1 //===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops
11 // and generates target-independent LLVM-IR.
12 // The vectorizer uses the TargetTransformInfo analysis to estimate the costs
13 // of instructions in order to estimate the profitability of vectorization.
15 // The loop vectorizer combines consecutive loop iterations into a single
16 // 'wide' iteration. After this transformation the index is incremented
17 // by the SIMD vector width, and not by one.
19 // This pass has three parts:
20 // 1. The main loop pass that drives the different parts.
21 // 2. LoopVectorizationLegality - A unit that checks for the legality
22 // of the vectorization.
23 // 3. InnerLoopVectorizer - A unit that performs the actual
24 // widening of instructions.
25 // 4. LoopVectorizationCostModel - A unit that checks for the profitability
26 // of vectorization. It decides on the optimal vector width, which
27 // can be one, if vectorization is not profitable.
29 //===----------------------------------------------------------------------===//
31 // The reduction-variable vectorization is based on the paper:
32 // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
34 // Variable uniformity checks are inspired by:
35 // Karrenberg, R. and Hack, S. Whole Function Vectorization.
37 // Other ideas/concepts are from:
38 // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
40 // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of
41 // Vectorizing Compilers.
43 //===----------------------------------------------------------------------===//
45 #include "llvm/Transforms/Vectorize.h"
46 #include "llvm/ADT/DenseMap.h"
47 #include "llvm/ADT/EquivalenceClasses.h"
48 #include "llvm/ADT/Hashing.h"
49 #include "llvm/ADT/MapVector.h"
50 #include "llvm/ADT/SetVector.h"
51 #include "llvm/ADT/SmallPtrSet.h"
52 #include "llvm/ADT/SmallSet.h"
53 #include "llvm/ADT/SmallVector.h"
54 #include "llvm/ADT/Statistic.h"
55 #include "llvm/ADT/StringExtras.h"
56 #include "llvm/Analysis/AliasAnalysis.h"
57 #include "llvm/Analysis/AliasSetTracker.h"
58 #include "llvm/Analysis/AssumptionCache.h"
59 #include "llvm/Analysis/BlockFrequencyInfo.h"
60 #include "llvm/Analysis/CodeMetrics.h"
61 #include "llvm/Analysis/LoopAccessAnalysis.h"
62 #include "llvm/Analysis/LoopInfo.h"
63 #include "llvm/Analysis/LoopIterator.h"
64 #include "llvm/Analysis/LoopPass.h"
65 #include "llvm/Analysis/ScalarEvolution.h"
66 #include "llvm/Analysis/ScalarEvolutionExpander.h"
67 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
68 #include "llvm/Analysis/TargetTransformInfo.h"
69 #include "llvm/Analysis/ValueTracking.h"
70 #include "llvm/IR/Constants.h"
71 #include "llvm/IR/DataLayout.h"
72 #include "llvm/IR/DebugInfo.h"
73 #include "llvm/IR/DerivedTypes.h"
74 #include "llvm/IR/DiagnosticInfo.h"
75 #include "llvm/IR/Dominators.h"
76 #include "llvm/IR/Function.h"
77 #include "llvm/IR/IRBuilder.h"
78 #include "llvm/IR/Instructions.h"
79 #include "llvm/IR/IntrinsicInst.h"
80 #include "llvm/IR/LLVMContext.h"
81 #include "llvm/IR/Module.h"
82 #include "llvm/IR/PatternMatch.h"
83 #include "llvm/IR/Type.h"
84 #include "llvm/IR/Value.h"
85 #include "llvm/IR/ValueHandle.h"
86 #include "llvm/IR/Verifier.h"
87 #include "llvm/Pass.h"
88 #include "llvm/Support/BranchProbability.h"
89 #include "llvm/Support/CommandLine.h"
90 #include "llvm/Support/Debug.h"
91 #include "llvm/Support/raw_ostream.h"
92 #include "llvm/Transforms/Scalar.h"
93 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
94 #include "llvm/Transforms/Utils/Local.h"
95 #include "llvm/Transforms/Utils/VectorUtils.h"
100 using namespace llvm;
101 using namespace llvm::PatternMatch;
103 #define LV_NAME "loop-vectorize"
104 #define DEBUG_TYPE LV_NAME
106 STATISTIC(LoopsVectorized, "Number of loops vectorized");
107 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
110 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
111 cl::desc("Enable if-conversion during vectorization."));
113 /// We don't vectorize loops with a known constant trip count below this number.
114 static cl::opt<unsigned>
115 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
117 cl::desc("Don't vectorize loops with a constant "
118 "trip count that is smaller than this "
121 /// This enables versioning on the strides of symbolically striding memory
122 /// accesses in code like the following.
123 /// for (i = 0; i < N; ++i)
124 /// A[i * Stride1] += B[i * Stride2] ...
126 /// Will be roughly translated to
127 /// if (Stride1 == 1 && Stride2 == 1) {
128 /// for (i = 0; i < N; i+=4)
132 static cl::opt<bool> EnableMemAccessVersioning(
133 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
134 cl::desc("Enable symblic stride memory access versioning"));
136 /// We don't unroll loops with a known constant trip count below this number.
137 static const unsigned TinyTripCountUnrollThreshold = 128;
139 static cl::opt<unsigned> ForceTargetNumScalarRegs(
140 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
141 cl::desc("A flag that overrides the target's number of scalar registers."));
143 static cl::opt<unsigned> ForceTargetNumVectorRegs(
144 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
145 cl::desc("A flag that overrides the target's number of vector registers."));
147 /// Maximum vectorization interleave count.
148 static const unsigned MaxInterleaveFactor = 16;
150 static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
151 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
152 cl::desc("A flag that overrides the target's max interleave factor for "
155 static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
156 "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
157 cl::desc("A flag that overrides the target's max interleave factor for "
158 "vectorized loops."));
160 static cl::opt<unsigned> ForceTargetInstructionCost(
161 "force-target-instruction-cost", cl::init(0), cl::Hidden,
162 cl::desc("A flag that overrides the target's expected cost for "
163 "an instruction to a single constant value. Mostly "
164 "useful for getting consistent testing."));
166 static cl::opt<unsigned> SmallLoopCost(
167 "small-loop-cost", cl::init(20), cl::Hidden,
168 cl::desc("The cost of a loop that is considered 'small' by the unroller."));
170 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
171 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
172 cl::desc("Enable the use of the block frequency analysis to access PGO "
173 "heuristics minimizing code growth in cold regions and being more "
174 "aggressive in hot regions."));
176 // Runtime unroll loops for load/store throughput.
177 static cl::opt<bool> EnableLoadStoreRuntimeUnroll(
178 "enable-loadstore-runtime-unroll", cl::init(true), cl::Hidden,
179 cl::desc("Enable runtime unrolling until load/store ports are saturated"));
181 /// The number of stores in a loop that are allowed to need predication.
182 static cl::opt<unsigned> NumberOfStoresToPredicate(
183 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
184 cl::desc("Max number of stores to be predicated behind an if."));
186 static cl::opt<bool> EnableIndVarRegisterHeur(
187 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
188 cl::desc("Count the induction variable only once when unrolling"));
190 static cl::opt<bool> EnableCondStoresVectorization(
191 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
192 cl::desc("Enable if predication of stores during vectorization."));
194 static cl::opt<unsigned> MaxNestedScalarReductionUF(
195 "max-nested-scalar-reduction-unroll", cl::init(2), cl::Hidden,
196 cl::desc("The maximum unroll factor to use when unrolling a scalar "
197 "reduction in a nested loop."));
201 // Forward declarations.
202 class LoopVectorizationLegality;
203 class LoopVectorizationCostModel;
204 class LoopVectorizeHints;
206 /// \brief This modifies LoopAccessReport to initialize message with
207 /// loop-vectorizer-specific part.
208 class VectorizationReport : public LoopAccessReport {
210 VectorizationReport(Instruction *I = nullptr)
211 : LoopAccessReport("loop not vectorized: ", I) {}
213 /// \brief This allows promotion of the loop-access analysis report into the
214 /// loop-vectorizer report. It modifies the message to add the
215 /// loop-vectorizer-specific part of the message.
216 explicit VectorizationReport(const LoopAccessReport &R)
217 : LoopAccessReport(Twine("loop not vectorized: ") + R.str(),
221 /// A helper function for converting Scalar types to vector types.
222 /// If the incoming type is void, we return void. If the VF is 1, we return
224 static Type* ToVectorTy(Type *Scalar, unsigned VF) {
225 if (Scalar->isVoidTy() || VF == 1)
227 return VectorType::get(Scalar, VF);
230 /// InnerLoopVectorizer vectorizes loops which contain only one basic
231 /// block to a specified vectorization factor (VF).
232 /// This class performs the widening of scalars into vectors, or multiple
233 /// scalars. This class also implements the following features:
234 /// * It inserts an epilogue loop for handling loops that don't have iteration
235 /// counts that are known to be a multiple of the vectorization factor.
236 /// * It handles the code generation for reduction variables.
237 /// * Scalarization (implementation using scalars) of un-vectorizable
239 /// InnerLoopVectorizer does not perform any vectorization-legality
240 /// checks, and relies on the caller to check for the different legality
241 /// aspects. The InnerLoopVectorizer relies on the
242 /// LoopVectorizationLegality class to provide information about the induction
243 /// and reduction variables that were found to a given vectorization factor.
244 class InnerLoopVectorizer {
246 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
247 DominatorTree *DT, const DataLayout *DL,
248 const TargetLibraryInfo *TLI, unsigned VecWidth,
249 unsigned UnrollFactor)
250 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
251 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
252 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
255 // Perform the actual loop widening (vectorization).
256 void vectorize(LoopVectorizationLegality *L) {
258 // Create a new empty loop. Unlink the old loop and connect the new one.
260 // Widen each instruction in the old loop to a new one in the new loop.
261 // Use the Legality module to find the induction and reduction variables.
263 // Register the new loop and update the analysis passes.
267 virtual ~InnerLoopVectorizer() {}
270 /// A small list of PHINodes.
271 typedef SmallVector<PHINode*, 4> PhiVector;
272 /// When we unroll loops we have multiple vector values for each scalar.
273 /// This data structure holds the unrolled and vectorized values that
274 /// originated from one scalar instruction.
275 typedef SmallVector<Value*, 2> VectorParts;
277 // When we if-convert we need create edge masks. We have to cache values so
278 // that we don't end up with exponential recursion/IR.
279 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
280 VectorParts> EdgeMaskCache;
282 /// \brief Add checks for strides that where assumed to be 1.
284 /// Returns the last check instruction and the first check instruction in the
285 /// pair as (first, last).
286 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
288 /// Create an empty loop, based on the loop ranges of the old loop.
289 void createEmptyLoop();
290 /// Copy and widen the instructions from the old loop.
291 virtual void vectorizeLoop();
293 /// \brief The Loop exit block may have single value PHI nodes where the
294 /// incoming value is 'Undef'. While vectorizing we only handled real values
295 /// that were defined inside the loop. Here we fix the 'undef case'.
299 /// A helper function that computes the predicate of the block BB, assuming
300 /// that the header block of the loop is set to True. It returns the *entry*
301 /// mask for the block BB.
302 VectorParts createBlockInMask(BasicBlock *BB);
303 /// A helper function that computes the predicate of the edge between SRC
305 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
307 /// A helper function to vectorize a single BB within the innermost loop.
308 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
310 /// Vectorize a single PHINode in a block. This method handles the induction
311 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
312 /// arbitrary length vectors.
313 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
314 unsigned UF, unsigned VF, PhiVector *PV);
316 /// Insert the new loop to the loop hierarchy and pass manager
317 /// and update the analysis passes.
318 void updateAnalysis();
320 /// This instruction is un-vectorizable. Implement it as a sequence
321 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
322 /// scalarized instruction behind an if block predicated on the control
323 /// dependence of the instruction.
324 virtual void scalarizeInstruction(Instruction *Instr,
325 bool IfPredicateStore=false);
327 /// Vectorize Load and Store instructions,
328 virtual void vectorizeMemoryInstruction(Instruction *Instr);
330 /// Create a broadcast instruction. This method generates a broadcast
331 /// instruction (shuffle) for loop invariant values and for the induction
332 /// value. If this is the induction variable then we extend it to N, N+1, ...
333 /// this is needed because each iteration in the loop corresponds to a SIMD
335 virtual Value *getBroadcastInstrs(Value *V);
337 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
338 /// to each vector element of Val. The sequence starts at StartIndex.
339 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step);
341 /// When we go over instructions in the basic block we rely on previous
342 /// values within the current basic block or on loop invariant values.
343 /// When we widen (vectorize) values we place them in the map. If the values
344 /// are not within the map, they have to be loop invariant, so we simply
345 /// broadcast them into a vector.
346 VectorParts &getVectorValue(Value *V);
348 /// Generate a shuffle sequence that will reverse the vector Vec.
349 virtual Value *reverseVector(Value *Vec);
351 /// This is a helper class that holds the vectorizer state. It maps scalar
352 /// instructions to vector instructions. When the code is 'unrolled' then
353 /// then a single scalar value is mapped to multiple vector parts. The parts
354 /// are stored in the VectorPart type.
356 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
358 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
360 /// \return True if 'Key' is saved in the Value Map.
361 bool has(Value *Key) const { return MapStorage.count(Key); }
363 /// Initializes a new entry in the map. Sets all of the vector parts to the
364 /// save value in 'Val'.
365 /// \return A reference to a vector with splat values.
366 VectorParts &splat(Value *Key, Value *Val) {
367 VectorParts &Entry = MapStorage[Key];
368 Entry.assign(UF, Val);
372 ///\return A reference to the value that is stored at 'Key'.
373 VectorParts &get(Value *Key) {
374 VectorParts &Entry = MapStorage[Key];
377 assert(Entry.size() == UF);
382 /// The unroll factor. Each entry in the map stores this number of vector
386 /// Map storage. We use std::map and not DenseMap because insertions to a
387 /// dense map invalidates its iterators.
388 std::map<Value *, VectorParts> MapStorage;
391 /// The original loop.
393 /// Scev analysis to use.
402 const DataLayout *DL;
403 /// Target Library Info.
404 const TargetLibraryInfo *TLI;
406 /// The vectorization SIMD factor to use. Each vector will have this many
411 /// The vectorization unroll factor to use. Each scalar is vectorized to this
412 /// many different vector instructions.
415 /// The builder that we use
418 // --- Vectorization state ---
420 /// The vector-loop preheader.
421 BasicBlock *LoopVectorPreHeader;
422 /// The scalar-loop preheader.
423 BasicBlock *LoopScalarPreHeader;
424 /// Middle Block between the vector and the scalar.
425 BasicBlock *LoopMiddleBlock;
426 ///The ExitBlock of the scalar loop.
427 BasicBlock *LoopExitBlock;
428 ///The vector loop body.
429 SmallVector<BasicBlock *, 4> LoopVectorBody;
430 ///The scalar loop body.
431 BasicBlock *LoopScalarBody;
432 /// A list of all bypass blocks. The first block is the entry of the loop.
433 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
435 /// The new Induction variable which was added to the new block.
437 /// The induction variable of the old basic block.
438 PHINode *OldInduction;
439 /// Holds the extended (to the widest induction type) start index.
441 /// Maps scalars to widened vectors.
443 EdgeMaskCache MaskCache;
445 LoopVectorizationLegality *Legal;
448 class InnerLoopUnroller : public InnerLoopVectorizer {
450 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
451 DominatorTree *DT, const DataLayout *DL,
452 const TargetLibraryInfo *TLI, unsigned UnrollFactor) :
453 InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { }
456 void scalarizeInstruction(Instruction *Instr,
457 bool IfPredicateStore = false) override;
458 void vectorizeMemoryInstruction(Instruction *Instr) override;
459 Value *getBroadcastInstrs(Value *V) override;
460 Value *getStepVector(Value *Val, int StartIdx, Value *Step) override;
461 Value *reverseVector(Value *Vec) override;
464 /// \brief Look for a meaningful debug location on the instruction or it's
466 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
471 if (I->getDebugLoc() != Empty)
474 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
475 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
476 if (OpInst->getDebugLoc() != Empty)
483 /// \brief Set the debug location in the builder using the debug location in the
485 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
486 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
487 B.SetCurrentDebugLocation(Inst->getDebugLoc());
489 B.SetCurrentDebugLocation(DebugLoc());
493 /// \return string containing a file name and a line # for the given loop.
494 static std::string getDebugLocString(const Loop *L) {
497 raw_string_ostream OS(Result);
498 const DebugLoc LoopDbgLoc = L->getStartLoc();
499 if (!LoopDbgLoc.isUnknown())
500 LoopDbgLoc.print(OS);
502 // Just print the module name.
503 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
510 /// \brief Propagate known metadata from one instruction to another.
511 static void propagateMetadata(Instruction *To, const Instruction *From) {
512 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
513 From->getAllMetadataOtherThanDebugLoc(Metadata);
515 for (auto M : Metadata) {
516 unsigned Kind = M.first;
518 // These are safe to transfer (this is safe for TBAA, even when we
519 // if-convert, because should that metadata have had a control dependency
520 // on the condition, and thus actually aliased with some other
521 // non-speculated memory access when the condition was false, this would be
522 // caught by the runtime overlap checks).
523 if (Kind != LLVMContext::MD_tbaa &&
524 Kind != LLVMContext::MD_alias_scope &&
525 Kind != LLVMContext::MD_noalias &&
526 Kind != LLVMContext::MD_fpmath)
529 To->setMetadata(Kind, M.second);
533 /// \brief Propagate known metadata from one instruction to a vector of others.
534 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
536 if (Instruction *I = dyn_cast<Instruction>(V))
537 propagateMetadata(I, From);
540 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
541 /// to what vectorization factor.
542 /// This class does not look at the profitability of vectorization, only the
543 /// legality. This class has two main kinds of checks:
544 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
545 /// will change the order of memory accesses in a way that will change the
546 /// correctness of the program.
547 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
548 /// checks for a number of different conditions, such as the availability of a
549 /// single induction variable, that all types are supported and vectorize-able,
550 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
551 /// This class is also used by InnerLoopVectorizer for identifying
552 /// induction variable and the different reduction variables.
