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 : NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
536 TLI(TLI), TheFunction(F), TTI(TTI), DT(DT), Induction(nullptr),
537 WidestIndTy(nullptr),
538 LAI(L, SE, DL, TLI, AA, DT),
539 HasFunNoNaNAttr(false) {}
541 /// This enum represents the kinds of reductions that we support.
543 RK_NoReduction, ///< Not a reduction.
544 RK_IntegerAdd, ///< Sum of integers.
545 RK_IntegerMult, ///< Product of integers.
546 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
547 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
548 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
549 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
550 RK_FloatAdd, ///< Sum of floats.
551 RK_FloatMult, ///< Product of floats.
552 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
555 /// This enum represents the kinds of inductions that we support.
557 IK_NoInduction, ///< Not an induction variable.
558 IK_IntInduction, ///< Integer induction variable. Step = C.
559 IK_PtrInduction ///< Pointer induction var. Step = C / sizeof(elem).
562 // This enum represents the kind of minmax reduction.
563 enum MinMaxReductionKind {
573 /// This struct holds information about reduction variables.
574 struct ReductionDescriptor {
575 ReductionDescriptor() : StartValue(nullptr), LoopExitInstr(nullptr),
576 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
578 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
579 MinMaxReductionKind MK)
580 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
582 // The starting value of the reduction.
583 // It does not have to be zero!
584 TrackingVH<Value> StartValue;
585 // The instruction who's value is used outside the loop.
586 Instruction *LoopExitInstr;
587 // The kind of the reduction.
589 // If this a min/max reduction the kind of reduction.
590 MinMaxReductionKind MinMaxKind;
593 /// This POD struct holds information about a potential reduction operation.
594 struct ReductionInstDesc {
595 ReductionInstDesc(bool IsRedux, Instruction *I) :
596 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
598 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
599 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
601 // Is this instruction a reduction candidate.
603 // The last instruction in a min/max pattern (select of the select(icmp())
604 // pattern), or the current reduction instruction otherwise.
605 Instruction *PatternLastInst;
606 // If this is a min/max pattern the comparison predicate.
607 MinMaxReductionKind MinMaxKind;
610 /// A struct for saving information about induction variables.
611 struct InductionInfo {
612 InductionInfo(Value *Start, InductionKind K, ConstantInt *Step)
613 : StartValue(Start), IK(K), StepValue(Step) {
614 assert(IK != IK_NoInduction && "Not an induction");
615 assert(StartValue && "StartValue is null");
616 assert(StepValue && !StepValue->isZero() && "StepValue is zero");
617 assert((IK != IK_PtrInduction || StartValue->getType()->isPointerTy()) &&
618 "StartValue is not a pointer for pointer induction");
619 assert((IK != IK_IntInduction || StartValue->getType()->isIntegerTy()) &&
620 "StartValue is not an integer for integer induction");
621 assert(StepValue->getType()->isIntegerTy() &&
622 "StepValue is not an integer");
625 : StartValue(nullptr), IK(IK_NoInduction), StepValue(nullptr) {}
627 /// Get the consecutive direction. Returns:
628 /// 0 - unknown or non-consecutive.
629 /// 1 - consecutive and increasing.
630 /// -1 - consecutive and decreasing.
631 int getConsecutiveDirection() const {
632 if (StepValue && (StepValue->isOne() || StepValue->isMinusOne()))
633 return StepValue->getSExtValue();
637 /// Compute the transformed value of Index at offset StartValue using step
639 /// For integer induction, returns StartValue + Index * StepValue.
640 /// For pointer induction, returns StartValue[Index * StepValue].
641 /// FIXME: The newly created binary instructions should contain nsw/nuw
642 /// flags, which can be found from the original scalar operations.
643 Value *transform(IRBuilder<> &B, Value *Index) const {
645 case IK_IntInduction:
646 assert(Index->getType() == StartValue->getType() &&
647 "Index type does not match StartValue type");
648 if (StepValue->isMinusOne())
649 return B.CreateSub(StartValue, Index);
650 if (!StepValue->isOne())
651 Index = B.CreateMul(Index, StepValue);
652 return B.CreateAdd(StartValue, Index);
654 case IK_PtrInduction:
655 if (StepValue->isMinusOne())
656 Index = B.CreateNeg(Index);
657 else if (!StepValue->isOne())
658 Index = B.CreateMul(Index, StepValue);
659 return B.CreateGEP(StartValue, Index);
664 llvm_unreachable("invalid enum");
668 TrackingVH<Value> StartValue;
672 ConstantInt *StepValue;
675 /// ReductionList contains the reduction descriptors for all
676 /// of the reductions that were found in the loop.
677 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
679 /// InductionList saves induction variables and maps them to the
680 /// induction descriptor.
681 typedef MapVector<PHINode*, InductionInfo> InductionList;
683 /// Returns true if it is legal to vectorize this loop.
684 /// This does not mean that it is profitable to vectorize this
685 /// loop, only that it is legal to do so.
688 /// Returns the Induction variable.
689 PHINode *getInduction() { return Induction; }
691 /// Returns the reduction variables found in the loop.
692 ReductionList *getReductionVars() { return &Reductions; }
694 /// Returns the induction variables found in the loop.
695 InductionList *getInductionVars() { return &Inductions; }
697 /// Returns the widest induction type.
698 Type *getWidestInductionType() { return WidestIndTy; }
700 /// Returns True if V is an induction variable in this loop.
701 bool isInductionVariable(const Value *V);
703 /// Return true if the block BB needs to be predicated in order for the loop
704 /// to be vectorized.
705 bool blockNeedsPredication(BasicBlock *BB);
707 /// Check if this pointer is consecutive when vectorizing. This happens
708 /// when the last index of the GEP is the induction variable, or that the
709 /// pointer itself is an induction variable.
710 /// This check allows us to vectorize A[idx] into a wide load/store.
712 /// 0 - Stride is unknown or non-consecutive.
713 /// 1 - Address is consecutive.
714 /// -1 - Address is consecutive, and decreasing.
715 int isConsecutivePtr(Value *Ptr);
717 /// Returns true if the value V is uniform within the loop.
718 bool isUniform(Value *V);
720 /// Returns true if this instruction will remain scalar after vectorization.
721 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
723 /// Returns the information that we collected about runtime memory check.
724 LoopAccessInfo::RuntimePointerCheck *getRuntimePointerCheck() {
725 return LAI.getRuntimePointerCheck();
728 LoopAccessInfo *getLAI() {
732 /// This function returns the identity element (or neutral element) for
734 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
736 unsigned getMaxSafeDepDistBytes() { return LAI.getMaxSafeDepDistBytes(); }
738 bool hasStride(Value *V) { return StrideSet.count(V); }
739 bool mustCheckStrides() { return !StrideSet.empty(); }
740 SmallPtrSet<Value *, 8>::iterator strides_begin() {
741 return StrideSet.begin();
743 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
745 /// Returns true if the target machine supports masked store operation
746 /// for the given \p DataType and kind of access to \p Ptr.
747 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
748 return TTI->isLegalMaskedStore(DataType, isConsecutivePtr(Ptr));
750 /// Returns true if the target machine supports masked load operation
751 /// for the given \p DataType and kind of access to \p Ptr.
752 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
753 return TTI->isLegalMaskedLoad(DataType, isConsecutivePtr(Ptr));
755 /// Returns true if vector representation of the instruction \p I
757 bool isMaskRequired(const Instruction* I) {
758 return (MaskedOp.count(I) != 0);
760 unsigned getNumStores() const {
761 return LAI.getNumStores();
763 unsigned getNumLoads() const {
764 return LAI.getNumLoads();
766 unsigned getNumPredStores() const {
767 return NumPredStores;
770 /// Check if a single basic block loop is vectorizable.
771 /// At this point we know that this is a loop with a constant trip count
772 /// and we only need to check individual instructions.
773 bool canVectorizeInstrs();
775 /// When we vectorize loops we may change the order in which
776 /// we read and write from memory. This method checks if it is
777 /// legal to vectorize the code, considering only memory constrains.
778 /// Returns true if the loop is vectorizable
779 bool canVectorizeMemory();
781 /// Return true if we can vectorize this loop using the IF-conversion
783 bool canVectorizeWithIfConvert();
785 /// Collect the variables that need to stay uniform after vectorization.
786 void collectLoopUniforms();
788 /// Return true if all of the instructions in the block can be speculatively
789 /// executed. \p SafePtrs is a list of addresses that are known to be legal
790 /// and we know that we can read from them without segfault.
791 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
793 /// Returns True, if 'Phi' is the kind of reduction variable for type
794 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
795 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
796 /// Returns a struct describing if the instruction 'I' can be a reduction
797 /// variable of type 'Kind'. If the reduction is a min/max pattern of
798 /// select(icmp()) this function advances the instruction pointer 'I' from the
799 /// compare instruction to the select instruction and stores this pointer in
800 /// 'PatternLastInst' member of the returned struct.
801 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
802 ReductionInstDesc &Desc);
803 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
804 /// pattern corresponding to a min(X, Y) or max(X, Y).
805 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
806 ReductionInstDesc &Prev);
807 /// Returns the induction kind of Phi and record the step. This function may
808 /// return NoInduction if the PHI is not an induction variable.
809 InductionKind isInductionVariable(PHINode *Phi, ConstantInt *&StepValue);
811 /// \brief Collect memory access with loop invariant strides.
813 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
815 void collectStridedAccess(Value *LoadOrStoreInst);
817 /// Report an analysis message to assist the user in diagnosing loops that are
819 void emitAnalysis(VectorizationReport &Message) {
820 VectorizationReport::emitAnalysis(Message, TheFunction, TheLoop);
823 unsigned NumPredStores;
825 /// The loop that we evaluate.
829 /// DataLayout analysis.
830 const DataLayout *DL;
831 /// Target Library Info.
832 TargetLibraryInfo *TLI;
834 Function *TheFunction;
835 /// Target Transform Info
836 const TargetTransformInfo *TTI;
840 // --- vectorization state --- //
842 /// Holds the integer induction variable. This is the counter of the
845 /// Holds the reduction variables.
846 ReductionList Reductions;
847 /// Holds all of the induction variables that we found in the loop.
848 /// Notice that inductions don't need to start at zero and that induction
849 /// variables can be pointers.
850 InductionList Inductions;
851 /// Holds the widest induction type encountered.
854 /// Allowed outside users. This holds the reduction
855 /// vars which can be accessed from outside the loop.
856 SmallPtrSet<Value*, 4> AllowedExit;
857 /// This set holds the variables which are known to be uniform after
859 SmallPtrSet<Instruction*, 4> Uniforms;
861 /// Can we assume the absence of NaNs.
862 bool HasFunNoNaNAttr;
864 ValueToValueMap Strides;
865 SmallPtrSet<Value *, 8> StrideSet;
867 /// While vectorizing these instructions we have to generate a
868 /// call to the appropriate masked intrinsic
869 SmallPtrSet<const Instruction*, 8> MaskedOp;
872 /// LoopVectorizationCostModel - estimates the expected speedups due to
874 /// In many cases vectorization is not profitable. This can happen because of
875 /// a number of reasons. In this class we mainly attempt to predict the
876 /// expected speedup/slowdowns due to the supported instruction set. We use the
877 /// TargetTransformInfo to query the different backends for the cost of
878 /// different operations.
879 class LoopVectorizationCostModel {
881 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
882 LoopVectorizationLegality *Legal,
883 const TargetTransformInfo &TTI,
884 const DataLayout *DL, const TargetLibraryInfo *TLI,
885 AssumptionCache *AC, const Function *F,
886 const LoopVectorizeHints *Hints)
887 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI),
888 TheFunction(F), Hints(Hints) {
889 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
892 /// Information about vectorization costs
893 struct VectorizationFactor {
894 unsigned Width; // Vector width with best cost
895 unsigned Cost; // Cost of the loop with that width
897 /// \return The most profitable vectorization factor and the cost of that VF.
898 /// This method checks every power of two up to VF. If UserVF is not ZERO
899 /// then this vectorization factor will be selected if vectorization is
901 VectorizationFactor selectVectorizationFactor(bool OptForSize);
903 /// \return The size (in bits) of the widest type in the code that
904 /// needs to be vectorized. We ignore values that remain scalar such as
905 /// 64 bit loop indices.
906 unsigned getWidestType();
908 /// \return The most profitable unroll factor.
909 /// If UserUF is non-zero then this method finds the best unroll-factor
910 /// based on register pressure and other parameters.
911 /// VF and LoopCost are the selected vectorization factor and the cost of the
913 unsigned selectUnrollFactor(bool OptForSize, unsigned VF, unsigned LoopCost);
915 /// \brief A struct that represents some properties of the register usage
917 struct RegisterUsage {
918 /// Holds the number of loop invariant values that are used in the loop.
919 unsigned LoopInvariantRegs;
920 /// Holds the maximum number of concurrent live intervals in the loop.
921 unsigned MaxLocalUsers;
922 /// Holds the number of instructions in the loop.
923 unsigned NumInstructions;
926 /// \return information about the register usage of the loop.
927 RegisterUsage calculateRegisterUsage();
930 /// Returns the expected execution cost. The unit of the cost does
931 /// not matter because we use the 'cost' units to compare different
932 /// vector widths. The cost that is returned is *not* normalized by
933 /// the factor width.
934 unsigned expectedCost(unsigned VF);
936 /// Returns the execution time cost of an instruction for a given vector
937 /// width. Vector width of one means scalar.
938 unsigned getInstructionCost(Instruction *I, unsigned VF);
940 /// A helper function for converting Scalar types to vector types.
941 /// If the incoming type is void, we return void. If the VF is 1, we return
943 static Type* ToVectorTy(Type *Scalar, unsigned VF);
945 /// Returns whether the instruction is a load or store and will be a emitted
946 /// as a vector operation.
947 bool isConsecutiveLoadOrStore(Instruction *I);
949 /// Report an analysis message to assist the user in diagnosing loops that are
951 void emitAnalysis(VectorizationReport &Message) {
952 VectorizationReport::emitAnalysis(Message, TheFunction, TheLoop);
955 /// Values used only by @llvm.assume calls.
956 SmallPtrSet<const Value *, 32> EphValues;
958 /// The loop that we evaluate.
962 /// Loop Info analysis.
964 /// Vectorization legality.
965 LoopVectorizationLegality *Legal;
966 /// Vector target information.
967 const TargetTransformInfo &TTI;
968 /// Target data layout information.
969 const DataLayout *DL;
970 /// Target Library Info.
971 const TargetLibraryInfo *TLI;
972 const Function *TheFunction;
973 // Loop Vectorize Hint.
974 const LoopVectorizeHints *Hints;
977 /// Utility class for getting and setting loop vectorizer hints in the form
978 /// of loop metadata.
979 /// This class keeps a number of loop annotations locally (as member variables)
980 /// and can, upon request, write them back as metadata on the loop. It will
981 /// initially scan the loop for existing metadata, and will update the local
982 /// values based on information in the loop.
983 /// We cannot write all values to metadata, as the mere presence of some info,
984 /// for example 'force', means a decision has been made. So, we need to be
985 /// careful NOT to add them if the user hasn't specifically asked so.
986 class LoopVectorizeHints {
993 /// Hint - associates name and validation with the hint value.
996 unsigned Value; // This may have to change for non-numeric values.
999 Hint(const char * Name, unsigned Value, HintKind Kind)
1000 : Name(Name), Value(Value), Kind(Kind) { }
1002 bool validate(unsigned Val) {
1005 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
1007 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1015 /// Vectorization width.
1017 /// Vectorization interleave factor.
1019 /// Vectorization forced
1022 /// Return the loop metadata prefix.
1023 static StringRef Prefix() { return "llvm.loop."; }
1027 FK_Undefined = -1, ///< Not selected.
1028 FK_Disabled = 0, ///< Forcing disabled.
1029 FK_Enabled = 1, ///< Forcing enabled.
1032 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
1033 : Width("vectorize.width", VectorizerParams::VectorizationFactor,
1035 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1036 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1038 // Populate values with existing loop metadata.
1039 getHintsFromMetadata();
1041 // force-vector-interleave overrides DisableInterleaving.
