one, if vectorization is not profitable.
29 //===----------------------------------------------------------------------===//
31 // The reduction-variable vectorization is based on the paper:
32 // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
34 // Variable uniformity checks are inspired by:
35 // Karrenberg, R. and Hack, S. Whole Function Vectorization.
37 // Other ideas/concepts are from:
38 // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
40 // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of
41 // Vectorizing Compilers.
43 //===----------------------------------------------------------------------===//
45 #include "llvm/Transforms/Vectorize.h"
46 #include "llvm/ADT/DenseMap.h"
47 #include "llvm/ADT/EquivalenceClasses.h"
48 #include "llvm/ADT/Hashing.h"
49 #include "llvm/ADT/MapVector.h"
50 #include "llvm/ADT/SetVector.h"
51 #include "llvm/ADT/SmallPtrSet.h"
52 #include "llvm/ADT/SmallSet.h"
53 #include "llvm/ADT/SmallVector.h"
54 #include "llvm/ADT/Statistic.h"
55 #include "llvm/ADT/StringExtras.h"
56 #include "llvm/Analysis/AliasAnalysis.h"
57 #include "llvm/Analysis/AliasSetTracker.h"
58 #include "llvm/Analysis/AssumptionCache.h"
59 #include "llvm/Analysis/BlockFrequencyInfo.h"
60 #include "llvm/Analysis/CodeMetrics.h"
61 #include "llvm/Analysis/LoopAccessAnalysis.h"
62 #include "llvm/Analysis/LoopInfo.h"
63 #include "llvm/Analysis/LoopIterator.h"
64 #include "llvm/Analysis/LoopPass.h"
65 #include "llvm/Analysis/ScalarEvolution.h"
66 #include "llvm/Analysis/ScalarEvolutionExpander.h"
67 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
68 #include "llvm/Analysis/TargetTransformInfo.h"
69 #include "llvm/Analysis/ValueTracking.h"
70 #include "llvm/IR/Constants.h"
71 #include "llvm/IR/DataLayout.h"
72 #include "llvm/IR/DebugInfo.h"
73 #include "llvm/IR/DerivedTypes.h"
74 #include "llvm/IR/DiagnosticInfo.h"
75 #include "llvm/IR/Dominators.h"
76 #include "llvm/IR/Function.h"
77 #include "llvm/IR/IRBuilder.h"
78 #include "llvm/IR/Instructions.h"
79 #include "llvm/IR/IntrinsicInst.h"
80 #include "llvm/IR/LLVMContext.h"
81 #include "llvm/IR/Module.h"
82 #include "llvm/IR/PatternMatch.h"
83 #include "llvm/IR/Type.h"
84 #include "llvm/IR/Value.h"
85 #include "llvm/IR/ValueHandle.h"
86 #include "llvm/IR/Verifier.h"
87 #include "llvm/Pass.h"
88 #include "llvm/Support/BranchProbability.h"
89 #include "llvm/Support/CommandLine.h"
90 #include "llvm/Support/Debug.h"
91 #include "llvm/Support/raw_ostream.h"
92 #include "llvm/Transforms/Scalar.h"
93 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
94 #include "llvm/Transforms/Utils/Local.h"
95 #include "llvm/Transforms/Utils/VectorUtils.h"
100 using namespace llvm;
101 using namespace llvm::PatternMatch;
103 #define LV_NAME "loop-vectorize"
104 #define DEBUG_TYPE LV_NAME
106 STATISTIC(LoopsVectorized, "Number of loops vectorized");
107 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
110 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
111 cl::desc("Enable if-conversion during vectorization."));
113 /// We don't vectorize loops with a known constant trip count below this number.
114 static cl::opt<unsigned>
115 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
117 cl::desc("Don't vectorize loops with a constant "
118 "trip count that is smaller than this "
121 /// This enables versioning on the strides of symbolically striding memory
122 /// accesses in code like the following.
123 /// for (i = 0; i < N; ++i)
124 /// A[i * Stride1] += B[i * Stride2] ...
126 /// Will be roughly translated to
127 /// if (Stride1 == 1 && Stride2 == 1) {
128 /// for (i = 0; i < N; i+=4)
132 static cl::opt<bool> EnableMemAccessVersioning(
133 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
134 cl::desc("Enable symblic stride memory access versioning"));
136 /// We don't unroll loops with a known constant trip count below this number.
137 static const unsigned TinyTripCountUnrollThreshold = 128;
139 static cl::opt<unsigned> ForceTargetNumScalarRegs(
140 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
141 cl::desc("A flag that overrides the target's number of scalar registers."));
143 static cl::opt<unsigned> ForceTargetNumVectorRegs(
144 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
145 cl::desc("A flag that overrides the target's number of vector registers."));
147 /// Maximum vectorization interleave count.
148 static const unsigned MaxInterleaveFactor = 16;
150 static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
151 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
152 cl::desc("A flag that overrides the target's max interleave factor for "
155 static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
156 "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
157 cl::desc("A flag that overrides the target's max interleave factor for "
158 "vectorized loops."));
160 static cl::opt<unsigned> ForceTargetInstructionCost(
161 "force-target-instruction-cost", cl::init(0), cl::Hidden,
162 cl::desc("A flag that overrides the target's expected cost for "
163 "an instruction to a single constant value. Mostly "
164 "useful for getting consistent testing."));
166 static cl::opt<unsigned> SmallLoopCost(
167 "small-loop-cost", cl::init(20), cl::Hidden,
168 cl::desc("The cost of a loop that is considered 'small' by the unroller."));
170 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
171 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
172 cl::desc("Enable the use of the block frequency analysis to access PGO "
173 "heuristics minimizing code growth in cold regions and being more "
174 "aggressive in hot regions."));
176 // Runtime unroll loops for load/store throughput.
177 static cl::opt<bool> EnableLoadStoreRuntimeUnroll(
178 "enable-loadstore-runtime-unroll", cl::init(true), cl::Hidden,
179 cl::desc("Enable runtime unrolling until load/store ports are saturated"));
181 /// The number of stores in a loop that are allowed to need predication.
182 static cl::opt<unsigned> NumberOfStoresToPredicate(
183 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
184 cl::desc("Max number of stores to be predicated behind an if."));
186 static cl::opt<bool> EnableIndVarRegisterHeur(
187 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
188 cl::desc("Count the induction variable only once when unrolling"));
190 static cl::opt<bool> EnableCondStoresVectorization(
191 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
192 cl::desc("Enable if predication of stores during vectorization."));
194 static cl::opt<unsigned> MaxNestedScalarReductionUF(
195 "max-nested-scalar-reduction-unroll", cl::init(2), cl::Hidden,
196 cl::desc("The maximum unroll factor to use when unrolling a scalar "
197 "reduction in a nested loop."));
201 // Forward declarations.
202 class LoopVectorizationLegality;
203 class LoopVectorizationCostModel;
204 class LoopVectorizeHints;
206 /// InnerLoopVectorizer vectorizes loops which contain only one basic
207 /// block to a specified vectorization factor (VF).
208 /// This class performs the widening of scalars into vectors, or multiple
209 /// scalars. This class also implements the following features:
210 /// * It inserts an epilogue loop for handling loops that don't have iteration
211 /// counts that are known to be a multiple of the vectorization factor.
212 /// * It handles the code generation for reduction variables.
213 /// * Scalarization (implementation using scalars) of un-vectorizable
215 /// InnerLoopVectorizer does not perform any vectorization-legality
216 /// checks, and relies on the caller to check for the different legality
217 /// aspects. The InnerLoopVectorizer relies on the
218 /// LoopVectorizationLegality class to provide information about the induction
219 /// and reduction variables that were found to a given vectorization factor.
220 class InnerLoopVectorizer {
222 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
223 DominatorTree *DT, const DataLayout *DL,
224 const TargetLibraryInfo *TLI, unsigned VecWidth,
225 unsigned UnrollFactor)
226 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
227 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
228 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
231 // Perform the actual loop widening (vectorization).
232 void vectorize(LoopVectorizationLegality *L) {
234 // Create a new empty loop. Unlink the old loop and connect the new one.
236 // Widen each instruction in the old loop to a new one in the new loop.
237 // Use the Legality module to find the induction and reduction variables.
239 // Register the new loop and update the analysis passes.
243 virtual ~InnerLoopVectorizer() {}
246 /// A small list of PHINodes.
247 typedef SmallVector<PHINode*, 4> PhiVector;
248 /// When we unroll loops we have multiple vector values for each scalar.
249 /// This data structure holds the unrolled and vectorized values that
250 /// originated from one scalar instruction.
251 typedef SmallVector<Value*, 2> VectorParts;
253 // When we if-convert we need create edge masks. We have to cache values so
254 // that we don't end up with exponential recursion/IR.
255 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
256 VectorParts> EdgeMaskCache;
258 /// \brief Add checks for strides that where assumed to be 1.
260 /// Returns the last check instruction and the first check instruction in the
261 /// pair as (first, last).
262 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
264 /// Create an empty loop, based on the loop ranges of the old loop.
265 void createEmptyLoop();
266 /// Copy and widen the instructions from the old loop.
267 virtual void vectorizeLoop();
269 /// \brief The Loop exit block may have single value PHI nodes where the
270 /// incoming value is 'Undef'. While vectorizing we only handled real values
271 /// that were defined inside the loop. Here we fix the 'undef case'.
275 /// A helper function that computes the predicate of the block BB, assuming
276 /// that the header block of the loop is set to True. It returns the *entry*
277 /// mask for the block BB.
278 VectorParts createBlockInMask(BasicBlock *BB);
279 /// A helper function that computes the predicate of the edge between SRC
281 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
283 /// A helper function to vectorize a single BB within the innermost loop.
284 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
286 /// Vectorize a single PHINode in a block. This method handles the induction
287 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
288 /// arbitrary length vectors.
289 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
290 unsigned UF, unsigned VF, PhiVector *PV);
292 /// Insert the new loop to the loop hierarchy and pass manager
293 /// and update the analysis passes.
294 void updateAnalysis();
296 /// This instruction is un-vectorizable. Implement it as a sequence
297 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
298 /// scalarized instruction behind an if block predicated on the control
299 /// dependence of the instruction.
300 virtual void scalarizeInstruction(Instruction *Instr,
301 bool IfPredicateStore=false);
303 /// Vectorize Load and Store instructions,
304 virtual void vectorizeMemoryInstruction(Instruction *Instr);
306 /// Create a broadcast instruction. This method generates a broadcast
307 /// instruction (shuffle) for loop invariant values and for the induction
308 /// value. If this is the induction variable then we extend it to N, N+1, ...
309 /// this is needed because each iteration in the loop corresponds to a SIMD
311 virtual Value *getBroadcastInstrs(Value *V);
313 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
314 /// to each vector element of Val. The sequence starts at StartIndex.
315 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step);
317 /// When we go over instructions in the basic block we rely on previous
318 /// values within the current basic block or on loop invariant values.
319 /// When we widen (vectorize) values we place them in the map. If the values
320 /// are not within the map, they have to be loop invariant, so we simply
321 /// broadcast them into a vector.
322 VectorParts &getVectorValue(Value *V);
324 /// Generate a shuffle sequence that will reverse the vector Vec.
325 virtual Value *reverseVector(Value *Vec);
327 /// This is a helper class that holds the vectorizer state. It maps scalar
328 /// instructions to vector instructions. When the code is 'unrolled' then
329 /// then a single scalar value is mapped to multiple vector parts. The parts
330 /// are stored in the VectorPart type.
332 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
334 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
336 /// \return True if 'Key' is saved in the Value Map.
337 bool has(Value *Key) const { return MapStorage.count(Key); }
339 /// Initializes a new entry in the map. Sets all of the vector parts to the
340 /// save value in 'Val'.
341 /// \return A reference to a vector with splat values.
342 VectorParts &splat(Value *Key, Value *Val) {
343 VectorParts &Entry = MapStorage[Key];
344 Entry.assign(UF, Val);
348 ///\return A reference to the value that is stored at 'Key'.
349 VectorParts &get(Value *Key) {
350 VectorParts &Entry = MapStorage[Key];
353 assert(Entry.size() == UF);
358 /// The unroll factor. Each entry in the map stores this number of vector
362 /// Map storage. We use std::map and not DenseMap because insertions to a
363 /// dense map invalidates its iterators.
364 std::map<Value *, VectorParts> MapStorage;
367 /// The original loop.
369 /// Scev analysis to use.
378 const DataLayout *DL;
379 /// Target Library Info.
380 const TargetLibraryInfo *TLI;
382 /// The vectorization SIMD factor to use. Each vector will have this many
387 /// The vectorization unroll factor to use. Each scalar is vectorized to this
388 /// many different vector instructions.
391 /// The builder that we use
394 // --- Vectorization state ---
396 /// The vector-loop preheader.
397 BasicBlock *LoopVectorPreHeader;
398 /// The scalar-loop preheader.
399 BasicBlock *LoopScalarPreHeader;
400 /// Middle Block between the vector and the scalar.
401 BasicBlock *LoopMiddleBlock;
402 ///The ExitBlock of the scalar loop.
403 BasicBlock *LoopExitBlock;
404 ///The vector loop body.
405 SmallVector<BasicBlock *, 4> LoopVectorBody;
406 ///The scalar loop body.
407 BasicBlock *LoopScalarBody;
408 /// A list of all bypass blocks. The first block is the entry of the loop.
409 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
411 /// The new Induction variable which was added to the new block.
413 /// The induction variable of the old basic block.
414 PHINode *OldInduction;
415 /// Holds the extended (to the widest induction type) start index.
417 /// Maps scalars to widened vectors.
419 EdgeMaskCache MaskCache;
421 LoopVectorizationLegality *Legal;
424 class InnerLoopUnroller : public InnerLoopVectorizer {
426 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
427 DominatorTree *DT, const DataLayout *DL,
428 const TargetLibraryInfo *TLI, unsigned UnrollFactor) :
429 InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { }
432 void scalarizeInstruction(Instruction *Instr,
433 bool IfPredicateStore = false) override;
434 void vectorizeMemoryInstruction(Instruction *Instr) override;
435 Value *getBroadcastInstrs(Value *V) override;
436 Value *getStepVector(Value *Val, int StartIdx, Value *Step) override;
437 Value *reverseVector(Value *Vec) override;
440 /// \brief Look for a meaningful debug location on the instruction or it's
442 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
447 if (I->getDebugLoc() != Empty)
450 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
451 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
452 if (OpInst->getDebugLoc() != Empty)
459 /// \brief Set the debug location in the builder using the debug location in the
461 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
462 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
463 B.SetCurrentDebugLocation(Inst->getDebugLoc());
465 B.SetCurrentDebugLocation(DebugLoc());
469 /// \return string containing a file name and a line # for the given loop.
470 static std::string getDebugLocString(const Loop *L) {
473 raw_string_ostream OS(Result);
474 const DebugLoc LoopDbgLoc = L->getStartLoc();
475 if (!LoopDbgLoc.isUnknown())
476 LoopDbgLoc.print(L->getHeader()->getContext(), OS);
478 // Just print the module name.
479 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
486 /// \brief Propagate known metadata from one instruction to another.
487 static void propagateMetadata(Instruction *To, const Instruction *From) {
488 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
489 From->getAllMetadataOtherThanDebugLoc(Metadata);
491 for (auto M : Metadata) {
492 unsigned Kind = M.first;
494 // These are safe to transfer (this is safe for TBAA, even when we
495 // if-convert, because should that metadata have had a control dependency
496 // on the condition, and thus actually aliased with some other
497 // non-speculated memory access when the condition was false, this would be
498 // caught by the runtime overlap checks).
499 if (Kind != LLVMContext::MD_tbaa &&
500 Kind != LLVMContext::MD_alias_scope &&
501 Kind != LLVMContext::MD_noalias &&
502 Kind != LLVMContext::MD_fpmath)
505 To->setMetadata(Kind, M.second);
509 /// \brief Propagate known metadata from one instruction to a vector of others.
510 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
512 if (Instruction *I = dyn_cast<Instruction>(V))
513 propagateMetadata(I, From);
516 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
517 /// to what vectorization factor.
518 /// This class does not look at the profitability of vectorization, only the
519 /// legality. This class has two main kinds of checks:
520 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
521 /// will change the order of memory accesses in a way that will change the
522 /// correctness of the program.
523 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
524 /// checks for a number of different conditions, such as the availability of a
525 /// single induction variable, that all types are supported and vectorize-able,
526 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
527 /// This class is also used by InnerLoopVectorizer for identifying
528 /// induction variable and the different reduction variables.
529 class LoopVectorizationLegality {
531 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, const DataLayout *DL,
532 DominatorTree *DT, TargetLibraryInfo *TLI,
533 AliasAnalysis *AA, Function *F,
534 const TargetTransformInfo *TTI,
535 LoopAccessAnalysis *LAA)
536 : NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
537 TLI(TLI), TheFunction(F), TTI(TTI), DT(DT), LAA(LAA), LAI(nullptr),
538 Induction(nullptr), WidestIndTy(nullptr), HasFunNoNaNAttr(false) {}
540 /// This enum represents the kinds of reductions that we support.
542 RK_NoReduction, ///< Not a reduction.
543 RK_IntegerAdd, ///< Sum of integers.
544 RK_IntegerMult, ///< Product of integers.
545 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
546 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
547 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
548 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
549 RK_FloatAdd, ///< Sum of floats.
550 RK_FloatMult, ///< Product of floats.
551 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
554 /// This enum represents the kinds of inductions that we support.
556 IK_NoInduction, ///< Not an induction variable.
557 IK_IntInduction, ///< Integer induction variable. Step = C.
558 IK_PtrInduction ///< Pointer induction var. Step = C / sizeof(elem).
561 // This enum represents the kind of minmax reduction.
562 enum MinMaxReductionKind {
572 /// This struct holds information about reduction variables.
573 struct ReductionDescriptor {
574 ReductionDescriptor() : StartValue(nullptr), LoopExitInstr(nullptr),
575 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
577 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
578 MinMaxReductionKind MK)
579 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
581 // The starting value of the reduction.
582 // It does not have to be zero!
583 TrackingVH<Value> StartValue;
584 // The instruction who's value is used outside the loop.
585 Instruction *LoopExitInstr;
586 // The kind of the reduction.
