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 /// InnerLoopVectorizer vectorizes loops which contain only one basic
222 /// block to a specified vectorization factor (VF).
223 /// This class performs the widening of scalars into vectors, or multiple
224 /// scalars. This class also implements the following features:
225 /// * It inserts an epilogue loop for handling loops that don't have iteration
226 /// counts that are known to be a multiple of the vectorization factor.
227 /// * It handles the code generation for reduction variables.
228 /// * Scalarization (implementation using scalars) of un-vectorizable
230 /// InnerLoopVectorizer does not perform any vectorization-legality
231 /// checks, and relies on the caller to check for the different legality
232 /// aspects. The InnerLoopVectorizer relies on the
233 /// LoopVectorizationLegality class to provide information about the induction
234 /// and reduction variables that were found to a given vectorization factor.
235 class InnerLoopVectorizer {
237 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
238 DominatorTree *DT, const DataLayout *DL,
239 const TargetLibraryInfo *TLI, unsigned VecWidth,
240 unsigned UnrollFactor)
241 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
242 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
243 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
246 // Perform the actual loop widening (vectorization).
247 void vectorize(LoopVectorizationLegality *L) {
249 // Create a new empty loop. Unlink the old loop and connect the new one.
251 // Widen each instruction in the old loop to a new one in the new loop.
252 // Use the Legality module to find the induction and reduction variables.
254 // Register the new loop and update the analysis passes.
258 virtual ~InnerLoopVectorizer() {}
261 /// A small list of PHINodes.
262 typedef SmallVector<PHINode*, 4> PhiVector;
263 /// When we unroll loops we have multiple vector values for each scalar.
264 /// This data structure holds the unrolled and vectorized values that
265 /// originated from one scalar instruction.
266 typedef SmallVector<Value*, 2> VectorParts;
268 // When we if-convert we need create edge masks. We have to cache values so
269 // that we don't end up with exponential recursion/IR.
270 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
271 VectorParts> EdgeMaskCache;
273 /// \brief Add checks for strides that where assumed to be 1.
275 /// Returns the last check instruction and the first check instruction in the
276 /// pair as (first, last).
277 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
279 /// Create an empty loop, based on the loop ranges of the old loop.
280 void createEmptyLoop();
281 /// Copy and widen the instructions from the old loop.
282 virtual void vectorizeLoop();
284 /// \brief The Loop exit block may have single value PHI nodes where the
285 /// incoming value is 'Undef'. While vectorizing we only handled real values
286 /// that were defined inside the loop. Here we fix the 'undef case'.
290 /// A helper function that computes the predicate of the block BB, assuming
291 /// that the header block of the loop is set to True. It returns the *entry*
292 /// mask for the block BB.
293 VectorParts createBlockInMask(BasicBlock *BB);
294 /// A helper function that computes the predicate of the edge between SRC
296 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
298 /// A helper function to vectorize a single BB within the innermost loop.
299 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
301 /// Vectorize a single PHINode in a block. This method handles the induction
302 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
303 /// arbitrary length vectors.
304 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
305 unsigned UF, unsigned VF, PhiVector *PV);
307 /// Insert the new loop to the loop hierarchy and pass manager
308 /// and update the analysis passes.
309 void updateAnalysis();
311 /// This instruction is un-vectorizable. Implement it as a sequence
312 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
313 /// scalarized instruction behind an if block predicated on the control
314 /// dependence of the instruction.
315 virtual void scalarizeInstruction(Instruction *Instr,
316 bool IfPredicateStore=false);
318 /// Vectorize Load and Store instructions,
319 virtual void vectorizeMemoryInstruction(Instruction *Instr);
321 /// Create a broadcast instruction. This method generates a broadcast
322 /// instruction (shuffle) for loop invariant values and for the induction
323 /// value. If this is the induction variable then we extend it to N, N+1, ...
324 /// this is needed because each iteration in the loop corresponds to a SIMD
326 virtual Value *getBroadcastInstrs(Value *V);
328 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
329 /// to each vector element of Val. The sequence starts at StartIndex.
330 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step);
332 /// When we go over instructions in the basic block we rely on previous
333 /// values within the current basic block or on loop invariant values.
334 /// When we widen (vectorize) values we place them in the map. If the values
335 /// are not within the map, they have to be loop invariant, so we simply
336 /// broadcast them into a vector.
337 VectorParts &getVectorValue(Value *V);
339 /// Generate a shuffle sequence that will reverse the vector Vec.
340 virtual Value *reverseVector(Value *Vec);
342 /// This is a helper class that holds the vectorizer state. It maps scalar
343 /// instructions to vector instructions. When the code is 'unrolled' then
344 /// then a single scalar value is mapped to multiple vector parts. The parts
345 /// are stored in the VectorPart type.
347 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
349 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
351 /// \return True if 'Key' is saved in the Value Map.
352 bool has(Value *Key) const { return MapStorage.count(Key); }
354 /// Initializes a new entry in the map. Sets all of the vector parts to the
355 /// save value in 'Val'.
356 /// \return A reference to a vector with splat values.
357 VectorParts &splat(Value *Key, Value *Val) {
358 VectorParts &Entry = MapStorage[Key];
359 Entry.assign(UF, Val);
363 ///\return A reference to the value that is stored at 'Key'.
364 VectorParts &get(Value *Key) {
365 VectorParts &Entry = MapStorage[Key];
368 assert(Entry.size() == UF);
373 /// The unroll factor. Each entry in the map stores this number of vector
377 /// Map storage. We use std::map and not DenseMap because insertions to a
378 /// dense map invalidates its iterators.
379 std::map<Value *, VectorParts> MapStorage;
382 /// The original loop.
384 /// Scev analysis to use.
393 const DataLayout *DL;
394 /// Target Library Info.
395 const TargetLibraryInfo *TLI;
397 /// The vectorization SIMD factor to use. Each vector will have this many
402 /// The vectorization unroll factor to use. Each scalar is vectorized to this
403 /// many different vector instructions.
406 /// The builder that we use
409 // --- Vectorization state ---
411 /// The vector-loop preheader.
412 BasicBlock *LoopVectorPreHeader;
413 /// The scalar-loop preheader.
414 BasicBlock *LoopScalarPreHeader;
415 /// Middle Block between the vector and the scalar.
416 BasicBlock *LoopMiddleBlock;
417 ///The ExitBlock of the scalar loop.
418 BasicBlock *LoopExitBlock;
419 ///The vector loop body.
420 SmallVector<BasicBlock *, 4> LoopVectorBody;
421 ///The scalar loop body.
422 BasicBlock *LoopScalarBody;
423 /// A list of all bypass blocks. The first block is the entry of the loop.
424 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
426 /// The new Induction variable which was added to the new block.
428 /// The induction variable of the old basic block.
429 PHINode *OldInduction;
430 /// Holds the extended (to the widest induction type) start index.
432 /// Maps scalars to widened vectors.
434 EdgeMaskCache MaskCache;
436 LoopVectorizationLegality *Legal;
439 class InnerLoopUnroller : public InnerLoopVectorizer {
441 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
442 DominatorTree *DT, const DataLayout *DL,
443 const TargetLibraryInfo *TLI, unsigned UnrollFactor) :
444 InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { }
447 void scalarizeInstruction(Instruction *Instr,
448 bool IfPredicateStore = false) override;
449 void vectorizeMemoryInstruction(Instruction *Instr) override;
450 Value *getBroadcastInstrs(Value *V) override;
451 Value *getStepVector(Value *Val, int StartIdx, Value *Step) override;
452 Value *reverseVector(Value *Vec) override;
455 /// \brief Look for a meaningful debug location on the instruction or it's
457 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
462 if (I->getDebugLoc() != Empty)
465 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
466 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
467 if (OpInst->getDebugLoc() != Empty)
474 /// \brief Set the debug location in the builder using the debug location in the
476 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
477 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
478 B.SetCurrentDebugLocation(Inst->getDebugLoc());
480 B.SetCurrentDebugLocation(DebugLoc());
484 /// \return string containing a file name and a line # for the given loop.
485 static std::string getDebugLocString(const Loop *L) {
488 raw_string_ostream OS(Result);
489 const DebugLoc LoopDbgLoc = L->getStartLoc();
490 if (!LoopDbgLoc.isUnknown())
491 LoopDbgLoc.print(L->getHeader()->getContext(), OS);
493 // Just print the module name.
494 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
501 /// \brief Propagate known metadata from one instruction to another.
502 static void propagateMetadata(Instruction *To, const Instruction *From) {
503 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
504 From->getAllMetadataOtherThanDebugLoc(Metadata);
506 for (auto M : Metadata) {
507 unsigned Kind = M.first;
509 // These are safe to transfer (this is safe for TBAA, even when we
510 // if-convert, because should that metadata have had a control dependency
511 // on the condition, and thus actually aliased with some other
512 // non-speculated memory access when the condition was false, this would be
513 // caught by the runtime overlap checks).
514 if (Kind != LLVMContext::MD_tbaa &&
515 Kind != LLVMContext::MD_alias_scope &&
516 Kind != LLVMContext::MD_noalias &&
517 Kind != LLVMContext::MD_fpmath)
520 To->setMetadata(Kind, M.second);
524 /// \brief Propagate known metadata from one instruction to a vector of others.
525 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
527 if (Instruction *I = dyn_cast<Instruction>(V))
528 propagateMetadata(I, From);
531 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
532 /// to what vectorization factor.
533 /// This class does not look at the profitability of vectorization, only the
534 /// legality. This class has two main kinds of checks:
535 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
536 /// will change the order of memory accesses in a way that will change the
537 /// correctness of the program.
538 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
539 /// checks for a number of different conditions, such as the availability of a
540 /// single induction variable, that all types are supported and vectorize-able,
541 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
542 /// This class is also used by InnerLoopVectorizer for identifying
543 /// induction variable and the different reduction variables.
544 class LoopVectorizationLegality {
546 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, const DataLayout *DL,
547 DominatorTree *DT, TargetLibraryInfo *TLI,
548 AliasAnalysis *AA, Function *F,
549 const TargetTransformInfo *TTI,
550 LoopAccessAnalysis *LAA)
551 : NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
552 TLI(TLI), TheFunction(F), TTI(TTI), DT(DT), LAA(LAA), LAI(nullptr),
553 Induction(nullptr), WidestIndTy(nullptr), HasFunNoNaNAttr(false) {}
555 /// This enum represents the kinds of reductions that we support.
557 RK_NoReduction, ///< Not a reduction.
558 RK_IntegerAdd, ///< Sum of integers.
559 RK_IntegerMult, ///< Product of integers.
560 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
561 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
562 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
563 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
564 RK_FloatAdd, ///< Sum of floats.
565 RK_FloatMult, ///< Product of floats.
566 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
569 /// This enum represents the kinds of inductions that we support.
571 IK_NoInduction, ///< Not an induction variable.
572 IK_IntInduction, ///< Integer induction variable. Step = C.
573 IK_PtrInduction ///< Pointer induction var. Step = C / sizeof(elem).
576 // This enum represents the kind of minmax reduction.
577 enum MinMaxReductionKind {
587 /// This struct holds information about reduction variables.
588 struct ReductionDescriptor {
589 ReductionDescriptor() : StartValue(nullptr), LoopExitInstr(nullptr),
590 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
592 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
593 MinMaxReductionKind MK)
594 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
596 // The starting value of the reduction.
597 // It does not have to be zero!
598 TrackingVH<Value> StartValue;
599 // The instruction who's value is used outside the loop.
600 Instruction *LoopExitInstr;
601 // The kind of the reduction.
603 // If this a min/max reduction the kind of reduction.
604 MinMaxReductionKind MinMaxKind;
607 /// This POD struct holds information about a potential reduction operation.
608 struct ReductionInstDesc {
609 ReductionInstDesc(bool IsRedux, Instruction *I) :
610 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
612 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
613 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
615 // Is this instruction a reduction candidate.
617 // The last instruction in a min/max pattern (select of the select(icmp())
618 // pattern), or the current reduction instruction otherwise.
619 Instruction *PatternLastInst;
620 // If this is a min/max pattern the comparison predicate.
621 MinMaxReductionKind MinMaxKind;
624 /// A struct for saving information about induction variables.
625 struct InductionInfo {
626 InductionInfo(Value *Start, InductionKind K, ConstantInt *Step)
627 : StartValue(Start), IK(K), StepValue(Step) {
628 assert(IK != IK_NoInduction && "Not an induction");
629 assert(StartValue && "StartValue is null");
630 assert(StepValue && !StepValue->isZero() && "StepValue is zero");
631 assert((IK != IK_PtrInduction || StartValue->getType()->isPointerTy()) &&
632 "StartValue is not a pointer for pointer induction");
633 assert((IK != IK_IntInduction || StartValue->getType()->isIntegerTy()) &&
634 "StartValue is not an integer for integer induction");
635 assert(StepValue->getType()->isIntegerTy() &&
636 "StepValue is not an integer");
639 : StartValue(nullptr), IK(IK_NoInduction), StepValue(nullptr) {}
641 /// Get the consecutive direction. Returns:
642 /// 0 - unknown or non-consecutive.
643 /// 1 - consecutive and increasing.
644 /// -1 - consecutive and decreasing.
645 int getConsecutiveDirection() const {
646 if (StepValue && (StepValue->isOne() || StepValue->isMinusOne()))
647 return StepValue->getSExtValue();
651 /// Compute the transformed value of Index at offset StartValue using step
653 /// For integer induction, returns StartValue + Index * StepValue.
654 /// For pointer induction, returns StartValue[Index * StepValue].
655 /// FIXME: The newly created binary instructions should contain nsw/nuw
656 /// flags, which can be found from the original scalar operations.
657 Value *transform(IRBuilder<> &B, Value *Index) const {
659 case IK_IntInduction:
660 assert(Index->getType() == StartValue->getType() &&
661 "Index type does not match StartValue type");
662 if (StepValue->isMinusOne())
663 return B.CreateSub(StartValue, Index);
664 if (!StepValue->isOne())
665 Index = B.CreateMul(Index, StepValue);
666 return B.CreateAdd(StartValue, Index);
668 case IK_PtrInduction:
669 if (StepValue->isMinusOne())
670 Index = B.CreateNeg(Index);
671 else if (!StepValue->isOne())
672 Index = B.CreateMul(Index, StepValue);
673 return B.CreateGEP(StartValue, Index);
678 llvm_unreachable("invalid enum");
682 TrackingVH<Value> StartValue;
686 ConstantInt *StepValue;
689 /// ReductionList contains the reduction descriptors for all
690 /// of the reductions that were found in the loop.
691 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
693 /// InductionList saves induction variables and maps them to the
694 /// induction descriptor.
695 typedef MapVector<PHINode*, InductionInfo> InductionList;
697 /// Returns true if it is legal to vectorize this loop.
698 /// This does not mean that it is profitable to vectorize this
699 /// loop, only that it is legal to do so.
702 /// Returns the Induction variable.
703 PHINode *getInduction() { return Induction; }
705 /// Returns the reduction variables found in the loop.
706 ReductionList *getReductionVars() { return &Reductions; }
708 /// Returns the induction variables found in the loop.
709 InductionList *getInductionVars() { return &Inductions; }
711 /// Returns the widest induction type.
712 Type *getWidestInductionType() { return WidestIndTy; }
714 /// Returns True if V is an induction variable in this loop.
715 bool isInductionVariable(const Value *V);
717 /// Return true if the block BB needs to be predicated in order for the loop
718 /// to be vectorized.
719 bool blockNeedsPredication(BasicBlock *BB);
721 /// Check if this pointer is consecutive when vectorizing. This happens
722 /// when the last index of the GEP is the induction variable, or that the
723 /// pointer itself is an induction variable.
724 /// This check allows us to vectorize A[idx] into a wide load/store.
726 /// 0 - Stride is unknown or non-consecutive.
727 /// 1 - Address is consecutive.
728 /// -1 - Address is consecutive, and decreasing.