1042 if (VectorizerParams::isInterleaveForced())
1043 Interleave.Value = VectorizerParams::VectorizationInterleave;
1045 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
1046 << "LV: Interleaving disabled by the pass manager\n");
1049 /// Mark the loop L as already vectorized by setting the width to 1.
1050 void setAlreadyVectorized() {
1051 Width.Value = Interleave.Value = 1;
1052 Hint Hints[] = {Width, Interleave};
1053 writeHintsToMetadata(Hints);
1056 /// Dumps all the hint information.
1057 std::string emitRemark() const {
1058 VectorizationReport R;
1059 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1060 R << "vectorization is explicitly disabled";
1062 R << "use -Rpass-analysis=loop-vectorize for more info";
1063 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1064 R << " (Force=true";
1065 if (Width.Value != 0)
1066 R << ", Vector Width=" << Width.Value;
1067 if (Interleave.Value != 0)
1068 R << ", Interleave Count=" << Interleave.Value;
1076 unsigned getWidth() const { return Width.Value; }
1077 unsigned getInterleave() const { return Interleave.Value; }
1078 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1081 /// Find hints specified in the loop metadata and update local values.
1082 void getHintsFromMetadata() {
1083 MDNode *LoopID = TheLoop->getLoopID();
1087 // First operand should refer to the loop id itself.
1088 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1089 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1091 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1092 const MDString *S = nullptr;
1093 SmallVector<Metadata *, 4> Args;
1095 // The expected hint is either a MDString or a MDNode with the first
1096 // operand a MDString.
1097 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1098 if (!MD || MD->getNumOperands() == 0)
1100 S = dyn_cast<MDString>(MD->getOperand(0));
1101 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1102 Args.push_back(MD->getOperand(i));
1104 S = dyn_cast<MDString>(LoopID->getOperand(i));
1105 assert(Args.size() == 0 && "too many arguments for MDString");
1111 // Check if the hint starts with the loop metadata prefix.
1112 StringRef Name = S->getString();
1113 if (Args.size() == 1)
1114 setHint(Name, Args[0]);
1118 /// Checks string hint with one operand and set value if valid.
1119 void setHint(StringRef Name, Metadata *Arg) {
1120 if (!Name.startswith(Prefix()))
1122 Name = Name.substr(Prefix().size(), StringRef::npos);
1124 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1126 unsigned Val = C->getZExtValue();
1128 Hint *Hints[] = {&Width, &Interleave, &Force};
1129 for (auto H : Hints) {
1130 if (Name == H->Name) {
1131 if (H->validate(Val))
1134 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1140 /// Create a new hint from name / value pair.
1141 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1142 LLVMContext &Context = TheLoop->getHeader()->getContext();
1143 Metadata *MDs[] = {MDString::get(Context, Name),
1144 ConstantAsMetadata::get(
1145 ConstantInt::get(Type::getInt32Ty(Context), V))};
1146 return MDNode::get(Context, MDs);
1149 /// Matches metadata with hint name.
1150 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1151 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1155 for (auto H : HintTypes)
1156 if (Name->getString().endswith(H.Name))
1161 /// Sets current hints into loop metadata, keeping other values intact.
1162 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1163 if (HintTypes.size() == 0)
1166 // Reserve the first element to LoopID (see below).
1167 SmallVector<Metadata *, 4> MDs(1);
1168 // If the loop already has metadata, then ignore the existing operands.
1169 MDNode *LoopID = TheLoop->getLoopID();
1171 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1172 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1173 // If node in update list, ignore old value.
1174 if (!matchesHintMetadataName(Node, HintTypes))
1175 MDs.push_back(Node);
1179 // Now, add the missing hints.
1180 for (auto H : HintTypes)
1181 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1183 // Replace current metadata node with new one.
1184 LLVMContext &Context = TheLoop->getHeader()->getContext();
1185 MDNode *NewLoopID = MDNode::get(Context, MDs);
1186 // Set operand 0 to refer to the loop id itself.
1187 NewLoopID->replaceOperandWith(0, NewLoopID);
1189 TheLoop->setLoopID(NewLoopID);
1192 /// The loop these hints belong to.
1193 const Loop *TheLoop;
1196 static void emitMissedWarning(Function *F, Loop *L,
1197 const LoopVectorizeHints &LH) {
1198 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1199 L->getStartLoc(), LH.emitRemark());
1201 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1202 if (LH.getWidth() != 1)
1203 emitLoopVectorizeWarning(
1204 F->getContext(), *F, L->getStartLoc(),
1205 "failed explicitly specified loop vectorization");
1206 else if (LH.getInterleave() != 1)
1207 emitLoopInterleaveWarning(
1208 F->getContext(), *F, L->getStartLoc(),
1209 "failed explicitly specified loop interleaving");
1213 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1215 return V.push_back(&L);
1217 for (Loop *InnerL : L)
1218 addInnerLoop(*InnerL, V);
1221 /// The LoopVectorize Pass.
1222 struct LoopVectorize : public FunctionPass {
1223 /// Pass identification, replacement for typeid
1226 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1228 DisableUnrolling(NoUnrolling),
1229 AlwaysVectorize(AlwaysVectorize) {
1230 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1233 ScalarEvolution *SE;
1234 const DataLayout *DL;
1236 TargetTransformInfo *TTI;
1238 BlockFrequencyInfo *BFI;
1239 TargetLibraryInfo *TLI;
1241 AssumptionCache *AC;
1242 bool DisableUnrolling;
1243 bool AlwaysVectorize;
1245 BlockFrequency ColdEntryFreq;
1247 bool runOnFunction(Function &F) override {
1248 SE = &getAnalysis<ScalarEvolution>();
1249 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1250 DL = DLP ? &DLP->getDataLayout() : nullptr;
1251 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1252 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1253 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1254 BFI = &getAnalysis<BlockFrequencyInfo>();
1255 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1256 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1257 AA = &getAnalysis<AliasAnalysis>();
1258 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1260 // Compute some weights outside of the loop over the loops. Compute this
1261 // using a BranchProbability to re-use its scaling math.
1262 const BranchProbability ColdProb(1, 5); // 20%
1263 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1265 // If the target claims to have no vector registers don't attempt
1267 if (!TTI->getNumberOfRegisters(true))
1271 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1272 << ": Missing data layout\n");
1276 // Build up a worklist of inner-loops to vectorize. This is necessary as
1277 // the act of vectorizing or partially unrolling a loop creates new loops
1278 // and can invalidate iterators across the loops.
1279 SmallVector<Loop *, 8> Worklist;
1282 addInnerLoop(*L, Worklist);
1284 LoopsAnalyzed += Worklist.size();
1286 // Now walk the identified inner loops.
1287 bool Changed = false;
1288 while (!Worklist.empty())
1289 Changed |= processLoop(Worklist.pop_back_val());
1291 // Process each loop nest in the function.
1295 bool processLoop(Loop *L) {
1296 assert(L->empty() && "Only process inner loops.");
1299 const std::string DebugLocStr = getDebugLocString(L);
1302 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1303 << L->getHeader()->getParent()->getName() << "\" from "
1304 << DebugLocStr << "\n");
1306 LoopVectorizeHints Hints(L, DisableUnrolling);
1308 DEBUG(dbgs() << "LV: Loop hints:"
1310 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1312 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1314 : "?")) << " width=" << Hints.getWidth()
1315 << " unroll=" << Hints.getInterleave() << "\n");
1317 // Function containing loop
1318 Function *F = L->getHeader()->getParent();
1320 // Looking at the diagnostic output is the only way to determine if a loop
1321 // was vectorized (other than looking at the IR or machine code), so it
1322 // is important to generate an optimization remark for each loop. Most of
1323 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1324 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1325 // less verbose reporting vectorized loops and unvectorized loops that may
1326 // benefit from vectorization, respectively.
1328 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1329 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1330 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1331 L->getStartLoc(), Hints.emitRemark());
1335 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1336 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1337 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1338 L->getStartLoc(), Hints.emitRemark());
1342 if (Hints.getWidth() == 1 && Hints.getInterleave() == 1) {
1343 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1344 emitOptimizationRemarkAnalysis(
1345 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1346 "loop not vectorized: vector width and interleave count are "
1347 "explicitly set to 1");
1351 // Check the loop for a trip count threshold:
1352 // do not vectorize loops with a tiny trip count.
1353 const unsigned TC = SE->getSmallConstantTripCount(L);
1354 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1355 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1356 << "This loop is not worth vectorizing.");
1357 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1358 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1360 DEBUG(dbgs() << "\n");
1361 emitOptimizationRemarkAnalysis(
1362 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1363 "vectorization is not beneficial and is not explicitly forced");
1368 // Check if it is legal to vectorize the loop.
1369 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, AA, F, TTI);
1370 if (!LVL.canVectorize()) {
1371 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1372 emitMissedWarning(F, L, Hints);
1376 // Use the cost model.
1377 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI, AC, F,
1380 // Check the function attributes to find out if this function should be
1381 // optimized for size.
1382 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1383 F->hasFnAttribute(Attribute::OptimizeForSize);
1385 // Compute the weighted frequency of this loop being executed and see if it
1386 // is less than 20% of the function entry baseline frequency. Note that we
1387 // always have a canonical loop here because we think we *can* vectoriez.
1388 // FIXME: This is hidden behind a flag due to pervasive problems with
1389 // exactly what block frequency models.
1390 if (LoopVectorizeWithBlockFrequency) {
1391 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1392 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1393 LoopEntryFreq < ColdEntryFreq)
1397 // Check the function attributes to see if implicit floats are allowed.a
1398 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1399 // an integer loop and the vector instructions selected are purely integer
1400 // vector instructions?
1401 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1402 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1403 "attribute is used.\n");
1404 emitOptimizationRemarkAnalysis(
1405 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1406 "loop not vectorized due to NoImplicitFloat attribute");
1407 emitMissedWarning(F, L, Hints);
1411 // Select the optimal vectorization factor.
1412 const LoopVectorizationCostModel::VectorizationFactor VF =
1413 CM.selectVectorizationFactor(OptForSize);
1415 // Select the unroll factor.
1417 CM.selectUnrollFactor(OptForSize, VF.Width, VF.Cost);
1419 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1420 << DebugLocStr << '\n');
1421 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1423 if (VF.Width == 1) {
1424 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
1427 emitOptimizationRemarkAnalysis(
1428 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1429 "not beneficial to vectorize and user disabled interleaving");
1432 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1434 // Report the unrolling decision.
1435 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1436 Twine("unrolled with interleaving factor " +
1438 " (vectorization not beneficial)"));
1440 // We decided not to vectorize, but we may want to unroll.
1442 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1443 Unroller.vectorize(&LVL);
1445 // If we decided that it is *legal* to vectorize the loop then do it.
1446 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1450 // Report the vectorization decision.
1451 emitOptimizationRemark(
1452 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1453 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1454 ", unrolling interleave factor: " + Twine(UF) + ")");
1457 // Mark the loop as already vectorized to avoid vectorizing again.
1458 Hints.setAlreadyVectorized();
1460 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1464 void getAnalysisUsage(AnalysisUsage &AU) const override {
1465 AU.addRequired<AssumptionCacheTracker>();
1466 AU.addRequiredID(LoopSimplifyID);
1467 AU.addRequiredID(LCSSAID);
1468 AU.addRequired<BlockFrequencyInfo>();
1469 AU.addRequired<DominatorTreeWrapperPass>();
1470 AU.addRequired<LoopInfoWrapperPass>();
1471 AU.addRequired<ScalarEvolution>();
1472 AU.addRequired<TargetTransformInfoWrapperPass>();
1473 AU.addRequired<AliasAnalysis>();
1474 AU.addPreserved<LoopInfoWrapperPass>();
1475 AU.addPreserved<DominatorTreeWrapperPass>();
1476 AU.addPreserved<AliasAnalysis>();
1481 } // end anonymous namespace
1483 //===----------------------------------------------------------------------===//
1484 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1485 // LoopVectorizationCostModel.
1486 //===----------------------------------------------------------------------===//
1488 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1489 // We need to place the broadcast of invariant variables outside the loop.
1490 Instruction *Instr = dyn_cast<Instruction>(V);
1492 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1493 Instr->getParent()) != LoopVectorBody.end());
1494 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1496 // Place the code for broadcasting invariant variables in the new preheader.
1497 IRBuilder<>::InsertPointGuard Guard(Builder);
1499 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1501 // Broadcast the scalar into all locations in the vector.
1502 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1507 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
1509 assert(Val->getType()->isVectorTy() && "Must be a vector");
1510 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1511 "Elem must be an integer");
1512 assert(Step->getType() == Val->getType()->getScalarType() &&
1513 "Step has wrong type");
1514 // Create the types.
1515 Type *ITy = Val->getType()->getScalarType();
1516 VectorType *Ty = cast<VectorType>(Val->getType());
1517 int VLen = Ty->getNumElements();
1518 SmallVector<Constant*, 8> Indices;
1520 // Create a vector of consecutive numbers from zero to VF.
1521 for (int i = 0; i < VLen; ++i)
1522 Indices.push_back(ConstantInt::get(ITy, StartIdx + i));
1524 // Add the consecutive indices to the vector value.
1525 Constant *Cv = ConstantVector::get(Indices);
1526 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1527 Step = Builder.CreateVectorSplat(VLen, Step);
1528 assert(Step->getType() == Val->getType() && "Invalid step vec");
1529 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
1530 // which can be found from the original scalar operations.
1531 Step = Builder.CreateMul(Cv, Step);
1532 return Builder.CreateAdd(Val, Step, "induction");
1535 /// \brief Find the operand of the GEP that should be checked for consecutive
1536 /// stores. This ignores trailing indices that have no effect on the final
1538 static unsigned getGEPInductionOperand(const DataLayout *DL,
1539 const GetElementPtrInst *Gep) {
1540 unsigned LastOperand = Gep->getNumOperands() - 1;
1541 unsigned GEPAllocSize = DL->getTypeAllocSize(
1542 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1544 // Walk backwards and try to peel off zeros.
1545 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1546 // Find the type we're currently indexing into.
1547 gep_type_iterator GEPTI = gep_type_begin(Gep);
1548 std::advance(GEPTI, LastOperand - 1);
1550 // If it's a type with the same allocation size as the result of the GEP we
1551 // can peel off the zero index.
1552 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1560 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1561 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1562 // Make sure that the pointer does not point to structs.
1563 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1566 // If this value is a pointer induction variable we know it is consecutive.
1567 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1568 if (Phi && Inductions.count(Phi)) {
1569 InductionInfo II = Inductions[Phi];
1570 return II.getConsecutiveDirection();
1573 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1577 unsigned NumOperands = Gep->getNumOperands();
1578 Value *GpPtr = Gep->getPointerOperand();
1579 // If this GEP value is a consecutive pointer induction variable and all of
1580 // the indices are constant then we know it is consecutive. We can
1581 Phi = dyn_cast<PHINode>(GpPtr);
1582 if (Phi && Inductions.count(Phi)) {
1584 // Make sure that the pointer does not point to structs.
1585 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1586 if (GepPtrType->getElementType()->isAggregateType())
1589 // Make sure that all of the index operands are loop invariant.
1590 for (unsigned i = 1; i < NumOperands; ++i)
1591 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1594 InductionInfo II = Inductions[Phi];
1595 return II.getConsecutiveDirection();
1598 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1600 // Check that all of the gep indices are uniform except for our induction
1602 for (unsigned i = 0; i != NumOperands; ++i)
1603 if (i != InductionOperand &&
1604 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1607 // We can emit wide load/stores only if the last non-zero index is the
1608 // induction variable.
1609 const SCEV *Last = nullptr;
1610 if (!Strides.count(Gep))
1611 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1613 // Because of the multiplication by a stride we can have a s/zext cast.
1614 // We are going to replace this stride by 1 so the cast is safe to ignore.
1616 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1617 // %0 = trunc i64 %indvars.iv to i32
1618 // %mul = mul i32 %0, %Stride1
1619 // %idxprom = zext i32 %mul to i64 << Safe cast.