553 class LoopVectorizationLegality {
555 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, const DataLayout *DL,
556 DominatorTree *DT, TargetLibraryInfo *TLI,
557 AliasAnalysis *AA, Function *F,
558 const TargetTransformInfo *TTI,
559 LoopAccessAnalysis *LAA)
560 : NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
561 TLI(TLI), TheFunction(F), TTI(TTI), DT(DT), LAA(LAA), LAI(nullptr),
562 Induction(nullptr), WidestIndTy(nullptr), HasFunNoNaNAttr(false) {}
564 /// This enum represents the kinds of reductions that we support.
566 RK_NoReduction, ///< Not a reduction.
567 RK_IntegerAdd, ///< Sum of integers.
568 RK_IntegerMult, ///< Product of integers.
569 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
570 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
571 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
572 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
573 RK_FloatAdd, ///< Sum of floats.
574 RK_FloatMult, ///< Product of floats.
575 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
578 /// This enum represents the kinds of inductions that we support.
580 IK_NoInduction, ///< Not an induction variable.
581 IK_IntInduction, ///< Integer induction variable. Step = C.
582 IK_PtrInduction ///< Pointer induction var. Step = C / sizeof(elem).
585 // This enum represents the kind of minmax reduction.
586 enum MinMaxReductionKind {
596 /// This struct holds information about reduction variables.
597 struct ReductionDescriptor {
598 ReductionDescriptor() : StartValue(nullptr), LoopExitInstr(nullptr),
599 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
601 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
602 MinMaxReductionKind MK)
603 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
605 // The starting value of the reduction.
606 // It does not have to be zero!
607 TrackingVH<Value> StartValue;
608 // The instruction who's value is used outside the loop.
609 Instruction *LoopExitInstr;
610 // The kind of the reduction.
612 // If this a min/max reduction the kind of reduction.
613 MinMaxReductionKind MinMaxKind;
616 /// This POD struct holds information about a potential reduction operation.
617 struct ReductionInstDesc {
618 ReductionInstDesc(bool IsRedux, Instruction *I) :
619 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
621 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
622 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
624 // Is this instruction a reduction candidate.
626 // The last instruction in a min/max pattern (select of the select(icmp())
627 // pattern), or the current reduction instruction otherwise.
628 Instruction *PatternLastInst;
629 // If this is a min/max pattern the comparison predicate.
630 MinMaxReductionKind MinMaxKind;
633 /// A struct for saving information about induction variables.
634 struct InductionInfo {
635 InductionInfo(Value *Start, InductionKind K, ConstantInt *Step)
636 : StartValue(Start), IK(K), StepValue(Step) {
637 assert(IK != IK_NoInduction && "Not an induction");
638 assert(StartValue && "StartValue is null");
639 assert(StepValue && !StepValue->isZero() && "StepValue is zero");
640 assert((IK != IK_PtrInduction || StartValue->getType()->isPointerTy()) &&
641 "StartValue is not a pointer for pointer induction");
642 assert((IK != IK_IntInduction || StartValue->getType()->isIntegerTy()) &&
643 "StartValue is not an integer for integer induction");
644 assert(StepValue->getType()->isIntegerTy() &&
645 "StepValue is not an integer");
648 : StartValue(nullptr), IK(IK_NoInduction), StepValue(nullptr) {}
650 /// Get the consecutive direction. Returns:
651 /// 0 - unknown or non-consecutive.
652 /// 1 - consecutive and increasing.
653 /// -1 - consecutive and decreasing.
654 int getConsecutiveDirection() const {
655 if (StepValue && (StepValue->isOne() || StepValue->isMinusOne()))
656 return StepValue->getSExtValue();
660 /// Compute the transformed value of Index at offset StartValue using step
662 /// For integer induction, returns StartValue + Index * StepValue.
663 /// For pointer induction, returns StartValue[Index * StepValue].
664 /// FIXME: The newly created binary instructions should contain nsw/nuw
665 /// flags, which can be found from the original scalar operations.
666 Value *transform(IRBuilder<> &B, Value *Index) const {
668 case IK_IntInduction:
669 assert(Index->getType() == StartValue->getType() &&
670 "Index type does not match StartValue type");
671 if (StepValue->isMinusOne())
672 return B.CreateSub(StartValue, Index);
673 if (!StepValue->isOne())
674 Index = B.CreateMul(Index, StepValue);
675 return B.CreateAdd(StartValue, Index);
677 case IK_PtrInduction:
678 if (StepValue->isMinusOne())
679 Index = B.CreateNeg(Index);
680 else if (!StepValue->isOne())
681 Index = B.CreateMul(Index, StepValue);
682 return B.CreateGEP(StartValue, Index);
687 llvm_unreachable("invalid enum");
691 TrackingVH<Value> StartValue;
695 ConstantInt *StepValue;
698 /// ReductionList contains the reduction descriptors for all
699 /// of the reductions that were found in the loop.
700 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
702 /// InductionList saves induction variables and maps them to the
703 /// induction descriptor.
704 typedef MapVector<PHINode*, InductionInfo> InductionList;
706 /// Returns true if it is legal to vectorize this loop.
707 /// This does not mean that it is profitable to vectorize this
708 /// loop, only that it is legal to do so.
711 /// Returns the Induction variable.
712 PHINode *getInduction() { return Induction; }
714 /// Returns the reduction variables found in the loop.
715 ReductionList *getReductionVars() { return &Reductions; }
717 /// Returns the induction variables found in the loop.
718 InductionList *getInductionVars() { return &Inductions; }
720 /// Returns the widest induction type.
721 Type *getWidestInductionType() { return WidestIndTy; }
723 /// Returns True if V is an induction variable in this loop.
724 bool isInductionVariable(const Value *V);
726 /// Return true if the block BB needs to be predicated in order for the loop
727 /// to be vectorized.
728 bool blockNeedsPredication(BasicBlock *BB);
730 /// Check if this pointer is consecutive when vectorizing. This happens
731 /// when the last index of the GEP is the induction variable, or that the
732 /// pointer itself is an induction variable.
733 /// This check allows us to vectorize A[idx] into a wide load/store.
735 /// 0 - Stride is unknown or non-consecutive.
736 /// 1 - Address is consecutive.
737 /// -1 - Address is consecutive, and decreasing.
738 int isConsecutivePtr(Value *Ptr);
740 /// Returns true if the value V is uniform within the loop.
741 bool isUniform(Value *V);
743 /// Returns true if this instruction will remain scalar after vectorization.
744 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
746 /// Returns the information that we collected about runtime memory check.
747 const LoopAccessInfo::RuntimePointerCheck *getRuntimePointerCheck() const {
748 return LAI->getRuntimePointerCheck();
751 const LoopAccessInfo *getLAI() const {
755 /// This function returns the identity element (or neutral element) for
757 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
759 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
761 bool hasStride(Value *V) { return StrideSet.count(V); }
762 bool mustCheckStrides() { return !StrideSet.empty(); }
763 SmallPtrSet<Value *, 8>::iterator strides_begin() {
764 return StrideSet.begin();
766 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
768 /// Returns true if the target machine supports masked store operation
769 /// for the given \p DataType and kind of access to \p Ptr.
770 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
771 return TTI->isLegalMaskedStore(DataType, isConsecutivePtr(Ptr));
773 /// Returns true if the target machine supports masked load operation
774 /// for the given \p DataType and kind of access to \p Ptr.
775 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
776 return TTI->isLegalMaskedLoad(DataType, isConsecutivePtr(Ptr));
778 /// Returns true if vector representation of the instruction \p I
780 bool isMaskRequired(const Instruction* I) {
781 return (MaskedOp.count(I) != 0);
783 unsigned getNumStores() const {
784 return LAI->getNumStores();
786 unsigned getNumLoads() const {
787 return LAI->getNumLoads();
789 unsigned getNumPredStores() const {
790 return NumPredStores;
793 /// Check if a single basic block loop is vectorizable.
794 /// At this point we know that this is a loop with a constant trip count
795 /// and we only need to check individual instructions.
796 bool canVectorizeInstrs();
798 /// When we vectorize loops we may change the order in which
799 /// we read and write from memory. This method checks if it is
800 /// legal to vectorize the code, considering only memory constrains.
801 /// Returns true if the loop is vectorizable
802 bool canVectorizeMemory();
804 /// Return true if we can vectorize this loop using the IF-conversion
806 bool canVectorizeWithIfConvert();
808 /// Collect the variables that need to stay uniform after vectorization.
809 void collectLoopUniforms();
811 /// Return true if all of the instructions in the block can be speculatively
812 /// executed. \p SafePtrs is a list of addresses that are known to be legal
813 /// and we know that we can read from them without segfault.
814 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
816 /// Returns True, if 'Phi' is the kind of reduction variable for type
817 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
818 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
819 /// Returns a struct describing if the instruction 'I' can be a reduction
820 /// variable of type 'Kind'. If the reduction is a min/max pattern of
821 /// select(icmp()) this function advances the instruction pointer 'I' from the
822 /// compare instruction to the select instruction and stores this pointer in
823 /// 'PatternLastInst' member of the returned struct.
824 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
825 ReductionInstDesc &Desc);
826 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
827 /// pattern corresponding to a min(X, Y) or max(X, Y).
828 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
829 ReductionInstDesc &Prev);
830 /// Returns the induction kind of Phi and record the step. This function may
831 /// return NoInduction if the PHI is not an induction variable.
832 InductionKind isInductionVariable(PHINode *Phi, ConstantInt *&StepValue);
834 /// \brief Collect memory access with loop invariant strides.
836 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
838 void collectStridedAccess(Value *LoadOrStoreInst);
840 /// Report an analysis message to assist the user in diagnosing loops that are
841 /// not vectorized. These are handled as LoopAccessReport rather than
842 /// VectorizationReport because the << operator of VectorizationReport returns
843 /// LoopAccessReport.
844 void emitAnalysis(const LoopAccessReport &Message) {
845 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, LV_NAME);
848 unsigned NumPredStores;
850 /// The loop that we evaluate.
854 /// DataLayout analysis.
855 const DataLayout *DL;
856 /// Target Library Info.
857 TargetLibraryInfo *TLI;
859 Function *TheFunction;
860 /// Target Transform Info
861 const TargetTransformInfo *TTI;
864 // LoopAccess analysis.
865 LoopAccessAnalysis *LAA;
866 // And the loop-accesses info corresponding to this loop. This pointer is
867 // null until canVectorizeMemory sets it up.
868 const LoopAccessInfo *LAI;
870 // --- vectorization state --- //
872 /// Holds the integer induction variable. This is the counter of the
875 /// Holds the reduction variables.
876 ReductionList Reductions;
877 /// Holds all of the induction variables that we found in the loop.
878 /// Notice that inductions don't need to start at zero and that induction
879 /// variables can be pointers.
880 InductionList Inductions;
881 /// Holds the widest induction type encountered.
884 /// Allowed outside users. This holds the reduction
885 /// vars which can be accessed from outside the loop.
886 SmallPtrSet<Value*, 4> AllowedExit;
887 /// This set holds the variables which are known to be uniform after
889 SmallPtrSet<Instruction*, 4> Uniforms;
891 /// Can we assume the absence of NaNs.
892 bool HasFunNoNaNAttr;
894 ValueToValueMap Strides;
895 SmallPtrSet<Value *, 8> StrideSet;
897 /// While vectorizing these instructions we have to generate a
898 /// call to the appropriate masked intrinsic
899 SmallPtrSet<const Instruction*, 8> MaskedOp;
902 /// LoopVectorizationCostModel - estimates the expected speedups due to
904 /// In many cases vectorization is not profitable. This can happen because of
905 /// a number of reasons. In this class we mainly attempt to predict the
906 /// expected speedup/slowdowns due to the supported instruction set. We use the
907 /// TargetTransformInfo to query the different backends for the cost of
908 /// different operations.
909 class LoopVectorizationCostModel {
911 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
912 LoopVectorizationLegality *Legal,
913 const TargetTransformInfo &TTI,
914 const DataLayout *DL, const TargetLibraryInfo *TLI,
915 AssumptionCache *AC, const Function *F,
916 const LoopVectorizeHints *Hints)
917 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI),
918 TheFunction(F), Hints(Hints) {
919 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
922 /// Information about vectorization costs
923 struct VectorizationFactor {
924 unsigned Width; // Vector width with best cost
925 unsigned Cost; // Cost of the loop with that width
927 /// \return The most profitable vectorization factor and the cost of that VF.
928 /// This method checks every power of two up to VF. If UserVF is not ZERO
929 /// then this vectorization factor will be selected if vectorization is
931 VectorizationFactor selectVectorizationFactor(bool OptForSize);
933 /// \return The size (in bits) of the widest type in the code that
934 /// needs to be vectorized. We ignore values that remain scalar such as
935 /// 64 bit loop indices.
936 unsigned getWidestType();
938 /// \return The most profitable unroll factor.
939 /// If UserUF is non-zero then this method finds the best unroll-factor
940 /// based on register pressure and other parameters.
941 /// VF and LoopCost are the selected vectorization factor and the cost of the
943 unsigned selectUnrollFactor(bool OptForSize, unsigned VF, unsigned LoopCost);
945 /// \brief A struct that represents some properties of the register usage
947 struct RegisterUsage {
948 /// Holds the number of loop invariant values that are used in the loop.
949 unsigned LoopInvariantRegs;
950 /// Holds the maximum number of concurrent live intervals in the loop.
951 unsigned MaxLocalUsers;
952 /// Holds the number of instructions in the loop.
953 unsigned NumInstructions;
956 /// \return information about the register usage of the loop.
957 RegisterUsage calculateRegisterUsage();
960 /// Returns the expected execution cost. The unit of the cost does
961 /// not matter because we use the 'cost' units to compare different
962 /// vector widths. The cost that is returned is *not* normalized by
963 /// the factor width.
964 unsigned expectedCost(unsigned VF);
966 /// Returns the execution time cost of an instruction for a given vector
967 /// width. Vector width of one means scalar.
968 unsigned getInstructionCost(Instruction *I, unsigned VF);
970 /// Returns whether the instruction is a load or store and will be a emitted
971 /// as a vector operation.
972 bool isConsecutiveLoadOrStore(Instruction *I);
974 /// Report an analysis message to assist the user in diagnosing loops that are
975 /// not vectorized. These are handled as LoopAccessReport rather than
976 /// VectorizationReport because the << operator of VectorizationReport returns
977 /// LoopAccessReport.
978 void emitAnalysis(const LoopAccessReport &Message) {
979 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, LV_NAME);
982 /// Values used only by @llvm.assume calls.
983 SmallPtrSet<const Value *, 32> EphValues;
985 /// The loop that we evaluate.
989 /// Loop Info analysis.
991 /// Vectorization legality.
992 LoopVectorizationLegality *Legal;
993 /// Vector target information.
994 const TargetTransformInfo &TTI;
995 /// Target data layout information.
996 const DataLayout *DL;
997 /// Target Library Info.
998 const TargetLibraryInfo *TLI;
999 const Function *TheFunction;
1000 // Loop Vectorize Hint.
1001 const LoopVectorizeHints *Hints;
1004 /// Utility class for getting and setting loop vectorizer hints in the form
1005 /// of loop metadata.
1006 /// This class keeps a number of loop annotations locally (as member variables)
1007 /// and can, upon request, write them back as metadata on the loop. It will
1008 /// initially scan the loop for existing metadata, and will update the local
1009 /// values based on information in the loop.
1010 /// We cannot write all values to metadata, as the mere presence of some info,
1011 /// for example 'force', means a decision has been made. So, we need to be
1012 /// careful NOT to add them if the user hasn't specifically asked so.
1013 class LoopVectorizeHints {
1020 /// Hint - associates name and validation with the hint value.
1023 unsigned Value; // This may have to change for non-numeric values.
1026 Hint(const char * Name, unsigned Value, HintKind Kind)
1027 : Name(Name), Value(Value), Kind(Kind) { }
1029 bool validate(unsigned Val) {
1032 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
1034 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1042 /// Vectorization width.
1044 /// Vectorization interleave factor.
1046 /// Vectorization forced
1049 /// Return the loop metadata prefix.
1050 static StringRef Prefix() { return "llvm.loop."; }
1054 FK_Undefined = -1, ///< Not selected.
1055 FK_Disabled = 0, ///< Forcing disabled.
1056 FK_Enabled = 1, ///< Forcing enabled.
1059 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
1060 : Width("vectorize.width", VectorizerParams::VectorizationFactor,
1062 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1063 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1065 // Populate values with existing loop metadata.
1066 getHintsFromMetadata();
1068 // force-vector-interleave overrides DisableInterleaving.
1069 if (VectorizerParams::isInterleaveForced())
1070 Interleave.Value = VectorizerParams::VectorizationInterleave;
1072 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
1073 << "LV: Interleaving disabled by the pass manager\n");
1076 /// Mark the loop L as already vectorized by setting the width to 1.
1077 void setAlreadyVectorized() {
1078 Width.Value = Interleave.Value = 1;
1079 Hint Hints[] = {Width, Interleave};
1080 writeHintsToMetadata(Hints);
1083 /// Dumps all the hint information.
1084 std::string emitRemark() const {
1085 VectorizationReport R;
1086 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1087 R << "vectorization is explicitly disabled";
1089 R << "use -Rpass-analysis=loop-vectorize for more info";
1090 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1091 R << " (Force=true";
1092 if (Width.Value != 0)
1093 R << ", Vector Width=" << Width.Value;
1094 if (Interleave.Value != 0)
1095 R << ", Interleave Count=" << Interleave.Value;
1103 unsigned getWidth() const { return Width.Value; }
1104 unsigned getInterleave() const { return Interleave.Value; }
1105 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1108 /// Find hints specified in the loop metadata and update local values.
1109 void getHintsFromMetadata() {
1110 MDNode *LoopID = TheLoop->getLoopID();
1114 // First operand should refer to the loop id itself.
1115 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1116 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1118 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1119 const MDString *S = nullptr;
1120 SmallVector<Metadata *, 4> Args;
1122 // The expected hint is either a MDString or a MDNode with the first
1123 // operand a MDString.
1124 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1125 if (!MD || MD->getNumOperands() == 0)
1127 S = dyn_cast<MDString>(MD->getOperand(0));
1128 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1129 Args.push_back(MD->getOperand(i));
1131 S = dyn_cast<MDString>(LoopID->getOperand(i));
1132 assert(Args.size() == 0 && "too many arguments for MDString");
1138 // Check if the hint starts with the loop metadata prefix.
1139 StringRef Name = S->getString();
1140 if (Args.size() == 1)
1141 setHint(Name, Args[0]);
1145 /// Checks string hint with one operand and set value if valid.