729 int isConsecutivePtr(Value *Ptr);
731 /// Returns true if the value V is uniform within the loop.
732 bool isUniform(Value *V);
734 /// Returns true if this instruction will remain scalar after vectorization.
735 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
737 /// Returns the information that we collected about runtime memory check.
738 const LoopAccessInfo::RuntimePointerCheck *getRuntimePointerCheck() const {
739 return LAI->getRuntimePointerCheck();
742 const LoopAccessInfo *getLAI() const {
746 /// This function returns the identity element (or neutral element) for
748 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
750 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
752 bool hasStride(Value *V) { return StrideSet.count(V); }
753 bool mustCheckStrides() { return !StrideSet.empty(); }
754 SmallPtrSet<Value *, 8>::iterator strides_begin() {
755 return StrideSet.begin();
757 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
759 /// Returns true if the target machine supports masked store operation
760 /// for the given \p DataType and kind of access to \p Ptr.
761 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
762 return TTI->isLegalMaskedStore(DataType, isConsecutivePtr(Ptr));
764 /// Returns true if the target machine supports masked load operation
765 /// for the given \p DataType and kind of access to \p Ptr.
766 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
767 return TTI->isLegalMaskedLoad(DataType, isConsecutivePtr(Ptr));
769 /// Returns true if vector representation of the instruction \p I
771 bool isMaskRequired(const Instruction* I) {
772 return (MaskedOp.count(I) != 0);
774 unsigned getNumStores() const {
775 return LAI->getNumStores();
777 unsigned getNumLoads() const {
778 return LAI->getNumLoads();
780 unsigned getNumPredStores() const {
781 return NumPredStores;
784 /// Check if a single basic block loop is vectorizable.
785 /// At this point we know that this is a loop with a constant trip count
786 /// and we only need to check individual instructions.
787 bool canVectorizeInstrs();
789 /// When we vectorize loops we may change the order in which
790 /// we read and write from memory. This method checks if it is
791 /// legal to vectorize the code, considering only memory constrains.
792 /// Returns true if the loop is vectorizable
793 bool canVectorizeMemory();
795 /// Return true if we can vectorize this loop using the IF-conversion
797 bool canVectorizeWithIfConvert();
799 /// Collect the variables that need to stay uniform after vectorization.
800 void collectLoopUniforms();
802 /// Return true if all of the instructions in the block can be speculatively
803 /// executed. \p SafePtrs is a list of addresses that are known to be legal
804 /// and we know that we can read from them without segfault.
805 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
807 /// Returns True, if 'Phi' is the kind of reduction variable for type
808 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
809 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
810 /// Returns a struct describing if the instruction 'I' can be a reduction
811 /// variable of type 'Kind'. If the reduction is a min/max pattern of
812 /// select(icmp()) this function advances the instruction pointer 'I' from the
813 /// compare instruction to the select instruction and stores this pointer in
814 /// 'PatternLastInst' member of the returned struct.
815 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
816 ReductionInstDesc &Desc);
817 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
818 /// pattern corresponding to a min(X, Y) or max(X, Y).
819 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
820 ReductionInstDesc &Prev);
821 /// Returns the induction kind of Phi and record the step. This function may
822 /// return NoInduction if the PHI is not an induction variable.
823 InductionKind isInductionVariable(PHINode *Phi, ConstantInt *&StepValue);
825 /// \brief Collect memory access with loop invariant strides.
827 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
829 void collectStridedAccess(Value *LoadOrStoreInst);
831 /// Report an analysis message to assist the user in diagnosing loops that are
832 /// not vectorized. These are handled as LoopAccessReport rather than
833 /// VectorizationReport because the << operator of VectorizationReport returns
834 /// LoopAccessReport.
835 void emitAnalysis(const LoopAccessReport &Message) {
836 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, LV_NAME);
839 unsigned NumPredStores;
841 /// The loop that we evaluate.
845 /// DataLayout analysis.
846 const DataLayout *DL;
847 /// Target Library Info.
848 TargetLibraryInfo *TLI;
850 Function *TheFunction;
851 /// Target Transform Info
852 const TargetTransformInfo *TTI;
855 // LoopAccess analysis.
856 LoopAccessAnalysis *LAA;
857 // And the loop-accesses info corresponding to this loop. This pointer is
858 // null until canVectorizeMemory sets it up.
859 const LoopAccessInfo *LAI;
861 // --- vectorization state --- //
863 /// Holds the integer induction variable. This is the counter of the
866 /// Holds the reduction variables.
867 ReductionList Reductions;
868 /// Holds all of the induction variables that we found in the loop.
869 /// Notice that inductions don't need to start at zero and that induction
870 /// variables can be pointers.
871 InductionList Inductions;
872 /// Holds the widest induction type encountered.
875 /// Allowed outside users. This holds the reduction
876 /// vars which can be accessed from outside the loop.
877 SmallPtrSet<Value*, 4> AllowedExit;
878 /// This set holds the variables which are known to be uniform after
880 SmallPtrSet<Instruction*, 4> Uniforms;
882 /// Can we assume the absence of NaNs.
883 bool HasFunNoNaNAttr;
885 ValueToValueMap Strides;
886 SmallPtrSet<Value *, 8> StrideSet;
888 /// While vectorizing these instructions we have to generate a
889 /// call to the appropriate masked intrinsic
890 SmallPtrSet<const Instruction*, 8> MaskedOp;
893 /// LoopVectorizationCostModel - estimates the expected speedups due to
895 /// In many cases vectorization is not profitable. This can happen because of
896 /// a number of reasons. In this class we mainly attempt to predict the
897 /// expected speedup/slowdowns due to the supported instruction set. We use the
898 /// TargetTransformInfo to query the different backends for the cost of
899 /// different operations.
900 class LoopVectorizationCostModel {
902 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
903 LoopVectorizationLegality *Legal,
904 const TargetTransformInfo &TTI,
905 const DataLayout *DL, const TargetLibraryInfo *TLI,
906 AssumptionCache *AC, const Function *F,
907 const LoopVectorizeHints *Hints)
908 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI),
909 TheFunction(F), Hints(Hints) {
910 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
913 /// Information about vectorization costs
914 struct VectorizationFactor {
915 unsigned Width; // Vector width with best cost
916 unsigned Cost; // Cost of the loop with that width
918 /// \return The most profitable vectorization factor and the cost of that VF.
919 /// This method checks every power of two up to VF. If UserVF is not ZERO
920 /// then this vectorization factor will be selected if vectorization is
922 VectorizationFactor selectVectorizationFactor(bool OptForSize);
924 /// \return The size (in bits) of the widest type in the code that
925 /// needs to be vectorized. We ignore values that remain scalar such as
926 /// 64 bit loop indices.
927 unsigned getWidestType();
929 /// \return The most profitable unroll factor.
930 /// If UserUF is non-zero then this method finds the best unroll-factor
931 /// based on register pressure and other parameters.
932 /// VF and LoopCost are the selected vectorization factor and the cost of the
934 unsigned selectUnrollFactor(bool OptForSize, unsigned VF, unsigned LoopCost);
936 /// \brief A struct that represents some properties of the register usage
938 struct RegisterUsage {
939 /// Holds the number of loop invariant values that are used in the loop.
940 unsigned LoopInvariantRegs;
941 /// Holds the maximum number of concurrent live intervals in the loop.
942 unsigned MaxLocalUsers;
943 /// Holds the number of instructions in the loop.
944 unsigned NumInstructions;
947 /// \return information about the register usage of the loop.
948 RegisterUsage calculateRegisterUsage();
951 /// Returns the expected execution cost. The unit of the cost does
952 /// not matter because we use the 'cost' units to compare different
953 /// vector widths. The cost that is returned is *not* normalized by
954 /// the factor width.
955 unsigned expectedCost(unsigned VF);
957 /// Returns the execution time cost of an instruction for a given vector
958 /// width. Vector width of one means scalar.
959 unsigned getInstructionCost(Instruction *I, unsigned VF);
961 /// A helper function for converting Scalar types to vector types.
962 /// If the incoming type is void, we return void. If the VF is 1, we return
964 static Type* ToVectorTy(Type *Scalar, unsigned VF);
966 /// Returns whether the instruction is a load or store and will be a emitted
967 /// as a vector operation.
968 bool isConsecutiveLoadOrStore(Instruction *I);
970 /// Report an analysis message to assist the user in diagnosing loops that are
971 /// not vectorized. These are handled as LoopAccessReport rather than
972 /// VectorizationReport because the << operator of VectorizationReport returns
973 /// LoopAccessReport.
974 void emitAnalysis(const LoopAccessReport &Message) {
975 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, LV_NAME);
978 /// Values used only by @llvm.assume calls.
979 SmallPtrSet<const Value *, 32> EphValues;
981 /// The loop that we evaluate.
985 /// Loop Info analysis.
987 /// Vectorization legality.
988 LoopVectorizationLegality *Legal;
989 /// Vector target information.
990 const TargetTransformInfo &TTI;
991 /// Target data layout information.
992 const DataLayout *DL;
993 /// Target Library Info.
994 const TargetLibraryInfo *TLI;
995 const Function *TheFunction;
996 // Loop Vectorize Hint.
997 const LoopVectorizeHints *Hints;
1000 /// Utility class for getting and setting loop vectorizer hints in the form
1001 /// of loop metadata.
1002 /// This class keeps a number of loop annotations locally (as member variables)
1003 /// and can, upon request, write them back as metadata on the loop. It will
1004 /// initially scan the loop for existing metadata, and will update the local
1005 /// values based on information in the loop.
1006 /// We cannot write all values to metadata, as the mere presence of some info,
1007 /// for example 'force', means a decision has been made. So, we need to be
1008 /// careful NOT to add them if the user hasn't specifically asked so.
1009 class LoopVectorizeHints {
1016 /// Hint - associates name and validation with the hint value.
1019 unsigned Value; // This may have to change for non-numeric values.
1022 Hint(const char * Name, unsigned Value, HintKind Kind)
1023 : Name(Name), Value(Value), Kind(Kind) { }
1025 bool validate(unsigned Val) {
1028 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
1030 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1038 /// Vectorization width.
1040 /// Vectorization interleave factor.
1042 /// Vectorization forced
1045 /// Return the loop metadata prefix.
1046 static StringRef Prefix() { return "llvm.loop."; }
1050 FK_Undefined = -1, ///< Not selected.
1051 FK_Disabled = 0, ///< Forcing disabled.
1052 FK_Enabled = 1, ///< Forcing enabled.
1055 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
1056 : Width("vectorize.width", VectorizerParams::VectorizationFactor,
1058 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1059 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1061 // Populate values with existing loop metadata.
1062 getHintsFromMetadata();
1064 // force-vector-interleave overrides DisableInterleaving.
1065 if (VectorizerParams::isInterleaveForced())
1066 Interleave.Value = VectorizerParams::VectorizationInterleave;
1068 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
1069 << "LV: Interleaving disabled by the pass manager\n");
1072 /// Mark the loop L as already vectorized by setting the width to 1.
1073 void setAlreadyVectorized() {
1074 Width.Value = Interleave.Value = 1;
1075 Hint Hints[] = {Width, Interleave};
1076 writeHintsToMetadata(Hints);
1079 /// Dumps all the hint information.
1080 std::string emitRemark() const {
1081 VectorizationReport R;
1082 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1083 R << "vectorization is explicitly disabled";
1085 R << "use -Rpass-analysis=loop-vectorize for more info";
1086 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1087 R << " (Force=true";
1088 if (Width.Value != 0)
1089 R << ", Vector Width=" << Width.Value;
1090 if (Interleave.Value != 0)
1091 R << ", Interleave Count=" << Interleave.Value;
1099 unsigned getWidth() const { return Width.Value; }
1100 unsigned getInterleave() const { return Interleave.Value; }
1101 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1104 /// Find hints specified in the loop metadata and update local values.
1105 void getHintsFromMetadata() {
1106 MDNode *LoopID = TheLoop->getLoopID();
1110 // First operand should refer to the loop id itself.
1111 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1112 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1114 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1115 const MDString *S = nullptr;
1116 SmallVector<Metadata *, 4> Args;
1118 // The expected hint is either a MDString or a MDNode with the first
1119 // operand a MDString.
1120 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1121 if (!MD || MD->getNumOperands() == 0)
1123 S = dyn_cast<MDString>(MD->getOperand(0));
1124 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1125 Args.push_back(MD->getOperand(i));
1127 S = dyn_cast<MDString>(LoopID->getOperand(i));
1128 assert(Args.size() == 0 && "too many arguments for MDString");
1134 // Check if the hint starts with the loop metadata prefix.
1135 StringRef Name = S->getString();
1136 if (Args.size() == 1)
1137 setHint(Name, Args[0]);
1141 /// Checks string hint with one operand and set value if valid.
1142 void setHint(StringRef Name, Metadata *Arg) {
1143 if (!Name.startswith(Prefix()))
1145 Name = Name.substr(Prefix().size(), StringRef::npos);
1147 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1149 unsigned Val = C->getZExtValue();
1151 Hint *Hints[] = {&Width, &Interleave, &Force};
1152 for (auto H : Hints) {
1153 if (Name == H->Name) {
1154 if (H->validate(Val))
1157 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1163 /// Create a new hint from name / value pair.
1164 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1165 LLVMContext &Context = TheLoop->getHeader()->getContext();
1166 Metadata *MDs[] = {MDString::get(Context, Name),
1167 ConstantAsMetadata::get(
1168 ConstantInt::get(Type::getInt32Ty(Context), V))};
1169 return MDNode::get(Context, MDs);
1172 /// Matches metadata with hint name.
1173 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1174 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1178 for (auto H : HintTypes)
1179 if (Name->getString().endswith(H.Name))
1184 /// Sets current hints into loop metadata, keeping other values intact.
1185 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1186 if (HintTypes.size() == 0)
1189 // Reserve the first element to LoopID (see below).
1190 SmallVector<Metadata *, 4> MDs(1);
1191 // If the loop already has metadata, then ignore the existing operands.
1192 MDNode *LoopID = TheLoop->getLoopID();
1194 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1195 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1196 // If node in update list, ignore old value.
1197 if (!matchesHintMetadataName(Node, HintTypes))
1198 MDs.push_back(Node);
1202 // Now, add the missing hints.
1203 for (auto H : HintTypes)
1204 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1206 // Replace current metadata node with new one.
1207 LLVMContext &Context = TheLoop->getHeader()->getContext();
1208 MDNode *NewLoopID = MDNode::get(Context, MDs);
1209 // Set operand 0 to refer to the loop id itself.
1210 NewLoopID->replaceOperandWith(0, NewLoopID);
1212 TheLoop->setLoopID(NewLoopID);
1215 /// The loop these hints belong to.
1216 const Loop *TheLoop;
1219 static void emitMissedWarning(Function *F, Loop *L,
1220 const LoopVectorizeHints &LH) {
1221 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1222 L->getStartLoc(), LH.emitRemark());
1224 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1225 if (LH.getWidth() != 1)
1226 emitLoopVectorizeWarning(
1227 F->getContext(), *F, L->getStartLoc(),
1228 "failed explicitly specified loop vectorization");
1229 else if (LH.getInterleave() != 1)
1230 emitLoopInterleaveWarning(
1231 F->getContext(), *F, L->getStartLoc(),
1232 "failed explicitly specified loop interleaving");
1236 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1238 return V.push_back(&L);
1240 for (Loop *InnerL : L)
1241 addInnerLoop(*InnerL, V);
1244 /// The LoopVectorize Pass.