1620 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1622 Last = replaceSymbolicStrideSCEV(SE, Strides,
1623 Gep->getOperand(InductionOperand), Gep);
1624 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1626 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1630 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1631 const SCEV *Step = AR->getStepRecurrence(*SE);
1633 // The memory is consecutive because the last index is consecutive
1634 // and all other indices are loop invariant.
1637 if (Step->isAllOnesValue())
1644 bool LoopVectorizationLegality::isUniform(Value *V) {
1645 return LAI.isUniform(V);
1648 InnerLoopVectorizer::VectorParts&
1649 InnerLoopVectorizer::getVectorValue(Value *V) {
1650 assert(V != Induction && "The new induction variable should not be used.");
1651 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1653 // If we have a stride that is replaced by one, do it here.
1654 if (Legal->hasStride(V))
1655 V = ConstantInt::get(V->getType(), 1);
1657 // If we have this scalar in the map, return it.
1658 if (WidenMap.has(V))
1659 return WidenMap.get(V);
1661 // If this scalar is unknown, assume that it is a constant or that it is
1662 // loop invariant. Broadcast V and save the value for future uses.
1663 Value *B = getBroadcastInstrs(V);
1664 return WidenMap.splat(V, B);
1667 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1668 assert(Vec->getType()->isVectorTy() && "Invalid type");
1669 SmallVector<Constant*, 8> ShuffleMask;
1670 for (unsigned i = 0; i < VF; ++i)
1671 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1673 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1674 ConstantVector::get(ShuffleMask),
1678 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1679 // Attempt to issue a wide load.
1680 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1681 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1683 assert((LI || SI) && "Invalid Load/Store instruction");
1685 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1686 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1687 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1688 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1689 // An alignment of 0 means target abi alignment. We need to use the scalar's
1690 // target abi alignment in such a case.
1692 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1693 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1694 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1695 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1697 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
1698 !Legal->isMaskRequired(SI))
1699 return scalarizeInstruction(Instr, true);
1701 if (ScalarAllocatedSize != VectorElementSize)
1702 return scalarizeInstruction(Instr);
1704 // If the pointer is loop invariant or if it is non-consecutive,
1705 // scalarize the load.
1706 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1707 bool Reverse = ConsecutiveStride < 0;
1708 bool UniformLoad = LI && Legal->isUniform(Ptr);
1709 if (!ConsecutiveStride || UniformLoad)
1710 return scalarizeInstruction(Instr);
1712 Constant *Zero = Builder.getInt32(0);
1713 VectorParts &Entry = WidenMap.get(Instr);
1715 // Handle consecutive loads/stores.
1716 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1717 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1718 setDebugLocFromInst(Builder, Gep);
1719 Value *PtrOperand = Gep->getPointerOperand();
1720 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1721 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1723 // Create the new GEP with the new induction variable.
1724 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1725 Gep2->setOperand(0, FirstBasePtr);
1726 Gep2->setName("gep.indvar.base");
1727 Ptr = Builder.Insert(Gep2);
1729 setDebugLocFromInst(Builder, Gep);
1730 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1731 OrigLoop) && "Base ptr must be invariant");
1733 // The last index does not have to be the induction. It can be
1734 // consecutive and be a function of the index. For example A[I+1];
1735 unsigned NumOperands = Gep->getNumOperands();
1736 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1737 // Create the new GEP with the new induction variable.
1738 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1740 for (unsigned i = 0; i < NumOperands; ++i) {
1741 Value *GepOperand = Gep->getOperand(i);
1742 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1744 // Update last index or loop invariant instruction anchored in loop.
1745 if (i == InductionOperand ||
1746 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1747 assert((i == InductionOperand ||
1748 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1749 "Must be last index or loop invariant");
1751 VectorParts &GEPParts = getVectorValue(GepOperand);
1752 Value *Index = GEPParts[0];
1753 Index = Builder.CreateExtractElement(Index, Zero);
1754 Gep2->setOperand(i, Index);
1755 Gep2->setName("gep.indvar.idx");
1758 Ptr = Builder.Insert(Gep2);
1760 // Use the induction element ptr.
1761 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1762 setDebugLocFromInst(Builder, Ptr);
1763 VectorParts &PtrVal = getVectorValue(Ptr);
1764 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1767 VectorParts Mask = createBlockInMask(Instr->getParent());
1770 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1771 "We do not allow storing to uniform addresses");
1772 setDebugLocFromInst(Builder, SI);
1773 // We don't want to update the value in the map as it might be used in
1774 // another expression. So don't use a reference type for "StoredVal".
1775 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1777 for (unsigned Part = 0; Part < UF; ++Part) {
1778 // Calculate the pointer for the specific unroll-part.
1779 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1782 // If we store to reverse consecutive memory locations then we need
1783 // to reverse the order of elements in the stored value.
1784 StoredVal[Part] = reverseVector(StoredVal[Part]);
1785 // If the address is consecutive but reversed, then the
1786 // wide store needs to start at the last vector element.
1787 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1788 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1789 Mask[Part] = reverseVector(Mask[Part]);
1792 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1793 DataTy->getPointerTo(AddressSpace));
1796 if (Legal->isMaskRequired(SI))
1797 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
1800 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
1801 propagateMetadata(NewSI, SI);
1807 assert(LI && "Must have a load instruction");
1808 setDebugLocFromInst(Builder, LI);
1809 for (unsigned Part = 0; Part < UF; ++Part) {
1810 // Calculate the pointer for the specific unroll-part.
1811 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1814 // If the address is consecutive but reversed, then the
1815 // wide load needs to start at the last vector element.
1816 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1817 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1818 Mask[Part] = reverseVector(Mask[Part]);
1822 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1823 DataTy->getPointerTo(AddressSpace));
1824 if (Legal->isMaskRequired(LI))
1825 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
1826 UndefValue::get(DataTy),
1827 "wide.masked.load");
1829 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
1830 propagateMetadata(NewLI, LI);
1831 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
1835 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1836 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1837 // Holds vector parameters or scalars, in case of uniform vals.
1838 SmallVector<VectorParts, 4> Params;
1840 setDebugLocFromInst(Builder, Instr);
1842 // Find all of the vectorized parameters.
1843 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1844 Value *SrcOp = Instr->getOperand(op);
1846 // If we are accessing the old induction variable, use the new one.
1847 if (SrcOp == OldInduction) {
1848 Params.push_back(getVectorValue(SrcOp));
1852 // Try using previously calculated values.
1853 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1855 // If the src is an instruction that appeared earlier in the basic block
1856 // then it should already be vectorized.
1857 if (SrcInst && OrigLoop->contains(SrcInst)) {
1858 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1859 // The parameter is a vector value from earlier.
1860 Params.push_back(WidenMap.get(SrcInst));
1862 // The parameter is a scalar from outside the loop. Maybe even a constant.
1863 VectorParts Scalars;
1864 Scalars.append(UF, SrcOp);
1865 Params.push_back(Scalars);
1869 assert(Params.size() == Instr->getNumOperands() &&
1870 "Invalid number of operands");
1872 // Does this instruction return a value ?
1873 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1875 Value *UndefVec = IsVoidRetTy ? nullptr :
1876 UndefValue::get(VectorType::get(Instr->getType(), VF));
1877 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1878 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1880 Instruction *InsertPt = Builder.GetInsertPoint();
1881 BasicBlock *IfBlock = Builder.GetInsertBlock();
1882 BasicBlock *CondBlock = nullptr;
1885 Loop *VectorLp = nullptr;
1886 if (IfPredicateStore) {
1887 assert(Instr->getParent()->getSinglePredecessor() &&
1888 "Only support single predecessor blocks");
1889 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1890 Instr->getParent());
1891 VectorLp = LI->getLoopFor(IfBlock);
1892 assert(VectorLp && "Must have a loop for this block");
1895 // For each vector unroll 'part':
1896 for (unsigned Part = 0; Part < UF; ++Part) {
1897 // For each scalar that we create:
1898 for (unsigned Width = 0; Width < VF; ++Width) {
1901 Value *Cmp = nullptr;
1902 if (IfPredicateStore) {
1903 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1904 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1905 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1906 LoopVectorBody.push_back(CondBlock);
1907 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
1908 // Update Builder with newly created basic block.
1909 Builder.SetInsertPoint(InsertPt);
1912 Instruction *Cloned = Instr->clone();
1914 Cloned->setName(Instr->getName() + ".cloned");
1915 // Replace the operands of the cloned instructions with extracted scalars.
1916 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1917 Value *Op = Params[op][Part];
1918 // Param is a vector. Need to extract the right lane.
1919 if (Op->getType()->isVectorTy())
1920 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1921 Cloned->setOperand(op, Op);
1924 // Place the cloned scalar in the new loop.
1925 Builder.Insert(Cloned);
1927 // If the original scalar returns a value we need to place it in a vector
1928 // so that future users will be able to use it.
1930 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1931 Builder.getInt32(Width));
1933 if (IfPredicateStore) {
1934 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1935 LoopVectorBody.push_back(NewIfBlock);
1936 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
1937 Builder.SetInsertPoint(InsertPt);
1938 Instruction *OldBr = IfBlock->getTerminator();
1939 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1940 OldBr->eraseFromParent();
1941 IfBlock = NewIfBlock;
1947 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1951 if (Instruction *I = dyn_cast<Instruction>(V))
1952 return I->getParent() == Loc->getParent() ? I : nullptr;
1956 std::pair<Instruction *, Instruction *>
1957 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
1958 Instruction *tnullptr = nullptr;
1959 if (!Legal->mustCheckStrides())
1960 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1962 IRBuilder<> ChkBuilder(Loc);
1965 Value *Check = nullptr;
1966 Instruction *FirstInst = nullptr;
1967 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
1968 SE = Legal->strides_end();
1970 Value *Ptr = stripIntegerCast(*SI);
1971 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
1973 // Store the first instruction we create.
1974 FirstInst = getFirstInst(FirstInst, C, Loc);
1976 Check = ChkBuilder.CreateOr(Check, C);
1981 // We have to do this trickery because the IRBuilder might fold the check to a
1982 // constant expression in which case there is no Instruction anchored in a
1984 LLVMContext &Ctx = Loc->getContext();
1985 Instruction *TheCheck =
1986 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
1987 ChkBuilder.Insert(TheCheck, "stride.not.one");
1988 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
1990 return std::make_pair(FirstInst, TheCheck);
1993 void InnerLoopVectorizer::createEmptyLoop() {
1995 In this function we generate a new loop. The new loop will contain
1996 the vectorized instructions while the old loop will continue to run the
1999 [ ] <-- Back-edge taken count overflow check.
2002 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2005 || [ ] <-- vector pre header.
2009 || [ ]_| <-- vector loop.
2012 | >[ ] <--- middle-block.
2015 -|- >[ ] <--- new preheader.
2019 | [ ]_| <-- old scalar loop to handle remainder.
2022 >[ ] <-- exit block.
2026 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2027 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2028 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2029 assert(BypassBlock && "Invalid loop structure");
2030 assert(ExitBlock && "Must have an exit block");
2032 // Some loops have a single integer induction variable, while other loops
2033 // don't. One example is c++ iterators that often have multiple pointer
2034 // induction variables. In the code below we also support a case where we
2035 // don't have a single induction variable.
2036 OldInduction = Legal->getInduction();
2037 Type *IdxTy = Legal->getWidestInductionType();
2039 // Find the loop boundaries.
2040 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2041 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2043 // The exit count might have the type of i64 while the phi is i32. This can
2044 // happen if we have an induction variable that is sign extended before the
2045 // compare. The only way that we get a backedge taken count is that the
2046 // induction variable was signed and as such will not overflow. In such a case
2047 // truncation is legal.
2048 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2049 IdxTy->getPrimitiveSizeInBits())
2050 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2052 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2053 // Get the total trip count from the count by adding 1.
2054 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2055 SE->getConstant(BackedgeTakeCount->getType(), 1));
2057 // Expand the trip count and place the new instructions in the preheader.
2058 // Notice that the pre-header does not change, only the loop body.
2059 SCEVExpander Exp(*SE, "induction");
2061 // We need to test whether the backedge-taken count is uint##_max. Adding one
2062 // to it will cause overflow and an incorrect loop trip count in the vector
2063 // body. In case of overflow we want to directly jump to the scalar remainder
2065 Value *BackedgeCount =
2066 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2067 BypassBlock->getTerminator());
2068 if (BackedgeCount->getType()->isPointerTy())
2069 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2070 "backedge.ptrcnt.to.int",
2071 BypassBlock->getTerminator());
2072 Instruction *CheckBCOverflow =
2073 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2074 Constant::getAllOnesValue(BackedgeCount->getType()),
2075 "backedge.overflow", BypassBlock->getTerminator());
2077 // The loop index does not have to start at Zero. Find the original start
2078 // value from the induction PHI node. If we don't have an induction variable
2079 // then we know that it starts at zero.
2080 Builder.SetInsertPoint(BypassBlock->getTerminator());
2081 Value *StartIdx = ExtendedIdx = OldInduction ?
2082 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2084 ConstantInt::get(IdxTy, 0);
2086 // We need an instruction to anchor the overflow check on. StartIdx needs to
2087 // be defined before the overflow check branch. Because the scalar preheader
2088 // is going to merge the start index and so the overflow branch block needs to
2089 // contain a definition of the start index.
2090 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2091 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2092 BypassBlock->getTerminator());
2094 // Count holds the overall loop count (N).
2095 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2096 BypassBlock->getTerminator());
2098 LoopBypassBlocks.push_back(BypassBlock);
2100 // Split the single block loop into the two loop structure described above.
2101 BasicBlock *VectorPH =
2102 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2103 BasicBlock *VecBody =
2104 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2105 BasicBlock *MiddleBlock =
2106 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2107 BasicBlock *ScalarPH =
2108 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2110 // Create and register the new vector loop.
2111 Loop* Lp = new Loop();
2112 Loop *ParentLoop = OrigLoop->getParentLoop();
2114 // Insert the new loop into the loop nest and register the new basic blocks
2115 // before calling any utilities such as SCEV that require valid LoopInfo.
2117 ParentLoop->addChildLoop(Lp);
2118 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
2119 ParentLoop->addBasicBlockToLoop(VectorPH, *LI);
2120 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
2122 LI->addTopLevelLoop(Lp);
2124 Lp->addBasicBlockToLoop(VecBody, *LI);
2126 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2128 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2130 // Generate the induction variable.
2131 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2132 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2133 // The loop step is equal to the vectorization factor (num of SIMD elements)
2134 // times the unroll factor (num of SIMD instructions).
2135 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2137 // This is the IR builder that we use to add all of the logic for bypassing
2138 // the new vector loop.
2139 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2140 setDebugLocFromInst(BypassBuilder,
2141 getDebugLocFromInstOrOperands(OldInduction));
2143 // We may need to extend the index in case there is a type mismatch.
2144 // We know that the count starts at zero and does not overflow.
2145 if (Count->getType() != IdxTy) {
2146 // The exit count can be of pointer type. Convert it to the correct
2148 if (ExitCount->getType()->isPointerTy())
2149 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2151 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2154 // Add the start index to the loop count to get the new end index.
2155 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2157 // Now we need to generate the expression for N - (N % VF), which is
2158 // the part that the vectorized body will execute.
2159 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2160 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2161 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2162 "end.idx.rnd.down");
2164 // Now, compare the new count to zero. If it is zero skip the vector loop and
2165 // jump to the scalar loop.
2167 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2169 BasicBlock *LastBypassBlock = BypassBlock;
2171 // Generate code to check that the loops trip count that we computed by adding
2172 // one to the backedge-taken count will not overflow.
2174 auto PastOverflowCheck =
2175 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2176 BasicBlock *CheckBlock =
2177 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2179 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2180 LoopBypassBlocks.push_back(CheckBlock);
2181 Instruction *OldTerm = LastBypassBlock->getTerminator();
2182 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2183 OldTerm->eraseFromParent();
2184 LastBypassBlock = CheckBlock;
2187 // Generate the code to check that the strides we assumed to be one are really
2188 // one. We want the new basic block to start at the first instruction in a
2189 // sequence of instructions that form a check.