588 // If this a min/max reduction the kind of reduction.
589 MinMaxReductionKind MinMaxKind;
592 /// This POD struct holds information about a potential reduction operation.
593 struct ReductionInstDesc {
594 ReductionInstDesc(bool IsRedux, Instruction *I) :
595 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
597 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
598 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
600 // Is this instruction a reduction candidate.
602 // The last instruction in a min/max pattern (select of the select(icmp())
603 // pattern), or the current reduction instruction otherwise.
604 Instruction *PatternLastInst;
605 // If this is a min/max pattern the comparison predicate.
606 MinMaxReductionKind MinMaxKind;
609 /// A struct for saving information about induction variables.
610 struct InductionInfo {
611 InductionInfo(Value *Start, InductionKind K, ConstantInt *Step)
612 : StartValue(Start), IK(K), StepValue(Step) {
613 assert(IK != IK_NoInduction && "Not an induction");
614 assert(StartValue && "StartValue is null");
615 assert(StepValue && !StepValue->isZero() && "StepValue is zero");
616 assert((IK != IK_PtrInduction || StartValue->getType()->isPointerTy()) &&
617 "StartValue is not a pointer for pointer induction");
618 assert((IK != IK_IntInduction || StartValue->getType()->isIntegerTy()) &&
619 "StartValue is not an integer for integer induction");
620 assert(StepValue->getType()->isIntegerTy() &&
621 "StepValue is not an integer");
624 : StartValue(nullptr), IK(IK_NoInduction), StepValue(nullptr) {}
626 /// Get the consecutive direction. Returns:
627 /// 0 - unknown or non-consecutive.
628 /// 1 - consecutive and increasing.
629 /// -1 - consecutive and decreasing.
630 int getConsecutiveDirection() const {
631 if (StepValue && (StepValue->isOne() || StepValue->isMinusOne()))
632 return StepValue->getSExtValue();
636 /// Compute the transformed value of Index at offset StartValue using step
638 /// For integer induction, returns StartValue + Index * StepValue.
639 /// For pointer induction, returns StartValue[Index * StepValue].
640 /// FIXME: The newly created binary instructions should contain nsw/nuw
641 /// flags, which can be found from the original scalar operations.
642 Value *transform(IRBuilder<> &B, Value *Index) const {
644 case IK_IntInduction:
645 assert(Index->getType() == StartValue->getType() &&
646 "Index type does not match StartValue type");
647 if (StepValue->isMinusOne())
648 return B.CreateSub(StartValue, Index);
649 if (!StepValue->isOne())
650 Index = B.CreateMul(Index, StepValue);
651 return B.CreateAdd(StartValue, Index);
653 case IK_PtrInduction:
654 if (StepValue->isMinusOne())
655 Index = B.CreateNeg(Index);
656 else if (!StepValue->isOne())
657 Index = B.CreateMul(Index, StepValue);
658 return B.CreateGEP(StartValue, Index);
663 llvm_unreachable("invalid enum");
667 TrackingVH<Value> StartValue;
671 ConstantInt *StepValue;
674 /// ReductionList contains the reduction descriptors for all
675 /// of the reductions that were found in the loop.
676 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
678 /// InductionList saves induction variables and maps them to the
679 /// induction descriptor.
680 typedef MapVector<PHINode*, InductionInfo> InductionList;
682 /// Returns true if it is legal to vectorize this loop.
683 /// This does not mean that it is profitable to vectorize this
684 /// loop, only that it is legal to do so.
687 /// Returns the Induction variable.
688 PHINode *getInduction() { return Induction; }
690 /// Returns the reduction variables found in the loop.
691 ReductionList *getReductionVars() { return &Reductions; }
693 /// Returns the induction variables found in the loop.
694 InductionList *getInductionVars() { return &Inductions; }
696 /// Returns the widest induction type.
697 Type *getWidestInductionType() { return WidestIndTy; }
699 /// Returns True if V is an induction variable in this loop.
700 bool isInductionVariable(const Value *V);
702 /// Return true if the block BB needs to be predicated in order for the loop
703 /// to be vectorized.
704 bool blockNeedsPredication(BasicBlock *BB);
706 /// Check if this pointer is consecutive when vectorizing. This happens
707 /// when the last index of the GEP is the induction variable, or that the
708 /// pointer itself is an induction variable.
709 /// This check allows us to vectorize A[idx] into a wide load/store.
711 /// 0 - Stride is unknown or non-consecutive.
712 /// 1 - Address is consecutive.
713 /// -1 - Address is consecutive, and decreasing.
714 int isConsecutivePtr(Value *Ptr);
716 /// Returns true if the value V is uniform within the loop.
717 bool isUniform(Value *V);
719 /// Returns true if this instruction will remain scalar after vectorization.
720 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
722 /// Returns the information that we collected about runtime memory check.
723 LoopAccessInfo::RuntimePointerCheck *getRuntimePointerCheck() {
724 return LAI->getRuntimePointerCheck();
727 LoopAccessInfo *getLAI() {
731 /// This function returns the identity element (or neutral element) for
733 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
735 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
737 bool hasStride(Value *V) { return StrideSet.count(V); }
738 bool mustCheckStrides() { return !StrideSet.empty(); }
739 SmallPtrSet<Value *, 8>::iterator strides_begin() {
740 return StrideSet.begin();
742 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
744 /// Returns true if the target machine supports masked store operation
745 /// for the given \p DataType and kind of access to \p Ptr.
746 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
747 return TTI->isLegalMaskedStore(DataType, isConsecutivePtr(Ptr));
749 /// Returns true if the target machine supports masked load operation
750 /// for the given \p DataType and kind of access to \p Ptr.
751 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
752 return TTI->isLegalMaskedLoad(DataType, isConsecutivePtr(Ptr));
754 /// Returns true if vector representation of the instruction \p I
756 bool isMaskRequired(const Instruction* I) {
757 return (MaskedOp.count(I) != 0);
759 unsigned getNumStores() const {
760 return LAI->getNumStores();
762 unsigned getNumLoads() const {
763 return LAI->getNumLoads();
765 unsigned getNumPredStores() const {
766 return NumPredStores;
769 /// Check if a single basic block loop is vectorizable.
770 /// At this point we know that this is a loop with a constant trip count
771 /// and we only need to check individual instructions.
772 bool canVectorizeInstrs();
774 /// When we vectorize loops we may change the order in which
775 /// we read and write from memory. This method checks if it is
776 /// legal to vectorize the code, considering only memory constrains.
777 /// Returns true if the loop is vectorizable
778 bool canVectorizeMemory();
780 /// Return true if we can vectorize this loop using the IF-conversion
782 bool canVectorizeWithIfConvert();
784 /// Collect the variables that need to stay uniform after vectorization.
785 void collectLoopUniforms();
787 /// Return true if all of the instructions in the block can be speculatively
788 /// executed. \p SafePtrs is a list of addresses that are known to be legal
789 /// and we know that we can read from them without segfault.
790 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
792 /// Returns True, if 'Phi' is the kind of reduction variable for type
793 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
794 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
795 /// Returns a struct describing if the instruction 'I' can be a reduction
796 /// variable of type 'Kind'. If the reduction is a min/max pattern of
797 /// select(icmp()) this function advances the instruction pointer 'I' from the
798 /// compare instruction to the select instruction and stores this pointer in
799 /// 'PatternLastInst' member of the returned struct.
800 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
801 ReductionInstDesc &Desc);
802 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
803 /// pattern corresponding to a min(X, Y) or max(X, Y).
804 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
805 ReductionInstDesc &Prev);
806 /// Returns the induction kind of Phi and record the step. This function may
807 /// return NoInduction if the PHI is not an induction variable.
808 InductionKind isInductionVariable(PHINode *Phi, ConstantInt *&StepValue);
810 /// \brief Collect memory access with loop invariant strides.
812 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
814 void collectStridedAccess(Value *LoadOrStoreInst);
816 /// Report an analysis message to assist the user in diagnosing loops that are
818 void emitAnalysis(VectorizationReport &Message) {
819 VectorizationReport::emitAnalysis(Message, TheFunction, TheLoop);
822 unsigned NumPredStores;
824 /// The loop that we evaluate.
828 /// DataLayout analysis.
829 const DataLayout *DL;
830 /// Target Library Info.
831 TargetLibraryInfo *TLI;
833 Function *TheFunction;
834 /// Target Transform Info
835 const TargetTransformInfo *TTI;
838 // LoopAccess analysis.
839 LoopAccessAnalysis *LAA;
840 // And the loop-accesses info corresponding to this loop. This pointer is
841 // null until canVectorizeMemory sets it up.
844 // --- vectorization state --- //
846 /// Holds the integer induction variable. This is the counter of the
849 /// Holds the reduction variables.
850 ReductionList Reductions;
851 /// Holds all of the induction variables that we found in the loop.
852 /// Notice that inductions don't need to start at zero and that induction
853 /// variables can be pointers.
854 InductionList Inductions;
855 /// Holds the widest induction type encountered.
858 /// Allowed outside users. This holds the reduction
859 /// vars which can be accessed from outside the loop.
860 SmallPtrSet<Value*, 4> AllowedExit;
861 /// This set holds the variables which are known to be uniform after
863 SmallPtrSet<Instruction*, 4> Uniforms;
865 /// Can we assume the absence of NaNs.
866 bool HasFunNoNaNAttr;
868 ValueToValueMap Strides;
869 SmallPtrSet<Value *, 8> StrideSet;
871 /// While vectorizing these instructions we have to generate a
872 /// call to the appropriate masked intrinsic
873 SmallPtrSet<const Instruction*, 8> MaskedOp;
876 /// LoopVectorizationCostModel - estimates the expected speedups due to
878 /// In many cases vectorization is not profitable. This can happen because of
879 /// a number of reasons. In this class we mainly attempt to predict the
880 /// expected speedup/slowdowns due to the supported instruction set. We use the
881 /// TargetTransformInfo to query the different backends for the cost of
882 /// different operations.
883 class LoopVectorizationCostModel {
885 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
886 LoopVectorizationLegality *Legal,
887 const TargetTransformInfo &TTI,
888 const DataLayout *DL, const TargetLibraryInfo *TLI,
889 AssumptionCache *AC, const Function *F,
890 const LoopVectorizeHints *Hints)
891 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI),
892 TheFunction(F), Hints(Hints) {
893 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
896 /// Information about vectorization costs
897 struct VectorizationFactor {
898 unsigned Width; // Vector width with best cost
899 unsigned Cost; // Cost of the loop with that width
901 /// \return The most profitable vectorization factor and the cost of that VF.
902 /// This method checks every power of two up to VF. If UserVF is not ZERO
903 /// then this vectorization factor will be selected if vectorization is
905 VectorizationFactor selectVectorizationFactor(bool OptForSize);
907 /// \return The size (in bits) of the widest type in the code that
908 /// needs to be vectorized. We ignore values that remain scalar such as
909 /// 64 bit loop indices.
910 unsigned getWidestType();
912 /// \return The most profitable unroll factor.
913 /// If UserUF is non-zero then this method finds the best unroll-factor
914 /// based on register pressure and other parameters.
915 /// VF and LoopCost are the selected vectorization factor and the cost of the
917 unsigned selectUnrollFactor(bool OptForSize, unsigned VF, unsigned LoopCost);
919 /// \brief A struct that represents some properties of the register usage
921 struct RegisterUsage {
922 /// Holds the number of loop invariant values that are used in the loop.
923 unsigned LoopInvariantRegs;
924 /// Holds the maximum number of concurrent live intervals in the loop.
925 unsigned MaxLocalUsers;
926 /// Holds the number of instructions in the loop.
927 unsigned NumInstructions;
930 /// \return information about the register usage of the loop.
931 RegisterUsage calculateRegisterUsage();
934 /// Returns the expected execution cost. The unit of the cost does
935 /// not matter because we use the 'cost' units to compare different
936 /// vector widths. The cost that is returned is *not* normalized by
937 /// the factor width.
938 unsigned expectedCost(unsigned VF);
940 /// Returns the execution time cost of an instruction for a given vector
941 /// width. Vector width of one means scalar.
942 unsigned getInstructionCost(Instruction *I, unsigned VF);
944 /// A helper function for converting Scalar types to vector types.
945 /// If the incoming type is void, we return void. If the VF is 1, we return
947 static Type* ToVectorTy(Type *Scalar, unsigned VF);
949 /// Returns whether the instruction is a load or store and will be a emitted
950 /// as a vector operation.
951 bool isConsecutiveLoadOrStore(Instruction *I);
953 /// Report an analysis message to assist the user in diagnosing loops that are
955 void emitAnalysis(VectorizationReport &Message) {
956 VectorizationReport::emitAnalysis(Message, TheFunction, TheLoop);
959 /// Values used only by @llvm.assume calls.
960 SmallPtrSet<const Value *, 32> EphValues;
962 /// The loop that we evaluate.
966 /// Loop Info analysis.
968 /// Vectorization legality.
969 LoopVectorizationLegality *Legal;
970 /// Vector target information.
971 const TargetTransformInfo &TTI;
972 /// Target data layout information.
973 const DataLayout *DL;
974 /// Target Library Info.
975 const TargetLibraryInfo *TLI;
976 const Function *TheFunction;
977 // Loop Vectorize Hint.
978 const LoopVectorizeHints *Hints;
981 /// Utility class for getting and setting loop vectorizer hints in the form
982 /// of loop metadata.
983 /// This class keeps a number of loop annotations locally (as member variables)
984 /// and can, upon request, write them back as metadata on the loop. It will
985 /// initially scan the loop for existing metadata, and will update the local
986 /// values based on information in the loop.
987 /// We cannot write all values to metadata, as the mere presence of some info,
988 /// for example 'force', means a decision has been made. So, we need to be
989 /// careful NOT to add them if the user hasn't specifically asked so.
990 class LoopVectorizeHints {
997 /// Hint - associates name and validation with the hint value.
1000 unsigned Value; // This may have to change for non-numeric values.
1003 Hint(const char * Name, unsigned Value, HintKind Kind)
1004 : Name(Name), Value(Value), Kind(Kind) { }
1006 bool validate(unsigned Val) {
1009 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
1011 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1019 /// Vectorization width.
1021 /// Vectorization interleave factor.
1023 /// Vectorization forced
1026 /// Return the loop metadata prefix.
1027 static StringRef Prefix() { return "llvm.loop."; }
1031 FK_Undefined = -1, ///< Not selected.
1032 FK_Disabled = 0, ///< Forcing disabled.
1033 FK_Enabled = 1, ///< Forcing enabled.
1036 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
1037 : Width("vectorize.width", VectorizerParams::VectorizationFactor,
1039 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1040 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1042 // Populate values with existing loop metadata.
1043 getHintsFromMetadata();
1045 // force-vector-interleave overrides DisableInterleaving.
1046 if (VectorizerParams::isInterleaveForced())
1047 Interleave.Value = VectorizerParams::VectorizationInterleave;
1049 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
1050 << "LV: Interleaving disabled by the pass manager\n");
1053 /// Mark the loop L as already vectorized by setting the width to 1.
1054 void setAlreadyVectorized() {
1055 Width.Value = Interleave.Value = 1;
1056 Hint Hints[] = {Width, Interleave};
1057 writeHintsToMetadata(Hints);
1060 /// Dumps all the hint information.
1061 std::string emitRemark() const {
1062 VectorizationReport R;
1063 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1064 R << "vectorization is explicitly disabled";
1066 R << "use -Rpass-analysis=loop-vectorize for more info";
1067 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1068 R << " (Force=true";
1069 if (Width.Value != 0)
1070 R << ", Vector Width=" << Width.Value;
1071 if (Interleave.Value != 0)
1072 R << ", Interleave Count=" << Interleave.Value;
1080 unsigned getWidth() const { return Width.Value; }
1081 unsigned getInterleave() const { return Interleave.Value; }
1082 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1085 /// Find hints specified in the loop metadata and update local values.
1086 void getHintsFromMetadata() {
1087 MDNode *LoopID = TheLoop->getLoopID();
1091 // First operand should refer to the loop id itself.
1092 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1093 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1095 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1096 const MDString *S = nullptr;
1097 SmallVector<Metadata *, 4> Args;
1099 // The expected hint is either a MDString or a MDNode with the first
1100 // operand a MDString.
1101 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1102 if (!MD || MD->getNumOperands() == 0)
1104 S = dyn_cast<MDString>(MD->getOperand(0));
1105 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1106 Args.push_back(MD->getOperand(i));
1108 S = dyn_cast<MDString>(LoopID->getOperand(i));
1109 assert(Args.size() == 0 && "too many arguments for MDString");
1115 // Check if the hint starts with the loop metadata prefix.
1116 StringRef Name = S->getString();
1117 if (Args.size() == 1)
1118 setHint(Name, Args[0]);
1122 /// Checks string hint with one operand and set value if valid.
1123 void setHint(StringRef Name, Metadata *Arg) {
1124 if (!Name.startswith(Prefix()))
1126 Name = Name.substr(Prefix().size(), StringRef::npos);
1128 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1130 unsigned Val = C->getZExtValue();
1132 Hint *Hints[] = {&Width, &Interleave, &Force};
1133 for (auto H : Hints) {
1134 if (Name == H->Name) {
1135 if (H->validate(Val))
1138 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1144 /// Create a new hint from name / value pair.
1145 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1146 LLVMContext &Context = TheLoop->getHeader()->getContext();
1147 Metadata *MDs[] = {MDString::get(Context, Name),
1148 ConstantAsMetadata::get(
1149 ConstantInt::get(Type::getInt32Ty(Context), V))};
1150 return MDNode::get(Context, MDs);
1153 /// Matches metadata with hint name.
1154 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1155 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1159 for (auto H : HintTypes)
1160 if (Name->getString().endswith(H.Name))
1165 /// Sets current hints into loop metadata, keeping other values intact.
1166 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1167 if (HintTypes.size() == 0)
1170 // Reserve the first element to LoopID (see below).
1171 SmallVector<Metadata *, 4> MDs(1);
1172 // If the loop already has metadata, then ignore the existing operands.
1173 MDNode *LoopID = TheLoop->getLoopID();
1175 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1176 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1177 // If node in update list, ignore old value.
1178 if (!matchesHintMetadataName(Node, HintTypes))
1179 MDs.push_back(Node);
1183 // Now, add the missing hints.