1245 struct LoopVectorize : public FunctionPass {
1246 /// Pass identification, replacement for typeid
1249 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1251 DisableUnrolling(NoUnrolling),
1252 AlwaysVectorize(AlwaysVectorize) {
1253 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1256 ScalarEvolution *SE;
1257 const DataLayout *DL;
1259 TargetTransformInfo *TTI;
1261 BlockFrequencyInfo *BFI;
1262 TargetLibraryInfo *TLI;
1264 AssumptionCache *AC;
1265 LoopAccessAnalysis *LAA;
1266 bool DisableUnrolling;
1267 bool AlwaysVectorize;
1269 BlockFrequency ColdEntryFreq;
1271 bool runOnFunction(Function &F) override {
1272 SE = &getAnalysis<ScalarEvolution>();
1273 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1274 DL = DLP ? &DLP->getDataLayout() : nullptr;
1275 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1276 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1277 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1278 BFI = &getAnalysis<BlockFrequencyInfo>();
1279 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1280 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1281 AA = &getAnalysis<AliasAnalysis>();
1282 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1283 LAA = &getAnalysis<LoopAccessAnalysis>();
1285 // Compute some weights outside of the loop over the loops. Compute this
1286 // using a BranchProbability to re-use its scaling math.
1287 const BranchProbability ColdProb(1, 5); // 20%
1288 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1290 // If the target claims to have no vector registers don't attempt
1292 if (!TTI->getNumberOfRegisters(true))
1296 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1297 << ": Missing data layout\n");
1301 // Build up a worklist of inner-loops to vectorize. This is necessary as
1302 // the act of vectorizing or partially unrolling a loop creates new loops
1303 // and can invalidate iterators across the loops.
1304 SmallVector<Loop *, 8> Worklist;
1307 addInnerLoop(*L, Worklist);
1309 LoopsAnalyzed += Worklist.size();
1311 // Now walk the identified inner loops.
1312 bool Changed = false;
1313 while (!Worklist.empty())
1314 Changed |= processLoop(Worklist.pop_back_val());
1316 // Process each loop nest in the function.
1320 bool processLoop(Loop *L) {
1321 assert(L->empty() && "Only process inner loops.");
1324 const std::string DebugLocStr = getDebugLocString(L);
1327 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1328 << L->getHeader()->getParent()->getName() << "\" from "
1329 << DebugLocStr << "\n");
1331 LoopVectorizeHints Hints(L, DisableUnrolling);
1333 DEBUG(dbgs() << "LV: Loop hints:"
1335 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1337 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1339 : "?")) << " width=" << Hints.getWidth()
1340 << " unroll=" << Hints.getInterleave() << "\n");
1342 // Function containing loop
1343 Function *F = L->getHeader()->getParent();
1345 // Looking at the diagnostic output is the only way to determine if a loop
1346 // was vectorized (other than looking at the IR or machine code), so it
1347 // is important to generate an optimization remark for each loop. Most of
1348 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1349 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1350 // less verbose reporting vectorized loops and unvectorized loops that may
1351 // benefit from vectorization, respectively.
1353 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1354 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1355 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1356 L->getStartLoc(), Hints.emitRemark());
1360 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1361 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1362 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1363 L->getStartLoc(), Hints.emitRemark());
1367 if (Hints.getWidth() == 1 && Hints.getInterleave() == 1) {
1368 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1369 emitOptimizationRemarkAnalysis(
1370 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1371 "loop not vectorized: vector width and interleave count are "
1372 "explicitly set to 1");
1376 // Check the loop for a trip count threshold:
1377 // do not vectorize loops with a tiny trip count.
1378 const unsigned TC = SE->getSmallConstantTripCount(L);
1379 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1380 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1381 << "This loop is not worth vectorizing.");
1382 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1383 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1385 DEBUG(dbgs() << "\n");
1386 emitOptimizationRemarkAnalysis(
1387 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1388 "vectorization is not beneficial and is not explicitly forced");
1393 // Check if it is legal to vectorize the loop.
1394 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, AA, F, TTI, LAA);
1395 if (!LVL.canVectorize()) {
1396 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1397 emitMissedWarning(F, L, Hints);
1401 // Use the cost model.
1402 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI, AC, F,
1405 // Check the function attributes to find out if this function should be
1406 // optimized for size.
1407 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1408 F->hasFnAttribute(Attribute::OptimizeForSize);
1410 // Compute the weighted frequency of this loop being executed and see if it
1411 // is less than 20% of the function entry baseline frequency. Note that we
1412 // always have a canonical loop here because we think we *can* vectoriez.
1413 // FIXME: This is hidden behind a flag due to pervasive problems with
1414 // exactly what block frequency models.
1415 if (LoopVectorizeWithBlockFrequency) {
1416 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1417 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1418 LoopEntryFreq < ColdEntryFreq)
1422 // Check the function attributes to see if implicit floats are allowed.a
1423 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1424 // an integer loop and the vector instructions selected are purely integer
1425 // vector instructions?
1426 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1427 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1428 "attribute is used.\n");
1429 emitOptimizationRemarkAnalysis(
1430 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1431 "loop not vectorized due to NoImplicitFloat attribute");
1432 emitMissedWarning(F, L, Hints);
1436 // Select the optimal vectorization factor.
1437 const LoopVectorizationCostModel::VectorizationFactor VF =
1438 CM.selectVectorizationFactor(OptForSize);
1440 // Select the unroll factor.
1442 CM.selectUnrollFactor(OptForSize, VF.Width, VF.Cost);
1444 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1445 << DebugLocStr << '\n');
1446 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1448 if (VF.Width == 1) {
1449 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
1452 emitOptimizationRemarkAnalysis(
1453 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1454 "not beneficial to vectorize and user disabled interleaving");
1457 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1459 // Report the unrolling decision.
1460 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1461 Twine("unrolled with interleaving factor " +
1463 " (vectorization not beneficial)"));
1465 // We decided not to vectorize, but we may want to unroll.
1467 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1468 Unroller.vectorize(&LVL);
1470 // If we decided that it is *legal* to vectorize the loop then do it.
1471 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1475 // Report the vectorization decision.
1476 emitOptimizationRemark(
1477 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1478 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1479 ", unrolling interleave factor: " + Twine(UF) + ")");
1482 // Mark the loop as already vectorized to avoid vectorizing again.
1483 Hints.setAlreadyVectorized();
1485 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1489 void getAnalysisUsage(AnalysisUsage &AU) const override {
1490 AU.addRequired<AssumptionCacheTracker>();
1491 AU.addRequiredID(LoopSimplifyID);
1492 AU.addRequiredID(LCSSAID);
1493 AU.addRequired<BlockFrequencyInfo>();
1494 AU.addRequired<DominatorTreeWrapperPass>();
1495 AU.addRequired<LoopInfoWrapperPass>();
1496 AU.addRequired<ScalarEvolution>();
1497 AU.addRequired<TargetTransformInfoWrapperPass>();
1498 AU.addRequired<AliasAnalysis>();
1499 AU.addRequired<LoopAccessAnalysis>();
1500 AU.addPreserved<LoopInfoWrapperPass>();
1501 AU.addPreserved<DominatorTreeWrapperPass>();
1502 AU.addPreserved<AliasAnalysis>();
1507 } // end anonymous namespace
1509 //===----------------------------------------------------------------------===//
1510 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1511 // LoopVectorizationCostModel.
1512 //===----------------------------------------------------------------------===//
1514 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1515 // We need to place the broadcast of invariant variables outside the loop.
1516 Instruction *Instr = dyn_cast<Instruction>(V);
1518 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1519 Instr->getParent()) != LoopVectorBody.end());
1520 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1522 // Place the code for broadcasting invariant variables in the new preheader.
1523 IRBuilder<>::InsertPointGuard Guard(Builder);
1525 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1527 // Broadcast the scalar into all locations in the vector.
1528 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1533 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
1535 assert(Val->getType()->isVectorTy() && "Must be a vector");
1536 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1537 "Elem must be an integer");
1538 assert(Step->getType() == Val->getType()->getScalarType() &&
1539 "Step has wrong type");
1540 // Create the types.
1541 Type *ITy = Val->getType()->getScalarType();
1542 VectorType *Ty = cast<VectorType>(Val->getType());
1543 int VLen = Ty->getNumElements();
1544 SmallVector<Constant*, 8> Indices;
1546 // Create a vector of consecutive numbers from zero to VF.
1547 for (int i = 0; i < VLen; ++i)
1548 Indices.push_back(ConstantInt::get(ITy, StartIdx + i));
1550 // Add the consecutive indices to the vector value.
1551 Constant *Cv = ConstantVector::get(Indices);
1552 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1553 Step = Builder.CreateVectorSplat(VLen, Step);
1554 assert(Step->getType() == Val->getType() && "Invalid step vec");
1555 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
1556 // which can be found from the original scalar operations.
1557 Step = Builder.CreateMul(Cv, Step);
1558 return Builder.CreateAdd(Val, Step, "induction");
1561 /// \brief Find the operand of the GEP that should be checked for consecutive
1562 /// stores. This ignores trailing indices that have no effect on the final
1564 static unsigned getGEPInductionOperand(const DataLayout *DL,
1565 const GetElementPtrInst *Gep) {
1566 unsigned LastOperand = Gep->getNumOperands() - 1;
1567 unsigned GEPAllocSize = DL->getTypeAllocSize(
1568 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1570 // Walk backwards and try to peel off zeros.
1571 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1572 // Find the type we're currently indexing into.
1573 gep_type_iterator GEPTI = gep_type_begin(Gep);
1574 std::advance(GEPTI, LastOperand - 1);
1576 // If it's a type with the same allocation size as the result of the GEP we
1577 // can peel off the zero index.
1578 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1586 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1587 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1588 // Make sure that the pointer does not point to structs.
1589 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1592 // If this value is a pointer induction variable we know it is consecutive.
1593 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1594 if (Phi && Inductions.count(Phi)) {
1595 InductionInfo II = Inductions[Phi];
1596 return II.getConsecutiveDirection();
1599 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1603 unsigned NumOperands = Gep->getNumOperands();
1604 Value *GpPtr = Gep->getPointerOperand();
1605 // If this GEP value is a consecutive pointer induction variable and all of
1606 // the indices are constant then we know it is consecutive. We can
1607 Phi = dyn_cast<PHINode>(GpPtr);
1608 if (Phi && Inductions.count(Phi)) {
1610 // Make sure that the pointer does not point to structs.
1611 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1612 if (GepPtrType->getElementType()->isAggregateType())
1615 // Make sure that all of the index operands are loop invariant.
1616 for (unsigned i = 1; i < NumOperands; ++i)
1617 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1620 InductionInfo II = Inductions[Phi];
1621 return II.getConsecutiveDirection();
1624 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1626 // Check that all of the gep indices are uniform except for our induction
1628 for (unsigned i = 0; i != NumOperands; ++i)
1629 if (i != InductionOperand &&
1630 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1633 // We can emit wide load/stores only if the last non-zero index is the
1634 // induction variable.
1635 const SCEV *Last = nullptr;
1636 if (!Strides.count(Gep))
1637 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1639 // Because of the multiplication by a stride we can have a s/zext cast.
1640 // We are going to replace this stride by 1 so the cast is safe to ignore.
1642 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1643 // %0 = trunc i64 %indvars.iv to i32
1644 // %mul = mul i32 %0, %Stride1
1645 // %idxprom = zext i32 %mul to i64 << Safe cast.
1646 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1648 Last = replaceSymbolicStrideSCEV(SE, Strides,
1649 Gep->getOperand(InductionOperand), Gep);
1650 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1652 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1656 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1657 const SCEV *Step = AR->getStepRecurrence(*SE);
1659 // The memory is consecutive because the last index is consecutive
1660 // and all other indices are loop invariant.
1663 if (Step->isAllOnesValue())
1670 bool LoopVectorizationLegality::isUniform(Value *V) {
1671 return LAI->isUniform(V);
1674 InnerLoopVectorizer::VectorParts&
1675 InnerLoopVectorizer::getVectorValue(Value *V) {
1676 assert(V != Induction && "The new induction variable should not be used.");
1677 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1679 // If we have a stride that is replaced by one, do it here.
1680 if (Legal->hasStride(V))
1681 V = ConstantInt::get(V->getType(), 1);
1683 // If we have this scalar in the map, return it.
1684 if (WidenMap.has(V))
1685 return WidenMap.get(V);
1687 // If this scalar is unknown, assume that it is a constant or that it is
1688 // loop invariant. Broadcast V and save the value for future uses.
1689 Value *B = getBroadcastInstrs(V);
1690 return WidenMap.splat(V, B);
1693 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1694 assert(Vec->getType()->isVectorTy() && "Invalid type");
1695 SmallVector<Constant*, 8> ShuffleMask;
1696 for (unsigned i = 0; i < VF; ++i)
1697 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1699 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1700 ConstantVector::get(ShuffleMask),
1704 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1705 // Attempt to issue a wide load.
1706 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1707 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1709 assert((LI || SI) && "Invalid Load/Store instruction");
1711 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1712 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1713 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1714 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1715 // An alignment of 0 means target abi alignment. We need to use the scalar's
1716 // target abi alignment in such a case.
1718 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1719 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1720 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1721 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1723 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
1724 !Legal->isMaskRequired(SI))
1725 return scalarizeInstruction(Instr, true);
1727 if (ScalarAllocatedSize != VectorElementSize)
1728 return scalarizeInstruction(Instr);
1730 // If the pointer is loop invariant or if it is non-consecutive,
1731 // scalarize the load.
1732 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1733 bool Reverse = ConsecutiveStride < 0;
1734 bool UniformLoad = LI && Legal->isUniform(Ptr);
1735 if (!ConsecutiveStride || UniformLoad)
1736 return scalarizeInstruction(Instr);
1738 Constant *Zero = Builder.getInt32(0);
1739 VectorParts &Entry = WidenMap.get(Instr);
1741 // Handle consecutive loads/stores.
1742 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1743 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1744 setDebugLocFromInst(Builder, Gep);
1745 Value *PtrOperand = Gep->getPointerOperand();
1746 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1747 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1749 // Create the new GEP with the new induction variable.
1750 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1751 Gep2->setOperand(0, FirstBasePtr);
1752 Gep2->setName("gep.indvar.base");
1753 Ptr = Builder.Insert(Gep2);
1755 setDebugLocFromInst(Builder, Gep);
1756 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1757 OrigLoop) && "Base ptr must be invariant");
1759 // The last index does not have to be the induction. It can be
1760 // consecutive and be a function of the index. For example A[I+1];
1761 unsigned NumOperands = Gep->getNumOperands();
1762 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1763 // Create the new GEP with the new induction variable.
1764 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1766 for (unsigned i = 0; i < NumOperands; ++i) {
1767 Value *GepOperand = Gep->getOperand(i);
1768 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1770 // Update last index or loop invariant instruction anchored in loop.
1771 if (i == InductionOperand ||
1772 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1773 assert((i == InductionOperand ||
1774 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1775 "Must be last index or loop invariant");
1777 VectorParts &GEPParts = getVectorValue(GepOperand);
1778 Value *Index = GEPParts[0];
1779 Index = Builder.CreateExtractElement(Index, Zero);
1780 Gep2->setOperand(i, Index);
1781 Gep2->setName("gep.indvar.idx");
1784 Ptr = Builder.Insert(Gep2);
1786 // Use the induction element ptr.
1787 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1788 setDebugLocFromInst(Builder, Ptr);
1789 VectorParts &PtrVal = getVectorValue(Ptr);
1790 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1793 VectorParts Mask = createBlockInMask(Instr->getParent());
1796 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1797 "We do not allow storing to uniform addresses");
1798 setDebugLocFromInst(Builder, SI);
1799 // We don't want to update the value in the map as it might be used in
1800 // another expression. So don't use a reference type for "StoredVal".
1801 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1803 for (unsigned Part = 0; Part < UF; ++Part) {
1804 // Calculate the pointer for the specific unroll-part.
1805 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1808 // If we store to reverse consecutive memory locations then we need
1809 // to reverse the order of elements in the stored value.
1810 StoredVal[Part] = reverseVector(StoredVal[Part]);
1811 // If the address is consecutive but reversed, then the
1812 // wide store needs to start at the last vector element.