2190 Instruction *StrideCheck;
2191 Instruction *FirstCheckInst;
2192 std::tie(FirstCheckInst, StrideCheck) =
2193 addStrideCheck(LastBypassBlock->getTerminator());
2195 // Create a new block containing the stride check.
2196 BasicBlock *CheckBlock =
2197 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2199 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2200 LoopBypassBlocks.push_back(CheckBlock);
2202 // Replace the branch into the memory check block with a conditional branch
2203 // for the "few elements case".
2204 Instruction *OldTerm = LastBypassBlock->getTerminator();
2205 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2206 OldTerm->eraseFromParent();
2209 LastBypassBlock = CheckBlock;
2212 // Generate the code that checks in runtime if arrays overlap. We put the
2213 // checks into a separate block to make the more common case of few elements
2215 Instruction *MemRuntimeCheck;
2216 std::tie(FirstCheckInst, MemRuntimeCheck) =
2217 Legal->getLAI()->addRuntimeCheck(LastBypassBlock->getTerminator());
2218 if (MemRuntimeCheck) {
2219 // Create a new block containing the memory check.
2220 BasicBlock *CheckBlock =
2221 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.memcheck");
2223 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2224 LoopBypassBlocks.push_back(CheckBlock);
2226 // Replace the branch into the memory check block with a conditional branch
2227 // for the "few elements case".
2228 Instruction *OldTerm = LastBypassBlock->getTerminator();
2229 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2230 OldTerm->eraseFromParent();
2232 Cmp = MemRuntimeCheck;
2233 LastBypassBlock = CheckBlock;
2236 LastBypassBlock->getTerminator()->eraseFromParent();
2237 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2240 // We are going to resume the execution of the scalar loop.
2241 // Go over all of the induction variables that we found and fix the
2242 // PHIs that are left in the scalar version of the loop.
2243 // The starting values of PHI nodes depend on the counter of the last
2244 // iteration in the vectorized loop.
2245 // If we come from a bypass edge then we need to start from the original
2248 // This variable saves the new starting index for the scalar loop.
2249 PHINode *ResumeIndex = nullptr;
2250 LoopVectorizationLegality::InductionList::iterator I, E;
2251 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2252 // Set builder to point to last bypass block.
2253 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2254 for (I = List->begin(), E = List->end(); I != E; ++I) {
2255 PHINode *OrigPhi = I->first;
2256 LoopVectorizationLegality::InductionInfo II = I->second;
2258 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2259 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2260 MiddleBlock->getTerminator());
2261 // We might have extended the type of the induction variable but we need a
2262 // truncated version for the scalar loop.
2263 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2264 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2265 MiddleBlock->getTerminator()) : nullptr;
2267 // Create phi nodes to merge from the backedge-taken check block.
2268 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2269 ScalarPH->getTerminator());
2270 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2272 PHINode *BCTruncResumeVal = nullptr;
2273 if (OrigPhi == OldInduction) {
2275 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2276 ScalarPH->getTerminator());
2277 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2280 Value *EndValue = nullptr;
2282 case LoopVectorizationLegality::IK_NoInduction:
2283 llvm_unreachable("Unknown induction");
2284 case LoopVectorizationLegality::IK_IntInduction: {
2285 // Handle the integer induction counter.
2286 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2288 // We have the canonical induction variable.
2289 if (OrigPhi == OldInduction) {
2290 // Create a truncated version of the resume value for the scalar loop,
2291 // we might have promoted the type to a larger width.
2293 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2294 // The new PHI merges the original incoming value, in case of a bypass,
2295 // or the value at the end of the vectorized loop.
2296 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2297 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2298 TruncResumeVal->addIncoming(EndValue, VecBody);
2300 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2302 // We know what the end value is.
2303 EndValue = IdxEndRoundDown;
2304 // We also know which PHI node holds it.
2305 ResumeIndex = ResumeVal;
2309 // Not the canonical induction variable - add the vector loop count to the
2311 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2312 II.StartValue->getType(),
2314 EndValue = II.transform(BypassBuilder, CRD);
2315 EndValue->setName("ind.end");
2318 case LoopVectorizationLegality::IK_PtrInduction: {
2319 EndValue = II.transform(BypassBuilder, CountRoundDown);
2320 EndValue->setName("ptr.ind.end");
2325 // The new PHI merges the original incoming value, in case of a bypass,
2326 // or the value at the end of the vectorized loop.
2327 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2328 if (OrigPhi == OldInduction)
2329 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2331 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2333 ResumeVal->addIncoming(EndValue, VecBody);
2335 // Fix the scalar body counter (PHI node).
2336 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2338 // The old induction's phi node in the scalar body needs the truncated
2340 if (OrigPhi == OldInduction) {
2341 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2342 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2344 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2345 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2349 // If we are generating a new induction variable then we also need to
2350 // generate the code that calculates the exit value. This value is not
2351 // simply the end of the counter because we may skip the vectorized body
2352 // in case of a runtime check.
2354 assert(!ResumeIndex && "Unexpected resume value found");
2355 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2356 MiddleBlock->getTerminator());
2357 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2358 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2359 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2362 // Make sure that we found the index where scalar loop needs to continue.
2363 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2364 "Invalid resume Index");
2366 // Add a check in the middle block to see if we have completed
2367 // all of the iterations in the first vector loop.
2368 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2369 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2370 ResumeIndex, "cmp.n",
2371 MiddleBlock->getTerminator());
2373 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2374 // Remove the old terminator.
2375 MiddleBlock->getTerminator()->eraseFromParent();
2377 // Create i+1 and fill the PHINode.
2378 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2379 Induction->addIncoming(StartIdx, VectorPH);
2380 Induction->addIncoming(NextIdx, VecBody);
2381 // Create the compare.
2382 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2383 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2385 // Now we have two terminators. Remove the old one from the block.
2386 VecBody->getTerminator()->eraseFromParent();
2388 // Get ready to start creating new instructions into the vectorized body.
2389 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2392 LoopVectorPreHeader = VectorPH;
2393 LoopScalarPreHeader = ScalarPH;
2394 LoopMiddleBlock = MiddleBlock;
2395 LoopExitBlock = ExitBlock;
2396 LoopVectorBody.push_back(VecBody);
2397 LoopScalarBody = OldBasicBlock;
2399 LoopVectorizeHints Hints(Lp, true);
2400 Hints.setAlreadyVectorized();
2403 /// This function returns the identity element (or neutral element) for
2404 /// the operation K.
2406 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2411 // Adding, Xoring, Oring zero to a number does not change it.
2412 return ConstantInt::get(Tp, 0);
2413 case RK_IntegerMult:
2414 // Multiplying a number by 1 does not change it.
2415 return ConstantInt::get(Tp, 1);
2417 // AND-ing a number with an all-1 value does not change it.
2418 return ConstantInt::get(Tp, -1, true);
2420 // Multiplying a number by 1 does not change it.
2421 return ConstantFP::get(Tp, 1.0L);
2423 // Adding zero to a number does not change it.
2424 return ConstantFP::get(Tp, 0.0L);
2426 llvm_unreachable("Unknown reduction kind");
2430 /// This function translates the reduction kind to an LLVM binary operator.
2432 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2434 case LoopVectorizationLegality::RK_IntegerAdd:
2435 return Instruction::Add;
2436 case LoopVectorizationLegality::RK_IntegerMult:
2437 return Instruction::Mul;
2438 case LoopVectorizationLegality::RK_IntegerOr:
2439 return Instruction::Or;
2440 case LoopVectorizationLegality::RK_IntegerAnd:
2441 return Instruction::And;
2442 case LoopVectorizationLegality::RK_IntegerXor:
2443 return Instruction::Xor;
2444 case LoopVectorizationLegality::RK_FloatMult:
2445 return Instruction::FMul;
2446 case LoopVectorizationLegality::RK_FloatAdd:
2447 return Instruction::FAdd;
2448 case LoopVectorizationLegality::RK_IntegerMinMax:
2449 return Instruction::ICmp;
2450 case LoopVectorizationLegality::RK_FloatMinMax:
2451 return Instruction::FCmp;
2453 llvm_unreachable("Unknown reduction operation");
2457 Value *createMinMaxOp(IRBuilder<> &Builder,
2458 LoopVectorizationLegality::MinMaxReductionKind RK,
2461 CmpInst::Predicate P = CmpInst::ICMP_NE;
2464 llvm_unreachable("Unknown min/max reduction kind");
2465 case LoopVectorizationLegality::MRK_UIntMin:
2466 P = CmpInst::ICMP_ULT;
2468 case LoopVectorizationLegality::MRK_UIntMax:
2469 P = CmpInst::ICMP_UGT;
2471 case LoopVectorizationLegality::MRK_SIntMin:
2472 P = CmpInst::ICMP_SLT;
2474 case LoopVectorizationLegality::MRK_SIntMax:
2475 P = CmpInst::ICMP_SGT;
2477 case LoopVectorizationLegality::MRK_FloatMin:
2478 P = CmpInst::FCMP_OLT;
2480 case LoopVectorizationLegality::MRK_FloatMax:
2481 P = CmpInst::FCMP_OGT;
2486 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2487 RK == LoopVectorizationLegality::MRK_FloatMax)
2488 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2490 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2492 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2497 struct CSEDenseMapInfo {
2498 static bool canHandle(Instruction *I) {
2499 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2500 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2502 static inline Instruction *getEmptyKey() {
2503 return DenseMapInfo<Instruction *>::getEmptyKey();
2505 static inline Instruction *getTombstoneKey() {
2506 return DenseMapInfo<Instruction *>::getTombstoneKey();
2508 static unsigned getHashValue(Instruction *I) {
2509 assert(canHandle(I) && "Unknown instruction!");
2510 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2511 I->value_op_end()));
2513 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2514 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2515 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2517 return LHS->isIdenticalTo(RHS);
2522 /// \brief Check whether this block is a predicated block.
2523 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2524 /// = ...; " blocks. We start with one vectorized basic block. For every
2525 /// conditional block we split this vectorized block. Therefore, every second
2526 /// block will be a predicated one.
2527 static bool isPredicatedBlock(unsigned BlockNum) {
2528 return BlockNum % 2;
2531 ///\brief Perform cse of induction variable instructions.
2532 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2533 // Perform simple cse.
2534 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2535 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2536 BasicBlock *BB = BBs[i];
2537 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2538 Instruction *In = I++;
2540 if (!CSEDenseMapInfo::canHandle(In))
2543 // Check if we can replace this instruction with any of the
2544 // visited instructions.
2545 if (Instruction *V = CSEMap.lookup(In)) {
2546 In->replaceAllUsesWith(V);
2547 In->eraseFromParent();
2550 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2551 // ...;" blocks for predicated stores. Every second block is a predicated
2553 if (isPredicatedBlock(i))
2561 /// \brief Adds a 'fast' flag to floating point operations.
2562 static Value *addFastMathFlag(Value *V) {
2563 if (isa<FPMathOperator>(V)){
2564 FastMathFlags Flags;
2565 Flags.setUnsafeAlgebra();
2566 cast<Instruction>(V)->setFastMathFlags(Flags);
2571 void InnerLoopVectorizer::vectorizeLoop() {
2572 //===------------------------------------------------===//
2574 // Notice: any optimization or new instruction that go
2575 // into the code below should be also be implemented in
2578 //===------------------------------------------------===//
2579 Constant *Zero = Builder.getInt32(0);
2581 // In order to support reduction variables we need to be able to vectorize
2582 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2583 // stages. First, we create a new vector PHI node with no incoming edges.
2584 // We use this value when we vectorize all of the instructions that use the
2585 // PHI. Next, after all of the instructions in the block are complete we
2586 // add the new incoming edges to the PHI. At this point all of the
2587 // instructions in the basic block are vectorized, so we can use them to
2588 // construct the PHI.
2589 PhiVector RdxPHIsToFix;
2591 // Scan the loop in a topological order to ensure that defs are vectorized
2593 LoopBlocksDFS DFS(OrigLoop);
2596 // Vectorize all of the blocks in the original loop.
2597 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2598 be = DFS.endRPO(); bb != be; ++bb)
2599 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2601 // At this point every instruction in the original loop is widened to
2602 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2603 // that we vectorized. The PHI nodes are currently empty because we did
2604 // not want to introduce cycles. Notice that the remaining PHI nodes
2605 // that we need to fix are reduction variables.
2607 // Create the 'reduced' values for each of the induction vars.
2608 // The reduced values are the vector values that we scalarize and combine
2609 // after the loop is finished.
2610 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2612 PHINode *RdxPhi = *it;
2613 assert(RdxPhi && "Unable to recover vectorized PHI");
2615 // Find the reduction variable descriptor.
2616 assert(Legal->getReductionVars()->count(RdxPhi) &&
2617 "Unable to find the reduction variable");
2618 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2619 (*Legal->getReductionVars())[RdxPhi];
2621 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2623 // We need to generate a reduction vector from the incoming scalar.
2624 // To do so, we need to generate the 'identity' vector and override
2625 // one of the elements with the incoming scalar reduction. We need
2626 // to do it in the vector-loop preheader.
2627 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
2629 // This is the vector-clone of the value that leaves the loop.
2630 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2631 Type *VecTy = VectorExit[0]->getType();
2633 // Find the reduction identity variable. Zero for addition, or, xor,
2634 // one for multiplication, -1 for And.
2637 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2638 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2639 // MinMax reduction have the start value as their identify.
2641 VectorStart = Identity = RdxDesc.StartValue;
2643 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2648 // Handle other reduction kinds:
2650 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2651 VecTy->getScalarType());
2654 // This vector is the Identity vector where the first element is the
2655 // incoming scalar reduction.
2656 VectorStart = RdxDesc.StartValue;
2658 Identity = ConstantVector::getSplat(VF, Iden);
2660 // This vector is the Identity vector where the first element is the
2661 // incoming scalar reduction.
2662 VectorStart = Builder.CreateInsertElement(Identity,
2663 RdxDesc.StartValue, Zero);
2667 // Fix the vector-loop phi.
2669 // Reductions do not have to start at zero. They can start with
2670 // any loop invariant values.
2671 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2672 BasicBlock *Latch = OrigLoop->getLoopLatch();
2673 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2674 VectorParts &Val = getVectorValue(LoopVal);
2675 for (unsigned part = 0; part < UF; ++part) {
2676 // Make sure to add the reduction stat value only to the
2677 // first unroll part.
2678 Value *StartVal = (part == 0) ? VectorStart : Identity;
2679 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
2680 LoopVectorPreHeader);
2681 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2682 LoopVectorBody.back());
2685 // Before each round, move the insertion point right between
2686 // the PHIs and the values we are going to write.
2687 // This allows us to write both PHINodes and the extractelement
2689 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2691 VectorParts RdxParts;
2692 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2693 for (unsigned part = 0; part < UF; ++part) {
2694 // This PHINode contains the vectorized reduction variable, or
2695 // the initial value vector, if we bypass the vector loop.
2696 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2697 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2698 Value *StartVal = (part == 0) ? VectorStart : Identity;
2699 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2700 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2701 NewPhi->addIncoming(RdxExitVal[part],
2702 LoopVectorBody.back());
2703 RdxParts.push_back(NewPhi);
2706 // Reduce all of the unrolled parts into a single vector.
2707 Value *ReducedPartRdx = RdxParts[0];
2708 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2709 setDebugLocFromInst(Builder, ReducedPartRdx);
2710 for (unsigned part = 1; part < UF; ++part) {
2711 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2712 // Floating point operations had to be 'fast' to enable the reduction.
2713 ReducedPartRdx = addFastMathFlag(
2714 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2715 ReducedPartRdx, "bin.rdx"));
2717 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2718 ReducedPartRdx, RdxParts[part]);
2722 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2723 // and vector ops, reducing the set of values being computed by half each
2725 assert(isPowerOf2_32(VF) &&
2726 "Reduction emission only supported for pow2 vectors!");
2727 Value *TmpVec = ReducedPartRdx;
2728 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2729 for (unsigned i = VF; i != 1; i >>= 1) {
2730 // Move the upper half of the vector to the lower half.
2731 for (unsigned j = 0; j != i/2; ++j)
2732 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2734 // Fill the rest of the mask with undef.