1146 void setHint(StringRef Name, Metadata *Arg) {
1147 if (!Name.startswith(Prefix()))
1149 Name = Name.substr(Prefix().size(), StringRef::npos);
1151 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1153 unsigned Val = C->getZExtValue();
1155 Hint *Hints[] = {&Width, &Interleave, &Force};
1156 for (auto H : Hints) {
1157 if (Name == H->Name) {
1158 if (H->validate(Val))
1161 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1167 /// Create a new hint from name / value pair.
1168 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1169 LLVMContext &Context = TheLoop->getHeader()->getContext();
1170 Metadata *MDs[] = {MDString::get(Context, Name),
1171 ConstantAsMetadata::get(
1172 ConstantInt::get(Type::getInt32Ty(Context), V))};
1173 return MDNode::get(Context, MDs);
1176 /// Matches metadata with hint name.
1177 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1178 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1182 for (auto H : HintTypes)
1183 if (Name->getString().endswith(H.Name))
1188 /// Sets current hints into loop metadata, keeping other values intact.
1189 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1190 if (HintTypes.size() == 0)
1193 // Reserve the first element to LoopID (see below).
1194 SmallVector<Metadata *, 4> MDs(1);
1195 // If the loop already has metadata, then ignore the existing operands.
1196 MDNode *LoopID = TheLoop->getLoopID();
1198 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1199 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1200 // If node in update list, ignore old value.
1201 if (!matchesHintMetadataName(Node, HintTypes))
1202 MDs.push_back(Node);
1206 // Now, add the missing hints.
1207 for (auto H : HintTypes)
1208 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1210 // Replace current metadata node with new one.
1211 LLVMContext &Context = TheLoop->getHeader()->getContext();
1212 MDNode *NewLoopID = MDNode::get(Context, MDs);
1213 // Set operand 0 to refer to the loop id itself.
1214 NewLoopID->replaceOperandWith(0, NewLoopID);
1216 TheLoop->setLoopID(NewLoopID);
1219 /// The loop these hints belong to.
1220 const Loop *TheLoop;
1223 static void emitMissedWarning(Function *F, Loop *L,
1224 const LoopVectorizeHints &LH) {
1225 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1226 L->getStartLoc(), LH.emitRemark());
1228 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1229 if (LH.getWidth() != 1)
1230 emitLoopVectorizeWarning(
1231 F->getContext(), *F, L->getStartLoc(),
1232 "failed explicitly specified loop vectorization");
1233 else if (LH.getInterleave() != 1)
1234 emitLoopInterleaveWarning(
1235 F->getContext(), *F, L->getStartLoc(),
1236 "failed explicitly specified loop interleaving");
1240 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1242 return V.push_back(&L);
1244 for (Loop *InnerL : L)
1245 addInnerLoop(*InnerL, V);
1248 /// The LoopVectorize Pass.
1249 struct LoopVectorize : public FunctionPass {
1250 /// Pass identification, replacement for typeid
1253 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1255 DisableUnrolling(NoUnrolling),
1256 AlwaysVectorize(AlwaysVectorize) {
1257 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1260 ScalarEvolution *SE;
1261 const DataLayout *DL;
1263 TargetTransformInfo *TTI;
1265 BlockFrequencyInfo *BFI;
1266 TargetLibraryInfo *TLI;
1268 AssumptionCache *AC;
1269 LoopAccessAnalysis *LAA;
1270 bool DisableUnrolling;
1271 bool AlwaysVectorize;
1273 BlockFrequency ColdEntryFreq;
1275 bool runOnFunction(Function &F) override {
1276 SE = &getAnalysis<ScalarEvolution>();
1277 DL = &F.getParent()->getDataLayout();
1278 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1279 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1280 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1281 BFI = &getAnalysis<BlockFrequencyInfo>();
1282 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1283 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1284 AA = &getAnalysis<AliasAnalysis>();
1285 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1286 LAA = &getAnalysis<LoopAccessAnalysis>();
1288 // Compute some weights outside of the loop over the loops. Compute this
1289 // using a BranchProbability to re-use its scaling math.
1290 const BranchProbability ColdProb(1, 5); // 20%
1291 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1293 // If the target claims to have no vector registers don't attempt
1295 if (!TTI->getNumberOfRegisters(true))
1299 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1300 << ": Missing data layout\n");
1304 // Build up a worklist of inner-loops to vectorize. This is necessary as
1305 // the act of vectorizing or partially unrolling a loop creates new loops
1306 // and can invalidate iterators across the loops.
1307 SmallVector<Loop *, 8> Worklist;
1310 addInnerLoop(*L, Worklist);
1312 LoopsAnalyzed += Worklist.size();
1314 // Now walk the identified inner loops.
1315 bool Changed = false;
1316 while (!Worklist.empty())
1317 Changed |= processLoop(Worklist.pop_back_val());
1319 // Process each loop nest in the function.
1323 bool processLoop(Loop *L) {
1324 assert(L->empty() && "Only process inner loops.");
1327 const std::string DebugLocStr = getDebugLocString(L);
1330 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1331 << L->getHeader()->getParent()->getName() << "\" from "
1332 << DebugLocStr << "\n");
1334 LoopVectorizeHints Hints(L, DisableUnrolling);
1336 DEBUG(dbgs() << "LV: Loop hints:"
1338 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1340 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1342 : "?")) << " width=" << Hints.getWidth()
1343 << " unroll=" << Hints.getInterleave() << "\n");
1345 // Function containing loop
1346 Function *F = L->getHeader()->getParent();
1348 // Looking at the diagnostic output is the only way to determine if a loop
1349 // was vectorized (other than looking at the IR or machine code), so it
1350 // is important to generate an optimization remark for each loop. Most of
1351 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1352 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1353 // less verbose reporting vectorized loops and unvectorized loops that may
1354 // benefit from vectorization, respectively.
1356 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1357 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1358 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1359 L->getStartLoc(), Hints.emitRemark());
1363 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1364 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1365 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1366 L->getStartLoc(), Hints.emitRemark());
1370 if (Hints.getWidth() == 1 && Hints.getInterleave() == 1) {
1371 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1372 emitOptimizationRemarkAnalysis(
1373 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1374 "loop not vectorized: vector width and interleave count are "
1375 "explicitly set to 1");
1379 // Check the loop for a trip count threshold:
1380 // do not vectorize loops with a tiny trip count.
1381 const unsigned TC = SE->getSmallConstantTripCount(L);
1382 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1383 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1384 << "This loop is not worth vectorizing.");
1385 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1386 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1388 DEBUG(dbgs() << "\n");
1389 emitOptimizationRemarkAnalysis(
1390 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1391 "vectorization is not beneficial and is not explicitly forced");
1396 // Check if it is legal to vectorize the loop.
1397 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, AA, F, TTI, LAA);
1398 if (!LVL.canVectorize()) {
1399 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1400 emitMissedWarning(F, L, Hints);
1404 // Use the cost model.
1405 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI, AC, F,
1408 // Check the function attributes to find out if this function should be
1409 // optimized for size.
1410 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1411 F->hasFnAttribute(Attribute::OptimizeForSize);
1413 // Compute the weighted frequency of this loop being executed and see if it
1414 // is less than 20% of the function entry baseline frequency. Note that we
1415 // always have a canonical loop here because we think we *can* vectoriez.
1416 // FIXME: This is hidden behind a flag due to pervasive problems with
1417 // exactly what block frequency models.
1418 if (LoopVectorizeWithBlockFrequency) {
1419 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1420 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1421 LoopEntryFreq < ColdEntryFreq)
1425 // Check the function attributes to see if implicit floats are allowed.a
1426 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1427 // an integer loop and the vector instructions selected are purely integer
1428 // vector instructions?
1429 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1430 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1431 "attribute is used.\n");
1432 emitOptimizationRemarkAnalysis(
1433 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1434 "loop not vectorized due to NoImplicitFloat attribute");
1435 emitMissedWarning(F, L, Hints);
1439 // Select the optimal vectorization factor.
1440 const LoopVectorizationCostModel::VectorizationFactor VF =
1441 CM.selectVectorizationFactor(OptForSize);
1443 // Select the unroll factor.
1445 CM.selectUnrollFactor(OptForSize, VF.Width, VF.Cost);
1447 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1448 << DebugLocStr << '\n');
1449 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1451 if (VF.Width == 1) {
1452 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
1455 emitOptimizationRemarkAnalysis(
1456 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1457 "not beneficial to vectorize and user disabled interleaving");
1460 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1462 // Report the unrolling decision.
1463 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1464 Twine("unrolled with interleaving factor " +
1466 " (vectorization not beneficial)"));
1468 // We decided not to vectorize, but we may want to unroll.
1470 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1471 Unroller.vectorize(&LVL);
1473 // If we decided that it is *legal* to vectorize the loop then do it.
1474 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1478 // Report the vectorization decision.
1479 emitOptimizationRemark(
1480 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1481 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1482 ", unrolling interleave factor: " + Twine(UF) + ")");
1485 // Mark the loop as already vectorized to avoid vectorizing again.
1486 Hints.setAlreadyVectorized();
1488 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1492 void getAnalysisUsage(AnalysisUsage &AU) const override {
1493 AU.addRequired<AssumptionCacheTracker>();
1494 AU.addRequiredID(LoopSimplifyID);
1495 AU.addRequiredID(LCSSAID);
1496 AU.addRequired<BlockFrequencyInfo>();
1497 AU.addRequired<DominatorTreeWrapperPass>();
1498 AU.addRequired<LoopInfoWrapperPass>();
1499 AU.addRequired<ScalarEvolution>();
1500 AU.addRequired<TargetTransformInfoWrapperPass>();
1501 AU.addRequired<AliasAnalysis>();
1502 AU.addRequired<LoopAccessAnalysis>();
1503 AU.addPreserved<LoopInfoWrapperPass>();
1504 AU.addPreserved<DominatorTreeWrapperPass>();
1505 AU.addPreserved<AliasAnalysis>();
1510 } // end anonymous namespace
1512 //===----------------------------------------------------------------------===//
1513 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1514 // LoopVectorizationCostModel.
1515 //===----------------------------------------------------------------------===//
1517 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1518 // We need to place the broadcast of invariant variables outside the loop.
1519 Instruction *Instr = dyn_cast<Instruction>(V);
1521 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1522 Instr->getParent()) != LoopVectorBody.end());
1523 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1525 // Place the code for broadcasting invariant variables in the new preheader.
1526 IRBuilder<>::InsertPointGuard Guard(Builder);
1528 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1530 // Broadcast the scalar into all locations in the vector.
1531 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1536 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
1538 assert(Val->getType()->isVectorTy() && "Must be a vector");
1539 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1540 "Elem must be an integer");
1541 assert(Step->getType() == Val->getType()->getScalarType() &&
1542 "Step has wrong type");
1543 // Create the types.
1544 Type *ITy = Val->getType()->getScalarType();
1545 VectorType *Ty = cast<VectorType>(Val->getType());
1546 int VLen = Ty->getNumElements();
1547 SmallVector<Constant*, 8> Indices;
1549 // Create a vector of consecutive numbers from zero to VF.
1550 for (int i = 0; i < VLen; ++i)
1551 Indices.push_back(ConstantInt::get(ITy, StartIdx + i));
1553 // Add the consecutive indices to the vector value.
1554 Constant *Cv = ConstantVector::get(Indices);
1555 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1556 Step = Builder.CreateVectorSplat(VLen, Step);
1557 assert(Step->getType() == Val->getType() && "Invalid step vec");
1558 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
1559 // which can be found from the original scalar operations.
1560 Step = Builder.CreateMul(Cv, Step);
1561 return Builder.CreateAdd(Val, Step, "induction");
1564 /// \brief Find the operand of the GEP that should be checked for consecutive
1565 /// stores. This ignores trailing indices that have no effect on the final
1567 static unsigned getGEPInductionOperand(const DataLayout *DL,
1568 const GetElementPtrInst *Gep) {
1569 unsigned LastOperand = Gep->getNumOperands() - 1;
1570 unsigned GEPAllocSize = DL->getTypeAllocSize(
1571 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1573 // Walk backwards and try to peel off zeros.
1574 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1575 // Find the type we're currently indexing into.
1576 gep_type_iterator GEPTI = gep_type_begin(Gep);
1577 std::advance(GEPTI, LastOperand - 1);
1579 // If it's a type with the same allocation size as the result of the GEP we
1580 // can peel off the zero index.
1581 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1589 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1590 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1591 // Make sure that the pointer does not point to structs.
1592 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1595 // If this value is a pointer induction variable we know it is consecutive.
1596 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1597 if (Phi && Inductions.count(Phi)) {
1598 InductionInfo II = Inductions[Phi];
1599 return II.getConsecutiveDirection();
1602 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1606 unsigned NumOperands = Gep->getNumOperands();
1607 Value *GpPtr = Gep->getPointerOperand();
1608 // If this GEP value is a consecutive pointer induction variable and all of
1609 // the indices are constant then we know it is consecutive. We can
1610 Phi = dyn_cast<PHINode>(GpPtr);
1611 if (Phi && Inductions.count(Phi)) {
1613 // Make sure that the pointer does not point to structs.
1614 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1615 if (GepPtrType->getElementType()->isAggregateType())
1618 // Make sure that all of the index operands are loop invariant.
1619 for (unsigned i = 1; i < NumOperands; ++i)
1620 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1623 InductionInfo II = Inductions[Phi];
1624 return II.getConsecutiveDirection();
1627 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1629 // Check that all of the gep indices are uniform except for our induction
1631 for (unsigned i = 0; i != NumOperands; ++i)
1632 if (i != InductionOperand &&
1633 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1636 // We can emit wide load/stores only if the last non-zero index is the
1637 // induction variable.
1638 const SCEV *Last = nullptr;
1639 if (!Strides.count(Gep))
1640 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1642 // Because of the multiplication by a stride we can have a s/zext cast.
1643 // We are going to replace this stride by 1 so the cast is safe to ignore.
1645 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1646 // %0 = trunc i64 %indvars.iv to i32
1647 // %mul = mul i32 %0, %Stride1
1648 // %idxprom = zext i32 %mul to i64 << Safe cast.
1649 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1651 Last = replaceSymbolicStrideSCEV(SE, Strides,
1652 Gep->getOperand(InductionOperand), Gep);
1653 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1655 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1659 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1660 const SCEV *Step = AR->getStepRecurrence(*SE);
1662 // The memory is consecutive because the last index is consecutive
1663 // and all other indices are loop invariant.
1666 if (Step->isAllOnesValue())
1673 bool LoopVectorizationLegality::isUniform(Value *V) {
1674 return LAI->isUniform(V);
1677 InnerLoopVectorizer::VectorParts&
1678 InnerLoopVectorizer::getVectorValue(Value *V) {
1679 assert(V != Induction && "The new induction variable should not be used.");
1680 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1682 // If we have a stride that is replaced by one, do it here.
1683 if (Legal->hasStride(V))
1684 V = ConstantInt::get(V->getType(), 1);
1686 // If we have this scalar in the map, return it.
1687 if (WidenMap.has(V))
1688 return WidenMap.get(V);
1690 // If this scalar is unknown, assume that it is a constant or that it is
1691 // loop invariant. Broadcast V and save the value for future uses.
1692 Value *B = getBroadcastInstrs(V);
1693 return WidenMap.splat(V, B);
1696 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1697 assert(Vec->getType()->isVectorTy() && "Invalid type");
1698 SmallVector<Constant*, 8> ShuffleMask;
1699 for (unsigned i = 0; i < VF; ++i)
1700 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1702 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1703 ConstantVector::get(ShuffleMask),
1707 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1708 // Attempt to issue a wide load.
1709 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1710 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1712 assert((LI || SI) && "Invalid Load/Store instruction");
1714 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1715 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1716 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1717 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1718 // An alignment of 0 means target abi alignment. We need to use the scalar's
1719 // target abi alignment in such a case.
1721 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1722 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1723 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1724 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1726 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
1727 !Legal->isMaskRequired(SI))
1728 return scalarizeInstruction(Instr, true);
1730 if (ScalarAllocatedSize != VectorElementSize)
1731 return scalarizeInstruction(Instr);
1733 // If the pointer is loop invariant or if it is non-consecutive,
1734 // scalarize the load.
1735 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1736 bool Reverse = ConsecutiveStride < 0;
1737 bool UniformLoad = LI && Legal->isUniform(Ptr);
1738 if (!ConsecutiveStride || UniformLoad)
1739 return scalarizeInstruction(Instr);
1741 Constant *Zero = Builder.getInt32(0);
1742 VectorParts &Entry = WidenMap.get(Instr);
1744 // Handle consecutive loads/stores.
1745 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1746 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1747 setDebugLocFromInst(Builder, Gep);
1748 Value *PtrOperand = Gep->getPointerOperand();
1749 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1750 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1752 // Create the new GEP with the new induction variable.
1753 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1754 Gep2->setOperand(0, FirstBasePtr);
1755 Gep2->setName("gep.indvar.base");
1756 Ptr = Builder.Insert(Gep2);
1758 setDebugLocFromInst(Builder, Gep);
1759 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1760 OrigLoop) && "Base ptr must be invariant");
1762 // The last index does not have to be the induction. It can be
1763 // consecutive and be a function of the index. For example A[I+1];
1764 unsigned NumOperands = Gep->getNumOperands();
1765 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1766 // Create the new GEP with the new induction variable.
1767 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1769 for (unsigned i = 0; i < NumOperands; ++i) {
1770 Value *GepOperand = Gep->getOperand(i);
1771 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1773 // Update last index or loop invariant instruction anchored in loop.
1774 if (i == InductionOperand ||
1775 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1776 assert((i == InductionOperand ||
1777 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1778 "Must be last index or loop invariant");
1780 VectorParts &GEPParts = getVectorValue(GepOperand);
1781 Value *Index = GEPParts[0];
1782 Index = Builder.CreateExtractElement(Index, Zero);
1783 Gep2->setOperand(i, Index);
1784 Gep2->setName("gep.indvar.idx");
1787 Ptr = Builder.Insert(Gep2);
1789 // Use the induction element ptr.
1790 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1791 setDebugLocFromInst(Builder, Ptr);
1792 VectorParts &PtrVal = getVectorValue(Ptr);
1793 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1796 VectorParts Mask = createBlockInMask(Instr->getParent());
1799 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1800 "We do not allow storing to uniform addresses");
1801 setDebugLocFromInst(Builder, SI);
1802 // We don't want to update the value in the map as it might be used in
1803 // another expression. So don't use a reference type for "StoredVal".