1184 for (auto H : HintTypes)
1185 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1187 // Replace current metadata node with new one.
1188 LLVMContext &Context = TheLoop->getHeader()->getContext();
1189 MDNode *NewLoopID = MDNode::get(Context, MDs);
1190 // Set operand 0 to refer to the loop id itself.
1191 NewLoopID->replaceOperandWith(0, NewLoopID);
1193 TheLoop->setLoopID(NewLoopID);
1196 /// The loop these hints belong to.
1197 const Loop *TheLoop;
1200 static void emitMissedWarning(Function *F, Loop *L,
1201 const LoopVectorizeHints &LH) {
1202 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1203 L->getStartLoc(), LH.emitRemark());
1205 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1206 if (LH.getWidth() != 1)
1207 emitLoopVectorizeWarning(
1208 F->getContext(), *F, L->getStartLoc(),
1209 "failed explicitly specified loop vectorization");
1210 else if (LH.getInterleave() != 1)
1211 emitLoopInterleaveWarning(
1212 F->getContext(), *F, L->getStartLoc(),
1213 "failed explicitly specified loop interleaving");
1217 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1219 return V.push_back(&L);
1221 for (Loop *InnerL : L)
1222 addInnerLoop(*InnerL, V);
1225 /// The LoopVectorize Pass.
1226 struct LoopVectorize : public FunctionPass {
1227 /// Pass identification, replacement for typeid
1230 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1232 DisableUnrolling(NoUnrolling),
1233 AlwaysVectorize(AlwaysVectorize) {
1234 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1237 ScalarEvolution *SE;
1238 const DataLayout *DL;
1240 TargetTransformInfo *TTI;
1242 BlockFrequencyInfo *BFI;
1243 TargetLibraryInfo *TLI;
1245 AssumptionCache *AC;
1246 LoopAccessAnalysis *LAA;
1247 bool DisableUnrolling;
1248 bool AlwaysVectorize;
1250 BlockFrequency ColdEntryFreq;
1252 bool runOnFunction(Function &F) override {
1253 SE = &getAnalysis<ScalarEvolution>();
1254 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1255 DL = DLP ? &DLP->getDataLayout() : nullptr;
1256 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1257 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1258 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1259 BFI = &getAnalysis<BlockFrequencyInfo>();
1260 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1261 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1262 AA = &getAnalysis<AliasAnalysis>();
1263 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1264 LAA = &getAnalysis<LoopAccessAnalysis>();
1266 // Compute some weights outside of the loop over the loops. Compute this
1267 // using a BranchProbability to re-use its scaling math.
1268 const BranchProbability ColdProb(1, 5); // 20%
1269 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1271 // If the target claims to have no vector registers don't attempt
1273 if (!TTI->getNumberOfRegisters(true))
1277 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1278 << ": Missing data layout\n");
1282 // Build up a worklist of inner-loops to vectorize. This is necessary as
1283 // the act of vectorizing or partially unrolling a loop creates new loops
1284 // and can invalidate iterators across the loops.
1285 SmallVector<Loop *, 8> Worklist;
1288 addInnerLoop(*L, Worklist);
1290 LoopsAnalyzed += Worklist.size();
1292 // Now walk the identified inner loops.
1293 bool Changed = false;
1294 while (!Worklist.empty())
1295 Changed |= processLoop(Worklist.pop_back_val());
1297 // Process each loop nest in the function.
1301 bool processLoop(Loop *L) {
1302 assert(L->empty() && "Only process inner loops.");
1305 const std::string DebugLocStr = getDebugLocString(L);
1308 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1309 << L->getHeader()->getParent()->getName() << "\" from "
1310 << DebugLocStr << "\n");
1312 LoopVectorizeHints Hints(L, DisableUnrolling);
1314 DEBUG(dbgs() << "LV: Loop hints:"
1316 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1318 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1320 : "?")) << " width=" << Hints.getWidth()
1321 << " unroll=" << Hints.getInterleave() << "\n");
1323 // Function containing loop
1324 Function *F = L->getHeader()->getParent();
1326 // Looking at the diagnostic output is the only way to determine if a loop
1327 // was vectorized (other than looking at the IR or machine code), so it
1328 // is important to generate an optimization remark for each loop. Most of
1329 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1330 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1331 // less verbose reporting vectorized loops and unvectorized loops that may
1332 // benefit from vectorization, respectively.
1334 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1335 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1336 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1337 L->getStartLoc(), Hints.emitRemark());
1341 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1342 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1343 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1344 L->getStartLoc(), Hints.emitRemark());
1348 if (Hints.getWidth() == 1 && Hints.getInterleave() == 1) {
1349 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1350 emitOptimizationRemarkAnalysis(
1351 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1352 "loop not vectorized: vector width and interleave count are "
1353 "explicitly set to 1");
1357 // Check the loop for a trip count threshold:
1358 // do not vectorize loops with a tiny trip count.
1359 const unsigned TC = SE->getSmallConstantTripCount(L);
1360 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1361 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1362 << "This loop is not worth vectorizing.");
1363 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1364 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1366 DEBUG(dbgs() << "\n");
1367 emitOptimizationRemarkAnalysis(
1368 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1369 "vectorization is not beneficial and is not explicitly forced");
1374 // Check if it is legal to vectorize the loop.
1375 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, AA, F, TTI, LAA);
1376 if (!LVL.canVectorize()) {
1377 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1378 emitMissedWarning(F, L, Hints);
1382 // Use the cost model.
1383 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI, AC, F,
1386 // Check the function attributes to find out if this function should be
1387 // optimized for size.
1388 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1389 F->hasFnAttribute(Attribute::OptimizeForSize);
1391 // Compute the weighted frequency of this loop being executed and see if it
1392 // is less than 20% of the function entry baseline frequency. Note that we
1393 // always have a canonical loop here because we think we *can* vectoriez.
1394 // FIXME: This is hidden behind a flag due to pervasive problems with
1395 // exactly what block frequency models.
1396 if (LoopVectorizeWithBlockFrequency) {
1397 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1398 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1399 LoopEntryFreq < ColdEntryFreq)
1403 // Check the function attributes to see if implicit floats are allowed.a
1404 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1405 // an integer loop and the vector instructions selected are purely integer
1406 // vector instructions?
1407 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1408 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1409 "attribute is used.\n");
1410 emitOptimizationRemarkAnalysis(
1411 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1412 "loop not vectorized due to NoImplicitFloat attribute");
1413 emitMissedWarning(F, L, Hints);
1417 // Select the optimal vectorization factor.
1418 const LoopVectorizationCostModel::VectorizationFactor VF =
1419 CM.selectVectorizationFactor(OptForSize);
1421 // Select the unroll factor.
1423 CM.selectUnrollFactor(OptForSize, VF.Width, VF.Cost);
1425 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1426 << DebugLocStr << '\n');
1427 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1429 if (VF.Width == 1) {
1430 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
1433 emitOptimizationRemarkAnalysis(
1434 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1435 "not beneficial to vectorize and user disabled interleaving");
1438 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1440 // Report the unrolling decision.
1441 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1442 Twine("unrolled with interleaving factor " +
1444 " (vectorization not beneficial)"));
1446 // We decided not to vectorize, but we may want to unroll.
1448 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1449 Unroller.vectorize(&LVL);
1451 // If we decided that it is *legal* to vectorize the loop then do it.
1452 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1456 // Report the vectorization decision.
1457 emitOptimizationRemark(
1458 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1459 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1460 ", unrolling interleave factor: " + Twine(UF) + ")");
1463 // Mark the loop as already vectorized to avoid vectorizing again.
1464 Hints.setAlreadyVectorized();
1466 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1470 void getAnalysisUsage(AnalysisUsage &AU) const override {
1471 AU.addRequired<AssumptionCacheTracker>();
1472 AU.addRequiredID(LoopSimplifyID);
1473 AU.addRequiredID(LCSSAID);
1474 AU.addRequired<BlockFrequencyInfo>();
1475 AU.addRequired<DominatorTreeWrapperPass>();
1476 AU.addRequired<LoopInfoWrapperPass>();
1477 AU.addRequired<ScalarEvolution>();
1478 AU.addRequired<TargetTransformInfoWrapperPass>();
1479 AU.addRequired<AliasAnalysis>();
1480 AU.addRequired<LoopAccessAnalysis>();
1481 AU.addPreserved<LoopInfoWrapperPass>();
1482 AU.addPreserved<DominatorTreeWrapperPass>();
1483 AU.addPreserved<AliasAnalysis>();
1488 } // end anonymous namespace
1490 //===----------------------------------------------------------------------===//
1491 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1492 // LoopVectorizationCostModel.
1493 //===----------------------------------------------------------------------===//
1495 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1496 // We need to place the broadcast of invariant variables outside the loop.
1497 Instruction *Instr = dyn_cast<Instruction>(V);
1499 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1500 Instr->getParent()) != LoopVectorBody.end());
1501 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1503 // Place the code for broadcasting invariant variables in the new preheader.
1504 IRBuilder<>::InsertPointGuard Guard(Builder);
1506 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1508 // Broadcast the scalar into all locations in the vector.
1509 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1514 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
1516 assert(Val->getType()->isVectorTy() && "Must be a vector");
1517 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1518 "Elem must be an integer");
1519 assert(Step->getType() == Val->getType()->getScalarType() &&
1520 "Step has wrong type");
1521 // Create the types.
1522 Type *ITy = Val->getType()->getScalarType();
1523 VectorType *Ty = cast<VectorType>(Val->getType());
1524 int VLen = Ty->getNumElements();
1525 SmallVector<Constant*, 8> Indices;
1527 // Create a vector of consecutive numbers from zero to VF.
1528 for (int i = 0; i < VLen; ++i)
1529 Indices.push_back(ConstantInt::get(ITy, StartIdx + i));
1531 // Add the consecutive indices to the vector value.
1532 Constant *Cv = ConstantVector::get(Indices);
1533 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1534 Step = Builder.CreateVectorSplat(VLen, Step);
1535 assert(Step->getType() == Val->getType() && "Invalid step vec");
1536 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
1537 // which can be found from the original scalar operations.
1538 Step = Builder.CreateMul(Cv, Step);
1539 return Builder.CreateAdd(Val, Step, "induction");
1542 /// \brief Find the operand of the GEP that should be checked for consecutive
1543 /// stores. This ignores trailing indices that have no effect on the final
1545 static unsigned getGEPInductionOperand(const DataLayout *DL,
1546 const GetElementPtrInst *Gep) {
1547 unsigned LastOperand = Gep->getNumOperands() - 1;
1548 unsigned GEPAllocSize = DL->getTypeAllocSize(
1549 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1551 // Walk backwards and try to peel off zeros.
1552 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1553 // Find the type we're currently indexing into.
1554 gep_type_iterator GEPTI = gep_type_begin(Gep);
1555 std::advance(GEPTI, LastOperand - 1);
1557 // If it's a type with the same allocation size as the result of the GEP we
1558 // can peel off the zero index.
1559 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1567 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1568 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1569 // Make sure that the pointer does not point to structs.
1570 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1573 // If this value is a pointer induction variable we know it is consecutive.
1574 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1575 if (Phi && Inductions.count(Phi)) {
1576 InductionInfo II = Inductions[Phi];
1577 return II.getConsecutiveDirection();
1580 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1584 unsigned NumOperands = Gep->getNumOperands();
1585 Value *GpPtr = Gep->getPointerOperand();
1586 // If this GEP value is a consecutive pointer induction variable and all of
1587 // the indices are constant then we know it is consecutive. We can
1588 Phi = dyn_cast<PHINode>(GpPtr);
1589 if (Phi && Inductions.count(Phi)) {
1591 // Make sure that the pointer does not point to structs.
1592 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1593 if (GepPtrType->getElementType()->isAggregateType())
1596 // Make sure that all of the index operands are loop invariant.
1597 for (unsigned i = 1; i < NumOperands; ++i)
1598 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1601 InductionInfo II = Inductions[Phi];
1602 return II.getConsecutiveDirection();
1605 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1607 // Check that all of the gep indices are uniform except for our induction
1609 for (unsigned i = 0; i != NumOperands; ++i)
1610 if (i != InductionOperand &&
1611 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1614 // We can emit wide load/stores only if the last non-zero index is the
1615 // induction variable.
1616 const SCEV *Last = nullptr;
1617 if (!Strides.count(Gep))
1618 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1620 // Because of the multiplication by a stride we can have a s/zext cast.
1621 // We are going to replace this stride by 1 so the cast is safe to ignore.
1623 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1624 // %0 = trunc i64 %indvars.iv to i32
1625 // %mul = mul i32 %0, %Stride1
1626 // %idxprom = zext i32 %mul to i64 << Safe cast.
1627 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1629 Last = replaceSymbolicStrideSCEV(SE, Strides,
1630 Gep->getOperand(InductionOperand), Gep);
1631 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1633 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1637 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1638 const SCEV *Step = AR->getStepRecurrence(*SE);
1640 // The memory is consecutive because the last index is consecutive
1641 // and all other indices are loop invariant.
1644 if (Step->isAllOnesValue())
1651 bool LoopVectorizationLegality::isUniform(Value *V) {
1652 return LAI->isUniform(V);
1655 InnerLoopVectorizer::VectorParts&
1656 InnerLoopVectorizer::getVectorValue(Value *V) {
1657 assert(V != Induction && "The new induction variable should not be used.");
1658 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1660 // If we have a stride that is replaced by one, do it here.
1661 if (Legal->hasStride(V))
1662 V = ConstantInt::get(V->getType(), 1);
1664 // If we have this scalar in the map, return it.
1665 if (WidenMap.has(V))
1666 return WidenMap.get(V);
1668 // If this scalar is unknown, assume that it is a constant or that it is
1669 // loop invariant. Broadcast V and save the value for future uses.
1670 Value *B = getBroadcastInstrs(V);
1671 return WidenMap.splat(V, B);
1674 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1675 assert(Vec->getType()->isVectorTy() && "Invalid type");
1676 SmallVector<Constant*, 8> ShuffleMask;
1677 for (unsigned i = 0; i < VF; ++i)
1678 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1680 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1681 ConstantVector::get(ShuffleMask),
1685 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1686 // Attempt to issue a wide load.
1687 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1688 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1690 assert((LI || SI) && "Invalid Load/Store instruction");
1692 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1693 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1694 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1695 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1696 // An alignment of 0 means target abi alignment. We need to use the scalar's
1697 // target abi alignment in such a case.
1699 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1700 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1701 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1702 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1704 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
1705 !Legal->isMaskRequired(SI))
1706 return scalarizeInstruction(Instr, true);
1708 if (ScalarAllocatedSize != VectorElementSize)
1709 return scalarizeInstruction(Instr);
1711 // If the pointer is loop invariant or if it is non-consecutive,
1712 // scalarize the load.
1713 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1714 bool Reverse = ConsecutiveStride < 0;
1715 bool UniformLoad = LI && Legal->isUniform(Ptr);
1716 if (!ConsecutiveStride || UniformLoad)
1717 return scalarizeInstruction(Instr);
1719 Constant *Zero = Builder.getInt32(0);
1720 VectorParts &Entry = WidenMap.get(Instr);
1722 // Handle consecutive loads/stores.
1723 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1724 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1725 setDebugLocFromInst(Builder, Gep);
1726 Value *PtrOperand = Gep->getPointerOperand();
1727 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1728 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1730 // Create the new GEP with the new induction variable.
1731 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1732 Gep2->setOperand(0, FirstBasePtr);
1733 Gep2->setName("gep.indvar.base");
1734 Ptr = Builder.Insert(Gep2);
1736 setDebugLocFromInst(Builder, Gep);
1737 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1738 OrigLoop) && "Base ptr must be invariant");
1740 // The last index does not have to be the induction. It can be
1741 // consecutive and be a function of the index. For example A[I+1];
1742 unsigned NumOperands = Gep->getNumOperands();
1743 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1744 // Create the new GEP with the new induction variable.
1745 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1747 for (unsigned i = 0; i < NumOperands; ++i) {
1748 Value *GepOperand = Gep->getOperand(i);
1749 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1751 // Update last index or loop invariant instruction anchored in loop.
1752 if (i == InductionOperand ||
1753 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1754 assert((i == InductionOperand ||
1755 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1756 "Must be last index or loop invariant");
1758 VectorParts &GEPParts = getVectorValue(GepOperand);
1759 Value *Index = GEPParts[0];
1760 Index = Builder.CreateExtractElement(Index, Zero);
1761 Gep2->setOperand(i, Index);
1762 Gep2->setName("gep.indvar.idx");
1765 Ptr = Builder.Insert(Gep2);
1767 // Use the induction element ptr.
1768 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1769 setDebugLocFromInst(Builder, Ptr);
1770 VectorParts &PtrVal = getVectorValue(Ptr);
1771 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1774 VectorParts Mask = createBlockInMask(Instr->getParent());
1777 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1778 "We do not allow storing to uniform addresses");
1779 setDebugLocFromInst(Builder, SI);
1780 // We don't want to update the value in the map as it might be used in
1781 // another expression. So don't use a reference type for "StoredVal".
1782 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1784 for (unsigned Part = 0; Part < UF; ++Part) {
1785 // Calculate the pointer for the specific unroll-part.
1786 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1789 // If we store to reverse consecutive memory locations then we need
1790 // to reverse the order of elements in the stored value.
1791 StoredVal[Part] = reverseVector(StoredVal[Part]);
1792 // If the address is consecutive but reversed, then the
1793 // wide store needs to start at the last vector element.
1794 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1795 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1796 Mask[Part] = reverseVector(Mask[Part]);
1799 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1800 DataTy->getPointerTo(AddressSpace));
1803 if (Legal->isMaskRequired(SI))
1804 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
1807 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
1808 propagateMetadata(NewSI, SI);
1814 assert(LI && "Must have a load instruction");
1815 setDebugLocFromInst(Builder, LI);
1816 for (unsigned Part = 0; Part < UF; ++Part) {
1817 // Calculate the pointer for the specific unroll-part.
1818 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1821 // If the address is consecutive but reversed, then the
1822 // wide load needs to start at the last vector element.