2735 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2736 UndefValue::get(Builder.getInt32Ty()));
2739 Builder.CreateShuffleVector(TmpVec,
2740 UndefValue::get(TmpVec->getType()),
2741 ConstantVector::get(ShuffleMask),
2744 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2745 // Floating point operations had to be 'fast' to enable the reduction.
2746 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2747 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2749 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2752 // The result is in the first element of the vector.
2753 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2754 Builder.getInt32(0));
2757 // Create a phi node that merges control-flow from the backedge-taken check
2758 // block and the middle block.
2759 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
2760 LoopScalarPreHeader->getTerminator());
2761 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
2762 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2764 // Now, we need to fix the users of the reduction variable
2765 // inside and outside of the scalar remainder loop.
2766 // We know that the loop is in LCSSA form. We need to update the
2767 // PHI nodes in the exit blocks.
2768 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2769 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2770 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2771 if (!LCSSAPhi) break;
2773 // All PHINodes need to have a single entry edge, or two if
2774 // we already fixed them.
2775 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2777 // We found our reduction value exit-PHI. Update it with the
2778 // incoming bypass edge.
2779 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2780 // Add an edge coming from the bypass.
2781 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2784 }// end of the LCSSA phi scan.
2786 // Fix the scalar loop reduction variable with the incoming reduction sum
2787 // from the vector body and from the backedge value.
2788 int IncomingEdgeBlockIdx =
2789 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2790 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2791 // Pick the other block.
2792 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2793 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
2794 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2795 }// end of for each redux variable.
2799 // Remove redundant induction instructions.
2800 cse(LoopVectorBody);
2803 void InnerLoopVectorizer::fixLCSSAPHIs() {
2804 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2805 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2806 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2807 if (!LCSSAPhi) break;
2808 if (LCSSAPhi->getNumIncomingValues() == 1)
2809 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2814 InnerLoopVectorizer::VectorParts
2815 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2816 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2819 // Look for cached value.
2820 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2821 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2822 if (ECEntryIt != MaskCache.end())
2823 return ECEntryIt->second;
2825 VectorParts SrcMask = createBlockInMask(Src);
2827 // The terminator has to be a branch inst!
2828 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2829 assert(BI && "Unexpected terminator found");
2831 if (BI->isConditional()) {
2832 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2834 if (BI->getSuccessor(0) != Dst)
2835 for (unsigned part = 0; part < UF; ++part)
2836 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2838 for (unsigned part = 0; part < UF; ++part)
2839 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2841 MaskCache[Edge] = EdgeMask;
2845 MaskCache[Edge] = SrcMask;
2849 InnerLoopVectorizer::VectorParts
2850 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2851 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2853 // Loop incoming mask is all-one.
2854 if (OrigLoop->getHeader() == BB) {
2855 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2856 return getVectorValue(C);
2859 // This is the block mask. We OR all incoming edges, and with zero.
2860 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2861 VectorParts BlockMask = getVectorValue(Zero);
2864 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2865 VectorParts EM = createEdgeMask(*it, BB);
2866 for (unsigned part = 0; part < UF; ++part)
2867 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2873 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2874 InnerLoopVectorizer::VectorParts &Entry,
2875 unsigned UF, unsigned VF, PhiVector *PV) {
2876 PHINode* P = cast<PHINode>(PN);
2877 // Handle reduction variables:
2878 if (Legal->getReductionVars()->count(P)) {
2879 for (unsigned part = 0; part < UF; ++part) {
2880 // This is phase one of vectorizing PHIs.
2881 Type *VecTy = (VF == 1) ? PN->getType() :
2882 VectorType::get(PN->getType(), VF);
2883 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2884 LoopVectorBody.back()-> getFirstInsertionPt());
2890 setDebugLocFromInst(Builder, P);
2891 // Check for PHI nodes that are lowered to vector selects.
2892 if (P->getParent() != OrigLoop->getHeader()) {
2893 // We know that all PHIs in non-header blocks are converted into
2894 // selects, so we don't have to worry about the insertion order and we
2895 // can just use the builder.
2896 // At this point we generate the predication tree. There may be
2897 // duplications since this is a simple recursive scan, but future
2898 // optimizations will clean it up.
2900 unsigned NumIncoming = P->getNumIncomingValues();
2902 // Generate a sequence of selects of the form:
2903 // SELECT(Mask3, In3,
2904 // SELECT(Mask2, In2,
2906 for (unsigned In = 0; In < NumIncoming; In++) {
2907 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2909 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2911 for (unsigned part = 0; part < UF; ++part) {
2912 // We might have single edge PHIs (blocks) - use an identity
2913 // 'select' for the first PHI operand.
2915 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2918 // Select between the current value and the previous incoming edge
2919 // based on the incoming mask.
2920 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2921 Entry[part], "predphi");
2927 // This PHINode must be an induction variable.
2928 // Make sure that we know about it.
2929 assert(Legal->getInductionVars()->count(P) &&
2930 "Not an induction variable");
2932 LoopVectorizationLegality::InductionInfo II =
2933 Legal->getInductionVars()->lookup(P);
2935 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
2936 // which can be found from the original scalar operations.
2938 case LoopVectorizationLegality::IK_NoInduction:
2939 llvm_unreachable("Unknown induction");
2940 case LoopVectorizationLegality::IK_IntInduction: {
2941 assert(P->getType() == II.StartValue->getType() && "Types must match");
2942 Type *PhiTy = P->getType();
2944 if (P == OldInduction) {
2945 // Handle the canonical induction variable. We might have had to
2947 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2949 // Handle other induction variables that are now based on the
2951 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2953 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2954 Broadcasted = II.transform(Builder, NormalizedIdx);
2955 Broadcasted->setName("offset.idx");
2957 Broadcasted = getBroadcastInstrs(Broadcasted);
2958 // After broadcasting the induction variable we need to make the vector
2959 // consecutive by adding 0, 1, 2, etc.
2960 for (unsigned part = 0; part < UF; ++part)
2961 Entry[part] = getStepVector(Broadcasted, VF * part, II.StepValue);
2964 case LoopVectorizationLegality::IK_PtrInduction:
2965 // Handle the pointer induction variable case.
2966 assert(P->getType()->isPointerTy() && "Unexpected type.");
2967 // This is the normalized GEP that starts counting at zero.
2968 Value *NormalizedIdx =
2969 Builder.CreateSub(Induction, ExtendedIdx, "normalized.idx");
2970 // This is the vector of results. Notice that we don't generate
2971 // vector geps because scalar geps result in better code.
2972 for (unsigned part = 0; part < UF; ++part) {
2974 int EltIndex = part;
2975 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2976 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
2977 Value *SclrGep = II.transform(Builder, GlobalIdx);
2978 SclrGep->setName("next.gep");
2979 Entry[part] = SclrGep;
2983 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
2984 for (unsigned int i = 0; i < VF; ++i) {
2985 int EltIndex = i + part * VF;
2986 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2987 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
2988 Value *SclrGep = II.transform(Builder, GlobalIdx);
2989 SclrGep->setName("next.gep");
2990 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
2991 Builder.getInt32(i),
2994 Entry[part] = VecVal;
3000 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3001 // For each instruction in the old loop.
3002 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3003 VectorParts &Entry = WidenMap.get(it);
3004 switch (it->getOpcode()) {
3005 case Instruction::Br:
3006 // Nothing to do for PHIs and BR, since we already took care of the
3007 // loop control flow instructions.
3009 case Instruction::PHI: {
3010 // Vectorize PHINodes.
3011 widenPHIInstruction(it, Entry, UF, VF, PV);
3015 case Instruction::Add:
3016 case Instruction::FAdd:
3017 case Instruction::Sub:
3018 case Instruction::FSub:
3019 case Instruction::Mul:
3020 case Instruction::FMul:
3021 case Instruction::UDiv:
3022 case Instruction::SDiv:
3023 case Instruction::FDiv:
3024 case Instruction::URem:
3025 case Instruction::SRem:
3026 case Instruction::FRem:
3027 case Instruction::Shl:
3028 case Instruction::LShr:
3029 case Instruction::AShr:
3030 case Instruction::And:
3031 case Instruction::Or:
3032 case Instruction::Xor: {
3033 // Just widen binops.
3034 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3035 setDebugLocFromInst(Builder, BinOp);
3036 VectorParts &A = getVectorValue(it->getOperand(0));
3037 VectorParts &B = getVectorValue(it->getOperand(1));
3039 // Use this vector value for all users of the original instruction.
3040 for (unsigned Part = 0; Part < UF; ++Part) {
3041 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3043 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3044 VecOp->copyIRFlags(BinOp);
3049 propagateMetadata(Entry, it);
3052 case Instruction::Select: {
3054 // If the selector is loop invariant we can create a select
3055 // instruction with a scalar condition. Otherwise, use vector-select.
3056 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3058 setDebugLocFromInst(Builder, it);
3060 // The condition can be loop invariant but still defined inside the
3061 // loop. This means that we can't just use the original 'cond' value.
3062 // We have to take the 'vectorized' value and pick the first lane.
3063 // Instcombine will make this a no-op.
3064 VectorParts &Cond = getVectorValue(it->getOperand(0));
3065 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3066 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3068 Value *ScalarCond = (VF == 1) ? Cond[0] :
3069 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3071 for (unsigned Part = 0; Part < UF; ++Part) {
3072 Entry[Part] = Builder.CreateSelect(
3073 InvariantCond ? ScalarCond : Cond[Part],
3078 propagateMetadata(Entry, it);
3082 case Instruction::ICmp:
3083 case Instruction::FCmp: {
3084 // Widen compares. Generate vector compares.
3085 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3086 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3087 setDebugLocFromInst(Builder, it);
3088 VectorParts &A = getVectorValue(it->getOperand(0));
3089 VectorParts &B = getVectorValue(it->getOperand(1));
3090 for (unsigned Part = 0; Part < UF; ++Part) {
3093 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3095 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3099 propagateMetadata(Entry, it);
3103 case Instruction::Store:
3104 case Instruction::Load:
3105 vectorizeMemoryInstruction(it);
3107 case Instruction::ZExt:
3108 case Instruction::SExt:
3109 case Instruction::FPToUI:
3110 case Instruction::FPToSI:
3111 case Instruction::FPExt:
3112 case Instruction::PtrToInt:
3113 case Instruction::IntToPtr:
3114 case Instruction::SIToFP:
3115 case Instruction::UIToFP:
3116 case Instruction::Trunc:
3117 case Instruction::FPTrunc:
3118 case Instruction::BitCast: {
3119 CastInst *CI = dyn_cast<CastInst>(it);
3120 setDebugLocFromInst(Builder, it);
3121 /// Optimize the special case where the source is the induction
3122 /// variable. Notice that we can only optimize the 'trunc' case
3123 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3124 /// c. other casts depend on pointer size.
3125 if (CI->getOperand(0) == OldInduction &&
3126 it->getOpcode() == Instruction::Trunc) {
3127 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3129 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3130 LoopVectorizationLegality::InductionInfo II =
3131 Legal->getInductionVars()->lookup(OldInduction);
3133 ConstantInt::getSigned(CI->getType(), II.StepValue->getSExtValue());
3134 for (unsigned Part = 0; Part < UF; ++Part)
3135 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
3136 propagateMetadata(Entry, it);
3139 /// Vectorize casts.
3140 Type *DestTy = (VF == 1) ? CI->getType() :
3141 VectorType::get(CI->getType(), VF);
3143 VectorParts &A = getVectorValue(it->getOperand(0));
3144 for (unsigned Part = 0; Part < UF; ++Part)
3145 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3146 propagateMetadata(Entry, it);
3150 case Instruction::Call: {
3151 // Ignore dbg intrinsics.
3152 if (isa<DbgInfoIntrinsic>(it))
3154 setDebugLocFromInst(Builder, it);
3156 Module *M = BB->getParent()->getParent();
3157 CallInst *CI = cast<CallInst>(it);
3158 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3159 assert(ID && "Not an intrinsic call!");
3161 case Intrinsic::assume:
3162 case Intrinsic::lifetime_end:
3163 case Intrinsic::lifetime_start:
3164 scalarizeInstruction(it);
3167 bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
3168 for (unsigned Part = 0; Part < UF; ++Part) {
3169 SmallVector<Value *, 4> Args;
3170 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3171 if (HasScalarOpd && i == 1) {
3172 Args.push_back(CI->getArgOperand(i));
3175 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3176 Args.push_back(Arg[Part]);
3178 Type *Tys[] = {CI->getType()};
3180 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3182 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3183 Entry[Part] = Builder.CreateCall(F, Args);
3186 propagateMetadata(Entry, it);
3193 // All other instructions are unsupported. Scalarize them.
3194 scalarizeInstruction(it);
3197 }// end of for_each instr.
3200 void InnerLoopVectorizer::updateAnalysis() {
3201 // Forget the original basic block.
3202 SE->forgetLoop(OrigLoop);
3204 // Update the dominator tree information.
3205 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3206 "Entry does not dominate exit.");
3208 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3209 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3210 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3212 // Due to if predication of stores we might create a sequence of "if(pred)
3213 // a[i] = ...; " blocks.
3214 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3216 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3217 else if (isPredicatedBlock(i)) {
3218 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3220 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3224 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3225 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3226 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3227 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3229 DEBUG(DT->verifyDomTree());
3232 /// \brief Check whether it is safe to if-convert this phi node.
3234 /// Phi nodes with constant expressions that can trap are not safe to if
3236 static bool canIfConvertPHINodes(BasicBlock *BB) {
3237 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3238 PHINode *Phi = dyn_cast<PHINode>(I);
3241 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3242 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3249 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3250 if (!EnableIfConversion) {
3251 emitAnalysis(VectorizationReport() << "if-conversion is disabled");
3255 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3257 // A list of pointers that we can safely read and write to.
3258 SmallPtrSet<Value *, 8> SafePointes;
3260 // Collect safe addresses.
3261 for (Loop::block_iterator BI = TheLoop->block_begin(),
3262 BE = TheLoop->block_end(); BI != BE; ++BI) {
3263 BasicBlock *BB = *BI;
3265 if (blockNeedsPredication(BB))
3268 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3269 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3270 SafePointes.insert(LI->getPointerOperand());
3271 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3272 SafePointes.insert(SI->getPointerOperand());
3276 // Collect the blocks that need predication.
3277 BasicBlock *Header = TheLoop->getHeader();
3278 for (Loop::block_iterator BI = TheLoop->block_begin(),
3279 BE = TheLoop->block_end(); BI != BE; ++BI) {
3280 BasicBlock *BB = *BI;
3282 // We don't support switch statements inside loops.
3283 if (!isa<BranchInst>(BB->getTerminator())) {
3284 emitAnalysis(VectorizationReport(BB->getTerminator())
3285 << "loop contains a switch statement");
3289 // We must be able to predicate all blocks that need to be predicated.
3290 if (blockNeedsPredication(BB)) {
3291 if (!blockCanBePredicated(BB, SafePointes)) {
3292 emitAnalysis(VectorizationReport(BB->getTerminator())
3293 << "control flow cannot be substituted for a select");
3296 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3297 emitAnalysis(VectorizationReport(BB->getTerminator())
3298 << "control flow cannot be substituted for a select");
3303 // We can if-convert this loop.
3307 bool LoopVectorizationLegality::canVectorize() {
3308 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3309 // be canonicalized.
3310 if (!TheLoop->getLoopPreheader()) {
3312 VectorizationReport() <<
3313 "loop control flow is not understood by vectorizer");
3317 // We can only vectorize innermost loops.
3318 if (!TheLoop->getSubLoopsVector().empty()) {
3319 emitAnalysis(VectorizationReport() << "loop is not the innermost loop");
3323 // We must have a single backedge.
3324 if (TheLoop->getNumBackEdges() != 1) {
3326 VectorizationReport() <<
3327 "loop control flow is not understood by vectorizer");
3331 // We must have a single exiting block.
3332 if (!TheLoop->getExitingBlock()) {
3334 VectorizationReport() <<
3335 "loop control flow is not understood by vectorizer");
3339 // We only handle bottom-tested loops, i.e. loop in which the condition is
3340 // checked at the end of each iteration. With that we can assume that all
3341 // instructions in the loop are executed the same number of times.