1804 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1806 for (unsigned Part = 0; Part < UF; ++Part) {
1807 // Calculate the pointer for the specific unroll-part.
1808 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1811 // If we store to reverse consecutive memory locations then we need
1812 // to reverse the order of elements in the stored value.
1813 StoredVal[Part] = reverseVector(StoredVal[Part]);
1814 // If the address is consecutive but reversed, then the
1815 // wide store 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]);
1821 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1822 DataTy->getPointerTo(AddressSpace));
1825 if (Legal->isMaskRequired(SI))
1826 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
1829 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
1830 propagateMetadata(NewSI, SI);
1836 assert(LI && "Must have a load instruction");
1837 setDebugLocFromInst(Builder, LI);
1838 for (unsigned Part = 0; Part < UF; ++Part) {
1839 // Calculate the pointer for the specific unroll-part.
1840 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1843 // If the address is consecutive but reversed, then the
1844 // wide load needs to start at the last vector element.
1845 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1846 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1847 Mask[Part] = reverseVector(Mask[Part]);
1851 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1852 DataTy->getPointerTo(AddressSpace));
1853 if (Legal->isMaskRequired(LI))
1854 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
1855 UndefValue::get(DataTy),
1856 "wide.masked.load");
1858 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
1859 propagateMetadata(NewLI, LI);
1860 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
1864 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1865 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1866 // Holds vector parameters or scalars, in case of uniform vals.
1867 SmallVector<VectorParts, 4> Params;
1869 setDebugLocFromInst(Builder, Instr);
1871 // Find all of the vectorized parameters.
1872 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1873 Value *SrcOp = Instr->getOperand(op);
1875 // If we are accessing the old induction variable, use the new one.
1876 if (SrcOp == OldInduction) {
1877 Params.push_back(getVectorValue(SrcOp));
1881 // Try using previously calculated values.
1882 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1884 // If the src is an instruction that appeared earlier in the basic block
1885 // then it should already be vectorized.
1886 if (SrcInst && OrigLoop->contains(SrcInst)) {
1887 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1888 // The parameter is a vector value from earlier.
1889 Params.push_back(WidenMap.get(SrcInst));
1891 // The parameter is a scalar from outside the loop. Maybe even a constant.
1892 VectorParts Scalars;
1893 Scalars.append(UF, SrcOp);
1894 Params.push_back(Scalars);
1898 assert(Params.size() == Instr->getNumOperands() &&
1899 "Invalid number of operands");
1901 // Does this instruction return a value ?
1902 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1904 Value *UndefVec = IsVoidRetTy ? nullptr :
1905 UndefValue::get(VectorType::get(Instr->getType(), VF));
1906 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1907 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1909 Instruction *InsertPt = Builder.GetInsertPoint();
1910 BasicBlock *IfBlock = Builder.GetInsertBlock();
1911 BasicBlock *CondBlock = nullptr;
1914 Loop *VectorLp = nullptr;
1915 if (IfPredicateStore) {
1916 assert(Instr->getParent()->getSinglePredecessor() &&
1917 "Only support single predecessor blocks");
1918 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1919 Instr->getParent());
1920 VectorLp = LI->getLoopFor(IfBlock);
1921 assert(VectorLp && "Must have a loop for this block");
1924 // For each vector unroll 'part':
1925 for (unsigned Part = 0; Part < UF; ++Part) {
1926 // For each scalar that we create:
1927 for (unsigned Width = 0; Width < VF; ++Width) {
1930 Value *Cmp = nullptr;
1931 if (IfPredicateStore) {
1932 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1933 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1934 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1935 LoopVectorBody.push_back(CondBlock);
1936 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
1937 // Update Builder with newly created basic block.
1938 Builder.SetInsertPoint(InsertPt);
1941 Instruction *Cloned = Instr->clone();
1943 Cloned->setName(Instr->getName() + ".cloned");
1944 // Replace the operands of the cloned instructions with extracted scalars.
1945 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1946 Value *Op = Params[op][Part];
1947 // Param is a vector. Need to extract the right lane.
1948 if (Op->getType()->isVectorTy())
1949 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1950 Cloned->setOperand(op, Op);
1953 // Place the cloned scalar in the new loop.
1954 Builder.Insert(Cloned);
1956 // If the original scalar returns a value we need to place it in a vector
1957 // so that future users will be able to use it.
1959 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1960 Builder.getInt32(Width));
1962 if (IfPredicateStore) {
1963 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1964 LoopVectorBody.push_back(NewIfBlock);
1965 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
1966 Builder.SetInsertPoint(InsertPt);
1967 Instruction *OldBr = IfBlock->getTerminator();
1968 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1969 OldBr->eraseFromParent();
1970 IfBlock = NewIfBlock;
1976 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1980 if (Instruction *I = dyn_cast<Instruction>(V))
1981 return I->getParent() == Loc->getParent() ? I : nullptr;
1985 std::pair<Instruction *, Instruction *>
1986 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
1987 Instruction *tnullptr = nullptr;
1988 if (!Legal->mustCheckStrides())
1989 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1991 IRBuilder<> ChkBuilder(Loc);
1994 Value *Check = nullptr;
1995 Instruction *FirstInst = nullptr;
1996 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
1997 SE = Legal->strides_end();
1999 Value *Ptr = stripIntegerCast(*SI);
2000 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
2002 // Store the first instruction we create.
2003 FirstInst = getFirstInst(FirstInst, C, Loc);
2005 Check = ChkBuilder.CreateOr(Check, C);
2010 // We have to do this trickery because the IRBuilder might fold the check to a
2011 // constant expression in which case there is no Instruction anchored in a
2013 LLVMContext &Ctx = Loc->getContext();
2014 Instruction *TheCheck =
2015 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2016 ChkBuilder.Insert(TheCheck, "stride.not.one");
2017 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2019 return std::make_pair(FirstInst, TheCheck);
2022 void InnerLoopVectorizer::createEmptyLoop() {
2024 In this function we generate a new loop. The new loop will contain
2025 the vectorized instructions while the old loop will continue to run the
2028 [ ] <-- Back-edge taken count overflow check.
2031 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2034 || [ ] <-- vector pre header.
2038 || [ ]_| <-- vector loop.
2041 | >[ ] <--- middle-block.
2044 -|- >[ ] <--- new preheader.
2048 | [ ]_| <-- old scalar loop to handle remainder.
2051 >[ ] <-- exit block.
2055 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2056 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2057 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2058 assert(BypassBlock && "Invalid loop structure");
2059 assert(ExitBlock && "Must have an exit block");
2061 // Some loops have a single integer induction variable, while other loops
2062 // don't. One example is c++ iterators that often have multiple pointer
2063 // induction variables. In the code below we also support a case where we
2064 // don't have a single induction variable.
2065 OldInduction = Legal->getInduction();
2066 Type *IdxTy = Legal->getWidestInductionType();
2068 // Find the loop boundaries.
2069 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2070 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2072 // The exit count might have the type of i64 while the phi is i32. This can
2073 // happen if we have an induction variable that is sign extended before the
2074 // compare. The only way that we get a backedge taken count is that the
2075 // induction variable was signed and as such will not overflow. In such a case
2076 // truncation is legal.
2077 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2078 IdxTy->getPrimitiveSizeInBits())
2079 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2081 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2082 // Get the total trip count from the count by adding 1.
2083 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2084 SE->getConstant(BackedgeTakeCount->getType(), 1));
2086 // Expand the trip count and place the new instructions in the preheader.
2087 // Notice that the pre-header does not change, only the loop body.
2088 SCEVExpander Exp(*SE, "induction");
2090 // We need to test whether the backedge-taken count is uint##_max. Adding one
2091 // to it will cause overflow and an incorrect loop trip count in the vector
2092 // body. In case of overflow we want to directly jump to the scalar remainder
2094 Value *BackedgeCount =
2095 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2096 BypassBlock->getTerminator());
2097 if (BackedgeCount->getType()->isPointerTy())
2098 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2099 "backedge.ptrcnt.to.int",
2100 BypassBlock->getTerminator());
2101 Instruction *CheckBCOverflow =
2102 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2103 Constant::getAllOnesValue(BackedgeCount->getType()),
2104 "backedge.overflow", BypassBlock->getTerminator());
2106 // The loop index does not have to start at Zero. Find the original start
2107 // value from the induction PHI node. If we don't have an induction variable
2108 // then we know that it starts at zero.
2109 Builder.SetInsertPoint(BypassBlock->getTerminator());
2110 Value *StartIdx = ExtendedIdx = OldInduction ?
2111 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2113 ConstantInt::get(IdxTy, 0);
2115 // We need an instruction to anchor the overflow check on. StartIdx needs to
2116 // be defined before the overflow check branch. Because the scalar preheader
2117 // is going to merge the start index and so the overflow branch block needs to
2118 // contain a definition of the start index.
2119 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2120 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2121 BypassBlock->getTerminator());
2123 // Count holds the overall loop count (N).
2124 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2125 BypassBlock->getTerminator());
2127 LoopBypassBlocks.push_back(BypassBlock);
2129 // Split the single block loop into the two loop structure described above.
2130 BasicBlock *VectorPH =
2131 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2132 BasicBlock *VecBody =
2133 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2134 BasicBlock *MiddleBlock =
2135 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2136 BasicBlock *ScalarPH =
2137 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2139 // Create and register the new vector loop.
2140 Loop* Lp = new Loop();
2141 Loop *ParentLoop = OrigLoop->getParentLoop();
2143 // Insert the new loop into the loop nest and register the new basic blocks
2144 // before calling any utilities such as SCEV that require valid LoopInfo.
2146 ParentLoop->addChildLoop(Lp);
2147 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
2148 ParentLoop->addBasicBlockToLoop(VectorPH, *LI);
2149 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
2151 LI->addTopLevelLoop(Lp);
2153 Lp->addBasicBlockToLoop(VecBody, *LI);
2155 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2157 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2159 // Generate the induction variable.
2160 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2161 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2162 // The loop step is equal to the vectorization factor (num of SIMD elements)
2163 // times the unroll factor (num of SIMD instructions).
2164 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2166 // This is the IR builder that we use to add all of the logic for bypassing
2167 // the new vector loop.
2168 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2169 setDebugLocFromInst(BypassBuilder,
2170 getDebugLocFromInstOrOperands(OldInduction));
2172 // We may need to extend the index in case there is a type mismatch.
2173 // We know that the count starts at zero and does not overflow.
2174 if (Count->getType() != IdxTy) {
2175 // The exit count can be of pointer type. Convert it to the correct
2177 if (ExitCount->getType()->isPointerTy())
2178 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2180 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2183 // Add the start index to the loop count to get the new end index.
2184 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2186 // Now we need to generate the expression for N - (N % VF), which is
2187 // the part that the vectorized body will execute.
2188 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2189 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2190 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2191 "end.idx.rnd.down");
2193 // Now, compare the new count to zero. If it is zero skip the vector loop and
2194 // jump to the scalar loop.
2196 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2198 BasicBlock *LastBypassBlock = BypassBlock;
2200 // Generate code to check that the loops trip count that we computed by adding
2201 // one to the backedge-taken count will not overflow.
2203 auto PastOverflowCheck =
2204 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2205 BasicBlock *CheckBlock =
2206 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2208 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2209 LoopBypassBlocks.push_back(CheckBlock);
2210 Instruction *OldTerm = LastBypassBlock->getTerminator();
2211 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2212 OldTerm->eraseFromParent();
2213 LastBypassBlock = CheckBlock;
2216 // Generate the code to check that the strides we assumed to be one are really
2217 // one. We want the new basic block to start at the first instruction in a
2218 // sequence of instructions that form a check.
2219 Instruction *StrideCheck;
2220 Instruction *FirstCheckInst;
2221 std::tie(FirstCheckInst, StrideCheck) =
2222 addStrideCheck(LastBypassBlock->getTerminator());
2224 // Create a new block containing the stride check.
2225 BasicBlock *CheckBlock =
2226 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2228 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2229 LoopBypassBlocks.push_back(CheckBlock);
2231 // Replace the branch into the memory check block with a conditional branch
2232 // for the "few elements case".
2233 Instruction *OldTerm = LastBypassBlock->getTerminator();
2234 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2235 OldTerm->eraseFromParent();
2238 LastBypassBlock = CheckBlock;
2241 // Generate the code that checks in runtime if arrays overlap. We put the
2242 // checks into a separate block to make the more common case of few elements
2244 Instruction *MemRuntimeCheck;
2245 std::tie(FirstCheckInst, MemRuntimeCheck) =
2246 Legal->getLAI()->addRuntimeCheck(LastBypassBlock->getTerminator());
2247 if (MemRuntimeCheck) {
2248 // Create a new block containing the memory check.
2249 BasicBlock *CheckBlock =
2250 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.memcheck");
2252 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2253 LoopBypassBlocks.push_back(CheckBlock);
2255 // Replace the branch into the memory check block with a conditional branch
2256 // for the "few elements case".
2257 Instruction *OldTerm = LastBypassBlock->getTerminator();
2258 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2259 OldTerm->eraseFromParent();
2261 Cmp = MemRuntimeCheck;
2262 LastBypassBlock = CheckBlock;
2265 LastBypassBlock->getTerminator()->eraseFromParent();
2266 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2269 // We are going to resume the execution of the scalar loop.
2270 // Go over all of the induction variables that we found and fix the
2271 // PHIs that are left in the scalar version of the loop.
2272 // The starting values of PHI nodes depend on the counter of the last
2273 // iteration in the vectorized loop.
2274 // If we come from a bypass edge then we need to start from the original
2277 // This variable saves the new starting index for the scalar loop.
2278 PHINode *ResumeIndex = nullptr;
2279 LoopVectorizationLegality::InductionList::iterator I, E;
2280 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2281 // Set builder to point to last bypass block.
2282 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2283 for (I = List->begin(), E = List->end(); I != E; ++I) {
2284 PHINode *OrigPhi = I->first;
2285 LoopVectorizationLegality::InductionInfo II = I->second;
2287 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2288 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2289 MiddleBlock->getTerminator());
2290 // We might have extended the type of the induction variable but we need a
2291 // truncated version for the scalar loop.
2292 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2293 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2294 MiddleBlock->getTerminator()) : nullptr;
2296 // Create phi nodes to merge from the backedge-taken check block.
2297 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2298 ScalarPH->getTerminator());
2299 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2301 PHINode *BCTruncResumeVal = nullptr;
2302 if (OrigPhi == OldInduction) {
2304 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2305 ScalarPH->getTerminator());
2306 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2309 Value *EndValue = nullptr;
2311 case LoopVectorizationLegality::IK_NoInduction:
2312 llvm_unreachable("Unknown induction");
2313 case LoopVectorizationLegality::IK_IntInduction: {
2314 // Handle the integer induction counter.
2315 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2317 // We have the canonical induction variable.
2318 if (OrigPhi == OldInduction) {
2319 // Create a truncated version of the resume value for the scalar loop,
2320 // we might have promoted the type to a larger width.
2322 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2323 // The new PHI merges the original incoming value, in case of a bypass,
2324 // or the value at the end of the vectorized loop.
2325 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2326 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2327 TruncResumeVal->addIncoming(EndValue, VecBody);
2329 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2331 // We know what the end value is.
2332 EndValue = IdxEndRoundDown;
2333 // We also know which PHI node holds it.
2334 ResumeIndex = ResumeVal;
2338 // Not the canonical induction variable - add the vector loop count to the
2340 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2341 II.StartValue->getType(),
2343 EndValue = II.transform(BypassBuilder, CRD);
2344 EndValue->setName("ind.end");
2347 case LoopVectorizationLegality::IK_PtrInduction: {
2348 EndValue = II.transform(BypassBuilder, CountRoundDown);
2349 EndValue->setName("ptr.ind.end");
2354 // The new PHI merges the original incoming value, in case of a bypass,
2355 // or the value at the end of the vectorized loop.
2356 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2357 if (OrigPhi == OldInduction)
2358 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2360 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2362 ResumeVal->addIncoming(EndValue, VecBody);
2364 // Fix the scalar body counter (PHI node).
2365 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2367 // The old induction's phi node in the scalar body needs the truncated
2369 if (OrigPhi == OldInduction) {
2370 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2371 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2373 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2374 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2378 // If we are generating a new induction variable then we also need to
2379 // generate the code that calculates the exit value. This value is not
2380 // simply the end of the counter because we may skip the vectorized body
2381 // in case of a runtime check.
2383 assert(!ResumeIndex && "Unexpected resume value found");
2384 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2385 MiddleBlock->getTerminator());
2386 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2387 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2388 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2391 // Make sure that we found the index where scalar loop needs to continue.
2392 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2393 "Invalid resume Index");
2395 // Add a check in the middle block to see if we have completed
2396 // all of the iterations in the first vector loop.
2397 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2398 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2399 ResumeIndex, "cmp.n",
2400 MiddleBlock->getTerminator());
2402 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2403 // Remove the old terminator.
2404 MiddleBlock->getTerminator()->eraseFromParent();
2406 // Create i+1 and fill the PHINode.
2407 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2408 Induction->addIncoming(StartIdx, VectorPH);
2409 Induction->addIncoming(NextIdx, VecBody);
2410 // Create the compare.
2411 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2412 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2414 // Now we have two terminators. Remove the old one from the block.
2415 VecBody->getTerminator()->eraseFromParent();
2417 // Get ready to start creating new instructions into the vectorized body.
2418 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2421 LoopVectorPreHeader = VectorPH;
2422 LoopScalarPreHeader = ScalarPH;
2423 LoopMiddleBlock = MiddleBlock;
2424 LoopExitBlock = ExitBlock;
2425 LoopVectorBody.push_back(VecBody);
2426 LoopScalarBody = OldBasicBlock;
2428 LoopVectorizeHints Hints(Lp, true);
2429 Hints.setAlreadyVectorized();
2432 /// This function returns the identity element (or neutral element) for
2433 /// the operation K.
2435 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2440 // Adding, Xoring, Oring zero to a number does not change it.
2441 return ConstantInt::get(Tp, 0);
2442 case RK_IntegerMult:
2443 // Multiplying a number by 1 does not change it.
2444 return ConstantInt::get(Tp, 1);
2446 // AND-ing a number with an all-1 value does not change it.
2447 return ConstantInt::get(Tp, -1, true);
2449 // Multiplying a number by 1 does not change it.