1823 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1824 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1825 Mask[Part] = reverseVector(Mask[Part]);
1829 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1830 DataTy->getPointerTo(AddressSpace));
1831 if (Legal->isMaskRequired(LI))
1832 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
1833 UndefValue::get(DataTy),
1834 "wide.masked.load");
1836 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
1837 propagateMetadata(NewLI, LI);
1838 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
1842 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1843 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1844 // Holds vector parameters or scalars, in case of uniform vals.
1845 SmallVector<VectorParts, 4> Params;
1847 setDebugLocFromInst(Builder, Instr);
1849 // Find all of the vectorized parameters.
1850 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1851 Value *SrcOp = Instr->getOperand(op);
1853 // If we are accessing the old induction variable, use the new one.
1854 if (SrcOp == OldInduction) {
1855 Params.push_back(getVectorValue(SrcOp));
1859 // Try using previously calculated values.
1860 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1862 // If the src is an instruction that appeared earlier in the basic block
1863 // then it should already be vectorized.
1864 if (SrcInst && OrigLoop->contains(SrcInst)) {
1865 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1866 // The parameter is a vector value from earlier.
1867 Params.push_back(WidenMap.get(SrcInst));
1869 // The parameter is a scalar from outside the loop. Maybe even a constant.
1870 VectorParts Scalars;
1871 Scalars.append(UF, SrcOp);
1872 Params.push_back(Scalars);
1876 assert(Params.size() == Instr->getNumOperands() &&
1877 "Invalid number of operands");
1879 // Does this instruction return a value ?
1880 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1882 Value *UndefVec = IsVoidRetTy ? nullptr :
1883 UndefValue::get(VectorType::get(Instr->getType(), VF));
1884 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1885 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1887 Instruction *InsertPt = Builder.GetInsertPoint();
1888 BasicBlock *IfBlock = Builder.GetInsertBlock();
1889 BasicBlock *CondBlock = nullptr;
1892 Loop *VectorLp = nullptr;
1893 if (IfPredicateStore) {
1894 assert(Instr->getParent()->getSinglePredecessor() &&
1895 "Only support single predecessor blocks");
1896 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1897 Instr->getParent());
1898 VectorLp = LI->getLoopFor(IfBlock);
1899 assert(VectorLp && "Must have a loop for this block");
1902 // For each vector unroll 'part':
1903 for (unsigned Part = 0; Part < UF; ++Part) {
1904 // For each scalar that we create:
1905 for (unsigned Width = 0; Width < VF; ++Width) {
1908 Value *Cmp = nullptr;
1909 if (IfPredicateStore) {
1910 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1911 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1912 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1913 LoopVectorBody.push_back(CondBlock);
1914 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
1915 // Update Builder with newly created basic block.
1916 Builder.SetInsertPoint(InsertPt);
1919 Instruction *Cloned = Instr->clone();
1921 Cloned->setName(Instr->getName() + ".cloned");
1922 // Replace the operands of the cloned instructions with extracted scalars.
1923 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1924 Value *Op = Params[op][Part];
1925 // Param is a vector. Need to extract the right lane.
1926 if (Op->getType()->isVectorTy())
1927 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1928 Cloned->setOperand(op, Op);
1931 // Place the cloned scalar in the new loop.
1932 Builder.Insert(Cloned);
1934 // If the original scalar returns a value we need to place it in a vector
1935 // so that future users will be able to use it.
1937 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1938 Builder.getInt32(Width));
1940 if (IfPredicateStore) {
1941 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1942 LoopVectorBody.push_back(NewIfBlock);
1943 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
1944 Builder.SetInsertPoint(InsertPt);
1945 Instruction *OldBr = IfBlock->getTerminator();
1946 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1947 OldBr->eraseFromParent();
1948 IfBlock = NewIfBlock;
1954 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1958 if (Instruction *I = dyn_cast<Instruction>(V))
1959 return I->getParent() == Loc->getParent() ? I : nullptr;
1963 std::pair<Instruction *, Instruction *>
1964 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
1965 Instruction *tnullptr = nullptr;
1966 if (!Legal->mustCheckStrides())
1967 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1969 IRBuilder<> ChkBuilder(Loc);
1972 Value *Check = nullptr;
1973 Instruction *FirstInst = nullptr;
1974 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
1975 SE = Legal->strides_end();
1977 Value *Ptr = stripIntegerCast(*SI);
1978 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
1980 // Store the first instruction we create.
1981 FirstInst = getFirstInst(FirstInst, C, Loc);
1983 Check = ChkBuilder.CreateOr(Check, C);
1988 // We have to do this trickery because the IRBuilder might fold the check to a
1989 // constant expression in which case there is no Instruction anchored in a
1991 LLVMContext &Ctx = Loc->getContext();
1992 Instruction *TheCheck =
1993 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
1994 ChkBuilder.Insert(TheCheck, "stride.not.one");
1995 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
1997 return std::make_pair(FirstInst, TheCheck);
2000 void InnerLoopVectorizer::createEmptyLoop() {
2002 In this function we generate a new loop. The new loop will contain
2003 the vectorized instructions while the old loop will continue to run the
2006 [ ] <-- Back-edge taken count overflow check.
2009 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2012 || [ ] <-- vector pre header.
2016 || [ ]_| <-- vector loop.
2019 | >[ ] <--- middle-block.
2022 -|- >[ ] <--- new preheader.
2026 | [ ]_| <-- old scalar loop to handle remainder.
2029 >[ ] <-- exit block.
2033 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2034 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2035 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2036 assert(BypassBlock && "Invalid loop structure");
2037 assert(ExitBlock && "Must have an exit block");
2039 // Some loops have a single integer induction variable, while other loops
2040 // don't. One example is c++ iterators that often have multiple pointer
2041 // induction variables. In the code below we also support a case where we
2042 // don't have a single induction variable.
2043 OldInduction = Legal->getInduction();
2044 Type *IdxTy = Legal->getWidestInductionType();
2046 // Find the loop boundaries.
2047 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2048 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2050 // The exit count might have the type of i64 while the phi is i32. This can
2051 // happen if we have an induction variable that is sign extended before the
2052 // compare. The only way that we get a backedge taken count is that the
2053 // induction variable was signed and as such will not overflow. In such a case
2054 // truncation is legal.
2055 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2056 IdxTy->getPrimitiveSizeInBits())
2057 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2059 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2060 // Get the total trip count from the count by adding 1.
2061 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2062 SE->getConstant(BackedgeTakeCount->getType(), 1));
2064 // Expand the trip count and place the new instructions in the preheader.
2065 // Notice that the pre-header does not change, only the loop body.
2066 SCEVExpander Exp(*SE, "induction");
2068 // We need to test whether the backedge-taken count is uint##_max. Adding one
2069 // to it will cause overflow and an incorrect loop trip count in the vector
2070 // body. In case of overflow we want to directly jump to the scalar remainder
2072 Value *BackedgeCount =
2073 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2074 BypassBlock->getTerminator());
2075 if (BackedgeCount->getType()->isPointerTy())
2076 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2077 "backedge.ptrcnt.to.int",
2078 BypassBlock->getTerminator());
2079 Instruction *CheckBCOverflow =
2080 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2081 Constant::getAllOnesValue(BackedgeCount->getType()),
2082 "backedge.overflow", BypassBlock->getTerminator());
2084 // The loop index does not have to start at Zero. Find the original start
2085 // value from the induction PHI node. If we don't have an induction variable
2086 // then we know that it starts at zero.
2087 Builder.SetInsertPoint(BypassBlock->getTerminator());
2088 Value *StartIdx = ExtendedIdx = OldInduction ?
2089 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2091 ConstantInt::get(IdxTy, 0);
2093 // We need an instruction to anchor the overflow check on. StartIdx needs to
2094 // be defined before the overflow check branch. Because the scalar preheader
2095 // is going to merge the start index and so the overflow branch block needs to
2096 // contain a definition of the start index.
2097 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2098 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2099 BypassBlock->getTerminator());
2101 // Count holds the overall loop count (N).
2102 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2103 BypassBlock->getTerminator());
2105 LoopBypassBlocks.push_back(BypassBlock);
2107 // Split the single block loop into the two loop structure described above.
2108 BasicBlock *VectorPH =
2109 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2110 BasicBlock *VecBody =
2111 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2112 BasicBlock *MiddleBlock =
2113 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2114 BasicBlock *ScalarPH =
2115 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2117 // Create and register the new vector loop.
2118 Loop* Lp = new Loop();
2119 Loop *ParentLoop = OrigLoop->getParentLoop();
2121 // Insert the new loop into the loop nest and register the new basic blocks
2122 // before calling any utilities such as SCEV that require valid LoopInfo.
2124 ParentLoop->addChildLoop(Lp);
2125 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
2126 ParentLoop->addBasicBlockToLoop(VectorPH, *LI);
2127 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
2129 LI->addTopLevelLoop(Lp);
2131 Lp->addBasicBlockToLoop(VecBody, *LI);
2133 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2135 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2137 // Generate the induction variable.
2138 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2139 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2140 // The loop step is equal to the vectorization factor (num of SIMD elements)
2141 // times the unroll factor (num of SIMD instructions).
2142 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2144 // This is the IR builder that we use to add all of the logic for bypassing
2145 // the new vector loop.
2146 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2147 setDebugLocFromInst(BypassBuilder,
2148 getDebugLocFromInstOrOperands(OldInduction));
2150 // We may need to extend the index in case there is a type mismatch.
2151 // We know that the count starts at zero and does not overflow.
2152 if (Count->getType() != IdxTy) {
2153 // The exit count can be of pointer type. Convert it to the correct
2155 if (ExitCount->getType()->isPointerTy())
2156 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2158 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2161 // Add the start index to the loop count to get the new end index.
2162 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2164 // Now we need to generate the expression for N - (N % VF), which is
2165 // the part that the vectorized body will execute.
2166 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2167 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2168 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2169 "end.idx.rnd.down");
2171 // Now, compare the new count to zero. If it is zero skip the vector loop and
2172 // jump to the scalar loop.
2174 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2176 BasicBlock *LastBypassBlock = BypassBlock;
2178 // Generate code to check that the loops trip count that we computed by adding
2179 // one to the backedge-taken count will not overflow.
2181 auto PastOverflowCheck =
2182 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2183 BasicBlock *CheckBlock =
2184 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2186 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2187 LoopBypassBlocks.push_back(CheckBlock);
2188 Instruction *OldTerm = LastBypassBlock->getTerminator();
2189 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2190 OldTerm->eraseFromParent();
2191 LastBypassBlock = CheckBlock;
2194 // Generate the code to check that the strides we assumed to be one are really
2195 // one. We want the new basic block to start at the first instruction in a
2196 // sequence of instructions that form a check.
2197 Instruction *StrideCheck;
2198 Instruction *FirstCheckInst;
2199 std::tie(FirstCheckInst, StrideCheck) =
2200 addStrideCheck(LastBypassBlock->getTerminator());
2202 // Create a new block containing the stride check.
2203 BasicBlock *CheckBlock =
2204 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2206 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2207 LoopBypassBlocks.push_back(CheckBlock);
2209 // Replace the branch into the memory check block with a conditional branch
2210 // for the "few elements case".
2211 Instruction *OldTerm = LastBypassBlock->getTerminator();
2212 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2213 OldTerm->eraseFromParent();
2216 LastBypassBlock = CheckBlock;
2219 // Generate the code that checks in runtime if arrays overlap. We put the
2220 // checks into a separate block to make the more common case of few elements
2222 Instruction *MemRuntimeCheck;
2223 std::tie(FirstCheckInst, MemRuntimeCheck) =
2224 Legal->getLAI()->addRuntimeCheck(LastBypassBlock->getTerminator());
2225 if (MemRuntimeCheck) {
2226 // Create a new block containing the memory check.
2227 BasicBlock *CheckBlock =
2228 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.memcheck");
2230 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2231 LoopBypassBlocks.push_back(CheckBlock);
2233 // Replace the branch into the memory check block with a conditional branch
2234 // for the "few elements case".
2235 Instruction *OldTerm = LastBypassBlock->getTerminator();
2236 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2237 OldTerm->eraseFromParent();
2239 Cmp = MemRuntimeCheck;
2240 LastBypassBlock = CheckBlock;
2243 LastBypassBlock->getTerminator()->eraseFromParent();
2244 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2247 // We are going to resume the execution of the scalar loop.
2248 // Go over all of the induction variables that we found and fix the
2249 // PHIs that are left in the scalar version of the loop.
2250 // The starting values of PHI nodes depend on the counter of the last
2251 // iteration in the vectorized loop.
2252 // If we come from a bypass edge then we need to start from the original
2255 // This variable saves the new starting index for the scalar loop.
2256 PHINode *ResumeIndex = nullptr;
2257 LoopVectorizationLegality::InductionList::iterator I, E;
2258 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2259 // Set builder to point to last bypass block.
2260 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2261 for (I = List->begin(), E = List->end(); I != E; ++I) {
2262 PHINode *OrigPhi = I->first;
2263 LoopVectorizationLegality::InductionInfo II = I->second;
2265 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2266 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2267 MiddleBlock->getTerminator());
2268 // We might have extended the type of the induction variable but we need a
2269 // truncated version for the scalar loop.
2270 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2271 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2272 MiddleBlock->getTerminator()) : nullptr;
2274 // Create phi nodes to merge from the backedge-taken check block.
2275 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2276 ScalarPH->getTerminator());
2277 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2279 PHINode *BCTruncResumeVal = nullptr;
2280 if (OrigPhi == OldInduction) {
2282 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2283 ScalarPH->getTerminator());
2284 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2287 Value *EndValue = nullptr;
2289 case LoopVectorizationLegality::IK_NoInduction:
2290 llvm_unreachable("Unknown induction");
2291 case LoopVectorizationLegality::IK_IntInduction: {
2292 // Handle the integer induction counter.
2293 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2295 // We have the canonical induction variable.
2296 if (OrigPhi == OldInduction) {
2297 // Create a truncated version of the resume value for the scalar loop,
2298 // we might have promoted the type to a larger width.
2300 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2301 // The new PHI merges the original incoming value, in case of a bypass,
2302 // or the value at the end of the vectorized loop.
2303 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2304 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2305 TruncResumeVal->addIncoming(EndValue, VecBody);
2307 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2309 // We know what the end value is.
2310 EndValue = IdxEndRoundDown;
2311 // We also know which PHI node holds it.
2312 ResumeIndex = ResumeVal;
2316 // Not the canonical induction variable - add the vector loop count to the
2318 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2319 II.StartValue->getType(),
2321 EndValue = II.transform(BypassBuilder, CRD);
2322 EndValue->setName("ind.end");
2325 case LoopVectorizationLegality::IK_PtrInduction: {
2326 EndValue = II.transform(BypassBuilder, CountRoundDown);
2327 EndValue->setName("ptr.ind.end");
2332 // The new PHI merges the original incoming value, in case of a bypass,
2333 // or the value at the end of the vectorized loop.
2334 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2335 if (OrigPhi == OldInduction)
2336 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2338 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2340 ResumeVal->addIncoming(EndValue, VecBody);
2342 // Fix the scalar body counter (PHI node).
2343 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2345 // The old induction's phi node in the scalar body needs the truncated
2347 if (OrigPhi == OldInduction) {
2348 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2349 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2351 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2352 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2356 // If we are generating a new induction variable then we also need to
2357 // generate the code that calculates the exit value. This value is not
2358 // simply the end of the counter because we may skip the vectorized body
2359 // in case of a runtime check.
2361 assert(!ResumeIndex && "Unexpected resume value found");
2362 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2363 MiddleBlock->getTerminator());
2364 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2365 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2366 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2369 // Make sure that we found the index where scalar loop needs to continue.
2370 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2371 "Invalid resume Index");
2373 // Add a check in the middle block to see if we have completed
2374 // all of the iterations in the first vector loop.
2375 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2376 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2377 ResumeIndex, "cmp.n",
2378 MiddleBlock->getTerminator());
2380 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2381 // Remove the old terminator.
2382 MiddleBlock->getTerminator()->eraseFromParent();
2384 // Create i+1 and fill the PHINode.
2385 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2386 Induction->addIncoming(StartIdx, VectorPH);
2387 Induction->addIncoming(NextIdx, VecBody);
2388 // Create the compare.
2389 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2390 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2392 // Now we have two terminators. Remove the old one from the block.
2393 VecBody->getTerminator()->eraseFromParent();
2395 // Get ready to start creating new instructions into the vectorized body.
2396 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2399 LoopVectorPreHeader = VectorPH;
2400 LoopScalarPreHeader = ScalarPH;
2401 LoopMiddleBlock = MiddleBlock;
2402 LoopExitBlock = ExitBlock;
2403 LoopVectorBody.push_back(VecBody);
2404 LoopScalarBody = OldBasicBlock;
2406 LoopVectorizeHints Hints(Lp, true);
2407 Hints.setAlreadyVectorized();
2410 /// This function returns the identity element (or neutral element) for
2411 /// the operation K.
2413 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2418 // Adding, Xoring, Oring zero to a number does not change it.
2419 return ConstantInt::get(Tp, 0);
2420 case RK_IntegerMult:
2421 // Multiplying a number by 1 does not change it.
2422 return ConstantInt::get(Tp, 1);
2424 // AND-ing a number with an all-1 value does not change it.
2425 return ConstantInt::get(Tp, -1, true);
2427 // Multiplying a number by 1 does not change it.
2428 return ConstantFP::get(Tp, 1.0L);
2430 // Adding zero to a number does not change it.
2431 return ConstantFP::get(Tp, 0.0L);
2433 llvm_unreachable("Unknown reduction kind");
2437 /// This function translates the reduction kind to an LLVM binary operator.