3342 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
3344 VectorizationReport() <<
3345 "loop control flow is not understood by vectorizer");
3349 // We need to have a loop header.
3350 DEBUG(dbgs() << "LV: Found a loop: " <<
3351 TheLoop->getHeader()->getName() << '\n');
3353 // Check if we can if-convert non-single-bb loops.
3354 unsigned NumBlocks = TheLoop->getNumBlocks();
3355 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3356 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3360 // ScalarEvolution needs to be able to find the exit count.
3361 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3362 if (ExitCount == SE->getCouldNotCompute()) {
3363 emitAnalysis(VectorizationReport() <<
3364 "could not determine number of loop iterations");
3365 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3369 // Check if we can vectorize the instructions and CFG in this loop.
3370 if (!canVectorizeInstrs()) {
3371 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3375 // Go over each instruction and look at memory deps.
3376 if (!canVectorizeMemory()) {
3377 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3381 // Collect all of the variables that remain uniform after vectorization.
3382 collectLoopUniforms();
3384 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3385 (LAI.getRuntimePointerCheck()->Need ? " (with a runtime bound check)" :
3389 // Okay! We can vectorize. At this point we don't have any other mem analysis
3390 // which may limit our maximum vectorization factor, so just return true with
3395 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3396 if (Ty->isPointerTy())
3397 return DL.getIntPtrType(Ty);
3399 // It is possible that char's or short's overflow when we ask for the loop's
3400 // trip count, work around this by changing the type size.
3401 if (Ty->getScalarSizeInBits() < 32)
3402 return Type::getInt32Ty(Ty->getContext());
3407 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3408 Ty0 = convertPointerToIntegerType(DL, Ty0);
3409 Ty1 = convertPointerToIntegerType(DL, Ty1);
3410 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3415 /// \brief Check that the instruction has outside loop users and is not an
3416 /// identified reduction variable.
3417 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3418 SmallPtrSetImpl<Value *> &Reductions) {
3419 // Reduction instructions are allowed to have exit users. All other
3420 // instructions must not have external users.
3421 if (!Reductions.count(Inst))
3422 //Check that all of the users of the loop are inside the BB.
3423 for (User *U : Inst->users()) {
3424 Instruction *UI = cast<Instruction>(U);
3425 // This user may be a reduction exit value.
3426 if (!TheLoop->contains(UI)) {
3427 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3434 bool LoopVectorizationLegality::canVectorizeInstrs() {
3435 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3436 BasicBlock *Header = TheLoop->getHeader();
3438 // Look for the attribute signaling the absence of NaNs.
3439 Function &F = *Header->getParent();
3440 if (F.hasFnAttribute("no-nans-fp-math"))
3442 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
3444 // For each block in the loop.
3445 for (Loop::block_iterator bb = TheLoop->block_begin(),
3446 be = TheLoop->block_end(); bb != be; ++bb) {
3448 // Scan the instructions in the block and look for hazards.
3449 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3452 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3453 Type *PhiTy = Phi->getType();
3454 // Check that this PHI type is allowed.
3455 if (!PhiTy->isIntegerTy() &&
3456 !PhiTy->isFloatingPointTy() &&
3457 !PhiTy->isPointerTy()) {
3458 emitAnalysis(VectorizationReport(it)
3459 << "loop control flow is not understood by vectorizer");
3460 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3464 // If this PHINode is not in the header block, then we know that we
3465 // can convert it to select during if-conversion. No need to check if
3466 // the PHIs in this block are induction or reduction variables.
3467 if (*bb != Header) {
3468 // Check that this instruction has no outside users or is an
3469 // identified reduction value with an outside user.
3470 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3472 emitAnalysis(VectorizationReport(it) <<
3473 "value could not be identified as "
3474 "an induction or reduction variable");
3478 // We only allow if-converted PHIs with exactly two incoming values.
3479 if (Phi->getNumIncomingValues() != 2) {
3480 emitAnalysis(VectorizationReport(it)
3481 << "control flow not understood by vectorizer");
3482 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3486 // This is the value coming from the preheader.
3487 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3488 ConstantInt *StepValue = nullptr;
3489 // Check if this is an induction variable.
3490 InductionKind IK = isInductionVariable(Phi, StepValue);
3492 if (IK_NoInduction != IK) {
3493 // Get the widest type.
3495 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3497 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3499 // Int inductions are special because we only allow one IV.
3500 if (IK == IK_IntInduction && StepValue->isOne()) {
3501 // Use the phi node with the widest type as induction. Use the last
3502 // one if there are multiple (no good reason for doing this other
3503 // than it is expedient).
3504 if (!Induction || PhiTy == WidestIndTy)
3508 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3509 Inductions[Phi] = InductionInfo(StartValue, IK, StepValue);
3511 // Until we explicitly handle the case of an induction variable with
3512 // an outside loop user we have to give up vectorizing this loop.
3513 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3514 emitAnalysis(VectorizationReport(it) <<
3515 "use of induction value outside of the "
3516 "loop is not handled by vectorizer");
3523 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3524 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3527 if (AddReductionVar(Phi, RK_IntegerMult)) {
3528 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3531 if (AddReductionVar(Phi, RK_IntegerOr)) {
3532 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3535 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3536 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3539 if (AddReductionVar(Phi, RK_IntegerXor)) {
3540 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3543 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3544 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3547 if (AddReductionVar(Phi, RK_FloatMult)) {
3548 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3551 if (AddReductionVar(Phi, RK_FloatAdd)) {
3552 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3555 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3556 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3561 emitAnalysis(VectorizationReport(it) <<
3562 "value that could not be identified as "
3563 "reduction is used outside the loop");
3564 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3566 }// end of PHI handling
3568 // We still don't handle functions. However, we can ignore dbg intrinsic
3569 // calls and we do handle certain intrinsic and libm functions.
3570 CallInst *CI = dyn_cast<CallInst>(it);
3571 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3572 emitAnalysis(VectorizationReport(it) <<
3573 "call instruction cannot be vectorized");
3574 DEBUG(dbgs() << "LV: Found a call site.\n");
3578 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
3579 // second argument is the same (i.e. loop invariant)
3581 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
3582 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
3583 emitAnalysis(VectorizationReport(it)
3584 << "intrinsic instruction cannot be vectorized");
3585 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
3590 // Check that the instruction return type is vectorizable.
3591 // Also, we can't vectorize extractelement instructions.
3592 if ((!VectorType::isValidElementType(it->getType()) &&
3593 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3594 emitAnalysis(VectorizationReport(it)
3595 << "instruction return type cannot be vectorized");
3596 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3600 // Check that the stored type is vectorizable.
3601 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3602 Type *T = ST->getValueOperand()->getType();
3603 if (!VectorType::isValidElementType(T)) {
3604 emitAnalysis(VectorizationReport(ST) <<
3605 "store instruction cannot be vectorized");
3608 if (EnableMemAccessVersioning)
3609 collectStridedAccess(ST);
3612 if (EnableMemAccessVersioning)
3613 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3614 collectStridedAccess(LI);
3616 // Reduction instructions are allowed to have exit users.
3617 // All other instructions must not have external users.
3618 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3619 emitAnalysis(VectorizationReport(it) <<
3620 "value cannot be used outside the loop");
3629 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3630 if (Inductions.empty()) {
3631 emitAnalysis(VectorizationReport()
3632 << "loop induction variable could not be identified");
3640 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3641 /// return the induction operand of the gep pointer.
3642 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3643 const DataLayout *DL, Loop *Lp) {
3644 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3648 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3650 // Check that all of the gep indices are uniform except for our induction
3652 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3653 if (i != InductionOperand &&
3654 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3656 return GEP->getOperand(InductionOperand);
3659 ///\brief Look for a cast use of the passed value.
3660 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3661 Value *UniqueCast = nullptr;
3662 for (User *U : Ptr->users()) {
3663 CastInst *CI = dyn_cast<CastInst>(U);
3664 if (CI && CI->getType() == Ty) {
3674 ///\brief Get the stride of a pointer access in a loop.
3675 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3676 /// pointer to the Value, or null otherwise.
3677 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3678 const DataLayout *DL, Loop *Lp) {
3679 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3680 if (!PtrTy || PtrTy->isAggregateType())
3683 // Try to remove a gep instruction to make the pointer (actually index at this
3684 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3685 // pointer, otherwise, we are analyzing the index.
3686 Value *OrigPtr = Ptr;
3688 // The size of the pointer access.
3689 int64_t PtrAccessSize = 1;
3691 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3692 const SCEV *V = SE->getSCEV(Ptr);
3696 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3697 V = C->getOperand();
3699 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3703 V = S->getStepRecurrence(*SE);
3707 // Strip off the size of access multiplication if we are still analyzing the
3709 if (OrigPtr == Ptr) {
3710 DL->getTypeAllocSize(PtrTy->getElementType());
3711 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3712 if (M->getOperand(0)->getSCEVType() != scConstant)
3715 const APInt &APStepVal =
3716 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3718 // Huge step value - give up.
3719 if (APStepVal.getBitWidth() > 64)
3722 int64_t StepVal = APStepVal.getSExtValue();
3723 if (PtrAccessSize != StepVal)
3725 V = M->getOperand(1);
3730 Type *StripedOffRecurrenceCast = nullptr;
3731 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3732 StripedOffRecurrenceCast = C->getType();
3733 V = C->getOperand();
3736 // Look for the loop invariant symbolic value.
3737 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3741 Value *Stride = U->getValue();
3742 if (!Lp->isLoopInvariant(Stride))
3745 // If we have stripped off the recurrence cast we have to make sure that we
3746 // return the value that is used in this loop so that we can replace it later.
3747 if (StripedOffRecurrenceCast)
3748 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3753 void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) {
3754 Value *Ptr = nullptr;
3755 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3756 Ptr = LI->getPointerOperand();
3757 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3758 Ptr = SI->getPointerOperand();
3762 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3766 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3767 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3768 Strides[Ptr] = Stride;
3769 StrideSet.insert(Stride);
3772 void LoopVectorizationLegality::collectLoopUniforms() {
3773 // We now know that the loop is vectorizable!
3774 // Collect variables that will remain uniform after vectorization.
3775 std::vector<Value*> Worklist;
3776 BasicBlock *Latch = TheLoop->getLoopLatch();
3778 // Start with the conditional branch and walk up the block.
3779 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3781 // Also add all consecutive pointer values; these values will be uniform
3782 // after vectorization (and subsequent cleanup) and, until revectorization is
3783 // supported, all dependencies must also be uniform.
3784 for (Loop::block_iterator B = TheLoop->block_begin(),
3785 BE = TheLoop->block_end(); B != BE; ++B)
3786 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
3788 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
3789 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3791 while (!Worklist.empty()) {
3792 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3793 Worklist.pop_back();
3795 // Look at instructions inside this loop.
3796 // Stop when reaching PHI nodes.
3797 // TODO: we need to follow values all over the loop, not only in this block.
3798 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3801 // This is a known uniform.
3804 // Insert all operands.
3805 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3809 bool LoopVectorizationLegality::canVectorizeMemory() {
3810 LAI.analyzeLoop(Strides);
3811 auto &OptionalReport = LAI.getReport();
3813 emitAnalysis(*OptionalReport);
3814 return LAI.canVectorizeMemory();
3817 static bool hasMultipleUsesOf(Instruction *I,
3818 SmallPtrSetImpl<Instruction *> &Insts) {
3819 unsigned NumUses = 0;
3820 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
3821 if (Insts.count(dyn_cast<Instruction>(*Use)))
3830 static bool areAllUsesIn(Instruction *I, SmallPtrSetImpl<Instruction *> &Set) {
3831 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
3832 if (!Set.count(dyn_cast<Instruction>(*Use)))
3837 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
3838 ReductionKind Kind) {
3839 if (Phi->getNumIncomingValues() != 2)
3842 // Reduction variables are only found in the loop header block.
3843 if (Phi->getParent() != TheLoop->getHeader())
3846 // Obtain the reduction start value from the value that comes from the loop
3848 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
3850 // ExitInstruction is the single value which is used outside the loop.
3851 // We only allow for a single reduction value to be used outside the loop.
3852 // This includes users of the reduction, variables (which form a cycle
3853 // which ends in the phi node).
3854 Instruction *ExitInstruction = nullptr;
3855 // Indicates that we found a reduction operation in our scan.
3856 bool FoundReduxOp = false;
3858 // We start with the PHI node and scan for all of the users of this
3859 // instruction. All users must be instructions that can be used as reduction
3860 // variables (such as ADD). We must have a single out-of-block user. The cycle
3861 // must include the original PHI.
3862 bool FoundStartPHI = false;
3864 // To recognize min/max patterns formed by a icmp select sequence, we store
3865 // the number of instruction we saw from the recognized min/max pattern,
3866 // to make sure we only see exactly the two instructions.
3867 unsigned NumCmpSelectPatternInst = 0;
3868 ReductionInstDesc ReduxDesc(false, nullptr);
3870 SmallPtrSet<Instruction *, 8> VisitedInsts;
3871 SmallVector<Instruction *, 8> Worklist;
3872 Worklist.push_back(Phi);
3873 VisitedInsts.insert(Phi);
3875 // A value in the reduction can be used:
3876 // - By the reduction:
3877 // - Reduction operation:
3878 // - One use of reduction value (safe).
3879 // - Multiple use of reduction value (not safe).
3881 // - All uses of the PHI must be the reduction (safe).
3882 // - Otherwise, not safe.
3883 // - By one instruction outside of the loop (safe).
3884 // - By further instructions outside of the loop (not safe).
3885 // - By an instruction that is not part of the reduction (not safe).
3887 // * An instruction type other than PHI or the reduction operation.
3888 // * A PHI in the header other than the initial PHI.
3889 while (!Worklist.empty()) {
3890 Instruction *Cur = Worklist.back();
3891 Worklist.pop_back();
3894 // If the instruction has no users then this is a broken chain and can't be
3895 // a reduction variable.
3896 if (Cur->use_empty())
3899 bool IsAPhi = isa<PHINode>(Cur);
3901 // A header PHI use other than the original PHI.
3902 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
3905 // Reductions of instructions such as Div, and Sub is only possible if the
3906 // LHS is the reduction variable.
3907 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
3908 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
3909 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
3912 // Any reduction instruction must be of one of the allowed kinds.
3913 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
3914 if (!ReduxDesc.IsReduction)
3917 // A reduction operation must only have one use of the reduction value.
3918 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
3919 hasMultipleUsesOf(Cur, VisitedInsts))
3922 // All inputs to a PHI node must be a reduction value.
3923 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
3926 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
3927 isa<SelectInst>(Cur)))
3928 ++NumCmpSelectPatternInst;
3929 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
3930 isa<SelectInst>(Cur)))
3931 ++NumCmpSelectPatternInst;
3933 // Check whether we found a reduction operator.
3934 FoundReduxOp |= !IsAPhi;
3936 // Process users of current instruction. Push non-PHI nodes after PHI nodes
3937 // onto the stack. This way we are going to have seen all inputs to PHI
3938 // nodes once we get to them.
3939 SmallVector<Instruction *, 8> NonPHIs;
3940 SmallVector<Instruction *, 8> PHIs;
3941 for (User *U : Cur->users()) {
3942 Instruction *UI = cast<Instruction>(U);
3944 // Check if we found the exit user.
3945 BasicBlock *Parent = UI->getParent();
3946 if (!TheLoop->contains(Parent)) {
3947 // Exit if you find multiple outside users or if the header phi node is
3948 // being used. In this case the user uses the value of the previous
3949 // iteration, in which case we would loose "VF-1" iterations of the
3950 // reduction operation if we vectorize.
3951 if (ExitInstruction != nullptr || Cur == Phi)
3954 // The instruction used by an outside user must be the last instruction
3955 // before we feed back to the reduction phi. Otherwise, we loose VF-1
3956 // operations on the value.
3957 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
3960 ExitInstruction = Cur;
3964 // Process instructions only once (termination). Each reduction cycle
3965 // value must only be used once, except by phi nodes and min/max
3966 // reductions which are represented as a cmp followed by a select.