2450 return ConstantFP::get(Tp, 1.0L);
2452 // Adding zero to a number does not change it.
2453 return ConstantFP::get(Tp, 0.0L);
2455 llvm_unreachable("Unknown reduction kind");
2459 /// This function translates the reduction kind to an LLVM binary operator.
2461 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2463 case LoopVectorizationLegality::RK_IntegerAdd:
2464 return Instruction::Add;
2465 case LoopVectorizationLegality::RK_IntegerMult:
2466 return Instruction::Mul;
2467 case LoopVectorizationLegality::RK_IntegerOr:
2468 return Instruction::Or;
2469 case LoopVectorizationLegality::RK_IntegerAnd:
2470 return Instruction::And;
2471 case LoopVectorizationLegality::RK_IntegerXor:
2472 return Instruction::Xor;
2473 case LoopVectorizationLegality::RK_FloatMult:
2474 return Instruction::FMul;
2475 case LoopVectorizationLegality::RK_FloatAdd:
2476 return Instruction::FAdd;
2477 case LoopVectorizationLegality::RK_IntegerMinMax:
2478 return Instruction::ICmp;
2479 case LoopVectorizationLegality::RK_FloatMinMax:
2480 return Instruction::FCmp;
2482 llvm_unreachable("Unknown reduction operation");
2486 Value *createMinMaxOp(IRBuilder<> &Builder,
2487 LoopVectorizationLegality::MinMaxReductionKind RK,
2490 CmpInst::Predicate P = CmpInst::ICMP_NE;
2493 llvm_unreachable("Unknown min/max reduction kind");
2494 case LoopVectorizationLegality::MRK_UIntMin:
2495 P = CmpInst::ICMP_ULT;
2497 case LoopVectorizationLegality::MRK_UIntMax:
2498 P = CmpInst::ICMP_UGT;
2500 case LoopVectorizationLegality::MRK_SIntMin:
2501 P = CmpInst::ICMP_SLT;
2503 case LoopVectorizationLegality::MRK_SIntMax:
2504 P = CmpInst::ICMP_SGT;
2506 case LoopVectorizationLegality::MRK_FloatMin:
2507 P = CmpInst::FCMP_OLT;
2509 case LoopVectorizationLegality::MRK_FloatMax:
2510 P = CmpInst::FCMP_OGT;
2515 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2516 RK == LoopVectorizationLegality::MRK_FloatMax)
2517 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2519 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2521 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2526 struct CSEDenseMapInfo {
2527 static bool canHandle(Instruction *I) {
2528 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2529 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2531 static inline Instruction *getEmptyKey() {
2532 return DenseMapInfo<Instruction *>::getEmptyKey();
2534 static inline Instruction *getTombstoneKey() {
2535 return DenseMapInfo<Instruction *>::getTombstoneKey();
2537 static unsigned getHashValue(Instruction *I) {
2538 assert(canHandle(I) && "Unknown instruction!");
2539 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2540 I->value_op_end()));
2542 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2543 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2544 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2546 return LHS->isIdenticalTo(RHS);
2551 /// \brief Check whether this block is a predicated block.
2552 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2553 /// = ...; " blocks. We start with one vectorized basic block. For every
2554 /// conditional block we split this vectorized block. Therefore, every second
2555 /// block will be a predicated one.
2556 static bool isPredicatedBlock(unsigned BlockNum) {
2557 return BlockNum % 2;
2560 ///\brief Perform cse of induction variable instructions.
2561 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2562 // Perform simple cse.
2563 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2564 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2565 BasicBlock *BB = BBs[i];
2566 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2567 Instruction *In = I++;
2569 if (!CSEDenseMapInfo::canHandle(In))
2572 // Check if we can replace this instruction with any of the
2573 // visited instructions.
2574 if (Instruction *V = CSEMap.lookup(In)) {
2575 In->replaceAllUsesWith(V);
2576 In->eraseFromParent();
2579 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2580 // ...;" blocks for predicated stores. Every second block is a predicated
2582 if (isPredicatedBlock(i))
2590 /// \brief Adds a 'fast' flag to floating point operations.
2591 static Value *addFastMathFlag(Value *V) {
2592 if (isa<FPMathOperator>(V)){
2593 FastMathFlags Flags;
2594 Flags.setUnsafeAlgebra();
2595 cast<Instruction>(V)->setFastMathFlags(Flags);
2600 void InnerLoopVectorizer::vectorizeLoop() {
2601 //===------------------------------------------------===//
2603 // Notice: any optimization or new instruction that go
2604 // into the code below should be also be implemented in
2607 //===------------------------------------------------===//
2608 Constant *Zero = Builder.getInt32(0);
2610 // In order to support reduction variables we need to be able to vectorize
2611 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2612 // stages. First, we create a new vector PHI node with no incoming edges.
2613 // We use this value when we vectorize all of the instructions that use the
2614 // PHI. Next, after all of the instructions in the block are complete we
2615 // add the new incoming edges to the PHI. At this point all of the
2616 // instructions in the basic block are vectorized, so we can use them to
2617 // construct the PHI.
2618 PhiVector RdxPHIsToFix;
2620 // Scan the loop in a topological order to ensure that defs are vectorized
2622 LoopBlocksDFS DFS(OrigLoop);
2625 // Vectorize all of the blocks in the original loop.
2626 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2627 be = DFS.endRPO(); bb != be; ++bb)
2628 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2630 // At this point every instruction in the original loop is widened to
2631 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2632 // that we vectorized. The PHI nodes are currently empty because we did
2633 // not want to introduce cycles. Notice that the remaining PHI nodes
2634 // that we need to fix are reduction variables.
2636 // Create the 'reduced' values for each of the induction vars.
2637 // The reduced values are the vector values that we scalarize and combine
2638 // after the loop is finished.
2639 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2641 PHINode *RdxPhi = *it;
2642 assert(RdxPhi && "Unable to recover vectorized PHI");
2644 // Find the reduction variable descriptor.
2645 assert(Legal->getReductionVars()->count(RdxPhi) &&
2646 "Unable to find the reduction variable");
2647 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2648 (*Legal->getReductionVars())[RdxPhi];
2650 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2652 // We need to generate a reduction vector from the incoming scalar.
2653 // To do so, we need to generate the 'identity' vector and override
2654 // one of the elements with the incoming scalar reduction. We need
2655 // to do it in the vector-loop preheader.
2656 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
2658 // This is the vector-clone of the value that leaves the loop.
2659 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2660 Type *VecTy = VectorExit[0]->getType();
2662 // Find the reduction identity variable. Zero for addition, or, xor,
2663 // one for multiplication, -1 for And.
2666 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2667 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2668 // MinMax reduction have the start value as their identify.
2670 VectorStart = Identity = RdxDesc.StartValue;
2672 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2677 // Handle other reduction kinds:
2679 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2680 VecTy->getScalarType());
2683 // This vector is the Identity vector where the first element is the
2684 // incoming scalar reduction.
2685 VectorStart = RdxDesc.StartValue;
2687 Identity = ConstantVector::getSplat(VF, Iden);
2689 // This vector is the Identity vector where the first element is the
2690 // incoming scalar reduction.
2691 VectorStart = Builder.CreateInsertElement(Identity,
2692 RdxDesc.StartValue, Zero);
2696 // Fix the vector-loop phi.
2698 // Reductions do not have to start at zero. They can start with
2699 // any loop invariant values.
2700 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2701 BasicBlock *Latch = OrigLoop->getLoopLatch();
2702 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2703 VectorParts &Val = getVectorValue(LoopVal);
2704 for (unsigned part = 0; part < UF; ++part) {
2705 // Make sure to add the reduction stat value only to the
2706 // first unroll part.
2707 Value *StartVal = (part == 0) ? VectorStart : Identity;
2708 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
2709 LoopVectorPreHeader);
2710 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2711 LoopVectorBody.back());
2714 // Before each round, move the insertion point right between
2715 // the PHIs and the values we are going to write.
2716 // This allows us to write both PHINodes and the extractelement
2718 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2720 VectorParts RdxParts;
2721 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2722 for (unsigned part = 0; part < UF; ++part) {
2723 // This PHINode contains the vectorized reduction variable, or
2724 // the initial value vector, if we bypass the vector loop.
2725 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2726 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2727 Value *StartVal = (part == 0) ? VectorStart : Identity;
2728 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2729 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2730 NewPhi->addIncoming(RdxExitVal[part],
2731 LoopVectorBody.back());
2732 RdxParts.push_back(NewPhi);
2735 // Reduce all of the unrolled parts into a single vector.
2736 Value *ReducedPartRdx = RdxParts[0];
2737 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2738 setDebugLocFromInst(Builder, ReducedPartRdx);
2739 for (unsigned part = 1; part < UF; ++part) {
2740 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2741 // Floating point operations had to be 'fast' to enable the reduction.
2742 ReducedPartRdx = addFastMathFlag(
2743 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2744 ReducedPartRdx, "bin.rdx"));
2746 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2747 ReducedPartRdx, RdxParts[part]);
2751 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2752 // and vector ops, reducing the set of values being computed by half each
2754 assert(isPowerOf2_32(VF) &&
2755 "Reduction emission only supported for pow2 vectors!");
2756 Value *TmpVec = ReducedPartRdx;
2757 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2758 for (unsigned i = VF; i != 1; i >>= 1) {
2759 // Move the upper half of the vector to the lower half.
2760 for (unsigned j = 0; j != i/2; ++j)
2761 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2763 // Fill the rest of the mask with undef.
2764 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2765 UndefValue::get(Builder.getInt32Ty()));
2768 Builder.CreateShuffleVector(TmpVec,
2769 UndefValue::get(TmpVec->getType()),
2770 ConstantVector::get(ShuffleMask),
2773 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2774 // Floating point operations had to be 'fast' to enable the reduction.
2775 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2776 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2778 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2781 // The result is in the first element of the vector.
2782 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2783 Builder.getInt32(0));
2786 // Create a phi node that merges control-flow from the backedge-taken check
2787 // block and the middle block.
2788 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
2789 LoopScalarPreHeader->getTerminator());
2790 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
2791 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2793 // Now, we need to fix the users of the reduction variable
2794 // inside and outside of the scalar remainder loop.
2795 // We know that the loop is in LCSSA form. We need to update the
2796 // PHI nodes in the exit blocks.
2797 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2798 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2799 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2800 if (!LCSSAPhi) break;
2802 // All PHINodes need to have a single entry edge, or two if
2803 // we already fixed them.
2804 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2806 // We found our reduction value exit-PHI. Update it with the
2807 // incoming bypass edge.
2808 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2809 // Add an edge coming from the bypass.
2810 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2813 }// end of the LCSSA phi scan.
2815 // Fix the scalar loop reduction variable with the incoming reduction sum
2816 // from the vector body and from the backedge value.
2817 int IncomingEdgeBlockIdx =
2818 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2819 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2820 // Pick the other block.
2821 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2822 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
2823 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2824 }// end of for each redux variable.
2828 // Remove redundant induction instructions.
2829 cse(LoopVectorBody);
2832 void InnerLoopVectorizer::fixLCSSAPHIs() {
2833 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2834 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2835 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2836 if (!LCSSAPhi) break;
2837 if (LCSSAPhi->getNumIncomingValues() == 1)
2838 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2843 InnerLoopVectorizer::VectorParts
2844 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2845 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2848 // Look for cached value.
2849 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2850 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2851 if (ECEntryIt != MaskCache.end())
2852 return ECEntryIt->second;
2854 VectorParts SrcMask = createBlockInMask(Src);
2856 // The terminator has to be a branch inst!
2857 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2858 assert(BI && "Unexpected terminator found");
2860 if (BI->isConditional()) {
2861 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2863 if (BI->getSuccessor(0) != Dst)
2864 for (unsigned part = 0; part < UF; ++part)
2865 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2867 for (unsigned part = 0; part < UF; ++part)
2868 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2870 MaskCache[Edge] = EdgeMask;
2874 MaskCache[Edge] = SrcMask;
2878 InnerLoopVectorizer::VectorParts
2879 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2880 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2882 // Loop incoming mask is all-one.
2883 if (OrigLoop->getHeader() == BB) {
2884 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2885 return getVectorValue(C);
2888 // This is the block mask. We OR all incoming edges, and with zero.
2889 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2890 VectorParts BlockMask = getVectorValue(Zero);
2893 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2894 VectorParts EM = createEdgeMask(*it, BB);
2895 for (unsigned part = 0; part < UF; ++part)
2896 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2902 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2903 InnerLoopVectorizer::VectorParts &Entry,
2904 unsigned UF, unsigned VF, PhiVector *PV) {
2905 PHINode* P = cast<PHINode>(PN);
2906 // Handle reduction variables:
2907 if (Legal->getReductionVars()->count(P)) {
2908 for (unsigned part = 0; part < UF; ++part) {
2909 // This is phase one of vectorizing PHIs.
2910 Type *VecTy = (VF == 1) ? PN->getType() :
2911 VectorType::get(PN->getType(), VF);
2912 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2913 LoopVectorBody.back()-> getFirstInsertionPt());
2919 setDebugLocFromInst(Builder, P);
2920 // Check for PHI nodes that are lowered to vector selects.
2921 if (P->getParent() != OrigLoop->getHeader()) {
2922 // We know that all PHIs in non-header blocks are converted into
2923 // selects, so we don't have to worry about the insertion order and we
2924 // can just use the builder.
2925 // At this point we generate the predication tree. There may be
2926 // duplications since this is a simple recursive scan, but future
2927 // optimizations will clean it up.
2929 unsigned NumIncoming = P->getNumIncomingValues();
2931 // Generate a sequence of selects of the form:
2932 // SELECT(Mask3, In3,
2933 // SELECT(Mask2, In2,
2935 for (unsigned In = 0; In < NumIncoming; In++) {
2936 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2938 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2940 for (unsigned part = 0; part < UF; ++part) {
2941 // We might have single edge PHIs (blocks) - use an identity
2942 // 'select' for the first PHI operand.
2944 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2947 // Select between the current value and the previous incoming edge
2948 // based on the incoming mask.
2949 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2950 Entry[part], "predphi");
2956 // This PHINode must be an induction variable.
2957 // Make sure that we know about it.
2958 assert(Legal->getInductionVars()->count(P) &&
2959 "Not an induction variable");
2961 LoopVectorizationLegality::InductionInfo II =
2962 Legal->getInductionVars()->lookup(P);
2964 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
2965 // which can be found from the original scalar operations.
2967 case LoopVectorizationLegality::IK_NoInduction:
2968 llvm_unreachable("Unknown induction");
2969 case LoopVectorizationLegality::IK_IntInduction: {
2970 assert(P->getType() == II.StartValue->getType() && "Types must match");
2971 Type *PhiTy = P->getType();
2973 if (P == OldInduction) {
2974 // Handle the canonical induction variable. We might have had to
2976 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2978 // Handle other induction variables that are now based on the
2980 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2982 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2983 Broadcasted = II.transform(Builder, NormalizedIdx);
2984 Broadcasted->setName("offset.idx");
2986 Broadcasted = getBroadcastInstrs(Broadcasted);
2987 // After broadcasting the induction variable we need to make the vector
2988 // consecutive by adding 0, 1, 2, etc.
2989 for (unsigned part = 0; part < UF; ++part)
2990 Entry[part] = getStepVector(Broadcasted, VF * part, II.StepValue);
2993 case LoopVectorizationLegality::IK_PtrInduction:
2994 // Handle the pointer induction variable case.
2995 assert(P->getType()->isPointerTy() && "Unexpected type.");
2996 // This is the normalized GEP that starts counting at zero.
2997 Value *NormalizedIdx =
2998 Builder.CreateSub(Induction, ExtendedIdx, "normalized.idx");
2999 // This is the vector of results. Notice that we don't generate
3000 // vector geps because scalar geps result in better code.
3001 for (unsigned part = 0; part < UF; ++part) {
3003 int EltIndex = part;
3004 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3005 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3006 Value *SclrGep = II.transform(Builder, GlobalIdx);
3007 SclrGep->setName("next.gep");
3008 Entry[part] = SclrGep;
3012 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3013 for (unsigned int i = 0; i < VF; ++i) {
3014 int EltIndex = i + part * VF;
3015 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3016 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3017 Value *SclrGep = II.transform(Builder, GlobalIdx);
3018 SclrGep->setName("next.gep");
3019 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3020 Builder.getInt32(i),
3023 Entry[part] = VecVal;
3029 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3030 // For each instruction in the old loop.
3031 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3032 VectorParts &Entry = WidenMap.get(it);
3033 switch (it->getOpcode()) {
3034 case Instruction::Br:
3035 // Nothing to do for PHIs and BR, since we already took care of the
3036 // loop control flow instructions.
3038 case Instruction::PHI: {
3039 // Vectorize PHINodes.
3040 widenPHIInstruction(it, Entry, UF, VF, PV);
3044 case Instruction::Add:
3045 case Instruction::FAdd:
3046 case Instruction::Sub:
3047 case Instruction::FSub:
3048 case Instruction::Mul:
3049 case Instruction::FMul:
3050 case Instruction::UDiv:
3051 case Instruction::SDiv:
3052 case Instruction::FDiv:
3053 case Instruction::URem:
3054 case Instruction::SRem:
3055 case Instruction::FRem:
3056 case Instruction::Shl:
3057 case Instruction::LShr:
3058 case Instruction::AShr:
3059 case Instruction::And:
3060 case Instruction::Or:
3061 case Instruction::Xor: {
3062 // Just widen binops.
3063 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3064 setDebugLocFromInst(Builder, BinOp);
3065 VectorParts &A = getVectorValue(it->getOperand(0));
3066 VectorParts &B = getVectorValue(it->getOperand(1));
3068 // Use this vector value for all users of the original instruction.
3069 for (unsigned Part = 0; Part < UF; ++Part) {
3070 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3072 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3073 VecOp->copyIRFlags(BinOp);
3078 propagateMetadata(Entry, it);
3081 case Instruction::Select: {
3083 // If the selector is loop invariant we can create a select
3084 // instruction with a scalar condition. Otherwise, use vector-select.
3085 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3087 setDebugLocFromInst(Builder, it);
3089 // The condition can be loop invariant but still defined inside the
3090 // loop. This means that we can't just use the original 'cond' value.
3091 // We have to take the 'vectorized' value and pick the first lane.
3092 // Instcombine will make this a no-op.