2439 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2441 case LoopVectorizationLegality::RK_IntegerAdd:
2442 return Instruction::Add;
2443 case LoopVectorizationLegality::RK_IntegerMult:
2444 return Instruction::Mul;
2445 case LoopVectorizationLegality::RK_IntegerOr:
2446 return Instruction::Or;
2447 case LoopVectorizationLegality::RK_IntegerAnd:
2448 return Instruction::And;
2449 case LoopVectorizationLegality::RK_IntegerXor:
2450 return Instruction::Xor;
2451 case LoopVectorizationLegality::RK_FloatMult:
2452 return Instruction::FMul;
2453 case LoopVectorizationLegality::RK_FloatAdd:
2454 return Instruction::FAdd;
2455 case LoopVectorizationLegality::RK_IntegerMinMax:
2456 return Instruction::ICmp;
2457 case LoopVectorizationLegality::RK_FloatMinMax:
2458 return Instruction::FCmp;
2460 llvm_unreachable("Unknown reduction operation");
2464 Value *createMinMaxOp(IRBuilder<> &Builder,
2465 LoopVectorizationLegality::MinMaxReductionKind RK,
2468 CmpInst::Predicate P = CmpInst::ICMP_NE;
2471 llvm_unreachable("Unknown min/max reduction kind");
2472 case LoopVectorizationLegality::MRK_UIntMin:
2473 P = CmpInst::ICMP_ULT;
2475 case LoopVectorizationLegality::MRK_UIntMax:
2476 P = CmpInst::ICMP_UGT;
2478 case LoopVectorizationLegality::MRK_SIntMin:
2479 P = CmpInst::ICMP_SLT;
2481 case LoopVectorizationLegality::MRK_SIntMax:
2482 P = CmpInst::ICMP_SGT;
2484 case LoopVectorizationLegality::MRK_FloatMin:
2485 P = CmpInst::FCMP_OLT;
2487 case LoopVectorizationLegality::MRK_FloatMax:
2488 P = CmpInst::FCMP_OGT;
2493 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2494 RK == LoopVectorizationLegality::MRK_FloatMax)
2495 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2497 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2499 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2504 struct CSEDenseMapInfo {
2505 static bool canHandle(Instruction *I) {
2506 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2507 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2509 static inline Instruction *getEmptyKey() {
2510 return DenseMapInfo<Instruction *>::getEmptyKey();
2512 static inline Instruction *getTombstoneKey() {
2513 return DenseMapInfo<Instruction *>::getTombstoneKey();
2515 static unsigned getHashValue(Instruction *I) {
2516 assert(canHandle(I) && "Unknown instruction!");
2517 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2518 I->value_op_end()));
2520 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2521 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2522 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2524 return LHS->isIdenticalTo(RHS);
2529 /// \brief Check whether this block is a predicated block.
2530 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2531 /// = ...; " blocks. We start with one vectorized basic block. For every
2532 /// conditional block we split this vectorized block. Therefore, every second
2533 /// block will be a predicated one.
2534 static bool isPredicatedBlock(unsigned BlockNum) {
2535 return BlockNum % 2;
2538 ///\brief Perform cse of induction variable instructions.
2539 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2540 // Perform simple cse.
2541 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2542 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2543 BasicBlock *BB = BBs[i];
2544 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2545 Instruction *In = I++;
2547 if (!CSEDenseMapInfo::canHandle(In))
2550 // Check if we can replace this instruction with any of the
2551 // visited instructions.
2552 if (Instruction *V = CSEMap.lookup(In)) {
2553 In->replaceAllUsesWith(V);
2554 In->eraseFromParent();
2557 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2558 // ...;" blocks for predicated stores. Every second block is a predicated
2560 if (isPredicatedBlock(i))
2568 /// \brief Adds a 'fast' flag to floating point operations.
2569 static Value *addFastMathFlag(Value *V) {
2570 if (isa<FPMathOperator>(V)){
2571 FastMathFlags Flags;
2572 Flags.setUnsafeAlgebra();
2573 cast<Instruction>(V)->setFastMathFlags(Flags);
2578 void InnerLoopVectorizer::vectorizeLoop() {
2579 //===------------------------------------------------===//
2581 // Notice: any optimization or new instruction that go
2582 // into the code below should be also be implemented in
2585 //===------------------------------------------------===//
2586 Constant *Zero = Builder.getInt32(0);
2588 // In order to support reduction variables we need to be able to vectorize
2589 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2590 // stages. First, we create a new vector PHI node with no incoming edges.
2591 // We use this value when we vectorize all of the instructions that use the
2592 // PHI. Next, after all of the instructions in the block are complete we
2593 // add the new incoming edges to the PHI. At this point all of the
2594 // instructions in the basic block are vectorized, so we can use them to
2595 // construct the PHI.
2596 PhiVector RdxPHIsToFix;
2598 // Scan the loop in a topological order to ensure that defs are vectorized
2600 LoopBlocksDFS DFS(OrigLoop);
2603 // Vectorize all of the blocks in the original loop.
2604 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2605 be = DFS.endRPO(); bb != be; ++bb)
2606 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2608 // At this point every instruction in the original loop is widened to
2609 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2610 // that we vectorized. The PHI nodes are currently empty because we did
2611 // not want to introduce cycles. Notice that the remaining PHI nodes
2612 // that we need to fix are reduction variables.
2614 // Create the 'reduced' values for each of the induction vars.
2615 // The reduced values are the vector values that we scalarize and combine
2616 // after the loop is finished.
2617 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2619 PHINode *RdxPhi = *it;
2620 assert(RdxPhi && "Unable to recover vectorized PHI");
2622 // Find the reduction variable descriptor.
2623 assert(Legal->getReductionVars()->count(RdxPhi) &&
2624 "Unable to find the reduction variable");
2625 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2626 (*Legal->getReductionVars())[RdxPhi];
2628 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2630 // We need to generate a reduction vector from the incoming scalar.
2631 // To do so, we need to generate the 'identity' vector and override
2632 // one of the elements with the incoming scalar reduction. We need
2633 // to do it in the vector-loop preheader.
2634 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
2636 // This is the vector-clone of the value that leaves the loop.
2637 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2638 Type *VecTy = VectorExit[0]->getType();
2640 // Find the reduction identity variable. Zero for addition, or, xor,
2641 // one for multiplication, -1 for And.
2644 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2645 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2646 // MinMax reduction have the start value as their identify.
2648 VectorStart = Identity = RdxDesc.StartValue;
2650 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2655 // Handle other reduction kinds:
2657 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2658 VecTy->getScalarType());
2661 // This vector is the Identity vector where the first element is the
2662 // incoming scalar reduction.
2663 VectorStart = RdxDesc.StartValue;
2665 Identity = ConstantVector::getSplat(VF, Iden);
2667 // This vector is the Identity vector where the first element is the
2668 // incoming scalar reduction.
2669 VectorStart = Builder.CreateInsertElement(Identity,
2670 RdxDesc.StartValue, Zero);
2674 // Fix the vector-loop phi.
2676 // Reductions do not have to start at zero. They can start with
2677 // any loop invariant values.
2678 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2679 BasicBlock *Latch = OrigLoop->getLoopLatch();
2680 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2681 VectorParts &Val = getVectorValue(LoopVal);
2682 for (unsigned part = 0; part < UF; ++part) {
2683 // Make sure to add the reduction stat value only to the
2684 // first unroll part.
2685 Value *StartVal = (part == 0) ? VectorStart : Identity;
2686 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
2687 LoopVectorPreHeader);
2688 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2689 LoopVectorBody.back());
2692 // Before each round, move the insertion point right between
2693 // the PHIs and the values we are going to write.
2694 // This allows us to write both PHINodes and the extractelement
2696 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2698 VectorParts RdxParts;
2699 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2700 for (unsigned part = 0; part < UF; ++part) {
2701 // This PHINode contains the vectorized reduction variable, or
2702 // the initial value vector, if we bypass the vector loop.
2703 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2704 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2705 Value *StartVal = (part == 0) ? VectorStart : Identity;
2706 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2707 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2708 NewPhi->addIncoming(RdxExitVal[part],
2709 LoopVectorBody.back());
2710 RdxParts.push_back(NewPhi);
2713 // Reduce all of the unrolled parts into a single vector.
2714 Value *ReducedPartRdx = RdxParts[0];
2715 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2716 setDebugLocFromInst(Builder, ReducedPartRdx);
2717 for (unsigned part = 1; part < UF; ++part) {
2718 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2719 // Floating point operations had to be 'fast' to enable the reduction.
2720 ReducedPartRdx = addFastMathFlag(
2721 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2722 ReducedPartRdx, "bin.rdx"));
2724 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2725 ReducedPartRdx, RdxParts[part]);
2729 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2730 // and vector ops, reducing the set of values being computed by half each
2732 assert(isPowerOf2_32(VF) &&
2733 "Reduction emission only supported for pow2 vectors!");
2734 Value *TmpVec = ReducedPartRdx;
2735 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2736 for (unsigned i = VF; i != 1; i >>= 1) {
2737 // Move the upper half of the vector to the lower half.
2738 for (unsigned j = 0; j != i/2; ++j)
2739 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2741 // Fill the rest of the mask with undef.
2742 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2743 UndefValue::get(Builder.getInt32Ty()));
2746 Builder.CreateShuffleVector(TmpVec,
2747 UndefValue::get(TmpVec->getType()),
2748 ConstantVector::get(ShuffleMask),
2751 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2752 // Floating point operations had to be 'fast' to enable the reduction.
2753 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2754 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2756 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2759 // The result is in the first element of the vector.
2760 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2761 Builder.getInt32(0));
2764 // Create a phi node that merges control-flow from the backedge-taken check
2765 // block and the middle block.
2766 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
2767 LoopScalarPreHeader->getTerminator());
2768 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
2769 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2771 // Now, we need to fix the users of the reduction variable
2772 // inside and outside of the scalar remainder loop.
2773 // We know that the loop is in LCSSA form. We need to update the
2774 // PHI nodes in the exit blocks.
2775 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2776 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2777 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2778 if (!LCSSAPhi) break;
2780 // All PHINodes need to have a single entry edge, or two if
2781 // we already fixed them.
2782 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2784 // We found our reduction value exit-PHI. Update it with the
2785 // incoming bypass edge.
2786 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2787 // Add an edge coming from the bypass.
2788 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2791 }// end of the LCSSA phi scan.
2793 // Fix the scalar loop reduction variable with the incoming reduction sum
2794 // from the vector body and from the backedge value.
2795 int IncomingEdgeBlockIdx =
2796 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2797 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2798 // Pick the other block.
2799 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2800 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
2801 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2802 }// end of for each redux variable.
2806 // Remove redundant induction instructions.
2807 cse(LoopVectorBody);
2810 void InnerLoopVectorizer::fixLCSSAPHIs() {
2811 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2812 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2813 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2814 if (!LCSSAPhi) break;
2815 if (LCSSAPhi->getNumIncomingValues() == 1)
2816 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2821 InnerLoopVectorizer::VectorParts
2822 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2823 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2826 // Look for cached value.
2827 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2828 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2829 if (ECEntryIt != MaskCache.end())
2830 return ECEntryIt->second;
2832 VectorParts SrcMask = createBlockInMask(Src);
2834 // The terminator has to be a branch inst!
2835 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2836 assert(BI && "Unexpected terminator found");
2838 if (BI->isConditional()) {
2839 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2841 if (BI->getSuccessor(0) != Dst)
2842 for (unsigned part = 0; part < UF; ++part)
2843 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2845 for (unsigned part = 0; part < UF; ++part)
2846 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2848 MaskCache[Edge] = EdgeMask;
2852 MaskCache[Edge] = SrcMask;
2856 InnerLoopVectorizer::VectorParts
2857 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2858 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2860 // Loop incoming mask is all-one.
2861 if (OrigLoop->getHeader() == BB) {
2862 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2863 return getVectorValue(C);
2866 // This is the block mask. We OR all incoming edges, and with zero.
2867 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2868 VectorParts BlockMask = getVectorValue(Zero);
2871 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2872 VectorParts EM = createEdgeMask(*it, BB);
2873 for (unsigned part = 0; part < UF; ++part)
2874 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2880 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2881 InnerLoopVectorizer::VectorParts &Entry,
2882 unsigned UF, unsigned VF, PhiVector *PV) {
2883 PHINode* P = cast<PHINode>(PN);
2884 // Handle reduction variables:
2885 if (Legal->getReductionVars()->count(P)) {
2886 for (unsigned part = 0; part < UF; ++part) {
2887 // This is phase one of vectorizing PHIs.
2888 Type *VecTy = (VF == 1) ? PN->getType() :
2889 VectorType::get(PN->getType(), VF);
2890 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2891 LoopVectorBody.back()-> getFirstInsertionPt());
2897 setDebugLocFromInst(Builder, P);
2898 // Check for PHI nodes that are lowered to vector selects.
2899 if (P->getParent() != OrigLoop->getHeader()) {
2900 // We know that all PHIs in non-header blocks are converted into
2901 // selects, so we don't have to worry about the insertion order and we
2902 // can just use the builder.
2903 // At this point we generate the predication tree. There may be
2904 // duplications since this is a simple recursive scan, but future
2905 // optimizations will clean it up.
2907 unsigned NumIncoming = P->getNumIncomingValues();
2909 // Generate a sequence of selects of the form:
2910 // SELECT(Mask3, In3,
2911 // SELECT(Mask2, In2,
2913 for (unsigned In = 0; In < NumIncoming; In++) {
2914 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2916 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2918 for (unsigned part = 0; part < UF; ++part) {
2919 // We might have single edge PHIs (blocks) - use an identity
2920 // 'select' for the first PHI operand.
2922 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2925 // Select between the current value and the previous incoming edge
2926 // based on the incoming mask.
2927 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2928 Entry[part], "predphi");
2934 // This PHINode must be an induction variable.
2935 // Make sure that we know about it.
2936 assert(Legal->getInductionVars()->count(P) &&
2937 "Not an induction variable");
2939 LoopVectorizationLegality::InductionInfo II =
2940 Legal->getInductionVars()->lookup(P);
2942 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
2943 // which can be found from the original scalar operations.
2945 case LoopVectorizationLegality::IK_NoInduction:
2946 llvm_unreachable("Unknown induction");
2947 case LoopVectorizationLegality::IK_IntInduction: {
2948 assert(P->getType() == II.StartValue->getType() && "Types must match");
2949 Type *PhiTy = P->getType();
2951 if (P == OldInduction) {
2952 // Handle the canonical induction variable. We might have had to
2954 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2956 // Handle other induction variables that are now based on the
2958 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2960 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2961 Broadcasted = II.transform(Builder, NormalizedIdx);
2962 Broadcasted->setName("offset.idx");
2964 Broadcasted = getBroadcastInstrs(Broadcasted);
2965 // After broadcasting the induction variable we need to make the vector
2966 // consecutive by adding 0, 1, 2, etc.
2967 for (unsigned part = 0; part < UF; ++part)
2968 Entry[part] = getStepVector(Broadcasted, VF * part, II.StepValue);
2971 case LoopVectorizationLegality::IK_PtrInduction:
2972 // Handle the pointer induction variable case.
2973 assert(P->getType()->isPointerTy() && "Unexpected type.");
2974 // This is the normalized GEP that starts counting at zero.
2975 Value *NormalizedIdx =
2976 Builder.CreateSub(Induction, ExtendedIdx, "normalized.idx");
2977 // This is the vector of results. Notice that we don't generate
2978 // vector geps because scalar geps result in better code.
2979 for (unsigned part = 0; part < UF; ++part) {
2981 int EltIndex = part;
2982 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2983 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
2984 Value *SclrGep = II.transform(Builder, GlobalIdx);
2985 SclrGep->setName("next.gep");
2986 Entry[part] = SclrGep;
2990 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2991 for (unsigned int i = 0; i < VF; ++i) {
2992 int EltIndex = i + part * VF;
2993 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2994 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
2995 Value *SclrGep = II.transform(Builder, GlobalIdx);
2996 SclrGep->setName("next.gep");
2997 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2998 Builder.getInt32(i),
3001 Entry[part] = VecVal;
3007 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3008 // For each instruction in the old loop.
3009 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3010 VectorParts &Entry = WidenMap.get(it);
3011 switch (it->getOpcode()) {
3012 case Instruction::Br:
3013 // Nothing to do for PHIs and BR, since we already took care of the
3014 // loop control flow instructions.
3016 case Instruction::PHI: {
3017 // Vectorize PHINodes.
3018 widenPHIInstruction(it, Entry, UF, VF, PV);
3022 case Instruction::Add:
3023 case Instruction::FAdd:
3024 case Instruction::Sub:
3025 case Instruction::FSub:
3026 case Instruction::Mul:
3027 case Instruction::FMul:
3028 case Instruction::UDiv:
3029 case Instruction::SDiv:
3030 case Instruction::FDiv:
3031 case Instruction::URem:
3032 case Instruction::SRem:
3033 case Instruction::FRem:
3034 case Instruction::Shl:
3035 case Instruction::LShr:
3036 case Instruction::AShr:
3037 case Instruction::And:
3038 case Instruction::Or:
3039 case Instruction::Xor: {
3040 // Just widen binops.
3041 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3042 setDebugLocFromInst(Builder, BinOp);
3043 VectorParts &A = getVectorValue(it->getOperand(0));
3044 VectorParts &B = getVectorValue(it->getOperand(1));
3046 // Use this vector value for all users of the original instruction.
3047 for (unsigned Part = 0; Part < UF; ++Part) {
3048 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3050 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3051 VecOp->copyIRFlags(BinOp);
3056 propagateMetadata(Entry, it);
3059 case Instruction::Select: {
3061 // If the selector is loop invariant we can create a select
3062 // instruction with a scalar condition. Otherwise, use vector-select.
3063 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3065 setDebugLocFromInst(Builder, it);
3067 // The condition can be loop invariant but still defined inside the
3068 // loop. This means that we can't just use the original 'cond' value.
3069 // We have to take the 'vectorized' value and pick the first lane.
3070 // Instcombine will make this a no-op.
3071 VectorParts &Cond = getVectorValue(it->getOperand(0));
3072 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3073 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3075 Value *ScalarCond = (VF == 1) ? Cond[0] :
3076 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3078 for (unsigned Part = 0; Part < UF; ++Part) {
3079 Entry[Part] = Builder.CreateSelect(
3080 InvariantCond ? ScalarCond : Cond[Part],
3085 propagateMetadata(Entry, it);
3089 case Instruction::ICmp:
3090 case Instruction::FCmp: {
3091 // Widen compares. Generate vector compares.