3967 ReductionInstDesc IgnoredVal(false, nullptr);
3968 if (VisitedInsts.insert(UI).second) {
3969 if (isa<PHINode>(UI))
3972 NonPHIs.push_back(UI);
3973 } else if (!isa<PHINode>(UI) &&
3974 ((!isa<FCmpInst>(UI) &&
3975 !isa<ICmpInst>(UI) &&
3976 !isa<SelectInst>(UI)) ||
3977 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
3980 // Remember that we completed the cycle.
3982 FoundStartPHI = true;
3984 Worklist.append(PHIs.begin(), PHIs.end());
3985 Worklist.append(NonPHIs.begin(), NonPHIs.end());
3988 // This means we have seen one but not the other instruction of the
3989 // pattern or more than just a select and cmp.
3990 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
3991 NumCmpSelectPatternInst != 2)
3994 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
3997 // We found a reduction var if we have reached the original phi node and we
3998 // only have a single instruction with out-of-loop users.
4000 // This instruction is allowed to have out-of-loop users.
4001 AllowedExit.insert(ExitInstruction);
4003 // Save the description of this reduction variable.
4004 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
4005 ReduxDesc.MinMaxKind);
4006 Reductions[Phi] = RD;
4007 // We've ended the cycle. This is a reduction variable if we have an
4008 // outside user and it has a binary op.
4013 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
4014 /// pattern corresponding to a min(X, Y) or max(X, Y).
4015 LoopVectorizationLegality::ReductionInstDesc
4016 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
4017 ReductionInstDesc &Prev) {
4019 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
4020 "Expect a select instruction");
4021 Instruction *Cmp = nullptr;
4022 SelectInst *Select = nullptr;
4024 // We must handle the select(cmp()) as a single instruction. Advance to the
4026 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
4027 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
4028 return ReductionInstDesc(false, I);
4029 return ReductionInstDesc(Select, Prev.MinMaxKind);
4032 // Only handle single use cases for now.
4033 if (!(Select = dyn_cast<SelectInst>(I)))
4034 return ReductionInstDesc(false, I);
4035 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
4036 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
4037 return ReductionInstDesc(false, I);
4038 if (!Cmp->hasOneUse())
4039 return ReductionInstDesc(false, I);
4044 // Look for a min/max pattern.
4045 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4046 return ReductionInstDesc(Select, MRK_UIntMin);
4047 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4048 return ReductionInstDesc(Select, MRK_UIntMax);
4049 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4050 return ReductionInstDesc(Select, MRK_SIntMax);
4051 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4052 return ReductionInstDesc(Select, MRK_SIntMin);
4053 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4054 return ReductionInstDesc(Select, MRK_FloatMin);
4055 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4056 return ReductionInstDesc(Select, MRK_FloatMax);
4057 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4058 return ReductionInstDesc(Select, MRK_FloatMin);
4059 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4060 return ReductionInstDesc(Select, MRK_FloatMax);
4062 return ReductionInstDesc(false, I);
4065 LoopVectorizationLegality::ReductionInstDesc
4066 LoopVectorizationLegality::isReductionInstr(Instruction *I,
4068 ReductionInstDesc &Prev) {
4069 bool FP = I->getType()->isFloatingPointTy();
4070 bool FastMath = FP && I->hasUnsafeAlgebra();
4071 switch (I->getOpcode()) {
4073 return ReductionInstDesc(false, I);
4074 case Instruction::PHI:
4075 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
4076 Kind != RK_FloatMinMax))
4077 return ReductionInstDesc(false, I);
4078 return ReductionInstDesc(I, Prev.MinMaxKind);
4079 case Instruction::Sub:
4080 case Instruction::Add:
4081 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4082 case Instruction::Mul:
4083 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4084 case Instruction::And:
4085 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4086 case Instruction::Or:
4087 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4088 case Instruction::Xor:
4089 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4090 case Instruction::FMul:
4091 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4092 case Instruction::FSub:
4093 case Instruction::FAdd:
4094 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4095 case Instruction::FCmp:
4096 case Instruction::ICmp:
4097 case Instruction::Select:
4098 if (Kind != RK_IntegerMinMax &&
4099 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4100 return ReductionInstDesc(false, I);
4101 return isMinMaxSelectCmpPattern(I, Prev);
4105 LoopVectorizationLegality::InductionKind
4106 LoopVectorizationLegality::isInductionVariable(PHINode *Phi,
4107 ConstantInt *&StepValue) {
4108 Type *PhiTy = Phi->getType();
4109 // We only handle integer and pointer inductions variables.
4110 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4111 return IK_NoInduction;
4113 // Check that the PHI is consecutive.
4114 const SCEV *PhiScev = SE->getSCEV(Phi);
4115 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4117 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4118 return IK_NoInduction;
4121 const SCEV *Step = AR->getStepRecurrence(*SE);
4122 // Calculate the pointer stride and check if it is consecutive.
4123 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4125 return IK_NoInduction;
4127 ConstantInt *CV = C->getValue();
4128 if (PhiTy->isIntegerTy()) {
4130 return IK_IntInduction;
4133 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4134 Type *PointerElementType = PhiTy->getPointerElementType();
4135 // The pointer stride cannot be determined if the pointer element type is not
4137 if (!PointerElementType->isSized())
4138 return IK_NoInduction;
4140 int64_t Size = static_cast<int64_t>(DL->getTypeAllocSize(PointerElementType));
4141 int64_t CVSize = CV->getSExtValue();
4143 return IK_NoInduction;
4144 StepValue = ConstantInt::getSigned(CV->getType(), CVSize / Size);
4145 return IK_PtrInduction;
4148 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4149 Value *In0 = const_cast<Value*>(V);
4150 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4154 return Inductions.count(PN);
4157 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4158 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4161 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4162 SmallPtrSetImpl<Value *> &SafePtrs) {
4164 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4165 // Check that we don't have a constant expression that can trap as operand.
4166 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4168 if (Constant *C = dyn_cast<Constant>(*OI))
4172 // We might be able to hoist the load.
4173 if (it->mayReadFromMemory()) {
4174 LoadInst *LI = dyn_cast<LoadInst>(it);
4177 if (!SafePtrs.count(LI->getPointerOperand())) {
4178 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
4179 MaskedOp.insert(LI);
4186 // We don't predicate stores at the moment.
4187 if (it->mayWriteToMemory()) {
4188 StoreInst *SI = dyn_cast<StoreInst>(it);
4189 // We only support predication of stores in basic blocks with one
4194 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
4195 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
4197 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
4198 !isSinglePredecessor) {
4199 // Build a masked store if it is legal for the target, otherwise scalarize
4201 bool isLegalMaskedOp =
4202 isLegalMaskedStore(SI->getValueOperand()->getType(),
4203 SI->getPointerOperand());
4204 if (isLegalMaskedOp) {
4206 MaskedOp.insert(SI);
4215 // The instructions below can trap.
4216 switch (it->getOpcode()) {
4218 case Instruction::UDiv:
4219 case Instruction::SDiv:
4220 case Instruction::URem:
4221 case Instruction::SRem:
4229 LoopVectorizationCostModel::VectorizationFactor
4230 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
4231 // Width 1 means no vectorize
4232 VectorizationFactor Factor = { 1U, 0U };
4233 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4234 emitAnalysis(VectorizationReport() <<
4235 "runtime pointer checks needed. Enable vectorization of this "
4236 "loop with '#pragma clang loop vectorize(enable)' when "
4237 "compiling with -Os");
4238 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4242 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
4243 emitAnalysis(VectorizationReport() <<
4244 "store that is conditionally executed prevents vectorization");
4245 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
4249 // Find the trip count.
4250 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4251 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4253 unsigned WidestType = getWidestType();
4254 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4255 unsigned MaxSafeDepDist = -1U;
4256 if (Legal->getMaxSafeDepDistBytes() != -1U)
4257 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4258 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4259 WidestRegister : MaxSafeDepDist);
4260 unsigned MaxVectorSize = WidestRegister / WidestType;
4261 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4262 DEBUG(dbgs() << "LV: The Widest register is: "
4263 << WidestRegister << " bits.\n");
4265 if (MaxVectorSize == 0) {
4266 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4270 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
4271 " into one vector!");
4273 unsigned VF = MaxVectorSize;
4275 // If we optimize the program for size, avoid creating the tail loop.
4277 // If we are unable to calculate the trip count then don't try to vectorize.
4280 (VectorizationReport() <<
4281 "unable to calculate the loop count due to complex control flow");
4282 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4286 // Find the maximum SIMD width that can fit within the trip count.
4287 VF = TC % MaxVectorSize;
4292 // If the trip count that we found modulo the vectorization factor is not
4293 // zero then we require a tail.
4295 emitAnalysis(VectorizationReport() <<
4296 "cannot optimize for size and vectorize at the "
4297 "same time. Enable vectorization of this loop "
4298 "with '#pragma clang loop vectorize(enable)' "
4299 "when compiling with -Os");
4300 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4305 int UserVF = Hints->getWidth();
4307 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4308 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4310 Factor.Width = UserVF;
4314 float Cost = expectedCost(1);
4316 const float ScalarCost = Cost;
4319 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
4321 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
4322 // Ignore scalar width, because the user explicitly wants vectorization.
4323 if (ForceVectorization && VF > 1) {
4325 Cost = expectedCost(Width) / (float)Width;
4328 for (unsigned i=2; i <= VF; i*=2) {
4329 // Notice that the vector loop needs to be executed less times, so
4330 // we need to divide the cost of the vector loops by the width of
4331 // the vector elements.
4332 float VectorCost = expectedCost(i) / (float)i;
4333 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
4334 (int)VectorCost << ".\n");
4335 if (VectorCost < Cost) {
4341 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
4342 << "LV: Vectorization seems to be not beneficial, "
4343 << "but was forced by a user.\n");
4344 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
4345 Factor.Width = Width;
4346 Factor.Cost = Width * Cost;
4350 unsigned LoopVectorizationCostModel::getWidestType() {
4351 unsigned MaxWidth = 8;
4354 for (Loop::block_iterator bb = TheLoop->block_begin(),
4355 be = TheLoop->block_end(); bb != be; ++bb) {
4356 BasicBlock *BB = *bb;
4358 // For each instruction in the loop.
4359 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4360 Type *T = it->getType();
4362 // Ignore ephemeral values.
4363 if (EphValues.count(it))
4366 // Only examine Loads, Stores and PHINodes.
4367 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4370 // Examine PHI nodes that are reduction variables.
4371 if (PHINode *PN = dyn_cast<PHINode>(it))
4372 if (!Legal->getReductionVars()->count(PN))
4375 // Examine the stored values.
4376 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4377 T = ST->getValueOperand()->getType();
4379 // Ignore loaded pointer types and stored pointer types that are not
4380 // consecutive. However, we do want to take consecutive stores/loads of
4381 // pointer vectors into account.
4382 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4385 MaxWidth = std::max(MaxWidth,
4386 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
4394 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
4396 unsigned LoopCost) {
4398 // -- The unroll heuristics --
4399 // We unroll the loop in order to expose ILP and reduce the loop overhead.
4400 // There are many micro-architectural considerations that we can't predict
4401 // at this level. For example, frontend pressure (on decode or fetch) due to
4402 // code size, or the number and capabilities of the execution ports.
4404 // We use the following heuristics to select the unroll factor:
4405 // 1. If the code has reductions, then we unroll in order to break the cross
4406 // iteration dependency.
4407 // 2. If the loop is really small, then we unroll in order to reduce the loop
4409 // 3. We don't unroll if we think that we will spill registers to memory due
4410 // to the increased register pressure.
4412 // Use the user preference, unless 'auto' is selected.
4413 int UserUF = Hints->getInterleave();
4417 // When we optimize for size, we don't unroll.
4421 // We used the distance for the unroll factor.
4422 if (Legal->getMaxSafeDepDistBytes() != -1U)
4425 // Do not unroll loops with a relatively small trip count.
4426 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4427 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
4430 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
4431 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
4435 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
4436 TargetNumRegisters = ForceTargetNumScalarRegs;
4438 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
4439 TargetNumRegisters = ForceTargetNumVectorRegs;
4442 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4443 // We divide by these constants so assume that we have at least one
4444 // instruction that uses at least one register.
4445 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4446 R.NumInstructions = std::max(R.NumInstructions, 1U);
4448 // We calculate the unroll factor using the following formula.
4449 // Subtract the number of loop invariants from the number of available
4450 // registers. These registers are used by all of the unrolled instances.
4451 // Next, divide the remaining registers by the number of registers that is
4452 // required by the loop, in order to estimate how many parallel instances
4453 // fit without causing spills. All of this is rounded down if necessary to be
4454 // a power of two. We want power of two unroll factors to simplify any
4455 // addressing operations or alignment considerations.
4456 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
4459 // Don't count the induction variable as unrolled.
4460 if (EnableIndVarRegisterHeur)
4461 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
4462 std::max(1U, (R.MaxLocalUsers - 1)));
4464 // Clamp the unroll factor ranges to reasonable factors.
4465 unsigned MaxInterleaveSize = TTI.getMaxInterleaveFactor();
4467 // Check if the user has overridden the unroll max.
4469 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
4470 MaxInterleaveSize = ForceTargetMaxScalarInterleaveFactor;
4472 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
4473 MaxInterleaveSize = ForceTargetMaxVectorInterleaveFactor;
4476 // If we did not calculate the cost for VF (because the user selected the VF)
4477 // then we calculate the cost of VF here.
4479 LoopCost = expectedCost(VF);
4481 // Clamp the calculated UF to be between the 1 and the max unroll factor
4482 // that the target allows.
4483 if (UF > MaxInterleaveSize)
4484 UF = MaxInterleaveSize;
4488 // Unroll if we vectorized this loop and there is a reduction that could
4489 // benefit from unrolling.
4490 if (VF > 1 && Legal->getReductionVars()->size()) {
4491 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
4495 // Note that if we've already vectorized the loop we will have done the
4496 // runtime check and so unrolling won't require further checks.
4497 bool UnrollingRequiresRuntimePointerCheck =
4498 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
4500 // We want to unroll small loops in order to reduce the loop overhead and
4501 // potentially expose ILP opportunities.
4502 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
4503 if (!UnrollingRequiresRuntimePointerCheck &&
4504 LoopCost < SmallLoopCost) {
4505 // We assume that the cost overhead is 1 and we use the cost model
4506 // to estimate the cost of the loop and unroll until the cost of the
4507 // loop overhead is about 5% of the cost of the loop.
4508 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
4510 // Unroll until store/load ports (estimated by max unroll factor) are
4512 unsigned NumStores = Legal->getNumStores();
4513 unsigned NumLoads = Legal->getNumLoads();
4514 unsigned StoresUF = UF / (NumStores ? NumStores : 1);
4515 unsigned LoadsUF = UF / (NumLoads ? NumLoads : 1);
4517 // If we have a scalar reduction (vector reductions are already dealt with
4518 // by this point), we can increase the critical path length if the loop
4519 // we're unrolling is inside another loop. Limit, by default to 2, so the
4520 // critical path only gets increased by one reduction operation.
4521 if (Legal->getReductionVars()->size() &&
4522 TheLoop->getLoopDepth() > 1) {
4523 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionUF);
4524 SmallUF = std::min(SmallUF, F);
4525 StoresUF = std::min(StoresUF, F);
4526 LoadsUF = std::min(LoadsUF, F);
4529 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
4530 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
4531 return std::max(StoresUF, LoadsUF);
4534 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
4538 DEBUG(dbgs() << "LV: Not Unrolling.\n");
4542 LoopVectorizationCostModel::RegisterUsage
4543 LoopVectorizationCostModel::calculateRegisterUsage() {
4544 // This function calculates the register usage by measuring the highest number
4545 // of values that are alive at a single location. Obviously, this is a very
4546 // rough estimation. We scan the loop in a topological order in order and
4547 // assign a number to each instruction. We use RPO to ensure that defs are
4548 // met before their users. We assume that each instruction that has in-loop
4549 // users starts an interval. We record every time that an in-loop value is
4550 // used, so we have a list of the first and last occurrences of each
4551 // instruction. Next, we transpose this data structure into a multi map that
4552 // holds the list of intervals that *end* at a specific location. This multi
4553 // map allows us to perform a linear search. We scan the instructions linearly
4554 // and record each time that a new interval starts, by placing it in a set.