3093 VectorParts &Cond = getVectorValue(it->getOperand(0));
3094 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3095 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3097 Value *ScalarCond = (VF == 1) ? Cond[0] :
3098 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3100 for (unsigned Part = 0; Part < UF; ++Part) {
3101 Entry[Part] = Builder.CreateSelect(
3102 InvariantCond ? ScalarCond : Cond[Part],
3107 propagateMetadata(Entry, it);
3111 case Instruction::ICmp:
3112 case Instruction::FCmp: {
3113 // Widen compares. Generate vector compares.
3114 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3115 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3116 setDebugLocFromInst(Builder, it);
3117 VectorParts &A = getVectorValue(it->getOperand(0));
3118 VectorParts &B = getVectorValue(it->getOperand(1));
3119 for (unsigned Part = 0; Part < UF; ++Part) {
3122 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3124 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3128 propagateMetadata(Entry, it);
3132 case Instruction::Store:
3133 case Instruction::Load:
3134 vectorizeMemoryInstruction(it);
3136 case Instruction::ZExt:
3137 case Instruction::SExt:
3138 case Instruction::FPToUI:
3139 case Instruction::FPToSI:
3140 case Instruction::FPExt:
3141 case Instruction::PtrToInt:
3142 case Instruction::IntToPtr:
3143 case Instruction::SIToFP:
3144 case Instruction::UIToFP:
3145 case Instruction::Trunc:
3146 case Instruction::FPTrunc:
3147 case Instruction::BitCast: {
3148 CastInst *CI = dyn_cast<CastInst>(it);
3149 setDebugLocFromInst(Builder, it);
3150 /// Optimize the special case where the source is the induction
3151 /// variable. Notice that we can only optimize the 'trunc' case
3152 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3153 /// c. other casts depend on pointer size.
3154 if (CI->getOperand(0) == OldInduction &&
3155 it->getOpcode() == Instruction::Trunc) {
3156 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3158 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3159 LoopVectorizationLegality::InductionInfo II =
3160 Legal->getInductionVars()->lookup(OldInduction);
3162 ConstantInt::getSigned(CI->getType(), II.StepValue->getSExtValue());
3163 for (unsigned Part = 0; Part < UF; ++Part)
3164 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
3165 propagateMetadata(Entry, it);
3168 /// Vectorize casts.
3169 Type *DestTy = (VF == 1) ? CI->getType() :
3170 VectorType::get(CI->getType(), VF);
3172 VectorParts &A = getVectorValue(it->getOperand(0));
3173 for (unsigned Part = 0; Part < UF; ++Part)
3174 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3175 propagateMetadata(Entry, it);
3179 case Instruction::Call: {
3180 // Ignore dbg intrinsics.
3181 if (isa<DbgInfoIntrinsic>(it))
3183 setDebugLocFromInst(Builder, it);
3185 Module *M = BB->getParent()->getParent();
3186 CallInst *CI = cast<CallInst>(it);
3187 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3188 assert(ID && "Not an intrinsic call!");
3190 case Intrinsic::assume:
3191 case Intrinsic::lifetime_end:
3192 case Intrinsic::lifetime_start:
3193 scalarizeInstruction(it);
3196 bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
3197 for (unsigned Part = 0; Part < UF; ++Part) {
3198 SmallVector<Value *, 4> Args;
3199 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3200 if (HasScalarOpd && i == 1) {
3201 Args.push_back(CI->getArgOperand(i));
3204 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3205 Args.push_back(Arg[Part]);
3207 Type *Tys[] = {CI->getType()};
3209 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3211 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3212 Entry[Part] = Builder.CreateCall(F, Args);
3215 propagateMetadata(Entry, it);
3222 // All other instructions are unsupported. Scalarize them.
3223 scalarizeInstruction(it);
3226 }// end of for_each instr.
3229 void InnerLoopVectorizer::updateAnalysis() {
3230 // Forget the original basic block.
3231 SE->forgetLoop(OrigLoop);
3233 // Update the dominator tree information.
3234 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3235 "Entry does not dominate exit.");
3237 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3238 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3239 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3241 // Due to if predication of stores we might create a sequence of "if(pred)
3242 // a[i] = ...; " blocks.
3243 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3245 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3246 else if (isPredicatedBlock(i)) {
3247 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3249 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3253 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3254 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3255 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3256 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3258 DEBUG(DT->verifyDomTree());
3261 /// \brief Check whether it is safe to if-convert this phi node.
3263 /// Phi nodes with constant expressions that can trap are not safe to if
3265 static bool canIfConvertPHINodes(BasicBlock *BB) {
3266 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3267 PHINode *Phi = dyn_cast<PHINode>(I);
3270 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3271 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3278 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3279 if (!EnableIfConversion) {
3280 emitAnalysis(VectorizationReport() << "if-conversion is disabled");
3284 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3286 // A list of pointers that we can safely read and write to.
3287 SmallPtrSet<Value *, 8> SafePointes;
3289 // Collect safe addresses.
3290 for (Loop::block_iterator BI = TheLoop->block_begin(),
3291 BE = TheLoop->block_end(); BI != BE; ++BI) {
3292 BasicBlock *BB = *BI;
3294 if (blockNeedsPredication(BB))
3297 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3298 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3299 SafePointes.insert(LI->getPointerOperand());
3300 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3301 SafePointes.insert(SI->getPointerOperand());
3305 // Collect the blocks that need predication.
3306 BasicBlock *Header = TheLoop->getHeader();
3307 for (Loop::block_iterator BI = TheLoop->block_begin(),
3308 BE = TheLoop->block_end(); BI != BE; ++BI) {
3309 BasicBlock *BB = *BI;
3311 // We don't support switch statements inside loops.
3312 if (!isa<BranchInst>(BB->getTerminator())) {
3313 emitAnalysis(VectorizationReport(BB->getTerminator())
3314 << "loop contains a switch statement");
3318 // We must be able to predicate all blocks that need to be predicated.
3319 if (blockNeedsPredication(BB)) {
3320 if (!blockCanBePredicated(BB, SafePointes)) {
3321 emitAnalysis(VectorizationReport(BB->getTerminator())
3322 << "control flow cannot be substituted for a select");
3325 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3326 emitAnalysis(VectorizationReport(BB->getTerminator())
3327 << "control flow cannot be substituted for a select");
3332 // We can if-convert this loop.
3336 bool LoopVectorizationLegality::canVectorize() {
3337 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3338 // be canonicalized.
3339 if (!TheLoop->getLoopPreheader()) {
3341 VectorizationReport() <<
3342 "loop control flow is not understood by vectorizer");
3346 // We can only vectorize innermost loops.
3347 if (!TheLoop->getSubLoopsVector().empty()) {
3348 emitAnalysis(VectorizationReport() << "loop is not the innermost loop");
3352 // We must have a single backedge.
3353 if (TheLoop->getNumBackEdges() != 1) {
3355 VectorizationReport() <<
3356 "loop control flow is not understood by vectorizer");
3360 // We must have a single exiting block.
3361 if (!TheLoop->getExitingBlock()) {
3363 VectorizationReport() <<
3364 "loop control flow is not understood by vectorizer");
3368 // We only handle bottom-tested loops, i.e. loop in which the condition is
3369 // checked at the end of each iteration. With that we can assume that all
3370 // instructions in the loop are executed the same number of times.
3371 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
3373 VectorizationReport() <<
3374 "loop control flow is not understood by vectorizer");
3378 // We need to have a loop header.
3379 DEBUG(dbgs() << "LV: Found a loop: " <<
3380 TheLoop->getHeader()->getName() << '\n');
3382 // Check if we can if-convert non-single-bb loops.
3383 unsigned NumBlocks = TheLoop->getNumBlocks();
3384 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3385 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3389 // ScalarEvolution needs to be able to find the exit count.
3390 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3391 if (ExitCount == SE->getCouldNotCompute()) {
3392 emitAnalysis(VectorizationReport() <<
3393 "could not determine number of loop iterations");
3394 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3398 // Check if we can vectorize the instructions and CFG in this loop.
3399 if (!canVectorizeInstrs()) {
3400 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3404 // Go over each instruction and look at memory deps.
3405 if (!canVectorizeMemory()) {
3406 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3410 // Collect all of the variables that remain uniform after vectorization.
3411 collectLoopUniforms();
3413 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3414 (LAI->getRuntimePointerCheck()->Need ? " (with a runtime bound check)" :
3418 // Okay! We can vectorize. At this point we don't have any other mem analysis
3419 // which may limit our maximum vectorization factor, so just return true with
3424 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3425 if (Ty->isPointerTy())
3426 return DL.getIntPtrType(Ty);
3428 // It is possible that char's or short's overflow when we ask for the loop's
3429 // trip count, work around this by changing the type size.
3430 if (Ty->getScalarSizeInBits() < 32)
3431 return Type::getInt32Ty(Ty->getContext());
3436 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3437 Ty0 = convertPointerToIntegerType(DL, Ty0);
3438 Ty1 = convertPointerToIntegerType(DL, Ty1);
3439 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3444 /// \brief Check that the instruction has outside loop users and is not an
3445 /// identified reduction variable.
3446 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3447 SmallPtrSetImpl<Value *> &Reductions) {
3448 // Reduction instructions are allowed to have exit users. All other
3449 // instructions must not have external users.
3450 if (!Reductions.count(Inst))
3451 //Check that all of the users of the loop are inside the BB.
3452 for (User *U : Inst->users()) {
3453 Instruction *UI = cast<Instruction>(U);
3454 // This user may be a reduction exit value.
3455 if (!TheLoop->contains(UI)) {
3456 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3463 bool LoopVectorizationLegality::canVectorizeInstrs() {
3464 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3465 BasicBlock *Header = TheLoop->getHeader();
3467 // Look for the attribute signaling the absence of NaNs.
3468 Function &F = *Header->getParent();
3469 if (F.hasFnAttribute("no-nans-fp-math"))
3471 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
3473 // For each block in the loop.
3474 for (Loop::block_iterator bb = TheLoop->block_begin(),
3475 be = TheLoop->block_end(); bb != be; ++bb) {
3477 // Scan the instructions in the block and look for hazards.
3478 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3481 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3482 Type *PhiTy = Phi->getType();
3483 // Check that this PHI type is allowed.
3484 if (!PhiTy->isIntegerTy() &&
3485 !PhiTy->isFloatingPointTy() &&
3486 !PhiTy->isPointerTy()) {
3487 emitAnalysis(VectorizationReport(it)
3488 << "loop control flow is not understood by vectorizer");
3489 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3493 // If this PHINode is not in the header block, then we know that we
3494 // can convert it to select during if-conversion. No need to check if
3495 // the PHIs in this block are induction or reduction variables.
3496 if (*bb != Header) {
3497 // Check that this instruction has no outside users or is an
3498 // identified reduction value with an outside user.
3499 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3501 emitAnalysis(VectorizationReport(it) <<
3502 "value could not be identified as "
3503 "an induction or reduction variable");
3507 // We only allow if-converted PHIs with exactly two incoming values.
3508 if (Phi->getNumIncomingValues() != 2) {
3509 emitAnalysis(VectorizationReport(it)
3510 << "control flow not understood by vectorizer");
3511 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3515 // This is the value coming from the preheader.
3516 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3517 ConstantInt *StepValue = nullptr;
3518 // Check if this is an induction variable.
3519 InductionKind IK = isInductionVariable(Phi, StepValue);
3521 if (IK_NoInduction != IK) {
3522 // Get the widest type.
3524 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3526 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3528 // Int inductions are special because we only allow one IV.
3529 if (IK == IK_IntInduction && StepValue->isOne()) {
3530 // Use the phi node with the widest type as induction. Use the last
3531 // one if there are multiple (no good reason for doing this other
3532 // than it is expedient).
3533 if (!Induction || PhiTy == WidestIndTy)
3537 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3538 Inductions[Phi] = InductionInfo(StartValue, IK, StepValue);
3540 // Until we explicitly handle the case of an induction variable with
3541 // an outside loop user we have to give up vectorizing this loop.
3542 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3543 emitAnalysis(VectorizationReport(it) <<
3544 "use of induction value outside of the "
3545 "loop is not handled by vectorizer");
3552 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3553 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3556 if (AddReductionVar(Phi, RK_IntegerMult)) {
3557 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3560 if (AddReductionVar(Phi, RK_IntegerOr)) {
3561 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3564 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3565 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3568 if (AddReductionVar(Phi, RK_IntegerXor)) {
3569 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3572 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3573 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3576 if (AddReductionVar(Phi, RK_FloatMult)) {
3577 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3580 if (AddReductionVar(Phi, RK_FloatAdd)) {
3581 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3584 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3585 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3590 emitAnalysis(VectorizationReport(it) <<
3591 "value that could not be identified as "
3592 "reduction is used outside the loop");
3593 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3595 }// end of PHI handling
3597 // We still don't handle functions. However, we can ignore dbg intrinsic
3598 // calls and we do handle certain intrinsic and libm functions.
3599 CallInst *CI = dyn_cast<CallInst>(it);
3600 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3601 emitAnalysis(VectorizationReport(it) <<
3602 "call instruction cannot be vectorized");
3603 DEBUG(dbgs() << "LV: Found a call site.\n");
3607 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
3608 // second argument is the same (i.e. loop invariant)
3610 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
3611 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
3612 emitAnalysis(VectorizationReport(it)
3613 << "intrinsic instruction cannot be vectorized");
3614 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
3619 // Check that the instruction return type is vectorizable.
3620 // Also, we can't vectorize extractelement instructions.
3621 if ((!VectorType::isValidElementType(it->getType()) &&
3622 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3623 emitAnalysis(VectorizationReport(it)
3624 << "instruction return type cannot be vectorized");
3625 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3629 // Check that the stored type is vectorizable.
3630 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3631 Type *T = ST->getValueOperand()->getType();
3632 if (!VectorType::isValidElementType(T)) {
3633 emitAnalysis(VectorizationReport(ST) <<
3634 "store instruction cannot be vectorized");
3637 if (EnableMemAccessVersioning)
3638 collectStridedAccess(ST);
3641 if (EnableMemAccessVersioning)
3642 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3643 collectStridedAccess(LI);
3645 // Reduction instructions are allowed to have exit users.
3646 // All other instructions must not have external users.
3647 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3648 emitAnalysis(VectorizationReport(it) <<
3649 "value cannot be used outside the loop");
3658 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3659 if (Inductions.empty()) {
3660 emitAnalysis(VectorizationReport()
3661 << "loop induction variable could not be identified");
3669 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3670 /// return the induction operand of the gep pointer.
3671 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3672 const DataLayout *DL, Loop *Lp) {
3673 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3677 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3679 // Check that all of the gep indices are uniform except for our induction
3681 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3682 if (i != InductionOperand &&
3683 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3685 return GEP->getOperand(InductionOperand);
3688 ///\brief Look for a cast use of the passed value.
3689 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3690 Value *UniqueCast = nullptr;
3691 for (User *U : Ptr->users()) {
3692 CastInst *CI = dyn_cast<CastInst>(U);
3693 if (CI && CI->getType() == Ty) {
3703 ///\brief Get the stride of a pointer access in a loop.
3704 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3705 /// pointer to the Value, or null otherwise.
3706 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3707 const DataLayout *DL, Loop *Lp) {
3708 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3709 if (!PtrTy || PtrTy->isAggregateType())
3712 // Try to remove a gep instruction to make the pointer (actually index at this
3713 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3714 // pointer, otherwise, we are analyzing the index.
3715 Value *OrigPtr = Ptr;
3717 // The size of the pointer access.
3718 int64_t PtrAccessSize = 1;
3720 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3721 const SCEV *V = SE->getSCEV(Ptr);
3725 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3726 V = C->getOperand();
3728 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3732 V = S->getStepRecurrence(*SE);
3736 // Strip off the size of access multiplication if we are still analyzing the
3738 if (OrigPtr == Ptr) {
3739 DL->getTypeAllocSize(PtrTy->getElementType());
3740 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3741 if (M->getOperand(0)->getSCEVType() != scConstant)
3744 const APInt &APStepVal =
3745 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3747 // Huge step value - give up.
3748 if (APStepVal.getBitWidth() > 64)
3751 int64_t StepVal = APStepVal.getSExtValue();
3752 if (PtrAccessSize != StepVal)
3754 V = M->getOperand(1);
3759 Type *StripedOffRecurrenceCast = nullptr;
3760 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3761 StripedOffRecurrenceCast = C->getType();
3762 V = C->getOperand();
3765 // Look for the loop invariant symbolic value.
3766 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3770 Value *Stride = U->getValue();
3771 if (!Lp->isLoopInvariant(Stride))
3774 // If we have stripped off the recurrence cast we have to make sure that we
3775 // return the value that is used in this loop so that we can replace it later.
3776 if (StripedOffRecurrenceCast)
3777 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3782 void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) {
3783 Value *Ptr = nullptr;
3784 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3785 Ptr = LI->getPointerOperand();
3786 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3787 Ptr = SI->getPointerOperand();
3791 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3795 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3796 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3797 Strides[Ptr] = Stride;
3798 StrideSet.insert(Stride);
3801 void LoopVectorizationLegality::collectLoopUniforms() {
3802 // We now know that the loop is vectorizable!
3803 // Collect variables that will remain uniform after vectorization.
3804 std::vector<Value*> Worklist;
3805 BasicBlock *Latch = TheLoop->getLoopLatch();
3807 // Start with the conditional branch and walk up the block.
3808 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3810 // Also add all consecutive pointer values; these values will be uniform
3811 // after vectorization (and subsequent cleanup) and, until revectorization is
3812 // supported, all dependencies must also be uniform.
3813 for (Loop::block_iterator B = TheLoop->block_begin(),
3814 BE = TheLoop->block_end(); B != BE; ++B)
3815 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
3817 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
3818 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3820 while (!Worklist.empty()) {
3821 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3822 Worklist.pop_back();
3824 // Look at instructions inside this loop.
3825 // Stop when reaching PHI nodes.
3826 // TODO: we need to follow values all over the loop, not only in this block.
3827 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3830 // This is a known uniform.
3833 // Insert all operands.
3834 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3838 bool LoopVectorizationLegality::canVectorizeMemory() {
3839 LAI = &LAA->getInfo(TheLoop, Strides);
3840 auto &OptionalReport = LAI->getReport();
3842 emitAnalysis(VectorizationReport(*OptionalReport));
3843 return LAI->canVectorizeMemory();
3846 static bool hasMultipleUsesOf(Instruction *I,
3847 SmallPtrSetImpl<Instruction *> &Insts) {
3848 unsigned NumUses = 0;
3849 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
3850 if (Insts.count(dyn_cast<Instruction>(*Use)))
3859 static bool areAllUsesIn(Instruction *I, SmallPtrSetImpl<Instruction *> &Set) {
3860 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
3861 if (!Set.count(dyn_cast<Instruction>(*Use)))
3866 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
3867 ReductionKind Kind) {
3868 if (Phi->getNumIncomingValues() != 2)
3871 // Reduction variables are only found in the loop header block.