1813 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1814 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1815 Mask[Part] = reverseVector(Mask[Part]);
1818 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1819 DataTy->getPointerTo(AddressSpace));
1822 if (Legal->isMaskRequired(SI))
1823 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
1826 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
1827 propagateMetadata(NewSI, SI);
1833 assert(LI && "Must have a load instruction");
1834 setDebugLocFromInst(Builder, LI);
1835 for (unsigned Part = 0; Part < UF; ++Part) {
1836 // Calculate the pointer for the specific unroll-part.
1837 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1840 // If the address is consecutive but reversed, then the
1841 // wide load needs to start at the last vector element.
1842 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1843 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1844 Mask[Part] = reverseVector(Mask[Part]);
1848 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1849 DataTy->getPointerTo(AddressSpace));
1850 if (Legal->isMaskRequired(LI))
1851 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
1852 UndefValue::get(DataTy),
1853 "wide.masked.load");
1855 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
1856 propagateMetadata(NewLI, LI);
1857 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
1861 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1862 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1863 // Holds vector parameters or scalars, in case of uniform vals.
1864 SmallVector<VectorParts, 4> Params;
1866 setDebugLocFromInst(Builder, Instr);
1868 // Find all of the vectorized parameters.
1869 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1870 Value *SrcOp = Instr->getOperand(op);
1872 // If we are accessing the old induction variable, use the new one.
1873 if (SrcOp == OldInduction) {
1874 Params.push_back(getVectorValue(SrcOp));
1878 // Try using previously calculated values.
1879 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1881 // If the src is an instruction that appeared earlier in the basic block
1882 // then it should already be vectorized.
1883 if (SrcInst && OrigLoop->contains(SrcInst)) {
1884 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1885 // The parameter is a vector value from earlier.
1886 Params.push_back(WidenMap.get(SrcInst));
1888 // The parameter is a scalar from outside the loop. Maybe even a constant.
1889 VectorParts Scalars;
1890 Scalars.append(UF, SrcOp);
1891 Params.push_back(Scalars);
1895 assert(Params.size() == Instr->getNumOperands() &&
1896 "Invalid number of operands");
1898 // Does this instruction return a value ?
1899 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1901 Value *UndefVec = IsVoidRetTy ? nullptr :
1902 UndefValue::get(VectorType::get(Instr->getType(), VF));
1903 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1904 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1906 Instruction *InsertPt = Builder.GetInsertPoint();
1907 BasicBlock *IfBlock = Builder.GetInsertBlock();
1908 BasicBlock *CondBlock = nullptr;
1911 Loop *VectorLp = nullptr;
1912 if (IfPredicateStore) {
1913 assert(Instr->getParent()->getSinglePredecessor() &&
1914 "Only support single predecessor blocks");
1915 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1916 Instr->getParent());
1917 VectorLp = LI->getLoopFor(IfBlock);
1918 assert(VectorLp && "Must have a loop for this block");
1921 // For each vector unroll 'part':
1922 for (unsigned Part = 0; Part < UF; ++Part) {
1923 // For each scalar that we create:
1924 for (unsigned Width = 0; Width < VF; ++Width) {
1927 Value *Cmp = nullptr;
1928 if (IfPredicateStore) {
1929 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1930 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1931 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1932 LoopVectorBody.push_back(CondBlock);
1933 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
1934 // Update Builder with newly created basic block.
1935 Builder.SetInsertPoint(InsertPt);
1938 Instruction *Cloned = Instr->clone();
1940 Cloned->setName(Instr->getName() + ".cloned");
1941 // Replace the operands of the cloned instructions with extracted scalars.
1942 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1943 Value *Op = Params[op][Part];
1944 // Param is a vector. Need to extract the right lane.
1945 if (Op->getType()->isVectorTy())
1946 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1947 Cloned->setOperand(op, Op);
1950 // Place the cloned scalar in the new loop.
1951 Builder.Insert(Cloned);
1953 // If the original scalar returns a value we need to place it in a vector
1954 // so that future users will be able to use it.
1956 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1957 Builder.getInt32(Width));
1959 if (IfPredicateStore) {
1960 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1961 LoopVectorBody.push_back(NewIfBlock);
1962 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
1963 Builder.SetInsertPoint(InsertPt);
1964 Instruction *OldBr = IfBlock->getTerminator();
1965 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1966 OldBr->eraseFromParent();
1967 IfBlock = NewIfBlock;
1973 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1977 if (Instruction *I = dyn_cast<Instruction>(V))
1978 return I->getParent() == Loc->getParent() ? I : nullptr;
1982 std::pair<Instruction *, Instruction *>
1983 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
1984 Instruction *tnullptr = nullptr;
1985 if (!Legal->mustCheckStrides())
1986 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1988 IRBuilder<> ChkBuilder(Loc);
1991 Value *Check = nullptr;
1992 Instruction *FirstInst = nullptr;
1993 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
1994 SE = Legal->strides_end();
1996 Value *Ptr = stripIntegerCast(*SI);
1997 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
1999 // Store the first instruction we create.
2000 FirstInst = getFirstInst(FirstInst, C, Loc);
2002 Check = ChkBuilder.CreateOr(Check, C);
2007 // We have to do this trickery because the IRBuilder might fold the check to a
2008 // constant expression in which case there is no Instruction anchored in a
2010 LLVMContext &Ctx = Loc->getContext();
2011 Instruction *TheCheck =
2012 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2013 ChkBuilder.Insert(TheCheck, "stride.not.one");
2014 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2016 return std::make_pair(FirstInst, TheCheck);
2019 void InnerLoopVectorizer::createEmptyLoop() {
2021 In this function we generate a new loop. The new loop will contain
2022 the vectorized instructions while the old loop will continue to run the
2025 [ ] <-- Back-edge taken count overflow check.
2028 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2031 || [ ] <-- vector pre header.
2035 || [ ]_| <-- vector loop.
2038 | >[ ] <--- middle-block.
2041 -|- >[ ] <--- new preheader.
2045 | [ ]_| <-- old scalar loop to handle remainder.
2048 >[ ] <-- exit block.
2052 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2053 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2054 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2055 assert(BypassBlock && "Invalid loop structure");
2056 assert(ExitBlock && "Must have an exit block");
2058 // Some loops have a single integer induction variable, while other loops
2059 // don't. One example is c++ iterators that often have multiple pointer
2060 // induction variables. In the code below we also support a case where we
2061 // don't have a single induction variable.
2062 OldInduction = Legal->getInduction();
2063 Type *IdxTy = Legal->getWidestInductionType();
2065 // Find the loop boundaries.
2066 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2067 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2069 // The exit count might have the type of i64 while the phi is i32. This can
2070 // happen if we have an induction variable that is sign extended before the
2071 // compare. The only way that we get a backedge taken count is that the
2072 // induction variable was signed and as such will not overflow. In such a case
2073 // truncation is legal.
2074 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2075 IdxTy->getPrimitiveSizeInBits())
2076 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2078 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2079 // Get the total trip count from the count by adding 1.
2080 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2081 SE->getConstant(BackedgeTakeCount->getType(), 1));
2083 // Expand the trip count and place the new instructions in the preheader.
2084 // Notice that the pre-header does not change, only the loop body.
2085 SCEVExpander Exp(*SE, "induction");
2087 // We need to test whether the backedge-taken count is uint##_max. Adding one
2088 // to it will cause overflow and an incorrect loop trip count in the vector
2089 // body. In case of overflow we want to directly jump to the scalar remainder
2091 Value *BackedgeCount =
2092 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2093 BypassBlock->getTerminator());
2094 if (BackedgeCount->getType()->isPointerTy())
2095 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2096 "backedge.ptrcnt.to.int",
2097 BypassBlock->getTerminator());
2098 Instruction *CheckBCOverflow =
2099 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2100 Constant::getAllOnesValue(BackedgeCount->getType()),
2101 "backedge.overflow", BypassBlock->getTerminator());
2103 // The loop index does not have to start at Zero. Find the original start
2104 // value from the induction PHI node. If we don't have an induction variable
2105 // then we know that it starts at zero.
2106 Builder.SetInsertPoint(BypassBlock->getTerminator());
2107 Value *StartIdx = ExtendedIdx = OldInduction ?
2108 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2110 ConstantInt::get(IdxTy, 0);
2112 // We need an instruction to anchor the overflow check on. StartIdx needs to
2113 // be defined before the overflow check branch. Because the scalar preheader
2114 // is going to merge the start index and so the overflow branch block needs to
2115 // contain a definition of the start index.
2116 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2117 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2118 BypassBlock->getTerminator());
2120 // Count holds the overall loop count (N).
2121 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2122 BypassBlock->getTerminator());
2124 LoopBypassBlocks.push_back(BypassBlock);
2126 // Split the single block loop into the two loop structure described above.
2127 BasicBlock *VectorPH =
2128 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2129 BasicBlock *VecBody =
2130 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2131 BasicBlock *MiddleBlock =
2132 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2133 BasicBlock *ScalarPH =
2134 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2136 // Create and register the new vector loop.
2137 Loop* Lp = new Loop();
2138 Loop *ParentLoop = OrigLoop->getParentLoop();
2140 // Insert the new loop into the loop nest and register the new basic blocks
2141 // before calling any utilities such as SCEV that require valid LoopInfo.
2143 ParentLoop->addChildLoop(Lp);
2144 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
2145 ParentLoop->addBasicBlockToLoop(VectorPH, *LI);
2146 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
2148 LI->addTopLevelLoop(Lp);
2150 Lp->addBasicBlockToLoop(VecBody, *LI);
2152 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2154 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2156 // Generate the induction variable.
2157 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2158 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2159 // The loop step is equal to the vectorization factor (num of SIMD elements)
2160 // times the unroll factor (num of SIMD instructions).
2161 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2163 // This is the IR builder that we use to add all of the logic for bypassing
2164 // the new vector loop.
2165 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2166 setDebugLocFromInst(BypassBuilder,
2167 getDebugLocFromInstOrOperands(OldInduction));
2169 // We may need to extend the index in case there is a type mismatch.
2170 // We know that the count starts at zero and does not overflow.
2171 if (Count->getType() != IdxTy) {
2172 // The exit count can be of pointer type. Convert it to the correct
2174 if (ExitCount->getType()->isPointerTy())
2175 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2177 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2180 // Add the start index to the loop count to get the new end index.
2181 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2183 // Now we need to generate the expression for N - (N % VF), which is
2184 // the part that the vectorized body will execute.
2185 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2186 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2187 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2188 "end.idx.rnd.down");
2190 // Now, compare the new count to zero. If it is zero skip the vector loop and
2191 // jump to the scalar loop.
2193 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2195 BasicBlock *LastBypassBlock = BypassBlock;
2197 // Generate code to check that the loops trip count that we computed by adding
2198 // one to the backedge-taken count will not overflow.
2200 auto PastOverflowCheck =
2201 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2202 BasicBlock *CheckBlock =
2203 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2205 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2206 LoopBypassBlocks.push_back(CheckBlock);
2207 Instruction *OldTerm = LastBypassBlock->getTerminator();
2208 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2209 OldTerm->eraseFromParent();
2210 LastBypassBlock = CheckBlock;
2213 // Generate the code to check that the strides we assumed to be one are really
2214 // one. We want the new basic block to start at the first instruction in a
2215 // sequence of instructions that form a check.
2216 Instruction *StrideCheck;
2217 Instruction *FirstCheckInst;
2218 std::tie(FirstCheckInst, StrideCheck) =
2219 addStrideCheck(LastBypassBlock->getTerminator());
2221 // Create a new block containing the stride check.
2222 BasicBlock *CheckBlock =
2223 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2225 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2226 LoopBypassBlocks.push_back(CheckBlock);
2228 // Replace the branch into the memory check block with a conditional branch
2229 // for the "few elements case".
2230 Instruction *OldTerm = LastBypassBlock->getTerminator();
2231 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2232 OldTerm->eraseFromParent();
2235 LastBypassBlock = CheckBlock;
2238 // Generate the code that checks in runtime if arrays overlap. We put the
2239 // checks into a separate block to make the more common case of few elements
2241 Instruction *MemRuntimeCheck;
2242 std::tie(FirstCheckInst, MemRuntimeCheck) =
2243 Legal->getLAI()->addRuntimeCheck(LastBypassBlock->getTerminator());
2244 if (MemRuntimeCheck) {
2245 // Create a new block containing the memory check.
2246 BasicBlock *CheckBlock =
2247 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.memcheck");
2249 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2250 LoopBypassBlocks.push_back(CheckBlock);
2252 // Replace the branch into the memory check block with a conditional branch
2253 // for the "few elements case".
2254 Instruction *OldTerm = LastBypassBlock->getTerminator();
2255 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2256 OldTerm->eraseFromParent();
2258 Cmp = MemRuntimeCheck;
2259 LastBypassBlock = CheckBlock;
2262 LastBypassBlock->getTerminator()->eraseFromParent();
2263 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2266 // We are going to resume the execution of the scalar loop.
2267 // Go over all of the induction variables that we found and fix the
2268 // PHIs that are left in the scalar version of the loop.
2269 // The starting values of PHI nodes depend on the counter of the last
2270 // iteration in the vectorized loop.
2271 // If we come from a bypass edge then we need to start from the original
2274 // This variable saves the new starting index for the scalar loop.
2275 PHINode *ResumeIndex = nullptr;
2276 LoopVectorizationLegality::InductionList::iterator I, E;
2277 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2278 // Set builder to point to last bypass block.
2279 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2280 for (I = List->begin(), E = List->end(); I != E; ++I) {
2281 PHINode *OrigPhi = I->first;
2282 LoopVectorizationLegality::InductionInfo II = I->second;
2284 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2285 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2286 MiddleBlock->getTerminator());
2287 // We might have extended the type of the induction variable but we need a
2288 // truncated version for the scalar loop.
2289 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2290 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2291 MiddleBlock->getTerminator()) : nullptr;
2293 // Create phi nodes to merge from the backedge-taken check block.
2294 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2295 ScalarPH->getTerminator());
2296 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2298 PHINode *BCTruncResumeVal = nullptr;
2299 if (OrigPhi == OldInduction) {
2301 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2302 ScalarPH->getTerminator());
2303 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2306 Value *EndValue = nullptr;
2308 case LoopVectorizationLegality::IK_NoInduction:
2309 llvm_unreachable("Unknown induction");
2310 case LoopVectorizationLegality::IK_IntInduction: {
2311 // Handle the integer induction counter.
2312 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2314 // We have the canonical induction variable.
2315 if (OrigPhi == OldInduction) {
2316 // Create a truncated version of the resume value for the scalar loop,
2317 // we might have promoted the type to a larger width.
2319 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2320 // The new PHI merges the original incoming value, in case of a bypass,
2321 // or the value at the end of the vectorized loop.
2322 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2323 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2324 TruncResumeVal->addIncoming(EndValue, VecBody);
2326 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2328 // We know what the end value is.
2329 EndValue = IdxEndRoundDown;
2330 // We also know which PHI node holds it.
2331 ResumeIndex = ResumeVal;
2335 // Not the canonical induction variable - add the vector loop count to the
2337 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2338 II.StartValue->getType(),
2340 EndValue = II.transform(BypassBuilder, CRD);
2341 EndValue->setName("ind.end");
2344 case LoopVectorizationLegality::IK_PtrInduction: {
2345 EndValue = II.transform(BypassBuilder, CountRoundDown);
2346 EndValue->setName("ptr.ind.end");
2351 // The new PHI merges the original incoming value, in case of a bypass,
2352 // or the value at the end of the vectorized loop.
2353 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2354 if (OrigPhi == OldInduction)
2355 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2357 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2359 ResumeVal->addIncoming(EndValue, VecBody);
2361 // Fix the scalar body counter (PHI node).
2362 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2364 // The old induction's phi node in the scalar body needs the truncated
2366 if (OrigPhi == OldInduction) {
2367 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2368 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2370 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2371 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2375 // If we are generating a new induction variable then we also need to
2376 // generate the code that calculates the exit value. This value is not
2377 // simply the end of the counter because we may skip the vectorized body
2378 // in case of a runtime check.