3092 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3093 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3094 setDebugLocFromInst(Builder, it);
3095 VectorParts &A = getVectorValue(it->getOperand(0));
3096 VectorParts &B = getVectorValue(it->getOperand(1));
3097 for (unsigned Part = 0; Part < UF; ++Part) {
3100 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3102 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3106 propagateMetadata(Entry, it);
3110 case Instruction::Store:
3111 case Instruction::Load:
3112 vectorizeMemoryInstruction(it);
3114 case Instruction::ZExt:
3115 case Instruction::SExt:
3116 case Instruction::FPToUI:
3117 case Instruction::FPToSI:
3118 case Instruction::FPExt:
3119 case Instruction::PtrToInt:
3120 case Instruction::IntToPtr:
3121 case Instruction::SIToFP:
3122 case Instruction::UIToFP:
3123 case Instruction::Trunc:
3124 case Instruction::FPTrunc:
3125 case Instruction::BitCast: {
3126 CastInst *CI = dyn_cast<CastInst>(it);
3127 setDebugLocFromInst(Builder, it);
3128 /// Optimize the special case where the source is the induction
3129 /// variable. Notice that we can only optimize the 'trunc' case
3130 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3131 /// c. other casts depend on pointer size.
3132 if (CI->getOperand(0) == OldInduction &&
3133 it->getOpcode() == Instruction::Trunc) {
3134 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3136 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3137 LoopVectorizationLegality::InductionInfo II =
3138 Legal->getInductionVars()->lookup(OldInduction);
3140 ConstantInt::getSigned(CI->getType(), II.StepValue->getSExtValue());
3141 for (unsigned Part = 0; Part < UF; ++Part)
3142 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
3143 propagateMetadata(Entry, it);
3146 /// Vectorize casts.
3147 Type *DestTy = (VF == 1) ? CI->getType() :
3148 VectorType::get(CI->getType(), VF);
3150 VectorParts &A = getVectorValue(it->getOperand(0));
3151 for (unsigned Part = 0; Part < UF; ++Part)
3152 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3153 propagateMetadata(Entry, it);
3157 case Instruction::Call: {
3158 // Ignore dbg intrinsics.
3159 if (isa<DbgInfoIntrinsic>(it))
3161 setDebugLocFromInst(Builder, it);
3163 Module *M = BB->getParent()->getParent();
3164 CallInst *CI = cast<CallInst>(it);
3165 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3166 assert(ID && "Not an intrinsic call!");
3168 case Intrinsic::assume:
3169 case Intrinsic::lifetime_end:
3170 case Intrinsic::lifetime_start:
3171 scalarizeInstruction(it);
3174 bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
3175 for (unsigned Part = 0; Part < UF; ++Part) {
3176 SmallVector<Value *, 4> Args;
3177 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3178 if (HasScalarOpd && i == 1) {
3179 Args.push_back(CI->getArgOperand(i));
3182 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3183 Args.push_back(Arg[Part]);
3185 Type *Tys[] = {CI->getType()};
3187 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3189 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3190 Entry[Part] = Builder.CreateCall(F, Args);
3193 propagateMetadata(Entry, it);
3200 // All other instructions are unsupported. Scalarize them.
3201 scalarizeInstruction(it);
3204 }// end of for_each instr.
3207 void InnerLoopVectorizer::updateAnalysis() {
3208 // Forget the original basic block.
3209 SE->forgetLoop(OrigLoop);
3211 // Update the dominator tree information.
3212 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3213 "Entry does not dominate exit.");
3215 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3216 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3217 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3219 // Due to if predication of stores we might create a sequence of "if(pred)
3220 // a[i] = ...; " blocks.
3221 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3223 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3224 else if (isPredicatedBlock(i)) {
3225 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3227 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3231 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3232 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3233 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3234 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3236 DEBUG(DT->verifyDomTree());
3239 /// \brief Check whether it is safe to if-convert this phi node.
3241 /// Phi nodes with constant expressions that can trap are not safe to if
3243 static bool canIfConvertPHINodes(BasicBlock *BB) {
3244 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3245 PHINode *Phi = dyn_cast<PHINode>(I);
3248 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3249 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3256 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3257 if (!EnableIfConversion) {
3258 emitAnalysis(VectorizationReport() << "if-conversion is disabled");
3262 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3264 // A list of pointers that we can safely read and write to.
3265 SmallPtrSet<Value *, 8> SafePointes;
3267 // Collect safe addresses.
3268 for (Loop::block_iterator BI = TheLoop->block_begin(),
3269 BE = TheLoop->block_end(); BI != BE; ++BI) {
3270 BasicBlock *BB = *BI;
3272 if (blockNeedsPredication(BB))
3275 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3276 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3277 SafePointes.insert(LI->getPointerOperand());
3278 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3279 SafePointes.insert(SI->getPointerOperand());
3283 // Collect the blocks that need predication.
3284 BasicBlock *Header = TheLoop->getHeader();
3285 for (Loop::block_iterator BI = TheLoop->block_begin(),
3286 BE = TheLoop->block_end(); BI != BE; ++BI) {
3287 BasicBlock *BB = *BI;
3289 // We don't support switch statements inside loops.
3290 if (!isa<BranchInst>(BB->getTerminator())) {
3291 emitAnalysis(VectorizationReport(BB->getTerminator())
3292 << "loop contains a switch statement");
3296 // We must be able to predicate all blocks that need to be predicated.
3297 if (blockNeedsPredication(BB)) {
3298 if (!blockCanBePredicated(BB, SafePointes)) {
3299 emitAnalysis(VectorizationReport(BB->getTerminator())
3300 << "control flow cannot be substituted for a select");
3303 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3304 emitAnalysis(VectorizationReport(BB->getTerminator())
3305 << "control flow cannot be substituted for a select");
3310 // We can if-convert this loop.
3314 bool LoopVectorizationLegality::canVectorize() {
3315 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3316 // be canonicalized.
3317 if (!TheLoop->getLoopPreheader()) {
3319 VectorizationReport() <<
3320 "loop control flow is not understood by vectorizer");
3324 // We can only vectorize innermost loops.
3325 if (!TheLoop->getSubLoopsVector().empty()) {
3326 emitAnalysis(VectorizationReport() << "loop is not the innermost loop");
3330 // We must have a single backedge.
3331 if (TheLoop->getNumBackEdges() != 1) {
3333 VectorizationReport() <<
3334 "loop control flow is not understood by vectorizer");
3338 // We must have a single exiting block.
3339 if (!TheLoop->getExitingBlock()) {
3341 VectorizationReport() <<
3342 "loop control flow is not understood by vectorizer");
3346 // We only handle bottom-tested loops, i.e. loop in which the condition is
3347 // checked at the end of each iteration. With that we can assume that all
3348 // instructions in the loop are executed the same number of times.
3349 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
3351 VectorizationReport() <<
3352 "loop control flow is not understood by vectorizer");
3356 // We need to have a loop header.
3357 DEBUG(dbgs() << "LV: Found a loop: " <<
3358 TheLoop->getHeader()->getName() << '\n');
3360 // Check if we can if-convert non-single-bb loops.
3361 unsigned NumBlocks = TheLoop->getNumBlocks();
3362 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3363 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3367 // ScalarEvolution needs to be able to find the exit count.
3368 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3369 if (ExitCount == SE->getCouldNotCompute()) {
3370 emitAnalysis(VectorizationReport() <<
3371 "could not determine number of loop iterations");
3372 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3376 // Check if we can vectorize the instructions and CFG in this loop.
3377 if (!canVectorizeInstrs()) {
3378 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3382 // Go over each instruction and look at memory deps.
3383 if (!canVectorizeMemory()) {
3384 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3388 // Collect all of the variables that remain uniform after vectorization.
3389 collectLoopUniforms();
3391 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3392 (LAI->getRuntimePointerCheck()->Need ? " (with a runtime bound check)" :
3396 // Okay! We can vectorize. At this point we don't have any other mem analysis
3397 // which may limit our maximum vectorization factor, so just return true with
3402 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3403 if (Ty->isPointerTy())
3404 return DL.getIntPtrType(Ty);
3406 // It is possible that char's or short's overflow when we ask for the loop's
3407 // trip count, work around this by changing the type size.
3408 if (Ty->getScalarSizeInBits() < 32)
3409 return Type::getInt32Ty(Ty->getContext());
3414 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3415 Ty0 = convertPointerToIntegerType(DL, Ty0);
3416 Ty1 = convertPointerToIntegerType(DL, Ty1);
3417 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3422 /// \brief Check that the instruction has outside loop users and is not an
3423 /// identified reduction variable.
3424 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3425 SmallPtrSetImpl<Value *> &Reductions) {
3426 // Reduction instructions are allowed to have exit users. All other
3427 // instructions must not have external users.
3428 if (!Reductions.count(Inst))
3429 //Check that all of the users of the loop are inside the BB.
3430 for (User *U : Inst->users()) {
3431 Instruction *UI = cast<Instruction>(U);
3432 // This user may be a reduction exit value.
3433 if (!TheLoop->contains(UI)) {
3434 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3441 bool LoopVectorizationLegality::canVectorizeInstrs() {
3442 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3443 BasicBlock *Header = TheLoop->getHeader();
3445 // Look for the attribute signaling the absence of NaNs.
3446 Function &F = *Header->getParent();
3447 if (F.hasFnAttribute("no-nans-fp-math"))
3449 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
3451 // For each block in the loop.
3452 for (Loop::block_iterator bb = TheLoop->block_begin(),
3453 be = TheLoop->block_end(); bb != be; ++bb) {
3455 // Scan the instructions in the block and look for hazards.
3456 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3459 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3460 Type *PhiTy = Phi->getType();
3461 // Check that this PHI type is allowed.
3462 if (!PhiTy->isIntegerTy() &&
3463 !PhiTy->isFloatingPointTy() &&
3464 !PhiTy->isPointerTy()) {
3465 emitAnalysis(VectorizationReport(it)
3466 << "loop control flow is not understood by vectorizer");
3467 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3471 // If this PHINode is not in the header block, then we know that we
3472 // can convert it to select during if-conversion. No need to check if
3473 // the PHIs in this block are induction or reduction variables.
3474 if (*bb != Header) {
3475 // Check that this instruction has no outside users or is an
3476 // identified reduction value with an outside user.
3477 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3479 emitAnalysis(VectorizationReport(it) <<
3480 "value could not be identified as "
3481 "an induction or reduction variable");
3485 // We only allow if-converted PHIs with exactly two incoming values.
3486 if (Phi->getNumIncomingValues() != 2) {
3487 emitAnalysis(VectorizationReport(it)
3488 << "control flow not understood by vectorizer");
3489 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3493 // This is the value coming from the preheader.
3494 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3495 ConstantInt *StepValue = nullptr;
3496 // Check if this is an induction variable.
3497 InductionKind IK = isInductionVariable(Phi, StepValue);
3499 if (IK_NoInduction != IK) {
3500 // Get the widest type.
3502 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3504 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3506 // Int inductions are special because we only allow one IV.
3507 if (IK == IK_IntInduction && StepValue->isOne()) {
3508 // Use the phi node with the widest type as induction. Use the last
3509 // one if there are multiple (no good reason for doing this other
3510 // than it is expedient).
3511 if (!Induction || PhiTy == WidestIndTy)
3515 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3516 Inductions[Phi] = InductionInfo(StartValue, IK, StepValue);
3518 // Until we explicitly handle the case of an induction variable with
3519 // an outside loop user we have to give up vectorizing this loop.
3520 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3521 emitAnalysis(VectorizationReport(it) <<
3522 "use of induction value outside of the "
3523 "loop is not handled by vectorizer");
3530 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3531 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3534 if (AddReductionVar(Phi, RK_IntegerMult)) {
3535 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3538 if (AddReductionVar(Phi, RK_IntegerOr)) {
3539 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3542 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3543 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3546 if (AddReductionVar(Phi, RK_IntegerXor)) {
3547 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3550 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3551 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3554 if (AddReductionVar(Phi, RK_FloatMult)) {
3555 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3558 if (AddReductionVar(Phi, RK_FloatAdd)) {
3559 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3562 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3563 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3568 emitAnalysis(VectorizationReport(it) <<
3569 "value that could not be identified as "
3570 "reduction is used outside the loop");
3571 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3573 }// end of PHI handling
3575 // We still don't handle functions. However, we can ignore dbg intrinsic
3576 // calls and we do handle certain intrinsic and libm functions.
3577 CallInst *CI = dyn_cast<CallInst>(it);
3578 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3579 emitAnalysis(VectorizationReport(it) <<
3580 "call instruction cannot be vectorized");
3581 DEBUG(dbgs() << "LV: Found a call site.\n");
3585 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
3586 // second argument is the same (i.e. loop invariant)
3588 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
3589 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
3590 emitAnalysis(VectorizationReport(it)
3591 << "intrinsic instruction cannot be vectorized");
3592 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
3597 // Check that the instruction return type is vectorizable.
3598 // Also, we can't vectorize extractelement instructions.
3599 if ((!VectorType::isValidElementType(it->getType()) &&
3600 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3601 emitAnalysis(VectorizationReport(it)
3602 << "instruction return type cannot be vectorized");
3603 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3607 // Check that the stored type is vectorizable.
3608 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3609 Type *T = ST->getValueOperand()->getType();
3610 if (!VectorType::isValidElementType(T)) {
3611 emitAnalysis(VectorizationReport(ST) <<
3612 "store instruction cannot be vectorized");
3615 if (EnableMemAccessVersioning)
3616 collectStridedAccess(ST);
3619 if (EnableMemAccessVersioning)
3620 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3621 collectStridedAccess(LI);
3623 // Reduction instructions are allowed to have exit users.
3624 // All other instructions must not have external users.
3625 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3626 emitAnalysis(VectorizationReport(it) <<
3627 "value cannot be used outside the loop");
3636 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3637 if (Inductions.empty()) {
3638 emitAnalysis(VectorizationReport()
3639 << "loop induction variable could not be identified");
3647 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3648 /// return the induction operand of the gep pointer.
3649 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3650 const DataLayout *DL, Loop *Lp) {
3651 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3655 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3657 // Check that all of the gep indices are uniform except for our induction
3659 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3660 if (i != InductionOperand &&
3661 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3663 return GEP->getOperand(InductionOperand);
3666 ///\brief Look for a cast use of the passed value.
3667 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3668 Value *UniqueCast = nullptr;
3669 for (User *U : Ptr->users()) {
3670 CastInst *CI = dyn_cast<CastInst>(U);
3671 if (CI && CI->getType() == Ty) {
3681 ///\brief Get the stride of a pointer access in a loop.
3682 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3683 /// pointer to the Value, or null otherwise.
3684 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3685 const DataLayout *DL, Loop *Lp) {
3686 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3687 if (!PtrTy || PtrTy->isAggregateType())
3690 // Try to remove a gep instruction to make the pointer (actually index at this
3691 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3692 // pointer, otherwise, we are analyzing the index.
3693 Value *OrigPtr = Ptr;
3695 // The size of the pointer access.
3696 int64_t PtrAccessSize = 1;
3698 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3699 const SCEV *V = SE->getSCEV(Ptr);
3703 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3704 V = C->getOperand();
3706 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3710 V = S->getStepRecurrence(*SE);
3714 // Strip off the size of access multiplication if we are still analyzing the
3716 if (OrigPtr == Ptr) {
3717 DL->getTypeAllocSize(PtrTy->getElementType());
3718 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3719 if (M->getOperand(0)->getSCEVType() != scConstant)
3722 const APInt &APStepVal =
3723 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3725 // Huge step value - give up.
3726 if (APStepVal.getBitWidth() > 64)
3729 int64_t StepVal = APStepVal.getSExtValue();
3730 if (PtrAccessSize != StepVal)
3732 V = M->getOperand(1);
3737 Type *StripedOffRecurrenceCast = nullptr;
3738 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3739 StripedOffRecurrenceCast = C->getType();
3740 V = C->getOperand();
3743 // Look for the loop invariant symbolic value.
3744 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3748 Value *Stride = U->getValue();
3749 if (!Lp->isLoopInvariant(Stride))
3752 // If we have stripped off the recurrence cast we have to make sure that we
3753 // return the value that is used in this loop so that we can replace it later.
3754 if (StripedOffRecurrenceCast)
3755 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3760 void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) {
3761 Value *Ptr = nullptr;
3762 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3763 Ptr = LI->getPointerOperand();
3764 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3765 Ptr = SI->getPointerOperand();
3769 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3773 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3774 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3775 Strides[Ptr] = Stride;
3776 StrideSet.insert(Stride);
3779 void LoopVectorizationLegality::collectLoopUniforms() {
3780 // We now know that the loop is vectorizable!
3781 // Collect variables that will remain uniform after vectorization.
3782 std::vector<Value*> Worklist;
3783 BasicBlock *Latch = TheLoop->getLoopLatch();
3785 // Start with the conditional branch and walk up the block.
3786 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3788 // Also add all consecutive pointer values; these values will be uniform
3789 // after vectorization (and subsequent cleanup) and, until revectorization is
3790 // supported, all dependencies must also be uniform.
3791 for (Loop::block_iterator B = TheLoop->block_begin(),
3792 BE = TheLoop->block_end(); B != BE; ++B)
3793 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
3795 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
3796 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3798 while (!Worklist.empty()) {
3799 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3800 Worklist.pop_back();
3802 // Look at instructions inside this loop.
3803 // Stop when reaching PHI nodes.
3804 // TODO: we need to follow values all over the loop, not only in this block.
3805 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3808 // This is a known uniform.
3811 // Insert all operands.
3812 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3816 bool LoopVectorizationLegality::canVectorizeMemory() {
3817 LAI = &LAA->getInfo(TheLoop, Strides);
3818 auto &OptionalReport = LAI->getReport();
3820 emitAnalysis(*OptionalReport);
3821 return LAI->canVectorizeMemory();
3824 static bool hasMultipleUsesOf(Instruction *I,
3825 SmallPtrSetImpl<Instruction *> &Insts) {
3826 unsigned NumUses = 0;
3827 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
3828 if (Insts.count(dyn_cast<Instruction>(*Use)))
3837 static bool areAllUsesIn(Instruction *I, SmallPtrSetImpl<Instruction *> &Set) {
3838 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
3839 if (!Set.count(dyn_cast<Instruction>(*Use)))
3844 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
3845 ReductionKind Kind) {
3846 if (Phi->getNumIncomingValues() != 2)
3849 // Reduction variables are only found in the loop header block.
3850 if (Phi->getParent() != TheLoop->getHeader())
3853 // Obtain the reduction start value from the value that comes from the loop
3855 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
3857 // ExitInstruction is the single value which is used outside the loop.