3872 if (Phi->getParent() != TheLoop->getHeader())
3875 // Obtain the reduction start value from the value that comes from the loop
3877 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
3879 // ExitInstruction is the single value which is used outside the loop.
3880 // We only allow for a single reduction value to be used outside the loop.
3881 // This includes users of the reduction, variables (which form a cycle
3882 // which ends in the phi node).
3883 Instruction *ExitInstruction = nullptr;
3884 // Indicates that we found a reduction operation in our scan.
3885 bool FoundReduxOp = false;
3887 // We start with the PHI node and scan for all of the users of this
3888 // instruction. All users must be instructions that can be used as reduction
3889 // variables (such as ADD). We must have a single out-of-block user. The cycle
3890 // must include the original PHI.
3891 bool FoundStartPHI = false;
3893 // To recognize min/max patterns formed by a icmp select sequence, we store
3894 // the number of instruction we saw from the recognized min/max pattern,
3895 // to make sure we only see exactly the two instructions.
3896 unsigned NumCmpSelectPatternInst = 0;
3897 ReductionInstDesc ReduxDesc(false, nullptr);
3899 SmallPtrSet<Instruction *, 8> VisitedInsts;
3900 SmallVector<Instruction *, 8> Worklist;
3901 Worklist.push_back(Phi);
3902 VisitedInsts.insert(Phi);
3904 // A value in the reduction can be used:
3905 // - By the reduction:
3906 // - Reduction operation:
3907 // - One use of reduction value (safe).
3908 // - Multiple use of reduction value (not safe).
3910 // - All uses of the PHI must be the reduction (safe).
3911 // - Otherwise, not safe.
3912 // - By one instruction outside of the loop (safe).
3913 // - By further instructions outside of the loop (not safe).
3914 // - By an instruction that is not part of the reduction (not safe).
3916 // * An instruction type other than PHI or the reduction operation.
3917 // * A PHI in the header other than the initial PHI.
3918 while (!Worklist.empty()) {
3919 Instruction *Cur = Worklist.back();
3920 Worklist.pop_back();
3923 // If the instruction has no users then this is a broken chain and can't be
3924 // a reduction variable.
3925 if (Cur->use_empty())
3928 bool IsAPhi = isa<PHINode>(Cur);
3930 // A header PHI use other than the original PHI.
3931 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
3934 // Reductions of instructions such as Div, and Sub is only possible if the
3935 // LHS is the reduction variable.
3936 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
3937 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
3938 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
3941 // Any reduction instruction must be of one of the allowed kinds.
3942 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
3943 if (!ReduxDesc.IsReduction)
3946 // A reduction operation must only have one use of the reduction value.
3947 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
3948 hasMultipleUsesOf(Cur, VisitedInsts))
3951 // All inputs to a PHI node must be a reduction value.
3952 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
3955 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
3956 isa<SelectInst>(Cur)))
3957 ++NumCmpSelectPatternInst;
3958 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
3959 isa<SelectInst>(Cur)))
3960 ++NumCmpSelectPatternInst;
3962 // Check whether we found a reduction operator.
3963 FoundReduxOp |= !IsAPhi;
3965 // Process users of current instruction. Push non-PHI nodes after PHI nodes
3966 // onto the stack. This way we are going to have seen all inputs to PHI
3967 // nodes once we get to them.
3968 SmallVector<Instruction *, 8> NonPHIs;
3969 SmallVector<Instruction *, 8> PHIs;
3970 for (User *U : Cur->users()) {
3971 Instruction *UI = cast<Instruction>(U);
3973 // Check if we found the exit user.
3974 BasicBlock *Parent = UI->getParent();
3975 if (!TheLoop->contains(Parent)) {
3976 // Exit if you find multiple outside users or if the header phi node is
3977 // being used. In this case the user uses the value of the previous
3978 // iteration, in which case we would loose "VF-1" iterations of the
3979 // reduction operation if we vectorize.
3980 if (ExitInstruction != nullptr || Cur == Phi)
3983 // The instruction used by an outside user must be the last instruction
3984 // before we feed back to the reduction phi. Otherwise, we loose VF-1
3985 // operations on the value.
3986 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
3989 ExitInstruction = Cur;
3993 // Process instructions only once (termination). Each reduction cycle
3994 // value must only be used once, except by phi nodes and min/max
3995 // reductions which are represented as a cmp followed by a select.
3996 ReductionInstDesc IgnoredVal(false, nullptr);
3997 if (VisitedInsts.insert(UI).second) {
3998 if (isa<PHINode>(UI))
4001 NonPHIs.push_back(UI);
4002 } else if (!isa<PHINode>(UI) &&
4003 ((!isa<FCmpInst>(UI) &&
4004 !isa<ICmpInst>(UI) &&
4005 !isa<SelectInst>(UI)) ||
4006 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
4009 // Remember that we completed the cycle.
4011 FoundStartPHI = true;
4013 Worklist.append(PHIs.begin(), PHIs.end());
4014 Worklist.append(NonPHIs.begin(), NonPHIs.end());
4017 // This means we have seen one but not the other instruction of the
4018 // pattern or more than just a select and cmp.
4019 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
4020 NumCmpSelectPatternInst != 2)
4023 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
4026 // We found a reduction var if we have reached the original phi node and we
4027 // only have a single instruction with out-of-loop users.
4029 // This instruction is allowed to have out-of-loop users.
4030 AllowedExit.insert(ExitInstruction);
4032 // Save the description of this reduction variable.
4033 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
4034 ReduxDesc.MinMaxKind);
4035 Reductions[Phi] = RD;
4036 // We've ended the cycle. This is a reduction variable if we have an
4037 // outside user and it has a binary op.
4042 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
4043 /// pattern corresponding to a min(X, Y) or max(X, Y).
4044 LoopVectorizationLegality::ReductionInstDesc
4045 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
4046 ReductionInstDesc &Prev) {
4048 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
4049 "Expect a select instruction");
4050 Instruction *Cmp = nullptr;
4051 SelectInst *Select = nullptr;
4053 // We must handle the select(cmp()) as a single instruction. Advance to the
4055 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
4056 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
4057 return ReductionInstDesc(false, I);
4058 return ReductionInstDesc(Select, Prev.MinMaxKind);
4061 // Only handle single use cases for now.
4062 if (!(Select = dyn_cast<SelectInst>(I)))
4063 return ReductionInstDesc(false, I);
4064 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
4065 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
4066 return ReductionInstDesc(false, I);
4067 if (!Cmp->hasOneUse())
4068 return ReductionInstDesc(false, I);
4073 // Look for a min/max pattern.
4074 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4075 return ReductionInstDesc(Select, MRK_UIntMin);
4076 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4077 return ReductionInstDesc(Select, MRK_UIntMax);
4078 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4079 return ReductionInstDesc(Select, MRK_SIntMax);
4080 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4081 return ReductionInstDesc(Select, MRK_SIntMin);
4082 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4083 return ReductionInstDesc(Select, MRK_FloatMin);
4084 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4085 return ReductionInstDesc(Select, MRK_FloatMax);
4086 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4087 return ReductionInstDesc(Select, MRK_FloatMin);
4088 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4089 return ReductionInstDesc(Select, MRK_FloatMax);
4091 return ReductionInstDesc(false, I);
4094 LoopVectorizationLegality::ReductionInstDesc
4095 LoopVectorizationLegality::isReductionInstr(Instruction *I,
4097 ReductionInstDesc &Prev) {
4098 bool FP = I->getType()->isFloatingPointTy();
4099 bool FastMath = FP && I->hasUnsafeAlgebra();
4100 switch (I->getOpcode()) {
4102 return ReductionInstDesc(false, I);
4103 case Instruction::PHI:
4104 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
4105 Kind != RK_FloatMinMax))
4106 return ReductionInstDesc(false, I);
4107 return ReductionInstDesc(I, Prev.MinMaxKind);
4108 case Instruction::Sub:
4109 case Instruction::Add:
4110 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4111 case Instruction::Mul:
4112 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4113 case Instruction::And:
4114 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4115 case Instruction::Or:
4116 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4117 case Instruction::Xor:
4118 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4119 case Instruction::FMul:
4120 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4121 case Instruction::FSub:
4122 case Instruction::FAdd:
4123 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4124 case Instruction::FCmp:
4125 case Instruction::ICmp:
4126 case Instruction::Select:
4127 if (Kind != RK_IntegerMinMax &&
4128 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4129 return ReductionInstDesc(false, I);
4130 return isMinMaxSelectCmpPattern(I, Prev);
4134 LoopVectorizationLegality::InductionKind
4135 LoopVectorizationLegality::isInductionVariable(PHINode *Phi,
4136 ConstantInt *&StepValue) {
4137 Type *PhiTy = Phi->getType();
4138 // We only handle integer and pointer inductions variables.
4139 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4140 return IK_NoInduction;
4142 // Check that the PHI is consecutive.
4143 const SCEV *PhiScev = SE->getSCEV(Phi);
4144 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4146 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4147 return IK_NoInduction;
4150 const SCEV *Step = AR->getStepRecurrence(*SE);
4151 // Calculate the pointer stride and check if it is consecutive.
4152 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4154 return IK_NoInduction;
4156 ConstantInt *CV = C->getValue();
4157 if (PhiTy->isIntegerTy()) {
4159 return IK_IntInduction;
4162 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4163 Type *PointerElementType = PhiTy->getPointerElementType();
4164 // The pointer stride cannot be determined if the pointer element type is not
4166 if (!PointerElementType->isSized())
4167 return IK_NoInduction;
4169 int64_t Size = static_cast<int64_t>(DL->getTypeAllocSize(PointerElementType));
4170 int64_t CVSize = CV->getSExtValue();
4172 return IK_NoInduction;
4173 StepValue = ConstantInt::getSigned(CV->getType(), CVSize / Size);
4174 return IK_PtrInduction;
4177 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4178 Value *In0 = const_cast<Value*>(V);
4179 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4183 return Inductions.count(PN);
4186 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4187 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4190 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4191 SmallPtrSetImpl<Value *> &SafePtrs) {
4193 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4194 // Check that we don't have a constant expression that can trap as operand.
4195 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4197 if (Constant *C = dyn_cast<Constant>(*OI))
4201 // We might be able to hoist the load.
4202 if (it->mayReadFromMemory()) {
4203 LoadInst *LI = dyn_cast<LoadInst>(it);
4206 if (!SafePtrs.count(LI->getPointerOperand())) {
4207 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
4208 MaskedOp.insert(LI);
4215 // We don't predicate stores at the moment.
4216 if (it->mayWriteToMemory()) {
4217 StoreInst *SI = dyn_cast<StoreInst>(it);
4218 // We only support predication of stores in basic blocks with one
4223 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
4224 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
4226 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
4227 !isSinglePredecessor) {
4228 // Build a masked store if it is legal for the target, otherwise scalarize
4230 bool isLegalMaskedOp =
4231 isLegalMaskedStore(SI->getValueOperand()->getType(),
4232 SI->getPointerOperand());
4233 if (isLegalMaskedOp) {
4235 MaskedOp.insert(SI);
4244 // The instructions below can trap.
4245 switch (it->getOpcode()) {
4247 case Instruction::UDiv:
4248 case Instruction::SDiv:
4249 case Instruction::URem:
4250 case Instruction::SRem:
4258 LoopVectorizationCostModel::VectorizationFactor
4259 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
4260 // Width 1 means no vectorize
4261 VectorizationFactor Factor = { 1U, 0U };
4262 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4263 emitAnalysis(VectorizationReport() <<
4264 "runtime pointer checks needed. Enable vectorization of this "
4265 "loop with '#pragma clang loop vectorize(enable)' when "
4266 "compiling with -Os");
4267 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4271 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
4272 emitAnalysis(VectorizationReport() <<
4273 "store that is conditionally executed prevents vectorization");
4274 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
4278 // Find the trip count.
4279 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4280 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4282 unsigned WidestType = getWidestType();
4283 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4284 unsigned MaxSafeDepDist = -1U;
4285 if (Legal->getMaxSafeDepDistBytes() != -1U)
4286 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4287 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4288 WidestRegister : MaxSafeDepDist);
4289 unsigned MaxVectorSize = WidestRegister / WidestType;
4290 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4291 DEBUG(dbgs() << "LV: The Widest register is: "
4292 << WidestRegister << " bits.\n");
4294 if (MaxVectorSize == 0) {
4295 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4299 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
4300 " into one vector!");
4302 unsigned VF = MaxVectorSize;
4304 // If we optimize the program for size, avoid creating the tail loop.
4306 // If we are unable to calculate the trip count then don't try to vectorize.
4309 (VectorizationReport() <<
4310 "unable to calculate the loop count due to complex control flow");
4311 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4315 // Find the maximum SIMD width that can fit within the trip count.
4316 VF = TC % MaxVectorSize;
4321 // If the trip count that we found modulo the vectorization factor is not
4322 // zero then we require a tail.
4324 emitAnalysis(VectorizationReport() <<
4325 "cannot optimize for size and vectorize at the "
4326 "same time. Enable vectorization of this loop "
4327 "with '#pragma clang loop vectorize(enable)' "
4328 "when compiling with -Os");
4329 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4334 int UserVF = Hints->getWidth();
4336 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4337 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4339 Factor.Width = UserVF;
4343 float Cost = expectedCost(1);
4345 const float ScalarCost = Cost;
4348 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
4350 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
4351 // Ignore scalar width, because the user explicitly wants vectorization.
4352 if (ForceVectorization && VF > 1) {
4354 Cost = expectedCost(Width) / (float)Width;
4357 for (unsigned i=2; i <= VF; i*=2) {
4358 // Notice that the vector loop needs to be executed less times, so
4359 // we need to divide the cost of the vector loops by the width of
4360 // the vector elements.
4361 float VectorCost = expectedCost(i) / (float)i;
4362 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
4363 (int)VectorCost << ".\n");
4364 if (VectorCost < Cost) {
4370 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
4371 << "LV: Vectorization seems to be not beneficial, "
4372 << "but was forced by a user.\n");
4373 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
4374 Factor.Width = Width;
4375 Factor.Cost = Width * Cost;
4379 unsigned LoopVectorizationCostModel::getWidestType() {
4380 unsigned MaxWidth = 8;
4383 for (Loop::block_iterator bb = TheLoop->block_begin(),
4384 be = TheLoop->block_end(); bb != be; ++bb) {
4385 BasicBlock *BB = *bb;
4387 // For each instruction in the loop.
4388 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4389 Type *T = it->getType();
4391 // Ignore ephemeral values.
4392 if (EphValues.count(it))
4395 // Only examine Loads, Stores and PHINodes.
4396 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4399 // Examine PHI nodes that are reduction variables.
4400 if (PHINode *PN = dyn_cast<PHINode>(it))
4401 if (!Legal->getReductionVars()->count(PN))
4404 // Examine the stored values.
4405 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4406 T = ST->getValueOperand()->getType();
4408 // Ignore loaded pointer types and stored pointer types that are not
4409 // consecutive. However, we do want to take consecutive stores/loads of
4410 // pointer vectors into account.
4411 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4414 MaxWidth = std::max(MaxWidth,
4415 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
4423 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
4425 unsigned LoopCost) {
4427 // -- The unroll heuristics --
4428 // We unroll the loop in order to expose ILP and reduce the loop overhead.
4429 // There are many micro-architectural considerations that we can't predict
4430 // at this level. For example, frontend pressure (on decode or fetch) due to
4431 // code size, or the number and capabilities of the execution ports.
4433 // We use the following heuristics to select the unroll factor:
4434 // 1. If the code has reductions, then we unroll in order to break the cross
4435 // iteration dependency.
4436 // 2. If the loop is really small, then we unroll in order to reduce the loop
4438 // 3. We don't unroll if we think that we will spill registers to memory due
4439 // to the increased register pressure.
4441 // Use the user preference, unless 'auto' is selected.
4442 int UserUF = Hints->getInterleave();
4446 // When we optimize for size, we don't unroll.
4450 // We used the distance for the unroll factor.
4451 if (Legal->getMaxSafeDepDistBytes() != -1U)
4454 // Do not unroll loops with a relatively small trip count.
4455 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4456 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
4459 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
4460 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
4464 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
4465 TargetNumRegisters = ForceTargetNumScalarRegs;
4467 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
4468 TargetNumRegisters = ForceTargetNumVectorRegs;
4471 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4472 // We divide by these constants so assume that we have at least one
4473 // instruction that uses at least one register.
4474 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4475 R.NumInstructions = std::max(R.NumInstructions, 1U);
4477 // We calculate the unroll factor using the following formula.
4478 // Subtract the number of loop invariants from the number of available
4479 // registers. These registers are used by all of the unrolled instances.
4480 // Next, divide the remaining registers by the number of registers that is
4481 // required by the loop, in order to estimate how many parallel instances
4482 // fit without causing spills. All of this is rounded down if necessary to be
4483 // a power of two. We want power of two unroll factors to simplify any
4484 // addressing operations or alignment considerations.
4485 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
4488 // Don't count the induction variable as unrolled.
4489 if (EnableIndVarRegisterHeur)
4490 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
4491 std::max(1U, (R.MaxLocalUsers - 1)));
4493 // Clamp the unroll factor ranges to reasonable factors.
4494 unsigned MaxInterleaveSize = TTI.getMaxInterleaveFactor();
4496 // Check if the user has overridden the unroll max.
4498 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
4499 MaxInterleaveSize = ForceTargetMaxScalarInterleaveFactor;
4501 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
4502 MaxInterleaveSize = ForceTargetMaxVectorInterleaveFactor;
4505 // If we did not calculate the cost for VF (because the user selected the VF)
4506 // then we calculate the cost of VF here.
4508 LoopCost = expectedCost(VF);
4510 // Clamp the calculated UF to be between the 1 and the max unroll factor
4511 // that the target allows.
4512 if (UF > MaxInterleaveSize)
4513 UF = MaxInterleaveSize;
4517 // Unroll if we vectorized this loop and there is a reduction that could
4518 // benefit from unrolling.