2380 assert(!ResumeIndex && "Unexpected resume value found");
2381 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2382 MiddleBlock->getTerminator());
2383 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2384 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2385 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2388 // Make sure that we found the index where scalar loop needs to continue.
2389 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2390 "Invalid resume Index");
2392 // Add a check in the middle block to see if we have completed
2393 // all of the iterations in the first vector loop.
2394 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2395 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2396 ResumeIndex, "cmp.n",
2397 MiddleBlock->getTerminator());
2399 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2400 // Remove the old terminator.
2401 MiddleBlock->getTerminator()->eraseFromParent();
2403 // Create i+1 and fill the PHINode.
2404 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2405 Induction->addIncoming(StartIdx, VectorPH);
2406 Induction->addIncoming(NextIdx, VecBody);
2407 // Create the compare.
2408 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2409 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2411 // Now we have two terminators. Remove the old one from the block.
2412 VecBody->getTerminator()->eraseFromParent();
2414 // Get ready to start creating new instructions into the vectorized body.
2415 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2418 LoopVectorPreHeader = VectorPH;
2419 LoopScalarPreHeader = ScalarPH;
2420 LoopMiddleBlock = MiddleBlock;
2421 LoopExitBlock = ExitBlock;
2422 LoopVectorBody.push_back(VecBody);
2423 LoopScalarBody = OldBasicBlock;
2425 LoopVectorizeHints Hints(Lp, true);
2426 Hints.setAlreadyVectorized();
2429 /// This function returns the identity element (or neutral element) for
2430 /// the operation K.
2432 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2437 // Adding, Xoring, Oring zero to a number does not change it.
2438 return ConstantInt::get(Tp, 0);
2439 case RK_IntegerMult:
2440 // Multiplying a number by 1 does not change it.
2441 return ConstantInt::get(Tp, 1);
2443 // AND-ing a number with an all-1 value does not change it.
2444 return ConstantInt::get(Tp, -1, true);
2446 // Multiplying a number by 1 does not change it.
2447 return ConstantFP::get(Tp, 1.0L);
2449 // Adding zero to a number does not change it.
2450 return ConstantFP::get(Tp, 0.0L);
2452 llvm_unreachable("Unknown reduction kind");
2456 /// This function translates the reduction kind to an LLVM binary operator.
2458 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2460 case LoopVectorizationLegality::RK_IntegerAdd:
2461 return Instruction::Add;
2462 case LoopVectorizationLegality::RK_IntegerMult:
2463 return Instruction::Mul;
2464 case LoopVectorizationLegality::RK_IntegerOr:
2465 return Instruction::Or;
2466 case LoopVectorizationLegality::RK_IntegerAnd:
2467 return Instruction::And;
2468 case LoopVectorizationLegality::RK_IntegerXor:
2469 return Instruction::Xor;
2470 case LoopVectorizationLegality::RK_FloatMult:
2471 return Instruction::FMul;
2472 case LoopVectorizationLegality::RK_FloatAdd:
2473 return Instruction::FAdd;
2474 case LoopVectorizationLegality::RK_IntegerMinMax:
2475 return Instruction::ICmp;
2476 case LoopVectorizationLegality::RK_FloatMinMax:
2477 return Instruction::FCmp;
2479 llvm_unreachable("Unknown reduction operation");
2483 Value *createMinMaxOp(IRBuilder<> &Builder,
2484 LoopVectorizationLegality::MinMaxReductionKind RK,
2487 CmpInst::Predicate P = CmpInst::ICMP_NE;
2490 llvm_unreachable("Unknown min/max reduction kind");
2491 case LoopVectorizationLegality::MRK_UIntMin:
2492 P = CmpInst::ICMP_ULT;
2494 case LoopVectorizationLegality::MRK_UIntMax:
2495 P = CmpInst::ICMP_UGT;
2497 case LoopVectorizationLegality::MRK_SIntMin:
2498 P = CmpInst::ICMP_SLT;
2500 case LoopVectorizationLegality::MRK_SIntMax:
2501 P = CmpInst::ICMP_SGT;
2503 case LoopVectorizationLegality::MRK_FloatMin:
2504 P = CmpInst::FCMP_OLT;
2506 case LoopVectorizationLegality::MRK_FloatMax:
2507 P = CmpInst::FCMP_OGT;
2512 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2513 RK == LoopVectorizationLegality::MRK_FloatMax)
2514 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2516 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2518 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2523 struct CSEDenseMapInfo {
2524 static bool canHandle(Instruction *I) {
2525 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2526 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2528 static inline Instruction *getEmptyKey() {
2529 return DenseMapInfo<Instruction *>::getEmptyKey();
2531 static inline Instruction *getTombstoneKey() {
2532 return DenseMapInfo<Instruction *>::getTombstoneKey();
2534 static unsigned getHashValue(Instruction *I) {
2535 assert(canHandle(I) && "Unknown instruction!");
2536 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2537 I->value_op_end()));
2539 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2540 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2541 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2543 return LHS->isIdenticalTo(RHS);
2548 /// \brief Check whether this block is a predicated block.
2549 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2550 /// = ...; " blocks. We start with one vectorized basic block. For every
2551 /// conditional block we split this vectorized block. Therefore, every second
2552 /// block will be a predicated one.
2553 static bool isPredicatedBlock(unsigned BlockNum) {
2554 return BlockNum % 2;
2557 ///\brief Perform cse of induction variable instructions.
2558 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2559 // Perform simple cse.
2560 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2561 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2562 BasicBlock *BB = BBs[i];
2563 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2564 Instruction *In = I++;
2566 if (!CSEDenseMapInfo::canHandle(In))
2569 // Check if we can replace this instruction with any of the
2570 // visited instructions.
2571 if (Instruction *V = CSEMap.lookup(In)) {
2572 In->replaceAllUsesWith(V);
2573 In->eraseFromParent();
2576 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2577 // ...;" blocks for predicated stores. Every second block is a predicated
2579 if (isPredicatedBlock(i))
2587 /// \brief Adds a 'fast' flag to floating point operations.
2588 static Value *addFastMathFlag(Value *V) {
2589 if (isa<FPMathOperator>(V)){
2590 FastMathFlags Flags;
2591 Flags.setUnsafeAlgebra();
2592 cast<Instruction>(V)->setFastMathFlags(Flags);
2597 void InnerLoopVectorizer::vectorizeLoop() {
2598 //===------------------------------------------------===//
2600 // Notice: any optimization or new instruction that go
2601 // into the code below should be also be implemented in
2604 //===------------------------------------------------===//
2605 Constant *Zero = Builder.getInt32(0);
2607 // In order to support reduction variables we need to be able to vectorize
2608 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2609 // stages. First, we create a new vector PHI node with no incoming edges.
2610 // We use this value when we vectorize all of the instructions that use the
2611 // PHI. Next, after all of the instructions in the block are complete we
2612 // add the new incoming edges to the PHI. At this point all of the
2613 // instructions in the basic block are vectorized, so we can use them to
2614 // construct the PHI.
2615 PhiVector RdxPHIsToFix;
2617 // Scan the loop in a topological order to ensure that defs are vectorized
2619 LoopBlocksDFS DFS(OrigLoop);
2622 // Vectorize all of the blocks in the original loop.
2623 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2624 be = DFS.endRPO(); bb != be; ++bb)
2625 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2627 // At this point every instruction in the original loop is widened to
2628 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2629 // that we vectorized. The PHI nodes are currently empty because we did
2630 // not want to introduce cycles. Notice that the remaining PHI nodes
2631 // that we need to fix are reduction variables.
2633 // Create the 'reduced' values for each of the induction vars.
2634 // The reduced values are the vector values that we scalarize and combine
2635 // after the loop is finished.
2636 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2638 PHINode *RdxPhi = *it;
2639 assert(RdxPhi && "Unable to recover vectorized PHI");
2641 // Find the reduction variable descriptor.
2642 assert(Legal->getReductionVars()->count(RdxPhi) &&
2643 "Unable to find the reduction variable");
2644 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2645 (*Legal->getReductionVars())[RdxPhi];
2647 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2649 // We need to generate a reduction vector from the incoming scalar.
2650 // To do so, we need to generate the 'identity' vector and override
2651 // one of the elements with the incoming scalar reduction. We need
2652 // to do it in the vector-loop preheader.
2653 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
2655 // This is the vector-clone of the value that leaves the loop.
2656 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2657 Type *VecTy = VectorExit[0]->getType();
2659 // Find the reduction identity variable. Zero for addition, or, xor,
2660 // one for multiplication, -1 for And.
2663 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2664 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2665 // MinMax reduction have the start value as their identify.
2667 VectorStart = Identity = RdxDesc.StartValue;
2669 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2674 // Handle other reduction kinds:
2676 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2677 VecTy->getScalarType());
2680 // This vector is the Identity vector where the first element is the
2681 // incoming scalar reduction.
2682 VectorStart = RdxDesc.StartValue;
2684 Identity = ConstantVector::getSplat(VF, Iden);
2686 // This vector is the Identity vector where the first element is the
2687 // incoming scalar reduction.
2688 VectorStart = Builder.CreateInsertElement(Identity,
2689 RdxDesc.StartValue, Zero);
2693 // Fix the vector-loop phi.
2695 // Reductions do not have to start at zero. They can start with
2696 // any loop invariant values.
2697 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2698 BasicBlock *Latch = OrigLoop->getLoopLatch();
2699 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2700 VectorParts &Val = getVectorValue(LoopVal);
2701 for (unsigned part = 0; part < UF; ++part) {
2702 // Make sure to add the reduction stat value only to the
2703 // first unroll part.
2704 Value *StartVal = (part == 0) ? VectorStart : Identity;
2705 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
2706 LoopVectorPreHeader);
2707 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2708 LoopVectorBody.back());
2711 // Before each round, move the insertion point right between
2712 // the PHIs and the values we are going to write.
2713 // This allows us to write both PHINodes and the extractelement
2715 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2717 VectorParts RdxParts;
2718 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2719 for (unsigned part = 0; part < UF; ++part) {
2720 // This PHINode contains the vectorized reduction variable, or
2721 // the initial value vector, if we bypass the vector loop.
2722 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2723 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2724 Value *StartVal = (part == 0) ? VectorStart : Identity;
2725 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2726 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2727 NewPhi->addIncoming(RdxExitVal[part],
2728 LoopVectorBody.back());
2729 RdxParts.push_back(NewPhi);
2732 // Reduce all of the unrolled parts into a single vector.
2733 Value *ReducedPartRdx = RdxParts[0];
2734 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2735 setDebugLocFromInst(Builder, ReducedPartRdx);
2736 for (unsigned part = 1; part < UF; ++part) {
2737 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2738 // Floating point operations had to be 'fast' to enable the reduction.
2739 ReducedPartRdx = addFastMathFlag(
2740 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2741 ReducedPartRdx, "bin.rdx"));
2743 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2744 ReducedPartRdx, RdxParts[part]);
2748 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2749 // and vector ops, reducing the set of values being computed by half each
2751 assert(isPowerOf2_32(VF) &&
2752 "Reduction emission only supported for pow2 vectors!");
2753 Value *TmpVec = ReducedPartRdx;
2754 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2755 for (unsigned i = VF; i != 1; i >>= 1) {
2756 // Move the upper half of the vector to the lower half.
2757 for (unsigned j = 0; j != i/2; ++j)
2758 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2760 // Fill the rest of the mask with undef.
2761 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2762 UndefValue::get(Builder.getInt32Ty()));
2765 Builder.CreateShuffleVector(TmpVec,
2766 UndefValue::get(TmpVec->getType()),
2767 ConstantVector::get(ShuffleMask),
2770 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2771 // Floating point operations had to be 'fast' to enable the reduction.
2772 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2773 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2775 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2778 // The result is in the first element of the vector.
2779 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2780 Builder.getInt32(0));
2783 // Create a phi node that merges control-flow from the backedge-taken check
2784 // block and the middle block.
2785 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
2786 LoopScalarPreHeader->getTerminator());
2787 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
2788 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2790 // Now, we need to fix the users of the reduction variable
2791 // inside and outside of the scalar remainder loop.
2792 // We know that the loop is in LCSSA form. We need to update the
2793 // PHI nodes in the exit blocks.
2794 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2795 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2796 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2797 if (!LCSSAPhi) break;
2799 // All PHINodes need to have a single entry edge, or two if
2800 // we already fixed them.
2801 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2803 // We found our reduction value exit-PHI. Update it with the
2804 // incoming bypass edge.
2805 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2806 // Add an edge coming from the bypass.
2807 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2810 }// end of the LCSSA phi scan.
2812 // Fix the scalar loop reduction variable with the incoming reduction sum
2813 // from the vector body and from the backedge value.
2814 int IncomingEdgeBlockIdx =
2815 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2816 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2817 // Pick the other block.
2818 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2819 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
2820 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2821 }// end of for each redux variable.
2825 // Remove redundant induction instructions.
2826 cse(LoopVectorBody);
2829 void InnerLoopVectorizer::fixLCSSAPHIs() {
2830 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2831 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2832 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2833 if (!LCSSAPhi) break;
2834 if (LCSSAPhi->getNumIncomingValues() == 1)
2835 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2840 InnerLoopVectorizer::VectorParts
2841 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2842 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2845 // Look for cached value.
2846 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2847 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2848 if (ECEntryIt != MaskCache.end())
2849 return ECEntryIt->second;
2851 VectorParts SrcMask = createBlockInMask(Src);
2853 // The terminator has to be a branch inst!
2854 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2855 assert(BI && "Unexpected terminator found");
2857 if (BI->isConditional()) {
2858 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2860 if (BI->getSuccessor(0) != Dst)
2861 for (unsigned part = 0; part < UF; ++part)
2862 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2864 for (unsigned part = 0; part < UF; ++part)
2865 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2867 MaskCache[Edge] = EdgeMask;
2871 MaskCache[Edge] = SrcMask;
2875 InnerLoopVectorizer::VectorParts
2876 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2877 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2879 // Loop incoming mask is all-one.
2880 if (OrigLoop->getHeader() == BB) {
2881 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2882 return getVectorValue(C);
2885 // This is the block mask. We OR all incoming edges, and with zero.
2886 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2887 VectorParts BlockMask = getVectorValue(Zero);
2890 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2891 VectorParts EM = createEdgeMask(*it, BB);
2892 for (unsigned part = 0; part < UF; ++part)
2893 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2899 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2900 InnerLoopVectorizer::VectorParts &Entry,
2901 unsigned UF, unsigned VF, PhiVector *PV) {
2902 PHINode* P = cast<PHINode>(PN);
2903 // Handle reduction variables:
2904 if (Legal->getReductionVars()->count(P)) {
2905 for (unsigned part = 0; part < UF; ++part) {
2906 // This is phase one of vectorizing PHIs.
2907 Type *VecTy = (VF == 1) ? PN->getType() :
2908 VectorType::get(PN->getType(), VF);
2909 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2910 LoopVectorBody.back()-> getFirstInsertionPt());
2916 setDebugLocFromInst(Builder, P);
2917 // Check for PHI nodes that are lowered to vector selects.
2918 if (P->getParent() != OrigLoop->getHeader()) {
2919 // We know that all PHIs in non-header blocks are converted into
2920 // selects, so we don't have to worry about the insertion order and we
2921 // can just use the builder.
2922 // At this point we generate the predication tree. There may be
2923 // duplications since this is a simple recursive scan, but future
2924 // optimizations will clean it up.
2926 unsigned NumIncoming = P->getNumIncomingValues();
2928 // Generate a sequence of selects of the form:
2929 // SELECT(Mask3, In3,
2930 // SELECT(Mask2, In2,
2932 for (unsigned In = 0; In < NumIncoming; In++) {
2933 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2935 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2937 for (unsigned part = 0; part < UF; ++part) {
2938 // We might have single edge PHIs (blocks) - use an identity
2939 // 'select' for the first PHI operand.