3858 // We only allow for a single reduction value to be used outside the loop.
3859 // This includes users of the reduction, variables (which form a cycle
3860 // which ends in the phi node).
3861 Instruction *ExitInstruction = nullptr;
3862 // Indicates that we found a reduction operation in our scan.
3863 bool FoundReduxOp = false;
3865 // We start with the PHI node and scan for all of the users of this
3866 // instruction. All users must be instructions that can be used as reduction
3867 // variables (such as ADD). We must have a single out-of-block user. The cycle
3868 // must include the original PHI.
3869 bool FoundStartPHI = false;
3871 // To recognize min/max patterns formed by a icmp select sequence, we store
3872 // the number of instruction we saw from the recognized min/max pattern,
3873 // to make sure we only see exactly the two instructions.
3874 unsigned NumCmpSelectPatternInst = 0;
3875 ReductionInstDesc ReduxDesc(false, nullptr);
3877 SmallPtrSet<Instruction *, 8> VisitedInsts;
3878 SmallVector<Instruction *, 8> Worklist;
3879 Worklist.push_back(Phi);
3880 VisitedInsts.insert(Phi);
3882 // A value in the reduction can be used:
3883 // - By the reduction:
3884 // - Reduction operation:
3885 // - One use of reduction value (safe).
3886 // - Multiple use of reduction value (not safe).
3888 // - All uses of the PHI must be the reduction (safe).
3889 // - Otherwise, not safe.
3890 // - By one instruction outside of the loop (safe).
3891 // - By further instructions outside of the loop (not safe).
3892 // - By an instruction that is not part of the reduction (not safe).
3894 // * An instruction type other than PHI or the reduction operation.
3895 // * A PHI in the header other than the initial PHI.
3896 while (!Worklist.empty()) {
3897 Instruction *Cur = Worklist.back();
3898 Worklist.pop_back();
3901 // If the instruction has no users then this is a broken chain and can't be
3902 // a reduction variable.
3903 if (Cur->use_empty())
3906 bool IsAPhi = isa<PHINode>(Cur);
3908 // A header PHI use other than the original PHI.
3909 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
3912 // Reductions of instructions such as Div, and Sub is only possible if the
3913 // LHS is the reduction variable.
3914 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
3915 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
3916 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
3919 // Any reduction instruction must be of one of the allowed kinds.
3920 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
3921 if (!ReduxDesc.IsReduction)
3924 // A reduction operation must only have one use of the reduction value.
3925 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
3926 hasMultipleUsesOf(Cur, VisitedInsts))
3929 // All inputs to a PHI node must be a reduction value.
3930 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
3933 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
3934 isa<SelectInst>(Cur)))
3935 ++NumCmpSelectPatternInst;
3936 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
3937 isa<SelectInst>(Cur)))
3938 ++NumCmpSelectPatternInst;
3940 // Check whether we found a reduction operator.
3941 FoundReduxOp |= !IsAPhi;
3943 // Process users of current instruction. Push non-PHI nodes after PHI nodes
3944 // onto the stack. This way we are going to have seen all inputs to PHI
3945 // nodes once we get to them.
3946 SmallVector<Instruction *, 8> NonPHIs;
3947 SmallVector<Instruction *, 8> PHIs;
3948 for (User *U : Cur->users()) {
3949 Instruction *UI = cast<Instruction>(U);
3951 // Check if we found the exit user.
3952 BasicBlock *Parent = UI->getParent();
3953 if (!TheLoop->contains(Parent)) {
3954 // Exit if you find multiple outside users or if the header phi node is
3955 // being used. In this case the user uses the value of the previous
3956 // iteration, in which case we would loose "VF-1" iterations of the
3957 // reduction operation if we vectorize.
3958 if (ExitInstruction != nullptr || Cur == Phi)
3961 // The instruction used by an outside user must be the last instruction
3962 // before we feed back to the reduction phi. Otherwise, we loose VF-1
3963 // operations on the value.
3964 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
3967 ExitInstruction = Cur;
3971 // Process instructions only once (termination). Each reduction cycle
3972 // value must only be used once, except by phi nodes and min/max
3973 // reductions which are represented as a cmp followed by a select.
3974 ReductionInstDesc IgnoredVal(false, nullptr);
3975 if (VisitedInsts.insert(UI).second) {
3976 if (isa<PHINode>(UI))
3979 NonPHIs.push_back(UI);
3980 } else if (!isa<PHINode>(UI) &&
3981 ((!isa<FCmpInst>(UI) &&
3982 !isa<ICmpInst>(UI) &&
3983 !isa<SelectInst>(UI)) ||
3984 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
3987 // Remember that we completed the cycle.
3989 FoundStartPHI = true;
3991 Worklist.append(PHIs.begin(), PHIs.end());
3992 Worklist.append(NonPHIs.begin(), NonPHIs.end());
3995 // This means we have seen one but not the other instruction of the
3996 // pattern or more than just a select and cmp.
3997 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
3998 NumCmpSelectPatternInst != 2)
4001 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
4004 // We found a reduction var if we have reached the original phi node and we
4005 // only have a single instruction with out-of-loop users.
4007 // This instruction is allowed to have out-of-loop users.
4008 AllowedExit.insert(ExitInstruction);
4010 // Save the description of this reduction variable.
4011 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
4012 ReduxDesc.MinMaxKind);
4013 Reductions[Phi] = RD;
4014 // We've ended the cycle. This is a reduction variable if we have an
4015 // outside user and it has a binary op.
4020 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
4021 /// pattern corresponding to a min(X, Y) or max(X, Y).
4022 LoopVectorizationLegality::ReductionInstDesc
4023 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
4024 ReductionInstDesc &Prev) {
4026 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
4027 "Expect a select instruction");
4028 Instruction *Cmp = nullptr;
4029 SelectInst *Select = nullptr;
4031 // We must handle the select(cmp()) as a single instruction. Advance to the
4033 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
4034 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
4035 return ReductionInstDesc(false, I);
4036 return ReductionInstDesc(Select, Prev.MinMaxKind);
4039 // Only handle single use cases for now.
4040 if (!(Select = dyn_cast<SelectInst>(I)))
4041 return ReductionInstDesc(false, I);
4042 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
4043 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
4044 return ReductionInstDesc(false, I);
4045 if (!Cmp->hasOneUse())
4046 return ReductionInstDesc(false, I);
4051 // Look for a min/max pattern.
4052 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4053 return ReductionInstDesc(Select, MRK_UIntMin);
4054 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4055 return ReductionInstDesc(Select, MRK_UIntMax);
4056 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4057 return ReductionInstDesc(Select, MRK_SIntMax);
4058 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4059 return ReductionInstDesc(Select, MRK_SIntMin);
4060 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4061 return ReductionInstDesc(Select, MRK_FloatMin);
4062 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4063 return ReductionInstDesc(Select, MRK_FloatMax);
4064 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4065 return ReductionInstDesc(Select, MRK_FloatMin);
4066 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4067 return ReductionInstDesc(Select, MRK_FloatMax);
4069 return ReductionInstDesc(false, I);
4072 LoopVectorizationLegality::ReductionInstDesc
4073 LoopVectorizationLegality::isReductionInstr(Instruction *I,
4075 ReductionInstDesc &Prev) {
4076 bool FP = I->getType()->isFloatingPointTy();
4077 bool FastMath = FP && I->hasUnsafeAlgebra();
4078 switch (I->getOpcode()) {
4080 return ReductionInstDesc(false, I);
4081 case Instruction::PHI:
4082 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
4083 Kind != RK_FloatMinMax))
4084 return ReductionInstDesc(false, I);
4085 return ReductionInstDesc(I, Prev.MinMaxKind);
4086 case Instruction::Sub:
4087 case Instruction::Add:
4088 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4089 case Instruction::Mul:
4090 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4091 case Instruction::And:
4092 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4093 case Instruction::Or:
4094 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4095 case Instruction::Xor:
4096 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4097 case Instruction::FMul:
4098 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4099 case Instruction::FSub:
4100 case Instruction::FAdd:
4101 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4102 case Instruction::FCmp:
4103 case Instruction::ICmp:
4104 case Instruction::Select:
4105 if (Kind != RK_IntegerMinMax &&
4106 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4107 return ReductionInstDesc(false, I);
4108 return isMinMaxSelectCmpPattern(I, Prev);
4112 LoopVectorizationLegality::InductionKind
4113 LoopVectorizationLegality::isInductionVariable(PHINode *Phi,
4114 ConstantInt *&StepValue) {
4115 Type *PhiTy = Phi->getType();
4116 // We only handle integer and pointer inductions variables.
4117 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4118 return IK_NoInduction;
4120 // Check that the PHI is consecutive.
4121 const SCEV *PhiScev = SE->getSCEV(Phi);
4122 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4124 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4125 return IK_NoInduction;
4128 const SCEV *Step = AR->getStepRecurrence(*SE);
4129 // Calculate the pointer stride and check if it is consecutive.
4130 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4132 return IK_NoInduction;
4134 ConstantInt *CV = C->getValue();
4135 if (PhiTy->isIntegerTy()) {
4137 return IK_IntInduction;
4140 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4141 Type *PointerElementType = PhiTy->getPointerElementType();
4142 // The pointer stride cannot be determined if the pointer element type is not
4144 if (!PointerElementType->isSized())
4145 return IK_NoInduction;
4147 int64_t Size = static_cast<int64_t>(DL->getTypeAllocSize(PointerElementType));
4148 int64_t CVSize = CV->getSExtValue();
4150 return IK_NoInduction;
4151 StepValue = ConstantInt::getSigned(CV->getType(), CVSize / Size);
4152 return IK_PtrInduction;
4155 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4156 Value *In0 = const_cast<Value*>(V);
4157 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4161 return Inductions.count(PN);
4164 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4165 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4168 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4169 SmallPtrSetImpl<Value *> &SafePtrs) {
4171 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4172 // Check that we don't have a constant expression that can trap as operand.
4173 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4175 if (Constant *C = dyn_cast<Constant>(*OI))
4179 // We might be able to hoist the load.
4180 if (it->mayReadFromMemory()) {
4181 LoadInst *LI = dyn_cast<LoadInst>(it);
4184 if (!SafePtrs.count(LI->getPointerOperand())) {
4185 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
4186 MaskedOp.insert(LI);
4193 // We don't predicate stores at the moment.
4194 if (it->mayWriteToMemory()) {
4195 StoreInst *SI = dyn_cast<StoreInst>(it);
4196 // We only support predication of stores in basic blocks with one
4201 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
4202 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
4204 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
4205 !isSinglePredecessor) {
4206 // Build a masked store if it is legal for the target, otherwise scalarize
4208 bool isLegalMaskedOp =
4209 isLegalMaskedStore(SI->getValueOperand()->getType(),
4210 SI->getPointerOperand());
4211 if (isLegalMaskedOp) {
4213 MaskedOp.insert(SI);
4222 // The instructions below can trap.
4223 switch (it->getOpcode()) {
4225 case Instruction::UDiv:
4226 case Instruction::SDiv:
4227 case Instruction::URem:
4228 case Instruction::SRem:
4236 LoopVectorizationCostModel::VectorizationFactor
4237 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
4238 // Width 1 means no vectorize
4239 VectorizationFactor Factor = { 1U, 0U };
4240 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4241 emitAnalysis(VectorizationReport() <<
4242 "runtime pointer checks needed. Enable vectorization of this "
4243 "loop with '#pragma clang loop vectorize(enable)' when "
4244 "compiling with -Os");
4245 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4249 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
4250 emitAnalysis(VectorizationReport() <<
4251 "store that is conditionally executed prevents vectorization");
4252 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
4256 // Find the trip count.
4257 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4258 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4260 unsigned WidestType = getWidestType();
4261 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4262 unsigned MaxSafeDepDist = -1U;
4263 if (Legal->getMaxSafeDepDistBytes() != -1U)
4264 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4265 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4266 WidestRegister : MaxSafeDepDist);
4267 unsigned MaxVectorSize = WidestRegister / WidestType;
4268 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4269 DEBUG(dbgs() << "LV: The Widest register is: "
4270 << WidestRegister << " bits.\n");
4272 if (MaxVectorSize == 0) {
4273 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4277 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
4278 " into one vector!");
4280 unsigned VF = MaxVectorSize;
4282 // If we optimize the program for size, avoid creating the tail loop.
4284 // If we are unable to calculate the trip count then don't try to vectorize.
4287 (VectorizationReport() <<
4288 "unable to calculate the loop count due to complex control flow");
4289 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4293 // Find the maximum SIMD width that can fit within the trip count.
4294 VF = TC % MaxVectorSize;
4299 // If the trip count that we found modulo the vectorization factor is not
4300 // zero then we require a tail.
4302 emitAnalysis(VectorizationReport() <<
4303 "cannot optimize for size and vectorize at the "
4304 "same time. Enable vectorization of this loop "
4305 "with '#pragma clang loop vectorize(enable)' "
4306 "when compiling with -Os");
4307 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4312 int UserVF = Hints->getWidth();
4314 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4315 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4317 Factor.Width = UserVF;
4321 float Cost = expectedCost(1);
4323 const float ScalarCost = Cost;
4326 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
4328 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
4329 // Ignore scalar width, because the user explicitly wants vectorization.
4330 if (ForceVectorization && VF > 1) {
4332 Cost = expectedCost(Width) / (float)Width;
4335 for (unsigned i=2; i <= VF; i*=2) {
4336 // Notice that the vector loop needs to be executed less times, so
4337 // we need to divide the cost of the vector loops by the width of
4338 // the vector elements.
4339 float VectorCost = expectedCost(i) / (float)i;
4340 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
4341 (int)VectorCost << ".\n");
4342 if (VectorCost < Cost) {
4348 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
4349 << "LV: Vectorization seems to be not beneficial, "
4350 << "but was forced by a user.\n");
4351 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
4352 Factor.Width = Width;
4353 Factor.Cost = Width * Cost;
4357 unsigned LoopVectorizationCostModel::getWidestType() {
4358 unsigned MaxWidth = 8;
4361 for (Loop::block_iterator bb = TheLoop->block_begin(),
4362 be = TheLoop->block_end(); bb != be; ++bb) {
4363 BasicBlock *BB = *bb;
4365 // For each instruction in the loop.
4366 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4367 Type *T = it->getType();
4369 // Ignore ephemeral values.
4370 if (EphValues.count(it))
4373 // Only examine Loads, Stores and PHINodes.
4374 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4377 // Examine PHI nodes that are reduction variables.
4378 if (PHINode *PN = dyn_cast<PHINode>(it))
4379 if (!Legal->getReductionVars()->count(PN))
4382 // Examine the stored values.
4383 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4384 T = ST->getValueOperand()->getType();
4386 // Ignore loaded pointer types and stored pointer types that are not
4387 // consecutive. However, we do want to take consecutive stores/loads of
4388 // pointer vectors into account.
4389 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4392 MaxWidth = std::max(MaxWidth,
4393 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
4401 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
4403 unsigned LoopCost) {
4405 // -- The unroll heuristics --
4406 // We unroll the loop in order to expose ILP and reduce the loop overhead.
4407 // There are many micro-architectural considerations that we can't predict
4408 // at this level. For example, frontend pressure (on decode or fetch) due to
4409 // code size, or the number and capabilities of the execution ports.
4411 // We use the following heuristics to select the unroll factor:
4412 // 1. If the code has reductions, then we unroll in order to break the cross
4413 // iteration dependency.
4414 // 2. If the loop is really small, then we unroll in order to reduce the loop
4416 // 3. We don't unroll if we think that we will spill registers to memory due
4417 // to the increased register pressure.
4419 // Use the user preference, unless 'auto' is selected.
4420 int UserUF = Hints->getInterleave();
4424 // When we optimize for size, we don't unroll.
4428 // We used the distance for the unroll factor.
4429 if (Legal->getMaxSafeDepDistBytes() != -1U)
4432 // Do not unroll loops with a relatively small trip count.
4433 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4434 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
4437 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
4438 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
4442 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
4443 TargetNumRegisters = ForceTargetNumScalarRegs;
4445 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
4446 TargetNumRegisters = ForceTargetNumVectorRegs;
4449 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4450 // We divide by these constants so assume that we have at least one
4451 // instruction that uses at least one register.
4452 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4453 R.NumInstructions = std::max(R.NumInstructions, 1U);
4455 // We calculate the unroll factor using the following formula.
4456 // Subtract the number of loop invariants from the number of available
4457 // registers. These registers are used by all of the unrolled instances.
4458 // Next, divide the remaining registers by the number of registers that is
4459 // required by the loop, in order to estimate how many parallel instances
4460 // fit without causing spills. All of this is rounded down if necessary to be
4461 // a power of two. We want power of two unroll factors to simplify any
4462 // addressing operations or alignment considerations.
4463 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
4466 // Don't count the induction variable as unrolled.
4467 if (EnableIndVarRegisterHeur)
4468 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
4469 std::max(1U, (R.MaxLocalUsers - 1)));
4471 // Clamp the unroll factor ranges to reasonable factors.
4472 unsigned MaxInterleaveSize = TTI.getMaxInterleaveFactor();
4474 // Check if the user has overridden the unroll max.
4476 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
4477 MaxInterleaveSize = ForceTargetMaxScalarInterleaveFactor;
4479 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
4480 MaxInterleaveSize = ForceTargetMaxVectorInterleaveFactor;
4483 // If we did not calculate the cost for VF (because the user selected the VF)
4484 // then we calculate the cost of VF here.
4486 LoopCost = expectedCost(VF);
4488 // Clamp the calculated UF to be between the 1 and the max unroll factor
4489 // that the target allows.
4490 if (UF > MaxInterleaveSize)
4491 UF = MaxInterleaveSize;
4495 // Unroll if we vectorized this loop and there is a reduction that could
4496 // benefit from unrolling.
4497 if (VF > 1 && Legal->getReductionVars()->size()) {
4498 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
4502 // Note that if we've already vectorized the loop we will have done the
4503 // runtime check and so unrolling won't require further checks.
4504 bool UnrollingRequiresRuntimePointerCheck =
4505 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
4507 // We want to unroll small loops in order to reduce the loop overhead and
4508 // potentially expose ILP opportunities.