4519 if (VF > 1 && Legal->getReductionVars()->size()) {
4520 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
4524 // Note that if we've already vectorized the loop we will have done the
4525 // runtime check and so unrolling won't require further checks.
4526 bool UnrollingRequiresRuntimePointerCheck =
4527 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
4529 // We want to unroll small loops in order to reduce the loop overhead and
4530 // potentially expose ILP opportunities.
4531 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
4532 if (!UnrollingRequiresRuntimePointerCheck &&
4533 LoopCost < SmallLoopCost) {
4534 // We assume that the cost overhead is 1 and we use the cost model
4535 // to estimate the cost of the loop and unroll until the cost of the
4536 // loop overhead is about 5% of the cost of the loop.
4537 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
4539 // Unroll until store/load ports (estimated by max unroll factor) are
4541 unsigned NumStores = Legal->getNumStores();
4542 unsigned NumLoads = Legal->getNumLoads();
4543 unsigned StoresUF = UF / (NumStores ? NumStores : 1);
4544 unsigned LoadsUF = UF / (NumLoads ? NumLoads : 1);
4546 // If we have a scalar reduction (vector reductions are already dealt with
4547 // by this point), we can increase the critical path length if the loop
4548 // we're unrolling is inside another loop. Limit, by default to 2, so the
4549 // critical path only gets increased by one reduction operation.
4550 if (Legal->getReductionVars()->size() &&
4551 TheLoop->getLoopDepth() > 1) {
4552 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionUF);
4553 SmallUF = std::min(SmallUF, F);
4554 StoresUF = std::min(StoresUF, F);
4555 LoadsUF = std::min(LoadsUF, F);
4558 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
4559 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
4560 return std::max(StoresUF, LoadsUF);
4563 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
4567 DEBUG(dbgs() << "LV: Not Unrolling.\n");
4571 LoopVectorizationCostModel::RegisterUsage
4572 LoopVectorizationCostModel::calculateRegisterUsage() {
4573 // This function calculates the register usage by measuring the highest number
4574 // of values that are alive at a single location. Obviously, this is a very
4575 // rough estimation. We scan the loop in a topological order in order and
4576 // assign a number to each instruction. We use RPO to ensure that defs are
4577 // met before their users. We assume that each instruction that has in-loop
4578 // users starts an interval. We record every time that an in-loop value is
4579 // used, so we have a list of the first and last occurrences of each
4580 // instruction. Next, we transpose this data structure into a multi map that
4581 // holds the list of intervals that *end* at a specific location. This multi
4582 // map allows us to perform a linear search. We scan the instructions linearly
4583 // and record each time that a new interval starts, by placing it in a set.
4584 // If we find this value in the multi-map then we remove it from the set.
4585 // The max register usage is the maximum size of the set.
4586 // We also search for instructions that are defined outside the loop, but are
4587 // used inside the loop. We need this number separately from the max-interval
4588 // usage number because when we unroll, loop-invariant values do not take
4590 LoopBlocksDFS DFS(TheLoop);
4594 R.NumInstructions = 0;
4596 // Each 'key' in the map opens a new interval. The values
4597 // of the map are the index of the 'last seen' usage of the
4598 // instruction that is the key.
4599 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4600 // Maps instruction to its index.
4601 DenseMap<unsigned, Instruction*> IdxToInstr;
4602 // Marks the end of each interval.
4603 IntervalMap EndPoint;
4604 // Saves the list of instruction indices that are used in the loop.
4605 SmallSet<Instruction*, 8> Ends;
4606 // Saves the list of values that are used in the loop but are
4607 // defined outside the loop, such as arguments and constants.
4608 SmallPtrSet<Value*, 8> LoopInvariants;
4611 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4612 be = DFS.endRPO(); bb != be; ++bb) {
4613 R.NumInstructions += (*bb)->size();
4614 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4616 Instruction *I = it;
4617 IdxToInstr[Index++] = I;
4619 // Save the end location of each USE.
4620 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4621 Value *U = I->getOperand(i);
4622 Instruction *Instr = dyn_cast<Instruction>(U);
4624 // Ignore non-instruction values such as arguments, constants, etc.
4625 if (!Instr) continue;
4627 // If this instruction is outside the loop then record it and continue.
4628 if (!TheLoop->contains(Instr)) {
4629 LoopInvariants.insert(Instr);
4633 // Overwrite previous end points.
4634 EndPoint[Instr] = Index;
4640 // Saves the list of intervals that end with the index in 'key'.
4641 typedef SmallVector<Instruction*, 2> InstrList;
4642 DenseMap<unsigned, InstrList> TransposeEnds;
4644 // Transpose the EndPoints to a list of values that end at each index.
4645 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4647 TransposeEnds[it->second].push_back(it->first);
4649 SmallSet<Instruction*, 8> OpenIntervals;
4650 unsigned MaxUsage = 0;
4653 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4654 for (unsigned int i = 0; i < Index; ++i) {
4655 Instruction *I = IdxToInstr[i];
4656 // Ignore instructions that are never used within the loop.
4657 if (!Ends.count(I)) continue;
4659 // Ignore ephemeral values.
4660 if (EphValues.count(I))
4663 // Remove all of the instructions that end at this location.
4664 InstrList &List = TransposeEnds[i];
4665 for (unsigned int j=0, e = List.size(); j < e; ++j)
4666 OpenIntervals.erase(List[j]);
4668 // Count the number of live interals.
4669 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4671 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4672 OpenIntervals.size() << '\n');
4674 // Add the current instruction to the list of open intervals.
4675 OpenIntervals.insert(I);
4678 unsigned Invariant = LoopInvariants.size();
4679 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
4680 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
4681 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
4683 R.LoopInvariantRegs = Invariant;
4684 R.MaxLocalUsers = MaxUsage;
4688 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4692 for (Loop::block_iterator bb = TheLoop->block_begin(),
4693 be = TheLoop->block_end(); bb != be; ++bb) {
4694 unsigned BlockCost = 0;
4695 BasicBlock *BB = *bb;
4697 // For each instruction in the old loop.
4698 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4699 // Skip dbg intrinsics.
4700 if (isa<DbgInfoIntrinsic>(it))
4703 // Ignore ephemeral values.
4704 if (EphValues.count(it))
4707 unsigned C = getInstructionCost(it, VF);
4709 // Check if we should override the cost.
4710 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
4711 C = ForceTargetInstructionCost;
4714 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
4715 VF << " For instruction: " << *it << '\n');
4718 // We assume that if-converted blocks have a 50% chance of being executed.
4719 // When the code is scalar then some of the blocks are avoided due to CF.
4720 // When the code is vectorized we execute all code paths.
4721 if (VF == 1 && Legal->blockNeedsPredication(*bb))
4730 /// \brief Check whether the address computation for a non-consecutive memory
4731 /// access looks like an unlikely candidate for being merged into the indexing
4734 /// We look for a GEP which has one index that is an induction variable and all
4735 /// other indices are loop invariant. If the stride of this access is also
4736 /// within a small bound we decide that this address computation can likely be
4737 /// merged into the addressing mode.
4738 /// In all other cases, we identify the address computation as complex.
4739 static bool isLikelyComplexAddressComputation(Value *Ptr,
4740 LoopVectorizationLegality *Legal,
4741 ScalarEvolution *SE,
4742 const Loop *TheLoop) {
4743 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
4747 // We are looking for a gep with all loop invariant indices except for one
4748 // which should be an induction variable.
4749 unsigned NumOperands = Gep->getNumOperands();
4750 for (unsigned i = 1; i < NumOperands; ++i) {
4751 Value *Opd = Gep->getOperand(i);
4752 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
4753 !Legal->isInductionVariable(Opd))
4757 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
4758 // can likely be merged into the address computation.
4759 unsigned MaxMergeDistance = 64;
4761 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
4765 // Check the step is constant.
4766 const SCEV *Step = AddRec->getStepRecurrence(*SE);
4767 // Calculate the pointer stride and check if it is consecutive.
4768 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4772 const APInt &APStepVal = C->getValue()->getValue();
4774 // Huge step value - give up.
4775 if (APStepVal.getBitWidth() > 64)
4778 int64_t StepVal = APStepVal.getSExtValue();
4780 return StepVal > MaxMergeDistance;
4783 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
4784 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
4790 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
4791 // If we know that this instruction will remain uniform, check the cost of
4792 // the scalar version.
4793 if (Legal->isUniformAfterVectorization(I))
4796 Type *RetTy = I->getType();
4797 Type *VectorTy = ToVectorTy(RetTy, VF);
4799 // TODO: We need to estimate the cost of intrinsic calls.
4800 switch (I->getOpcode()) {
4801 case Instruction::GetElementPtr:
4802 // We mark this instruction as zero-cost because the cost of GEPs in
4803 // vectorized code depends on whether the corresponding memory instruction
4804 // is scalarized or not. Therefore, we handle GEPs with the memory
4805 // instruction cost.
4807 case Instruction::Br: {
4808 return TTI.getCFInstrCost(I->getOpcode());
4810 case Instruction::PHI:
4811 //TODO: IF-converted IFs become selects.
4813 case Instruction::Add:
4814 case Instruction::FAdd:
4815 case Instruction::Sub:
4816 case Instruction::FSub:
4817 case Instruction::Mul:
4818 case Instruction::FMul:
4819 case Instruction::UDiv:
4820 case Instruction::SDiv:
4821 case Instruction::FDiv:
4822 case Instruction::URem:
4823 case Instruction::SRem:
4824 case Instruction::FRem:
4825 case Instruction::Shl:
4826 case Instruction::LShr:
4827 case Instruction::AShr:
4828 case Instruction::And:
4829 case Instruction::Or:
4830 case Instruction::Xor: {
4831 // Since we will replace the stride by 1 the multiplication should go away.
4832 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
4834 // Certain instructions can be cheaper to vectorize if they have a constant
4835 // second vector operand. One example of this are shifts on x86.
4836 TargetTransformInfo::OperandValueKind Op1VK =
4837 TargetTransformInfo::OK_AnyValue;
4838 TargetTransformInfo::OperandValueKind Op2VK =
4839 TargetTransformInfo::OK_AnyValue;
4840 TargetTransformInfo::OperandValueProperties Op1VP =
4841 TargetTransformInfo::OP_None;
4842 TargetTransformInfo::OperandValueProperties Op2VP =
4843 TargetTransformInfo::OP_None;
4844 Value *Op2 = I->getOperand(1);
4846 // Check for a splat of a constant or for a non uniform vector of constants.
4847 if (isa<ConstantInt>(Op2)) {
4848 ConstantInt *CInt = cast<ConstantInt>(Op2);
4849 if (CInt && CInt->getValue().isPowerOf2())
4850 Op2VP = TargetTransformInfo::OP_PowerOf2;
4851 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4852 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
4853 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
4854 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
4856 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
4857 if (CInt && CInt->getValue().isPowerOf2())
4858 Op2VP = TargetTransformInfo::OP_PowerOf2;
4859 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4863 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
4866 case Instruction::Select: {
4867 SelectInst *SI = cast<SelectInst>(I);
4868 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
4869 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
4870 Type *CondTy = SI->getCondition()->getType();
4872 CondTy = VectorType::get(CondTy, VF);
4874 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
4876 case Instruction::ICmp:
4877 case Instruction::FCmp: {
4878 Type *ValTy = I->getOperand(0)->getType();
4879 VectorTy = ToVectorTy(ValTy, VF);
4880 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
4882 case Instruction::Store:
4883 case Instruction::Load: {
4884 StoreInst *SI = dyn_cast<StoreInst>(I);
4885 LoadInst *LI = dyn_cast<LoadInst>(I);
4886 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
4888 VectorTy = ToVectorTy(ValTy, VF);
4890 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
4891 unsigned AS = SI ? SI->getPointerAddressSpace() :
4892 LI->getPointerAddressSpace();
4893 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
4894 // We add the cost of address computation here instead of with the gep
4895 // instruction because only here we know whether the operation is
4898 return TTI.getAddressComputationCost(VectorTy) +
4899 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4901 // Scalarized loads/stores.
4902 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
4903 bool Reverse = ConsecutiveStride < 0;
4904 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
4905 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
4906 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
4907 bool IsComplexComputation =
4908 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
4910 // The cost of extracting from the value vector and pointer vector.
4911 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
4912 for (unsigned i = 0; i < VF; ++i) {
4913 // The cost of extracting the pointer operand.
4914 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
4915 // In case of STORE, the cost of ExtractElement from the vector.
4916 // In case of LOAD, the cost of InsertElement into the returned
4918 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
4919 Instruction::InsertElement,
4923 // The cost of the scalar loads/stores.
4924 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
4925 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
4930 // Wide load/stores.
4931 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
4932 if (Legal->isMaskRequired(I))
4933 Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment,
4936 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4939 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
4943 case Instruction::ZExt:
4944 case Instruction::SExt:
4945 case Instruction::FPToUI:
4946 case Instruction::FPToSI:
4947 case Instruction::FPExt:
4948 case Instruction::PtrToInt:
4949 case Instruction::IntToPtr:
4950 case Instruction::SIToFP:
4951 case Instruction::UIToFP:
4952 case Instruction::Trunc:
4953 case Instruction::FPTrunc:
4954 case Instruction::BitCast: {
4955 // We optimize the truncation of induction variable.
4956 // The cost of these is the same as the scalar operation.
4957 if (I->getOpcode() == Instruction::Trunc &&
4958 Legal->isInductionVariable(I->getOperand(0)))
4959 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
4960 I->getOperand(0)->getType());
4962 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
4963 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
4965 case Instruction::Call: {
4966 CallInst *CI = cast<CallInst>(I);
4967 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
4968 assert(ID && "Not an intrinsic call!");
4969 Type *RetTy = ToVectorTy(CI->getType(), VF);
4970 SmallVector<Type*, 4> Tys;
4971 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
4972 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
4973 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
4976 // We are scalarizing the instruction. Return the cost of the scalar
4977 // instruction, plus the cost of insert and extract into vector
4978 // elements, times the vector width.
4981 if (!RetTy->isVoidTy() && VF != 1) {
4982 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
4984 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
4987 // The cost of inserting the results plus extracting each one of the
4989 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
4992 // The cost of executing VF copies of the scalar instruction. This opcode
4993 // is unknown. Assume that it is the same as 'mul'.
4994 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5000 char LoopVectorize::ID = 0;
5001 static const char lv_name[] = "Loop Vectorization";
5002 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5003 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
5004 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
5005 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
5006 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
5007 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5008 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5009 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5010 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
5011 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5012 INITIALIZE_PASS_DEPENDENCY(LoopAccessAnalysis)
5013 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5016 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5017 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5021 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5022 // Check for a store.
5023 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5024 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5026 // Check for a load.
5027 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5028 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5034 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5035 bool IfPredicateStore) {
5036 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5037 // Holds vector parameters or scalars, in case of uniform vals.
5038 SmallVector<VectorParts, 4> Params;
5040 setDebugLocFromInst(Builder, Instr);
5042 // Find all of the vectorized parameters.
5043 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5044 Value *SrcOp = Instr->getOperand(op);
5046 // If we are accessing the old induction variable, use the new one.
5047 if (SrcOp == OldInduction) {
5048 Params.push_back(getVectorValue(SrcOp));
5052 // Try using previously calculated values.
5053 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5055 // If the src is an instruction that appeared earlier in the basic block
5056 // then it should already be vectorized.
5057 if (SrcInst && OrigLoop->contains(SrcInst)) {
5058 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5059 // The parameter is a vector value from earlier.
5060 Params.push_back(WidenMap.get(SrcInst));
5062 // The parameter is a scalar from outside the loop. Maybe even a constant.
5063 VectorParts Scalars;
5064 Scalars.append(UF, SrcOp);
5065 Params.push_back(Scalars);
5069 assert(Params.size() == Instr->getNumOperands() &&
5070 "Invalid number of operands");
5072 // Does this instruction return a value ?
5073 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5075 Value *UndefVec = IsVoidRetTy ? nullptr :
5076 UndefValue::get(Instr->getType());
5077 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5078 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5080 Instruction *InsertPt = Builder.GetInsertPoint();
5081 BasicBlock *IfBlock = Builder.GetInsertBlock();
5082 BasicBlock *CondBlock = nullptr;
5085 Loop *VectorLp = nullptr;
5086 if (IfPredicateStore) {
5087 assert(Instr->getParent()->getSinglePredecessor() &&
5088 "Only support single predecessor blocks");
5089 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5090 Instr->getParent());
5091 VectorLp = LI->getLoopFor(IfBlock);
5092 assert(VectorLp && "Must have a loop for this block");
5095 // For each vector unroll 'part':
5096 for (unsigned Part = 0; Part < UF; ++Part) {
5097 // For each scalar that we create:
5099 // Start an "if (pred) a[i] = ..." block.
5100 Value *Cmp = nullptr;
5101 if (IfPredicateStore) {
5102 if (Cond[Part]->getType()->isVectorTy())
5104 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5105 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5106 ConstantInt::get(Cond[Part]->getType(), 1));
5107 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5108 LoopVectorBody.push_back(CondBlock);
5109 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
5110 // Update Builder with newly created basic block.
5111 Builder.SetInsertPoint(InsertPt);
5114 Instruction *Cloned = Instr->clone();
5116 Cloned->setName(Instr->getName() + ".cloned");
5117 // Replace the operands of the cloned instructions with extracted scalars.
5118 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5119 Value *Op = Params[op][Part];
5120 Cloned->setOperand(op, Op);
5123 // Place the cloned scalar in the new loop.
5124 Builder.Insert(Cloned);
5126 // If the original scalar returns a value we need to place it in a vector
5127 // so that future users will be able to use it.
5129 VecResults[Part] = Cloned;
5132 if (IfPredicateStore) {
5133 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
5134 LoopVectorBody.push_back(NewIfBlock);
5135 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
5136 Builder.SetInsertPoint(InsertPt);
5137 Instruction *OldBr = IfBlock->getTerminator();
5138 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
5139 OldBr->eraseFromParent();
5140 IfBlock = NewIfBlock;
5145 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5146 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5147 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5149 return scalarizeInstruction(Instr, IfPredicateStore);
5152 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5156 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5160 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) {
5161 // When unrolling and the VF is 1, we only need to add a simple scalar.
5162 Type *ITy = Val->getType();
5163 assert(!ITy->isVectorTy() && "Val must be a scalar");
5164 Constant *C = ConstantInt::get(ITy, StartIdx);
5165 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");