2941 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2944 // Select between the current value and the previous incoming edge
2945 // based on the incoming mask.
2946 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2947 Entry[part], "predphi");
2953 // This PHINode must be an induction variable.
2954 // Make sure that we know about it.
2955 assert(Legal->getInductionVars()->count(P) &&
2956 "Not an induction variable");
2958 LoopVectorizationLegality::InductionInfo II =
2959 Legal->getInductionVars()->lookup(P);
2961 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
2962 // which can be found from the original scalar operations.
2964 case LoopVectorizationLegality::IK_NoInduction:
2965 llvm_unreachable("Unknown induction");
2966 case LoopVectorizationLegality::IK_IntInduction: {
2967 assert(P->getType() == II.StartValue->getType() && "Types must match");
2968 Type *PhiTy = P->getType();
2970 if (P == OldInduction) {
2971 // Handle the canonical induction variable. We might have had to
2973 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2975 // Handle other induction variables that are now based on the
2977 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2979 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2980 Broadcasted = II.transform(Builder, NormalizedIdx);
2981 Broadcasted->setName("offset.idx");
2983 Broadcasted = getBroadcastInstrs(Broadcasted);
2984 // After broadcasting the induction variable we need to make the vector
2985 // consecutive by adding 0, 1, 2, etc.
2986 for (unsigned part = 0; part < UF; ++part)
2987 Entry[part] = getStepVector(Broadcasted, VF * part, II.StepValue);
2990 case LoopVectorizationLegality::IK_PtrInduction:
2991 // Handle the pointer induction variable case.
2992 assert(P->getType()->isPointerTy() && "Unexpected type.");
2993 // This is the normalized GEP that starts counting at zero.
2994 Value *NormalizedIdx =
2995 Builder.CreateSub(Induction, ExtendedIdx, "normalized.idx");
2996 // This is the vector of results. Notice that we don't generate
2997 // vector geps because scalar geps result in better code.
2998 for (unsigned part = 0; part < UF; ++part) {
3000 int EltIndex = part;
3001 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3002 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3003 Value *SclrGep = II.transform(Builder, GlobalIdx);
3004 SclrGep->setName("next.gep");
3005 Entry[part] = SclrGep;
3009 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3010 for (unsigned int i = 0; i < VF; ++i) {
3011 int EltIndex = i + part * VF;
3012 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3013 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3014 Value *SclrGep = II.transform(Builder, GlobalIdx);
3015 SclrGep->setName("next.gep");
3016 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3017 Builder.getInt32(i),
3020 Entry[part] = VecVal;
3026 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3027 // For each instruction in the old loop.
3028 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3029 VectorParts &Entry = WidenMap.get(it);
3030 switch (it->getOpcode()) {
3031 case Instruction::Br:
3032 // Nothing to do for PHIs and BR, since we already took care of the
3033 // loop control flow instructions.
3035 case Instruction::PHI: {
3036 // Vectorize PHINodes.
3037 widenPHIInstruction(it, Entry, UF, VF, PV);
3041 case Instruction::Add:
3042 case Instruction::FAdd:
3043 case Instruction::Sub:
3044 case Instruction::FSub:
3045 case Instruction::Mul:
3046 case Instruction::FMul:
3047 case Instruction::UDiv:
3048 case Instruction::SDiv:
3049 case Instruction::FDiv:
3050 case Instruction::URem:
3051 case Instruction::SRem:
3052 case Instruction::FRem:
3053 case Instruction::Shl:
3054 case Instruction::LShr:
3055 case Instruction::AShr:
3056 case Instruction::And:
3057 case Instruction::Or:
3058 case Instruction::Xor: {
3059 // Just widen binops.
3060 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3061 setDebugLocFromInst(Builder, BinOp);
3062 VectorParts &A = getVectorValue(it->getOperand(0));
3063 VectorParts &B = getVectorValue(it->getOperand(1));
3065 // Use this vector value for all users of the original instruction.
3066 for (unsigned Part = 0; Part < UF; ++Part) {
3067 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3069 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3070 VecOp->copyIRFlags(BinOp);
3075 propagateMetadata(Entry, it);
3078 case Instruction::Select: {
3080 // If the selector is loop invariant we can create a select
3081 // instruction with a scalar condition. Otherwise, use vector-select.
3082 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3084 setDebugLocFromInst(Builder, it);
3086 // The condition can be loop invariant but still defined inside the
3087 // loop. This means that we can't just use the original 'cond' value.
3088 // We have to take the 'vectorized' value and pick the first lane.
3089 // Instcombine will make this a no-op.
3090 VectorParts &Cond = getVectorValue(it->getOperand(0));
3091 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3092 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3094 Value *ScalarCond = (VF == 1) ? Cond[0] :
3095 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3097 for (unsigned Part = 0; Part < UF; ++Part) {
3098 Entry[Part] = Builder.CreateSelect(
3099 InvariantCond ? ScalarCond : Cond[Part],
3104 propagateMetadata(Entry, it);
3108 case Instruction::ICmp:
3109 case Instruction::FCmp: {
3110 // Widen compares. Generate vector compares.
3111 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3112 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3113 setDebugLocFromInst(Builder, it);
3114 VectorParts &A = getVectorValue(it->getOperand(0));
3115 VectorParts &B = getVectorValue(it->getOperand(1));
3116 for (unsigned Part = 0; Part < UF; ++Part) {
3119 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3121 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3125 propagateMetadata(Entry, it);
3129 case Instruction::Store:
3130 case Instruction::Load:
3131 vectorizeMemoryInstruction(it);
3133 case Instruction::ZExt:
3134 case Instruction::SExt:
3135 case Instruction::FPToUI:
3136 case Instruction::FPToSI:
3137 case Instruction::FPExt:
3138 case Instruction::PtrToInt:
3139 case Instruction::IntToPtr:
3140 case Instruction::SIToFP:
3141 case Instruction::UIToFP:
3142 case Instruction::Trunc:
3143 case Instruction::FPTrunc:
3144 case Instruction::BitCast: {
3145 CastInst *CI = dyn_cast<CastInst>(it);
3146 setDebugLocFromInst(Builder, it);
3147 /// Optimize the special case where the source is the induction
3148 /// variable. Notice that we can only optimize the 'trunc' case
3149 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3150 /// c. other casts depend on pointer size.
3151 if (CI->getOperand(0) == OldInduction &&
3152 it->getOpcode() == Instruction::Trunc) {
3153 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3155 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3156 LoopVectorizationLegality::InductionInfo II =
3157 Legal->getInductionVars()->lookup(OldInduction);
3159 ConstantInt::getSigned(CI->getType(), II.StepValue->getSExtValue());
3160 for (unsigned Part = 0; Part < UF; ++Part)
3161 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
3162 propagateMetadata(Entry, it);
3165 /// Vectorize casts.
3166 Type *DestTy = (VF == 1) ? CI->getType() :
3167 VectorType::get(CI->getType(), VF);
3169 VectorParts &A = getVectorValue(it->getOperand(0));
3170 for (unsigned Part = 0; Part < UF; ++Part)
3171 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3172 propagateMetadata(Entry, it);
3176 case Instruction::Call: {
3177 // Ignore dbg intrinsics.
3178 if (isa<DbgInfoIntrinsic>(it))
3180 setDebugLocFromInst(Builder, it);
3182 Module *M = BB->getParent()->getParent();
3183 CallInst *CI = cast<CallInst>(it);
3184 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3185 assert(ID && "Not an intrinsic call!");
3187 case Intrinsic::assume:
3188 case Intrinsic::lifetime_end:
3189 case Intrinsic::lifetime_start:
3190 scalarizeInstruction(it);
3193 bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
3194 for (unsigned Part = 0; Part < UF; ++Part) {
3195 SmallVector<Value *, 4> Args;
3196 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3197 if (HasScalarOpd && i == 1) {
3198 Args.push_back(CI->getArgOperand(i));
3201 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3202 Args.push_back(Arg[Part]);
3204 Type *Tys[] = {CI->getType()};
3206 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3208 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3209 Entry[Part] = Builder.CreateCall(F, Args);
3212 propagateMetadata(Entry, it);
3219 // All other instructions are unsupported. Scalarize them.
3220 scalarizeInstruction(it);
3223 }// end of for_each instr.
3226 void InnerLoopVectorizer::updateAnalysis() {
3227 // Forget the original basic block.
3228 SE->forgetLoop(OrigLoop);
3230 // Update the dominator tree information.
3231 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3232 "Entry does not dominate exit.");
3234 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3235 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3236 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3238 // Due to if predication of stores we might create a sequence of "if(pred)
3239 // a[i] = ...; " blocks.
3240 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3242 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3243 else if (isPredicatedBlock(i)) {
3244 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3246 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3250 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3251 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3252 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3253 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3255 DEBUG(DT->verifyDomTree());
3258 /// \brief Check whether it is safe to if-convert this phi node.
3260 /// Phi nodes with constant expressions that can trap are not safe to if
3262 static bool canIfConvertPHINodes(BasicBlock *BB) {
3263 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3264 PHINode *Phi = dyn_cast<PHINode>(I);
3267 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3268 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3275 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3276 if (!EnableIfConversion) {
3277 emitAnalysis(VectorizationReport() << "if-conversion is disabled");
3281 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3283 // A list of pointers that we can safely read and write to.
3284 SmallPtrSet<Value *, 8> SafePointes;
3286 // Collect safe addresses.
3287 for (Loop::block_iterator BI = TheLoop->block_begin(),
3288 BE = TheLoop->block_end(); BI != BE; ++BI) {
3289 BasicBlock *BB = *BI;
3291 if (blockNeedsPredication(BB))
3294 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3295 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3296 SafePointes.insert(LI->getPointerOperand());
3297 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3298 SafePointes.insert(SI->getPointerOperand());
3302 // Collect the blocks that need predication.
3303 BasicBlock *Header = TheLoop->getHeader();
3304 for (Loop::block_iterator BI = TheLoop->block_begin(),
3305 BE = TheLoop->block_end(); BI != BE; ++BI) {
3306 BasicBlock *BB = *BI;
3308 // We don't support switch statements inside loops.
3309 if (!isa<BranchInst>(BB->getTerminator())) {
3310 emitAnalysis(VectorizationReport(BB->getTerminator())
3311 << "loop contains a switch statement");
3315 // We must be able to predicate all blocks that need to be predicated.
3316 if (blockNeedsPredication(BB)) {
3317 if (!blockCanBePredicated(BB, SafePointes)) {
3318 emitAnalysis(VectorizationReport(BB->getTerminator())
3319 << "control flow cannot be substituted for a select");
3322 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3323 emitAnalysis(VectorizationReport(BB->getTerminator())
3324 << "control flow cannot be substituted for a select");
3329 // We can if-convert this loop.
3333 bool LoopVectorizationLegality::canVectorize() {
3334 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3335 // be canonicalized.
3336 if (!TheLoop->getLoopPreheader()) {
3338 VectorizationReport() <<
3339 "loop control flow is not understood by vectorizer");
3343 // We can only vectorize innermost loops.
3344 if (!TheLoop->getSubLoopsVector().empty()) {
3345 emitAnalysis(VectorizationReport() << "loop is not the innermost loop");
3349 // We must have a single backedge.
3350 if (TheLoop->getNumBackEdges() != 1) {
3352 VectorizationReport() <<
3353 "loop control flow is not understood by vectorizer");
3357 // We must have a single exiting block.
3358 if (!TheLoop->getExitingBlock()) {
3360 VectorizationReport() <<
3361 "loop control flow is not understood by vectorizer");
3365 // We only handle bottom-tested loops, i.e. loop in which the condition is
3366 // checked at the end of each iteration. With that we can assume that all
3367 // instructions in the loop are executed the same number of times.
3368 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
3370 VectorizationReport() <<
3371 "loop control flow is not understood by vectorizer");
3375 // We need to have a loop header.
3376 DEBUG(dbgs() << "LV: Found a loop: " <<
3377 TheLoop->getHeader()->getName() << '\n');
3379 // Check if we can if-convert non-single-bb loops.
3380 unsigned NumBlocks = TheLoop->getNumBlocks();
3381 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3382 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3386 // ScalarEvolution needs to be able to find the exit count.
3387 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3388 if (ExitCount == SE->getCouldNotCompute()) {
3389 emitAnalysis(VectorizationReport() <<
3390 "could not determine number of loop iterations");
3391 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3395 // Check if we can vectorize the instructions and CFG in this loop.
3396 if (!canVectorizeInstrs()) {
3397 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3401 // Go over each instruction and look at memory deps.
3402 if (!canVectorizeMemory()) {
3403 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3407 // Collect all of the variables that remain uniform after vectorization.
3408 collectLoopUniforms();
3410 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3411 (LAI->getRuntimePointerCheck()->Need ? " (with a runtime bound check)" :
3415 // Okay! We can vectorize. At this point we don't have any other mem analysis
3416 // which may limit our maximum vectorization factor, so just return true with
3421 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3422 if (Ty->isPointerTy())
3423 return DL.getIntPtrType(Ty);
3425 // It is possible that char's or short's overflow when we ask for the loop's
3426 // trip count, work around this by changing the type size.
3427 if (Ty->getScalarSizeInBits() < 32)
3428 return Type::getInt32Ty(Ty->getContext());
3433 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3434 Ty0 = convertPointerToIntegerType(DL, Ty0);
3435 Ty1 = convertPointerToIntegerType(DL, Ty1);
3436 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3441 /// \brief Check that the instruction has outside loop users and is not an
3442 /// identified reduction variable.
3443 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3444 SmallPtrSetImpl<Value *> &Reductions) {
3445 // Reduction instructions are allowed to have exit users. All other
3446 // instructions must not have external users.
3447 if (!Reductions.count(Inst))
3448 //Check that all of the users of the loop are inside the BB.
3449 for (User *U : Inst->users()) {
3450 Instruction *UI = cast<Instruction>(U);
3451 // This user may be a reduction exit value.
3452 if (!TheLoop->contains(UI)) {
3453 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3460 bool LoopVectorizationLegality::canVectorizeInstrs() {
3461 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3462 BasicBlock *Header = TheLoop->getHeader();
3464 // Look for the attribute signaling the absence of NaNs.
3465 Function &F = *Header->getParent();
3466 if (F.hasFnAttribute("no-nans-fp-math"))
3468 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
3470 // For each block in the loop.
3471 for (Loop::block_iterator bb = TheLoop->block_begin(),
3472 be = TheLoop->block_end(); bb != be; ++bb) {
3474 // Scan the instructions in the block and look for hazards.
3475 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3478 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3479 Type *PhiTy = Phi->getType();
3480 // Check that this PHI type is allowed.
3481 if (!PhiTy->isIntegerTy() &&
3482 !PhiTy->isFloatingPointTy() &&
3483 !PhiTy->isPointerTy()) {
3484 emitAnalysis(VectorizationReport(it)
3485 << "loop control flow is not understood by vectorizer");
3486 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3490 // If this PHINode is not in the header block, then we know that we
3491 // can convert it to select during if-conversion. No need to check if
3492 // the PHIs in this block are induction or reduction variables.
3493 if (*bb != Header) {
3494 // Check that this instruction has no outside users or is an
3495 // identified reduction value with an outside user.
3496 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3498 emitAnalysis(VectorizationReport(it) <<
3499 "value could not be identified as "
3500 "an induction or reduction variable");
3504 // We only allow if-converted PHIs with exactly two incoming values.
3505 if (Phi->getNumIncomingValues() != 2) {
3506 emitAnalysis(VectorizationReport(it)
3507 << "control flow not understood by vectorizer");
3508 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3512 // This is the value coming from the preheader.
3513 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3514 ConstantInt *StepValue = nullptr;
3515 // Check if this is an induction variable.
3516 InductionKind IK = isInductionVariable(Phi, StepValue);
3518 if (IK_NoInduction != IK) {
3519 // Get the widest type.