4509 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
4510 if (!UnrollingRequiresRuntimePointerCheck &&
4511 LoopCost < SmallLoopCost) {
4512 // We assume that the cost overhead is 1 and we use the cost model
4513 // to estimate the cost of the loop and unroll until the cost of the
4514 // loop overhead is about 5% of the cost of the loop.
4515 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
4517 // Unroll until store/load ports (estimated by max unroll factor) are
4519 unsigned NumStores = Legal->getNumStores();
4520 unsigned NumLoads = Legal->getNumLoads();
4521 unsigned StoresUF = UF / (NumStores ? NumStores : 1);
4522 unsigned LoadsUF = UF / (NumLoads ? NumLoads : 1);
4524 // If we have a scalar reduction (vector reductions are already dealt with
4525 // by this point), we can increase the critical path length if the loop
4526 // we're unrolling is inside another loop. Limit, by default to 2, so the
4527 // critical path only gets increased by one reduction operation.
4528 if (Legal->getReductionVars()->size() &&
4529 TheLoop->getLoopDepth() > 1) {
4530 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionUF);
4531 SmallUF = std::min(SmallUF, F);
4532 StoresUF = std::min(StoresUF, F);
4533 LoadsUF = std::min(LoadsUF, F);
4536 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
4537 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
4538 return std::max(StoresUF, LoadsUF);
4541 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
4545 DEBUG(dbgs() << "LV: Not Unrolling.\n");
4549 LoopVectorizationCostModel::RegisterUsage
4550 LoopVectorizationCostModel::calculateRegisterUsage() {
4551 // This function calculates the register usage by measuring the highest number
4552 // of values that are alive at a single location. Obviously, this is a very
4553 // rough estimation. We scan the loop in a topological order in order and
4554 // assign a number to each instruction. We use RPO to ensure that defs are
4555 // met before their users. We assume that each instruction that has in-loop
4556 // users starts an interval. We record every time that an in-loop value is
4557 // used, so we have a list of the first and last occurrences of each
4558 // instruction. Next, we transpose this data structure into a multi map that
4559 // holds the list of intervals that *end* at a specific location. This multi
4560 // map allows us to perform a linear search. We scan the instructions linearly
4561 // and record each time that a new interval starts, by placing it in a set.
4562 // If we find this value in the multi-map then we remove it from the set.
4563 // The max register usage is the maximum size of the set.
4564 // We also search for instructions that are defined outside the loop, but are
4565 // used inside the loop. We need this number separately from the max-interval
4566 // usage number because when we unroll, loop-invariant values do not take
4568 LoopBlocksDFS DFS(TheLoop);
4572 R.NumInstructions = 0;
4574 // Each 'key' in the map opens a new interval. The values
4575 // of the map are the index of the 'last seen' usage of the
4576 // instruction that is the key.
4577 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4578 // Maps instruction to its index.
4579 DenseMap<unsigned, Instruction*> IdxToInstr;
4580 // Marks the end of each interval.
4581 IntervalMap EndPoint;
4582 // Saves the list of instruction indices that are used in the loop.
4583 SmallSet<Instruction*, 8> Ends;
4584 // Saves the list of values that are used in the loop but are
4585 // defined outside the loop, such as arguments and constants.
4586 SmallPtrSet<Value*, 8> LoopInvariants;
4589 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4590 be = DFS.endRPO(); bb != be; ++bb) {
4591 R.NumInstructions += (*bb)->size();
4592 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4594 Instruction *I = it;
4595 IdxToInstr[Index++] = I;
4597 // Save the end location of each USE.
4598 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4599 Value *U = I->getOperand(i);
4600 Instruction *Instr = dyn_cast<Instruction>(U);
4602 // Ignore non-instruction values such as arguments, constants, etc.
4603 if (!Instr) continue;
4605 // If this instruction is outside the loop then record it and continue.
4606 if (!TheLoop->contains(Instr)) {
4607 LoopInvariants.insert(Instr);
4611 // Overwrite previous end points.
4612 EndPoint[Instr] = Index;
4618 // Saves the list of intervals that end with the index in 'key'.
4619 typedef SmallVector<Instruction*, 2> InstrList;
4620 DenseMap<unsigned, InstrList> TransposeEnds;
4622 // Transpose the EndPoints to a list of values that end at each index.
4623 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4625 TransposeEnds[it->second].push_back(it->first);
4627 SmallSet<Instruction*, 8> OpenIntervals;
4628 unsigned MaxUsage = 0;
4631 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4632 for (unsigned int i = 0; i < Index; ++i) {
4633 Instruction *I = IdxToInstr[i];
4634 // Ignore instructions that are never used within the loop.
4635 if (!Ends.count(I)) continue;
4637 // Ignore ephemeral values.
4638 if (EphValues.count(I))
4641 // Remove all of the instructions that end at this location.
4642 InstrList &List = TransposeEnds[i];
4643 for (unsigned int j=0, e = List.size(); j < e; ++j)
4644 OpenIntervals.erase(List[j]);
4646 // Count the number of live interals.
4647 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4649 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4650 OpenIntervals.size() << '\n');
4652 // Add the current instruction to the list of open intervals.
4653 OpenIntervals.insert(I);
4656 unsigned Invariant = LoopInvariants.size();
4657 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
4658 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
4659 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
4661 R.LoopInvariantRegs = Invariant;
4662 R.MaxLocalUsers = MaxUsage;
4666 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4670 for (Loop::block_iterator bb = TheLoop->block_begin(),
4671 be = TheLoop->block_end(); bb != be; ++bb) {
4672 unsigned BlockCost = 0;
4673 BasicBlock *BB = *bb;
4675 // For each instruction in the old loop.
4676 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4677 // Skip dbg intrinsics.
4678 if (isa<DbgInfoIntrinsic>(it))
4681 // Ignore ephemeral values.
4682 if (EphValues.count(it))
4685 unsigned C = getInstructionCost(it, VF);
4687 // Check if we should override the cost.
4688 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
4689 C = ForceTargetInstructionCost;
4692 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
4693 VF << " For instruction: " << *it << '\n');
4696 // We assume that if-converted blocks have a 50% chance of being executed.
4697 // When the code is scalar then some of the blocks are avoided due to CF.
4698 // When the code is vectorized we execute all code paths.
4699 if (VF == 1 && Legal->blockNeedsPredication(*bb))
4708 /// \brief Check whether the address computation for a non-consecutive memory
4709 /// access looks like an unlikely candidate for being merged into the indexing
4712 /// We look for a GEP which has one index that is an induction variable and all
4713 /// other indices are loop invariant. If the stride of this access is also
4714 /// within a small bound we decide that this address computation can likely be
4715 /// merged into the addressing mode.
4716 /// In all other cases, we identify the address computation as complex.
4717 static bool isLikelyComplexAddressComputation(Value *Ptr,
4718 LoopVectorizationLegality *Legal,
4719 ScalarEvolution *SE,
4720 const Loop *TheLoop) {
4721 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
4725 // We are looking for a gep with all loop invariant indices except for one
4726 // which should be an induction variable.
4727 unsigned NumOperands = Gep->getNumOperands();
4728 for (unsigned i = 1; i < NumOperands; ++i) {
4729 Value *Opd = Gep->getOperand(i);
4730 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
4731 !Legal->isInductionVariable(Opd))
4735 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
4736 // can likely be merged into the address computation.
4737 unsigned MaxMergeDistance = 64;
4739 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
4743 // Check the step is constant.
4744 const SCEV *Step = AddRec->getStepRecurrence(*SE);
4745 // Calculate the pointer stride and check if it is consecutive.
4746 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4750 const APInt &APStepVal = C->getValue()->getValue();
4752 // Huge step value - give up.
4753 if (APStepVal.getBitWidth() > 64)
4756 int64_t StepVal = APStepVal.getSExtValue();
4758 return StepVal > MaxMergeDistance;
4761 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
4762 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
4768 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
4769 // If we know that this instruction will remain uniform, check the cost of
4770 // the scalar version.
4771 if (Legal->isUniformAfterVectorization(I))
4774 Type *RetTy = I->getType();
4775 Type *VectorTy = ToVectorTy(RetTy, VF);
4777 // TODO: We need to estimate the cost of intrinsic calls.
4778 switch (I->getOpcode()) {
4779 case Instruction::GetElementPtr:
4780 // We mark this instruction as zero-cost because the cost of GEPs in
4781 // vectorized code depends on whether the corresponding memory instruction
4782 // is scalarized or not. Therefore, we handle GEPs with the memory
4783 // instruction cost.
4785 case Instruction::Br: {
4786 return TTI.getCFInstrCost(I->getOpcode());
4788 case Instruction::PHI:
4789 //TODO: IF-converted IFs become selects.
4791 case Instruction::Add:
4792 case Instruction::FAdd:
4793 case Instruction::Sub:
4794 case Instruction::FSub:
4795 case Instruction::Mul:
4796 case Instruction::FMul:
4797 case Instruction::UDiv:
4798 case Instruction::SDiv:
4799 case Instruction::FDiv:
4800 case Instruction::URem:
4801 case Instruction::SRem:
4802 case Instruction::FRem:
4803 case Instruction::Shl:
4804 case Instruction::LShr:
4805 case Instruction::AShr:
4806 case Instruction::And:
4807 case Instruction::Or:
4808 case Instruction::Xor: {
4809 // Since we will replace the stride by 1 the multiplication should go away.
4810 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
4812 // Certain instructions can be cheaper to vectorize if they have a constant
4813 // second vector operand. One example of this are shifts on x86.
4814 TargetTransformInfo::OperandValueKind Op1VK =
4815 TargetTransformInfo::OK_AnyValue;
4816 TargetTransformInfo::OperandValueKind Op2VK =
4817 TargetTransformInfo::OK_AnyValue;
4818 TargetTransformInfo::OperandValueProperties Op1VP =
4819 TargetTransformInfo::OP_None;
4820 TargetTransformInfo::OperandValueProperties Op2VP =
4821 TargetTransformInfo::OP_None;
4822 Value *Op2 = I->getOperand(1);
4824 // Check for a splat of a constant or for a non uniform vector of constants.
4825 if (isa<ConstantInt>(Op2)) {
4826 ConstantInt *CInt = cast<ConstantInt>(Op2);
4827 if (CInt && CInt->getValue().isPowerOf2())
4828 Op2VP = TargetTransformInfo::OP_PowerOf2;
4829 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4830 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
4831 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
4832 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
4834 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
4835 if (CInt && CInt->getValue().isPowerOf2())
4836 Op2VP = TargetTransformInfo::OP_PowerOf2;
4837 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4841 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
4844 case Instruction::Select: {
4845 SelectInst *SI = cast<SelectInst>(I);
4846 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
4847 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
4848 Type *CondTy = SI->getCondition()->getType();
4850 CondTy = VectorType::get(CondTy, VF);
4852 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
4854 case Instruction::ICmp:
4855 case Instruction::FCmp: {
4856 Type *ValTy = I->getOperand(0)->getType();
4857 VectorTy = ToVectorTy(ValTy, VF);
4858 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
4860 case Instruction::Store:
4861 case Instruction::Load: {
4862 StoreInst *SI = dyn_cast<StoreInst>(I);
4863 LoadInst *LI = dyn_cast<LoadInst>(I);
4864 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
4866 VectorTy = ToVectorTy(ValTy, VF);
4868 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
4869 unsigned AS = SI ? SI->getPointerAddressSpace() :
4870 LI->getPointerAddressSpace();
4871 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
4872 // We add the cost of address computation here instead of with the gep
4873 // instruction because only here we know whether the operation is
4876 return TTI.getAddressComputationCost(VectorTy) +
4877 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4879 // Scalarized loads/stores.
4880 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
4881 bool Reverse = ConsecutiveStride < 0;
4882 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
4883 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
4884 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
4885 bool IsComplexComputation =
4886 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
4888 // The cost of extracting from the value vector and pointer vector.
4889 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
4890 for (unsigned i = 0; i < VF; ++i) {
4891 // The cost of extracting the pointer operand.
4892 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
4893 // In case of STORE, the cost of ExtractElement from the vector.
4894 // In case of LOAD, the cost of InsertElement into the returned
4896 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
4897 Instruction::InsertElement,
4901 // The cost of the scalar loads/stores.
4902 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
4903 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
4908 // Wide load/stores.
4909 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
4910 if (Legal->isMaskRequired(I))
4911 Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment,
4914 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4917 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
4921 case Instruction::ZExt:
4922 case Instruction::SExt:
4923 case Instruction::FPToUI:
4924 case Instruction::FPToSI:
4925 case Instruction::FPExt:
4926 case Instruction::PtrToInt:
4927 case Instruction::IntToPtr:
4928 case Instruction::SIToFP:
4929 case Instruction::UIToFP:
4930 case Instruction::Trunc:
4931 case Instruction::FPTrunc:
4932 case Instruction::BitCast: {
4933 // We optimize the truncation of induction variable.
4934 // The cost of these is the same as the scalar operation.
4935 if (I->getOpcode() == Instruction::Trunc &&
4936 Legal->isInductionVariable(I->getOperand(0)))
4937 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
4938 I->getOperand(0)->getType());
4940 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
4941 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
4943 case Instruction::Call: {
4944 CallInst *CI = cast<CallInst>(I);
4945 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
4946 assert(ID && "Not an intrinsic call!");
4947 Type *RetTy = ToVectorTy(CI->getType(), VF);
4948 SmallVector<Type*, 4> Tys;
4949 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
4950 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
4951 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
4954 // We are scalarizing the instruction. Return the cost of the scalar
4955 // instruction, plus the cost of insert and extract into vector
4956 // elements, times the vector width.
4959 if (!RetTy->isVoidTy() && VF != 1) {
4960 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
4962 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
4965 // The cost of inserting the results plus extracting each one of the
4967 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
4970 // The cost of executing VF copies of the scalar instruction. This opcode
4971 // is unknown. Assume that it is the same as 'mul'.
4972 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
4978 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
4979 if (Scalar->isVoidTy() || VF == 1)
4981 return VectorType::get(Scalar, VF);
4984 char LoopVectorize::ID = 0;
4985 static const char lv_name[] = "Loop Vectorization";
4986 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
4987 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
4988 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
4989 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
4990 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
4991 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
4992 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4993 INITIALIZE_PASS_DEPENDENCY(LCSSA)
4994 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
4995 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4996 INITIALIZE_PASS_DEPENDENCY(LoopAccessAnalysis)
4997 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5000 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5001 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5005 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5006 // Check for a store.
5007 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5008 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5010 // Check for a load.
5011 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5012 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5018 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5019 bool IfPredicateStore) {
5020 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5021 // Holds vector parameters or scalars, in case of uniform vals.
5022 SmallVector<VectorParts, 4> Params;
5024 setDebugLocFromInst(Builder, Instr);
5026 // Find all of the vectorized parameters.
5027 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5028 Value *SrcOp = Instr->getOperand(op);
5030 // If we are accessing the old induction variable, use the new one.
5031 if (SrcOp == OldInduction) {
5032 Params.push_back(getVectorValue(SrcOp));
5036 // Try using previously calculated values.
5037 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5039 // If the src is an instruction that appeared earlier in the basic block
5040 // then it should already be vectorized.
5041 if (SrcInst && OrigLoop->contains(SrcInst)) {
5042 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5043 // The parameter is a vector value from earlier.
5044 Params.push_back(WidenMap.get(SrcInst));
5046 // The parameter is a scalar from outside the loop. Maybe even a constant.
5047 VectorParts Scalars;
5048 Scalars.append(UF, SrcOp);
5049 Params.push_back(Scalars);
5053 assert(Params.size() == Instr->getNumOperands() &&
5054 "Invalid number of operands");
5056 // Does this instruction return a value ?
5057 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5059 Value *UndefVec = IsVoidRetTy ? nullptr :
5060 UndefValue::get(Instr->getType());
5061 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5062 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5064 Instruction *InsertPt = Builder.GetInsertPoint();
5065 BasicBlock *IfBlock = Builder.GetInsertBlock();
5066 BasicBlock *CondBlock = nullptr;
5069 Loop *VectorLp = nullptr;
5070 if (IfPredicateStore) {
5071 assert(Instr->getParent()->getSinglePredecessor() &&
5072 "Only support single predecessor blocks");
5073 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5074 Instr->getParent());
5075 VectorLp = LI->getLoopFor(IfBlock);
5076 assert(VectorLp && "Must have a loop for this block");
5079 // For each vector unroll 'part':
5080 for (unsigned Part = 0; Part < UF; ++Part) {
5081 // For each scalar that we create:
5083 // Start an "if (pred) a[i] = ..." block.
5084 Value *Cmp = nullptr;
5085 if (IfPredicateStore) {
5086 if (Cond[Part]->getType()->isVectorTy())
5088 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5089 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5090 ConstantInt::get(Cond[Part]->getType(), 1));
5091 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5092 LoopVectorBody.push_back(CondBlock);
5093 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
5094 // Update Builder with newly created basic block.
5095 Builder.SetInsertPoint(InsertPt);
5098 Instruction *Cloned = Instr->clone();
5100 Cloned->setName(Instr->getName() + ".cloned");
5101 // Replace the operands of the cloned instructions with extracted scalars.
5102 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5103 Value *Op = Params[op][Part];
5104 Cloned->setOperand(op, Op);
5107 // Place the cloned scalar in the new loop.
5108 Builder.Insert(Cloned);
5110 // If the original scalar returns a value we need to place it in a vector
5111 // so that future users will be able to use it.
5113 VecResults[Part] = Cloned;
5116 if (IfPredicateStore) {
5117 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
5118 LoopVectorBody.push_back(NewIfBlock);
5119 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
5120 Builder.SetInsertPoint(InsertPt);
5121 Instruction *OldBr = IfBlock->getTerminator();
5122 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
5123 OldBr->eraseFromParent();
5124 IfBlock = NewIfBlock;
5129 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5130 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5131 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5133 return scalarizeInstruction(Instr, IfPredicateStore);
5136 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5140 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5144 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) {
5145 // When unrolling and the VF is 1, we only need to add a simple scalar.
5146 Type *ITy = Val->getType();
5147 assert(!ITy->isVectorTy() && "Val must be a scalar");
5148 Constant *C = ConstantInt::get(ITy, StartIdx);
5149 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");