4555 // If we find this value in the multi-map then we remove it from the set.
4556 // The max register usage is the maximum size of the set.
4557 // We also search for instructions that are defined outside the loop, but are
4558 // used inside the loop. We need this number separately from the max-interval
4559 // usage number because when we unroll, loop-invariant values do not take
4561 LoopBlocksDFS DFS(TheLoop);
4565 R.NumInstructions = 0;
4567 // Each 'key' in the map opens a new interval. The values
4568 // of the map are the index of the 'last seen' usage of the
4569 // instruction that is the key.
4570 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4571 // Maps instruction to its index.
4572 DenseMap<unsigned, Instruction*> IdxToInstr;
4573 // Marks the end of each interval.
4574 IntervalMap EndPoint;
4575 // Saves the list of instruction indices that are used in the loop.
4576 SmallSet<Instruction*, 8> Ends;
4577 // Saves the list of values that are used in the loop but are
4578 // defined outside the loop, such as arguments and constants.
4579 SmallPtrSet<Value*, 8> LoopInvariants;
4582 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4583 be = DFS.endRPO(); bb != be; ++bb) {
4584 R.NumInstructions += (*bb)->size();
4585 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4587 Instruction *I = it;
4588 IdxToInstr[Index++] = I;
4590 // Save the end location of each USE.
4591 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4592 Value *U = I->getOperand(i);
4593 Instruction *Instr = dyn_cast<Instruction>(U);
4595 // Ignore non-instruction values such as arguments, constants, etc.
4596 if (!Instr) continue;
4598 // If this instruction is outside the loop then record it and continue.
4599 if (!TheLoop->contains(Instr)) {
4600 LoopInvariants.insert(Instr);
4604 // Overwrite previous end points.
4605 EndPoint[Instr] = Index;
4611 // Saves the list of intervals that end with the index in 'key'.
4612 typedef SmallVector<Instruction*, 2> InstrList;
4613 DenseMap<unsigned, InstrList> TransposeEnds;
4615 // Transpose the EndPoints to a list of values that end at each index.
4616 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4618 TransposeEnds[it->second].push_back(it->first);
4620 SmallSet<Instruction*, 8> OpenIntervals;
4621 unsigned MaxUsage = 0;
4624 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4625 for (unsigned int i = 0; i < Index; ++i) {
4626 Instruction *I = IdxToInstr[i];
4627 // Ignore instructions that are never used within the loop.
4628 if (!Ends.count(I)) continue;
4630 // Ignore ephemeral values.
4631 if (EphValues.count(I))
4634 // Remove all of the instructions that end at this location.
4635 InstrList &List = TransposeEnds[i];
4636 for (unsigned int j=0, e = List.size(); j < e; ++j)
4637 OpenIntervals.erase(List[j]);
4639 // Count the number of live interals.
4640 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4642 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4643 OpenIntervals.size() << '\n');
4645 // Add the current instruction to the list of open intervals.
4646 OpenIntervals.insert(I);
4649 unsigned Invariant = LoopInvariants.size();
4650 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
4651 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
4652 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
4654 R.LoopInvariantRegs = Invariant;
4655 R.MaxLocalUsers = MaxUsage;
4659 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4663 for (Loop::block_iterator bb = TheLoop->block_begin(),
4664 be = TheLoop->block_end(); bb != be; ++bb) {
4665 unsigned BlockCost = 0;
4666 BasicBlock *BB = *bb;
4668 // For each instruction in the old loop.
4669 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4670 // Skip dbg intrinsics.
4671 if (isa<DbgInfoIntrinsic>(it))
4674 // Ignore ephemeral values.
4675 if (EphValues.count(it))
4678 unsigned C = getInstructionCost(it, VF);
4680 // Check if we should override the cost.
4681 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
4682 C = ForceTargetInstructionCost;
4685 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
4686 VF << " For instruction: " << *it << '\n');
4689 // We assume that if-converted blocks have a 50% chance of being executed.
4690 // When the code is scalar then some of the blocks are avoided due to CF.
4691 // When the code is vectorized we execute all code paths.
4692 if (VF == 1 && Legal->blockNeedsPredication(*bb))
4701 /// \brief Check whether the address computation for a non-consecutive memory
4702 /// access looks like an unlikely candidate for being merged into the indexing
4705 /// We look for a GEP which has one index that is an induction variable and all
4706 /// other indices are loop invariant. If the stride of this access is also
4707 /// within a small bound we decide that this address computation can likely be
4708 /// merged into the addressing mode.
4709 /// In all other cases, we identify the address computation as complex.
4710 static bool isLikelyComplexAddressComputation(Value *Ptr,
4711 LoopVectorizationLegality *Legal,
4712 ScalarEvolution *SE,
4713 const Loop *TheLoop) {
4714 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
4718 // We are looking for a gep with all loop invariant indices except for one
4719 // which should be an induction variable.
4720 unsigned NumOperands = Gep->getNumOperands();
4721 for (unsigned i = 1; i < NumOperands; ++i) {
4722 Value *Opd = Gep->getOperand(i);
4723 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
4724 !Legal->isInductionVariable(Opd))
4728 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
4729 // can likely be merged into the address computation.
4730 unsigned MaxMergeDistance = 64;
4732 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
4736 // Check the step is constant.
4737 const SCEV *Step = AddRec->getStepRecurrence(*SE);
4738 // Calculate the pointer stride and check if it is consecutive.
4739 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4743 const APInt &APStepVal = C->getValue()->getValue();
4745 // Huge step value - give up.
4746 if (APStepVal.getBitWidth() > 64)
4749 int64_t StepVal = APStepVal.getSExtValue();
4751 return StepVal > MaxMergeDistance;
4754 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
4755 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
4761 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
4762 // If we know that this instruction will remain uniform, check the cost of
4763 // the scalar version.
4764 if (Legal->isUniformAfterVectorization(I))
4767 Type *RetTy = I->getType();
4768 Type *VectorTy = ToVectorTy(RetTy, VF);
4770 // TODO: We need to estimate the cost of intrinsic calls.
4771 switch (I->getOpcode()) {
4772 case Instruction::GetElementPtr:
4773 // We mark this instruction as zero-cost because the cost of GEPs in
4774 // vectorized code depends on whether the corresponding memory instruction
4775 // is scalarized or not. Therefore, we handle GEPs with the memory
4776 // instruction cost.
4778 case Instruction::Br: {
4779 return TTI.getCFInstrCost(I->getOpcode());
4781 case Instruction::PHI:
4782 //TODO: IF-converted IFs become selects.
4784 case Instruction::Add:
4785 case Instruction::FAdd:
4786 case Instruction::Sub:
4787 case Instruction::FSub:
4788 case Instruction::Mul:
4789 case Instruction::FMul:
4790 case Instruction::UDiv:
4791 case Instruction::SDiv:
4792 case Instruction::FDiv:
4793 case Instruction::URem:
4794 case Instruction::SRem:
4795 case Instruction::FRem:
4796 case Instruction::Shl:
4797 case Instruction::LShr:
4798 case Instruction::AShr:
4799 case Instruction::And:
4800 case Instruction::Or:
4801 case Instruction::Xor: {
4802 // Since we will replace the stride by 1 the multiplication should go away.
4803 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
4805 // Certain instructions can be cheaper to vectorize if they have a constant
4806 // second vector operand. One example of this are shifts on x86.
4807 TargetTransformInfo::OperandValueKind Op1VK =
4808 TargetTransformInfo::OK_AnyValue;
4809 TargetTransformInfo::OperandValueKind Op2VK =
4810 TargetTransformInfo::OK_AnyValue;
4811 TargetTransformInfo::OperandValueProperties Op1VP =
4812 TargetTransformInfo::OP_None;
4813 TargetTransformInfo::OperandValueProperties Op2VP =
4814 TargetTransformInfo::OP_None;
4815 Value *Op2 = I->getOperand(1);
4817 // Check for a splat of a constant or for a non uniform vector of constants.
4818 if (isa<ConstantInt>(Op2)) {
4819 ConstantInt *CInt = cast<ConstantInt>(Op2);
4820 if (CInt && CInt->getValue().isPowerOf2())
4821 Op2VP = TargetTransformInfo::OP_PowerOf2;
4822 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4823 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
4824 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
4825 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
4827 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
4828 if (CInt && CInt->getValue().isPowerOf2())
4829 Op2VP = TargetTransformInfo::OP_PowerOf2;
4830 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4834 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
4837 case Instruction::Select: {
4838 SelectInst *SI = cast<SelectInst>(I);
4839 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
4840 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
4841 Type *CondTy = SI->getCondition()->getType();
4843 CondTy = VectorType::get(CondTy, VF);
4845 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
4847 case Instruction::ICmp:
4848 case Instruction::FCmp: {
4849 Type *ValTy = I->getOperand(0)->getType();
4850 VectorTy = ToVectorTy(ValTy, VF);
4851 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
4853 case Instruction::Store:
4854 case Instruction::Load: {
4855 StoreInst *SI = dyn_cast<StoreInst>(I);
4856 LoadInst *LI = dyn_cast<LoadInst>(I);
4857 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
4859 VectorTy = ToVectorTy(ValTy, VF);
4861 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
4862 unsigned AS = SI ? SI->getPointerAddressSpace() :
4863 LI->getPointerAddressSpace();
4864 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
4865 // We add the cost of address computation here instead of with the gep
4866 // instruction because only here we know whether the operation is
4869 return TTI.getAddressComputationCost(VectorTy) +
4870 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4872 // Scalarized loads/stores.
4873 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
4874 bool Reverse = ConsecutiveStride < 0;
4875 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
4876 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
4877 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
4878 bool IsComplexComputation =
4879 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
4881 // The cost of extracting from the value vector and pointer vector.
4882 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
4883 for (unsigned i = 0; i < VF; ++i) {
4884 // The cost of extracting the pointer operand.
4885 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
4886 // In case of STORE, the cost of ExtractElement from the vector.
4887 // In case of LOAD, the cost of InsertElement into the returned
4889 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
4890 Instruction::InsertElement,
4894 // The cost of the scalar loads/stores.
4895 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
4896 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
4901 // Wide load/stores.
4902 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
4903 if (Legal->isMaskRequired(I))
4904 Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment,
4907 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4910 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
4914 case Instruction::ZExt:
4915 case Instruction::SExt:
4916 case Instruction::FPToUI:
4917 case Instruction::FPToSI:
4918 case Instruction::FPExt:
4919 case Instruction::PtrToInt:
4920 case Instruction::IntToPtr:
4921 case Instruction::SIToFP:
4922 case Instruction::UIToFP:
4923 case Instruction::Trunc:
4924 case Instruction::FPTrunc:
4925 case Instruction::BitCast: {
4926 // We optimize the truncation of induction variable.
4927 // The cost of these is the same as the scalar operation.
4928 if (I->getOpcode() == Instruction::Trunc &&
4929 Legal->isInductionVariable(I->getOperand(0)))
4930 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
4931 I->getOperand(0)->getType());
4933 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
4934 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
4936 case Instruction::Call: {
4937 CallInst *CI = cast<CallInst>(I);
4938 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
4939 assert(ID && "Not an intrinsic call!");
4940 Type *RetTy = ToVectorTy(CI->getType(), VF);
4941 SmallVector<Type*, 4> Tys;
4942 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
4943 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
4944 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
4947 // We are scalarizing the instruction. Return the cost of the scalar
4948 // instruction, plus the cost of insert and extract into vector
4949 // elements, times the vector width.
4952 if (!RetTy->isVoidTy() && VF != 1) {
4953 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
4955 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
4958 // The cost of inserting the results plus extracting each one of the
4960 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
4963 // The cost of executing VF copies of the scalar instruction. This opcode
4964 // is unknown. Assume that it is the same as 'mul'.
4965 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
4971 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
4972 if (Scalar->isVoidTy() || VF == 1)
4974 return VectorType::get(Scalar, VF);
4977 char LoopVectorize::ID = 0;
4978 static const char lv_name[] = "Loop Vectorization";
4979 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
4980 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
4981 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
4982 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
4983 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
4984 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
4985 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4986 INITIALIZE_PASS_DEPENDENCY(LCSSA)
4987 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
4988 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4989 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
4992 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
4993 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
4997 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
4998 // Check for a store.
4999 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5000 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5002 // Check for a load.
5003 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5004 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5010 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5011 bool IfPredicateStore) {
5012 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5013 // Holds vector parameters or scalars, in case of uniform vals.
5014 SmallVector<VectorParts, 4> Params;
5016 setDebugLocFromInst(Builder, Instr);
5018 // Find all of the vectorized parameters.
5019 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5020 Value *SrcOp = Instr->getOperand(op);
5022 // If we are accessing the old induction variable, use the new one.
5023 if (SrcOp == OldInduction) {
5024 Params.push_back(getVectorValue(SrcOp));
5028 // Try using previously calculated values.
5029 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5031 // If the src is an instruction that appeared earlier in the basic block
5032 // then it should already be vectorized.
5033 if (SrcInst && OrigLoop->contains(SrcInst)) {
5034 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5035 // The parameter is a vector value from earlier.
5036 Params.push_back(WidenMap.get(SrcInst));
5038 // The parameter is a scalar from outside the loop. Maybe even a constant.
5039 VectorParts Scalars;
5040 Scalars.append(UF, SrcOp);
5041 Params.push_back(Scalars);
5045 assert(Params.size() == Instr->getNumOperands() &&
5046 "Invalid number of operands");
5048 // Does this instruction return a value ?
5049 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5051 Value *UndefVec = IsVoidRetTy ? nullptr :
5052 UndefValue::get(Instr->getType());
5053 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5054 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5056 Instruction *InsertPt = Builder.GetInsertPoint();
5057 BasicBlock *IfBlock = Builder.GetInsertBlock();
5058 BasicBlock *CondBlock = nullptr;
5061 Loop *VectorLp = nullptr;
5062 if (IfPredicateStore) {
5063 assert(Instr->getParent()->getSinglePredecessor() &&
5064 "Only support single predecessor blocks");
5065 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5066 Instr->getParent());
5067 VectorLp = LI->getLoopFor(IfBlock);
5068 assert(VectorLp && "Must have a loop for this block");
5071 // For each vector unroll 'part':
5072 for (unsigned Part = 0; Part < UF; ++Part) {
5073 // For each scalar that we create:
5075 // Start an "if (pred) a[i] = ..." block.
5076 Value *Cmp = nullptr;
5077 if (IfPredicateStore) {
5078 if (Cond[Part]->getType()->isVectorTy())
5080 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5081 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5082 ConstantInt::get(Cond[Part]->getType(), 1));
5083 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5084 LoopVectorBody.push_back(CondBlock);
5085 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
5086 // Update Builder with newly created basic block.
5087 Builder.SetInsertPoint(InsertPt);
5090 Instruction *Cloned = Instr->clone();
5092 Cloned->setName(Instr->getName() + ".cloned");
5093 // Replace the operands of the cloned instructions with extracted scalars.
5094 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5095 Value *Op = Params[op][Part];
5096 Cloned->setOperand(op, Op);
5099 // Place the cloned scalar in the new loop.
5100 Builder.Insert(Cloned);
5102 // If the original scalar returns a value we need to place it in a vector
5103 // so that future users will be able to use it.
5105 VecResults[Part] = Cloned;
5108 if (IfPredicateStore) {
5109 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
5110 LoopVectorBody.push_back(NewIfBlock);
5111 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
5112 Builder.SetInsertPoint(InsertPt);
5113 Instruction *OldBr = IfBlock->getTerminator();
5114 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
5115 OldBr->eraseFromParent();
5116 IfBlock = NewIfBlock;
5121 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5122 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5123 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5125 return scalarizeInstruction(Instr, IfPredicateStore);
5128 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5132 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5136 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) {
5137 // When unrolling and the VF is 1, we only need to add a simple scalar.
5138 Type *ITy = Val->getType();
5139 assert(!ITy->isVectorTy() && "Val must be a scalar");
5140 Constant *C = ConstantInt::get(ITy, StartIdx);
5141 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");