3521 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3523 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3525 // Int inductions are special because we only allow one IV.
3526 if (IK == IK_IntInduction && StepValue->isOne()) {
3527 // Use the phi node with the widest type as induction. Use the last
3528 // one if there are multiple (no good reason for doing this other
3529 // than it is expedient).
3530 if (!Induction || PhiTy == WidestIndTy)
3534 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3535 Inductions[Phi] = InductionInfo(StartValue, IK, StepValue);
3537 // Until we explicitly handle the case of an induction variable with
3538 // an outside loop user we have to give up vectorizing this loop.
3539 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3540 emitAnalysis(VectorizationReport(it) <<
3541 "use of induction value outside of the "
3542 "loop is not handled by vectorizer");
3549 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3550 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3553 if (AddReductionVar(Phi, RK_IntegerMult)) {
3554 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3557 if (AddReductionVar(Phi, RK_IntegerOr)) {
3558 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3561 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3562 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3565 if (AddReductionVar(Phi, RK_IntegerXor)) {
3566 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3569 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3570 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3573 if (AddReductionVar(Phi, RK_FloatMult)) {
3574 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3577 if (AddReductionVar(Phi, RK_FloatAdd)) {
3578 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3581 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3582 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3587 emitAnalysis(VectorizationReport(it) <<
3588 "value that could not be identified as "
3589 "reduction is used outside the loop");
3590 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3592 }// end of PHI handling
3594 // We still don't handle functions. However, we can ignore dbg intrinsic
3595 // calls and we do handle certain intrinsic and libm functions.
3596 CallInst *CI = dyn_cast<CallInst>(it);
3597 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3598 emitAnalysis(VectorizationReport(it) <<
3599 "call instruction cannot be vectorized");
3600 DEBUG(dbgs() << "LV: Found a call site.\n");
3604 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
3605 // second argument is the same (i.e. loop invariant)
3607 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
3608 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
3609 emitAnalysis(VectorizationReport(it)
3610 << "intrinsic instruction cannot be vectorized");
3611 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
3616 // Check that the instruction return type is vectorizable.
3617 // Also, we can't vectorize extractelement instructions.
3618 if ((!VectorType::isValidElementType(it->getType()) &&
3619 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3620 emitAnalysis(VectorizationReport(it)
3621 << "instruction return type cannot be vectorized");
3622 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3626 // Check that the stored type is vectorizable.
3627 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3628 Type *T = ST->getValueOperand()->getType();
3629 if (!VectorType::isValidElementType(T)) {
3630 emitAnalysis(VectorizationReport(ST) <<
3631 "store instruction cannot be vectorized");
3634 if (EnableMemAccessVersioning)
3635 collectStridedAccess(ST);
3638 if (EnableMemAccessVersioning)
3639 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3640 collectStridedAccess(LI);
3642 // Reduction instructions are allowed to have exit users.
3643 // All other instructions must not have external users.
3644 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3645 emitAnalysis(VectorizationReport(it) <<
3646 "value cannot be used outside the loop");
3655 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3656 if (Inductions.empty()) {
3657 emitAnalysis(VectorizationReport()
3658 << "loop induction variable could not be identified");
3666 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3667 /// return the induction operand of the gep pointer.
3668 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3669 const DataLayout *DL, Loop *Lp) {
3670 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3674 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3676 // Check that all of the gep indices are uniform except for our induction
3678 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3679 if (i != InductionOperand &&
3680 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3682 return GEP->getOperand(InductionOperand);
3685 ///\brief Look for a cast use of the passed value.
3686 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3687 Value *UniqueCast = nullptr;
3688 for (User *U : Ptr->users()) {
3689 CastInst *CI = dyn_cast<CastInst>(U);
3690 if (CI && CI->getType() == Ty) {
3700 ///\brief Get the stride of a pointer access in a loop.
3701 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3702 /// pointer to the Value, or null otherwise.
3703 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3704 const DataLayout *DL, Loop *Lp) {
3705 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3706 if (!PtrTy || PtrTy->isAggregateType())
3709 // Try to remove a gep instruction to make the pointer (actually index at this
3710 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3711 // pointer, otherwise, we are analyzing the index.
3712 Value *OrigPtr = Ptr;
3714 // The size of the pointer access.
3715 int64_t PtrAccessSize = 1;
3717 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3718 const SCEV *V = SE->getSCEV(Ptr);
3722 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3723 V = C->getOperand();
3725 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3729 V = S->getStepRecurrence(*SE);
3733 // Strip off the size of access multiplication if we are still analyzing the
3735 if (OrigPtr == Ptr) {
3736 DL->getTypeAllocSize(PtrTy->getElementType());
3737 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3738 if (M->getOperand(0)->getSCEVType() != scConstant)
3741 const APInt &APStepVal =
3742 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3744 // Huge step value - give up.
3745 if (APStepVal.getBitWidth() > 64)
3748 int64_t StepVal = APStepVal.getSExtValue();
3749 if (PtrAccessSize != StepVal)
3751 V = M->getOperand(1);
3756 Type *StripedOffRecurrenceCast = nullptr;
3757 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3758 StripedOffRecurrenceCast = C->getType();
3759 V = C->getOperand();
3762 // Look for the loop invariant symbolic value.
3763 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3767 Value *Stride = U->getValue();
3768 if (!Lp->isLoopInvariant(Stride))
3771 // If we have stripped off the recurrence cast we have to make sure that we
3772 // return the value that is used in this loop so that we can replace it later.
3773 if (StripedOffRecurrenceCast)
3774 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3779 void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) {
3780 Value *Ptr = nullptr;
3781 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3782 Ptr = LI->getPointerOperand();
3783 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3784 Ptr = SI->getPointerOperand();
3788 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3792 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3793 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3794 Strides[Ptr] = Stride;
3795 StrideSet.insert(Stride);
3798 void LoopVectorizationLegality::collectLoopUniforms() {
3799 // We now know that the loop is vectorizable!
3800 // Collect variables that will remain uniform after vectorization.
3801 std::vector<Value*> Worklist;
3802 BasicBlock *Latch = TheLoop->getLoopLatch();
3804 // Start with the conditional branch and walk up the block.
3805 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3807 // Also add all consecutive pointer values; these values will be uniform
3808 // after vectorization (and subsequent cleanup) and, until revectorization is
3809 // supported, all dependencies must also be uniform.
3810 for (Loop::block_iterator B = TheLoop->block_begin(),
3811 BE = TheLoop->block_end(); B != BE; ++B)
3812 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
3814 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
3815 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3817 while (!Worklist.empty()) {
3818 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3819 Worklist.pop_back();
3821 // Look at instructions inside this loop.
3822 // Stop when reaching PHI nodes.
3823 // TODO: we need to follow values all over the loop, not only in this block.
3824 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3827 // This is a known uniform.
3830 // Insert all operands.
3831 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3835 bool LoopVectorizationLegality::canVectorizeMemory() {
3836 LAI = &LAA->getInfo(TheLoop, Strides);
3837 auto &OptionalReport = LAI->getReport();
3839 emitAnalysis(VectorizationReport(*OptionalReport));
3840 return LAI->canVectorizeMemory();
3843 static bool hasMultipleUsesOf(Instruction *I,
3844 SmallPtrSetImpl<Instruction *> &Insts) {
3845 unsigned NumUses = 0;
3846 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
3847 if (Insts.count(dyn_cast<Instruction>(*Use)))
3856 static bool areAllUsesIn(Instruction *I, SmallPtrSetImpl<Instruction *> &Set) {
3857 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
3858 if (!Set.count(dyn_cast<Instruction>(*Use)))
3863 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
3864 ReductionKind Kind) {
3865 if (Phi->getNumIncomingValues() != 2)
3868 // Reduction variables are only found in the loop header block.
3869 if (Phi->getParent() != TheLoop->getHeader())
3872 // Obtain the reduction start value from the value that comes from the loop
3874 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
3876 // ExitInstruction is the single value which is used outside the loop.
3877 // We only allow for a single reduction value to be used outside the loop.
3878 // This includes users of the reduction, variables (which form a cycle
3879 // which ends in the phi node).
3880 Instruction *ExitInstruction = nullptr;
3881 // Indicates that we found a reduction operation in our scan.
3882 bool FoundReduxOp = false;
3884 // We start with the PHI node and scan for all of the users of this
3885 // instruction. All users must be instructions that can be used as reduction
3886 // variables (such as ADD). We must have a single out-of-block user. The cycle
3887 // must include the original PHI.
3888 bool FoundStartPHI = false;
3890 // To recognize min/max patterns formed by a icmp select sequence, we store
3891 // the number of instruction we saw from the recognized min/max pattern,
3892 // to make sure we only see exactly the two instructions.
3893 unsigned NumCmpSelectPatternInst = 0;
3894 ReductionInstDesc ReduxDesc(false, nullptr);
3896 SmallPtrSet<Instruction *, 8> VisitedInsts;
3897 SmallVector<Instruction *, 8> Worklist;
3898 Worklist.push_back(Phi);
3899 VisitedInsts.insert(Phi);
3901 // A value in the reduction can be used:
3902 // - By the reduction:
3903 // - Reduction operation:
3904 // - One use of reduction value (safe).
3905 // - Multiple use of reduction value (not safe).
3907 // - All uses of the PHI must be the reduction (safe).
3908 // - Otherwise, not safe.
3909 // - By one instruction outside of the loop (safe).
3910 // - By further instructions outside of the loop (not safe).
3911 // - By an instruction that is not part of the reduction (not safe).
3913 // * An instruction type other than PHI or the reduction operation.
3914 // * A PHI in the header other than the initial PHI.
3915 while (!Worklist.empty()) {
3916 Instruction *Cur = Worklist.back();
3917 Worklist.pop_back();
3920 // If the instruction has no users then this is a broken chain and can't be
3921 // a reduction variable.
3922 if (Cur->use_empty())
3925 bool IsAPhi = isa<PHINode>(Cur);
3927 // A header PHI use other than the original PHI.
3928 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
3931 // Reductions of instructions such as Div, and Sub is only possible if the
3932 // LHS is the reduction variable.
3933 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
3934 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
3935 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
3938 // Any reduction instruction must be of one of the allowed kinds.
3939 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
3940 if (!ReduxDesc.IsReduction)
3943 // A reduction operation must only have one use of the reduction value.
3944 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
3945 hasMultipleUsesOf(Cur, VisitedInsts))
3948 // All inputs to a PHI node must be a reduction value.
3949 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
3952 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
3953 isa<SelectInst>(Cur)))
3954 ++NumCmpSelectPatternInst;
3955 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
3956 isa<SelectInst>(Cur)))
3957 ++NumCmpSelectPatternInst;
3959 // Check whether we found a reduction operator.
3960 FoundReduxOp |= !IsAPhi;
3962 // Process users of current instruction. Push non-PHI nodes after PHI nodes
3963 // onto the stack. This way we are going to have seen all inputs to PHI
3964 // nodes once we get to them.
3965 SmallVector<Instruction *, 8> NonPHIs;
3966 SmallVector<Instruction *, 8> PHIs;
3967 for (User *U : Cur->users()) {
3968 Instruction *UI = cast<Instruction>(U);
3970 // Check if we found the exit user.
3971 BasicBlock *Parent = UI->getParent();
3972 if (!TheLoop->contains(Parent)) {
3973 // Exit if you find multiple outside users or if the header phi node is
3974 // being used. In this case the user uses the value of the previous
3975 // iteration, in which case we would loose "VF-1" iterations of the
3976 // reduction operation if we vectorize.
3977 if (ExitInstruction != nullptr || Cur == Phi)
3980 // The instruction used by an outside user must be the last instruction
3981 // before we feed back to the reduction phi. Otherwise, we loose VF-1
3982 // operations on the value.
3983 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
3986 ExitInstruction = Cur;
3990 // Process instructions only once (termination). Each reduction cycle
3991 // value must only be used once, except by phi nodes and min/max
3992 // reductions which are represented as a cmp followed by a select.
3993 ReductionInstDesc IgnoredVal(false, nullptr);
3994 if (VisitedInsts.insert(UI).second) {
3995 if (isa<PHINode>(UI))
3998 NonPHIs.push_back(UI);
3999 } else if (!isa<PHINode>(UI) &&
4000 ((!isa<FCmpInst>(UI) &&
4001 !isa<ICmpInst>(UI) &&
4002 !isa<SelectInst>(UI)) ||
4003 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
4006 // Remember that we completed the cycle.
4008 FoundStartPHI = true;
4010 Worklist.append(PHIs.begin(), PHIs.end());
4011 Worklist.append(NonPHIs.begin(), NonPHIs.end());
4014 // This means we have seen one but not the other instruction of the
4015 // pattern or more than just a select and cmp.
4016 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
4017 NumCmpSelectPatternInst != 2)
4020 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
4023 // We found a reduction var if we have reached the original phi node and we
4024 // only have a single instruction with out-of-loop users.
4026 // This instruction is allowed to have out-of-loop users.
4027 AllowedExit.insert(ExitInstruction);
4029 // Save the description of this reduction variable.
4030 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
4031 ReduxDesc.MinMaxKind);
4032 Reductions[Phi] = RD;
4033 // We've ended the cycle. This is a reduction variable if we have an
4034 // outside user and it has a binary op.
4039 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
4040 /// pattern corresponding to a min(X, Y) or max(X, Y).
4041 LoopVectorizationLegality::ReductionInstDesc
4042 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
4043 ReductionInstDesc &Prev) {
4045 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
4046 "Expect a select instruction");
4047 Instruction *Cmp = nullptr;
4048 SelectInst *Select = nullptr;
4050 // We must handle the select(cmp()) as a single instruction. Advance to the
4052 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
4053 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
4054 return ReductionInstDesc(false, I);
4055 return ReductionInstDesc(Select, Prev.MinMaxKind);
4058 // Only handle single use cases for now.
4059 if (!(Select = dyn_cast<SelectInst>(I)))
4060 return ReductionInstDesc(false, I);
4061 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
4062 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
4063 return ReductionInstDesc(false, I);
4064 if (!Cmp->hasOneUse())
4065 return ReductionInstDesc(false, I);
4070 // Look for a min/max pattern.
4071 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4072 return ReductionInstDesc(Select, MRK_UIntMin);
4073 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4074 return ReductionInstDesc(Select, MRK_UIntMax);
4075 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4076 return ReductionInstDesc(Select, MRK_SIntMax);
4077 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4078 return ReductionInstDesc(Select, MRK_SIntMin);
4079 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4080 return ReductionInstDesc(Select, MRK_FloatMin);
4081 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4082 return ReductionInstDesc(Select, MRK_FloatMax);
4083 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4084 return ReductionInstDesc(Select, MRK_FloatMin);
4085 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4086 return ReductionInstDesc(Select, MRK_FloatMax);
4088 return ReductionInstDesc(false, I);
4091 LoopVectorizationLegality::ReductionInstDesc
4092 LoopVectorizationLegality::isReductionInstr(Instruction *I,
4094 ReductionInstDesc &Prev) {
4095 bool FP = I->getType()->isFloatingPointTy();
4096 bool FastMath = FP && I->hasUnsafeAlgebra();
4097 switch (I->getOpcode()) {
4099 return ReductionInstDesc(false, I);
4100 case Instruction::PHI:
4101 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
4102 Kind != RK_FloatMinMax))
4103 return ReductionInstDesc(false, I);
4104 return ReductionInstDesc(I, Prev.MinMaxKind);
4105 case Instruction::Sub:
4106 case Instruction::Add:
4107 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4108 case Instruction::Mul:
4109 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4110 case Instruction::And:
4111 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4112 case Instruction::Or:
4113 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4114 case Instruction::Xor:
4115 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4116 case Instruction::FMul:
4117 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4118 case Instruction::FSub:
4119 case Instruction::FAdd:
4120 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4121 case Instruction::FCmp:
4122 case Instruction::ICmp:
4123 case Instruction::Select:
4124 if (Kind != RK_IntegerMinMax &&
4125 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4126 return ReductionInstDesc(false, I);
4127 return isMinMaxSelectCmpPattern(I, Prev);
4131 LoopVectorizationLegality::InductionKind
4132 LoopVectorizationLegality::isInductionVariable(PHINode *Phi,
4133 ConstantInt *&StepValue) {
4134 Type *PhiTy = Phi->getType();
4135 // We only handle integer and pointer inductions variables.
4136 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4137 return IK_NoInduction;
4139 // Check that the PHI is consecutive.
4140 const SCEV *PhiScev = SE->getSCEV(Phi);
4141 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4143 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4144 return IK_NoInduction;
4147 const SCEV *Step = AR->getStepRecurrence(*SE);
4148 // Calculate the pointer stride and check if it is consecutive.
4149 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4151 return IK_NoInduction;
4153 ConstantInt *CV = C->getValue();
4154 if (PhiTy->isIntegerTy()) {
4156 return IK_IntInduction;
4159 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4160 Type *PointerElementType = PhiTy->getPointerElementType();
4161 // The pointer stride cannot be determined if the pointer element type is not
4163 if (!PointerElementType->isSized())
4164 return IK_NoInduction;
4166 int64_t Size = static_cast<int64_t>(DL->getTypeAllocSize(PointerElementType));
4167 int64_t CVSize = CV->getSExtValue();
4169 return IK_NoInduction;
4170 StepValue = ConstantInt::getSigned(CV->getType(), CVSize / Size);
4171 return IK_PtrInduction;
4174 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4175 Value *In0 = const_cast<Value*>(V);
4176 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4180 return Inductions.count(PN);
4183 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4184 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4187 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4188 SmallPtrSetImpl<Value *> &SafePtrs) {
4190 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4191 // Check that we don't have a constant expression that can trap as operand.
4192 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4194 if (Constant *C = dyn_cast<Constant>(*OI))
4198 // We might be able to hoist the load.
4199 if (it->mayReadFromMemory()) {
4200 LoadInst *LI = dyn_cast<LoadInst>(it);
4203 if (!SafePtrs.count(LI->getPointerOperand())) {
4204 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
4205 MaskedOp.insert(LI);
4212 // We don't predicate stores at the moment.
4213 if (it->mayWriteToMemory()) {
4214 StoreInst *SI = dyn_cast<StoreInst>(it);
4215 // We only support predication of stores in basic blocks with one
4220 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
4221 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
4223 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
4224 !isSinglePredecessor) {
4225 // Build a masked store if it is legal for the target, otherwise scalarize
4227 bool isLegalMaskedOp =
4228 isLegalMaskedStore(SI->getValueOperand()->getType(),
4229 SI->getPointerOperand());
4230 if (isLegalMaskedOp) {
4232 MaskedOp.insert(SI);
4241 // The instructions below can trap.
4242 switch (it->getOpcode()) {
4244 case Instruction::UDiv:
4245 case Instruction::SDiv:
4246 case Instruction::URem:
4247 case Instruction::SRem:
4255 LoopVectorizationCostModel::VectorizationFactor
4256 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
4257 // Width 1 means no vectorize
4258 VectorizationFactor Factor = { 1U, 0U };
4259 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4260 emitAnalysis(VectorizationReport() <<
4261 "runtime pointer checks needed. Enable vectorization of this "
4262 "loop with '#pragma clang loop vectorize(enable)' when "
4263 "compiling with -Os");
4264 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4268 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
4269 emitAnalysis(VectorizationReport() <<
4270 "store that is conditionally executed prevents vectorization");
4271 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
4275 // Find the trip count.
4276 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4277 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4279 unsigned WidestType = getWidestType();
4280 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4281 unsigned MaxSafeDepDist = -1U;
4282 if (Legal->getMaxSafeDepDistBytes() != -1U)
4283 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4284 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4285 WidestRegister : MaxSafeDepDist);
4286 unsigned MaxVectorSize = WidestRegister / WidestType;
4287 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4288 DEBUG(dbgs() << "LV: The Widest register is: "
4289 << WidestRegister << " bits.\n");
4291 if (MaxVectorSize == 0) {
4292 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4296 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
4297 " into one vector!");
4299 unsigned VF = MaxVectorSize;
4301 // If we optimize the program for size, avoid creating the tail loop.
4303 // If we are unable to calculate the trip count then don't try to vectorize.
4306 (VectorizationReport() <<
4307 "unable to calculate the loop count due to complex control flow");
4308 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4312 // Find the maximum SIMD width that can fit within the trip count.
4313 VF = TC % MaxVectorSize;
4318 // If the trip count that we found modulo the vectorization factor is not
4319 // zero then we require a tail.
4321 emitAnalysis(VectorizationReport() <<
4322 "cannot optimize for size and vectorize at the "
4323 "same time. Enable vectorization of this loop "
4324 "with '#pragma clang loop vectorize(enable)' "
4325 "when compiling with -Os");
4326 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4331 int UserVF = Hints->getWidth();
4333 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4334 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4336 Factor.Width = UserVF;
4340 float Cost = expectedCost(1);
4342 const float ScalarCost = Cost;
4345 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
4347 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
4348 // Ignore scalar width, because the user explicitly wants vectorization.
4349 if (ForceVectorization && VF > 1) {
4351 Cost = expectedCost(Width) / (float)Width;
4354 for (unsigned i=2; i <= VF; i*=2) {
4355 // Notice that the vector loop needs to be executed less times, so
4356 // we need to divide the cost of the vector loops by the width of
4357 // the vector elements.
4358 float VectorCost = expectedCost(i) / (float)i;
4359 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
4360 (int)VectorCost << ".\n");
4361 if (VectorCost < Cost) {
4367 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
4368 << "LV: Vectorization seems to be not beneficial, "
4369 << "but was forced by a user.\n");
4370 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
4371 Factor.Width = Width;
4372 Factor.Cost = Width * Cost;
4376 unsigned LoopVectorizationCostModel::getWidestType() {
4377 unsigned MaxWidth = 8;
4380 for (Loop::block_iterator bb = TheLoop->block_begin(),
4381 be = TheLoop->block_end(); bb != be; ++bb) {
4382 BasicBlock *BB = *bb;
4384 // For each instruction in the loop.
4385 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4386 Type *T = it->getType();
4388 // Ignore ephemeral values.
4389 if (EphValues.count(it))
4392 // Only examine Loads, Stores and PHINodes.
4393 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4396 // Examine PHI nodes that are reduction variables.
4397 if (PHINode *PN = dyn_cast<PHINode>(it))
4398 if (!Legal->getReductionVars()->count(PN))
4401 // Examine the stored values.
4402 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4403 T = ST->getValueOperand()->getType();
4405 // Ignore loaded pointer types and stored pointer types that are not
4406 // consecutive. However, we do want to take consecutive stores/loads of
4407 // pointer vectors into account.
4408 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4411 MaxWidth = std::max(MaxWidth,
4412 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
4420 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
4422 unsigned LoopCost) {
4424 // -- The unroll heuristics --
4425 // We unroll the loop in order to expose ILP and reduce the loop overhead.
4426 // There are many micro-architectural considerations that we can't predict
4427 // at this level. For example, frontend pressure (on decode or fetch) due to
4428 // code size, or the number and capabilities of the execution ports.
4430 // We use the following heuristics to select the unroll factor:
4431 // 1. If the code has reductions, then we unroll in order to break the cross
4432 // iteration dependency.
4433 // 2. If the loop is really small, then we unroll in order to reduce the loop
4435 // 3. We don't unroll if we think that we will spill registers to memory due
4436 // to the increased register pressure.
4438 // Use the user preference, unless 'auto' is selected.
4439 int UserUF = Hints->getInterleave();
4443 // When we optimize for size, we don't unroll.
4447 // We used the distance for the unroll factor.
4448 if (Legal->getMaxSafeDepDistBytes() != -1U)
4451 // Do not unroll loops with a relatively small trip count.
4452 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4453 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
4456 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
4457 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
4461 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
4462 TargetNumRegisters = ForceTargetNumScalarRegs;
4464 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
4465 TargetNumRegisters = ForceTargetNumVectorRegs;
4468 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4469 // We divide by these constants so assume that we have at least one
4470 // instruction that uses at least one register.
4471 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4472 R.NumInstructions = std::max(R.NumInstructions, 1U);
4474 // We calculate the unroll factor using the following formula.
4475 // Subtract the number of loop invariants from the number of available
4476 // registers. These registers are used by all of the unrolled instances.
4477 // Next, divide the remaining registers by the number of registers that is
4478 // required by the loop, in order to estimate how many parallel instances
4479 // fit without causing spills. All of this is rounded down if necessary to be
4480 // a power of two. We want power of two unroll factors to simplify any
4481 // addressing operations or alignment considerations.
4482 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
4485 // Don't count the induction variable as unrolled.
4486 if (EnableIndVarRegisterHeur)
4487 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
4488 std::max(1U, (R.MaxLocalUsers - 1)));
4490 // Clamp the unroll factor ranges to reasonable factors.
4491 unsigned MaxInterleaveSize = TTI.getMaxInterleaveFactor();
4493 // Check if the user has overridden the unroll max.
4495 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
4496 MaxInterleaveSize = ForceTargetMaxScalarInterleaveFactor;
4498 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
4499 MaxInterleaveSize = ForceTargetMaxVectorInterleaveFactor;
4502 // If we did not calculate the cost for VF (because the user selected the VF)
4503 // then we calculate the cost of VF here.
4505 LoopCost = expectedCost(VF);
4507 // Clamp the calculated UF to be between the 1 and the max unroll factor
4508 // that the target allows.
4509 if (UF > MaxInterleaveSize)
4510 UF = MaxInterleaveSize;
4514 // Unroll if we vectorized this loop and there is a reduction that could
4515 // benefit from unrolling.
4516 if (VF > 1 && Legal->getReductionVars()->size()) {
4517 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
4521 // Note that if we've already vectorized the loop we will have done the
4522 // runtime check and so unrolling won't require further checks.
4523 bool UnrollingRequiresRuntimePointerCheck =
4524 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
4526 // We want to unroll small loops in order to reduce the loop overhead and
4527 // potentially expose ILP opportunities.
4528 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
4529 if (!UnrollingRequiresRuntimePointerCheck &&
4530 LoopCost < SmallLoopCost) {
4531 // We assume that the cost overhead is 1 and we use the cost model
4532 // to estimate the cost of the loop and unroll until the cost of the
4533 // loop overhead is about 5% of the cost of the loop.
4534 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
4536 // Unroll until store/load ports (estimated by max unroll factor) are
4538 unsigned NumStores = Legal->getNumStores();
4539 unsigned NumLoads = Legal->getNumLoads();
4540 unsigned StoresUF = UF / (NumStores ? NumStores : 1);
4541 unsigned LoadsUF = UF / (NumLoads ? NumLoads : 1);
4543 // If we have a scalar reduction (vector reductions are already dealt with
4544 // by this point), we can increase the critical path length if the loop
4545 // we're unrolling is inside another loop. Limit, by default to 2, so the
4546 // critical path only gets increased by one reduction operation.
4547 if (Legal->getReductionVars()->size() &&
4548 TheLoop->getLoopDepth() > 1) {
4549 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionUF);
4550 SmallUF = std::min(SmallUF, F);
4551 StoresUF = std::min(StoresUF, F);
4552 LoadsUF = std::min(LoadsUF, F);
4555 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
4556 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
4557 return std::max(StoresUF, LoadsUF);
4560 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
4564 DEBUG(dbgs() << "LV: Not Unrolling.\n");
4568 LoopVectorizationCostModel::RegisterUsage
4569 LoopVectorizationCostModel::calculateRegisterUsage() {
4570 // This function calculates the register usage by measuring the highest number
4571 // of values that are alive at a single location. Obviously, this is a very
4572 // rough estimation. We scan the loop in a topological order in order and
4573 // assign a number to each instruction. We use RPO to ensure that defs are
4574 // met before their users. We assume that each instruction that has in-loop
4575 // users starts an interval. We record every time that an in-loop value is
4576 // used, so we have a list of the first and last occurrences of each
4577 // instruction. Next, we transpose this data structure into a multi map that
4578 // holds the list of intervals that *end* at a specific location. This multi
4579 // map allows us to perform a linear search. We scan the instructions linearly
4580 // and record each time that a new interval starts, by placing it in a set.
4581 // If we find this value in the multi-map then we remove it from the set.
4582 // The max register usage is the maximum size of the set.
4583 // We also search for instructions that are defined outside the loop, but are
4584 // used inside the loop. We need this number separately from the max-interval
4585 // usage number because when we unroll, loop-invariant values do not take
4587 LoopBlocksDFS DFS(TheLoop);
4591 R.NumInstructions = 0;
4593 // Each 'key' in the map opens a new interval. The values
4594 // of the map are the index of the 'last seen' usage of the
4595 // instruction that is the key.
4596 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4597 // Maps instruction to its index.
4598 DenseMap<unsigned, Instruction*> IdxToInstr;
4599 // Marks the end of each interval.
4600 IntervalMap EndPoint;
4601 // Saves the list of instruction indices that are used in the loop.
4602 SmallSet<Instruction*, 8> Ends;
4603 // Saves the list of values that are used in the loop but are
4604 // defined outside the loop, such as arguments and constants.
4605 SmallPtrSet<Value*, 8> LoopInvariants;
4608 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4609 be = DFS.endRPO(); bb != be; ++bb) {
4610 R.NumInstructions += (*bb)->size();
4611 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4613 Instruction *I = it;
4614 IdxToInstr[Index++] = I;
4616 // Save the end location of each USE.
4617 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4618 Value *U = I->getOperand(i);
4619 Instruction *Instr = dyn_cast<Instruction>(U);
4621 // Ignore non-instruction values such as arguments, constants, etc.
4622 if (!Instr) continue;
4624 // If this instruction is outside the loop then record it and continue.
4625 if (!TheLoop->contains(Instr)) {
4626 LoopInvariants.insert(Instr);
4630 // Overwrite previous end points.
4631 EndPoint[Instr] = Index;
4637 // Saves the list of intervals that end with the index in 'key'.
4638 typedef SmallVector<Instruction*, 2> InstrList;
4639 DenseMap<unsigned, InstrList> TransposeEnds;
4641 // Transpose the EndPoints to a list of values that end at each index.
4642 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4644 TransposeEnds[it->second].push_back(it->first);
4646 SmallSet<Instruction*, 8> OpenIntervals;
4647 unsigned MaxUsage = 0;
4650 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4651 for (unsigned int i = 0; i < Index; ++i) {
4652 Instruction *I = IdxToInstr[i];
4653 // Ignore instructions that are never used within the loop.
4654 if (!Ends.count(I)) continue;
4656 // Ignore ephemeral values.
4657 if (EphValues.count(I))
4660 // Remove all of the instructions that end at this location.
4661 InstrList &List = TransposeEnds[i];
4662 for (unsigned int j=0, e = List.size(); j < e; ++j)
4663 OpenIntervals.erase(List[j]);
4665 // Count the number of live interals.
4666 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4668 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4669 OpenIntervals.size() << '\n');
4671 // Add the current instruction to the list of open intervals.
4672 OpenIntervals.insert(I);
4675 unsigned Invariant = LoopInvariants.size();
4676 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
4677 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
4678 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
4680 R.LoopInvariantRegs = Invariant;
4681 R.MaxLocalUsers = MaxUsage;
4685 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4689 for (Loop::block_iterator bb = TheLoop->block_begin(),
4690 be = TheLoop->block_end(); bb != be; ++bb) {
4691 unsigned BlockCost = 0;
4692 BasicBlock *BB = *bb;
4694 // For each instruction in the old loop.
4695 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4696 // Skip dbg intrinsics.
4697 if (isa<DbgInfoIntrinsic>(it))
4700 // Ignore ephemeral values.
4701 if (EphValues.count(it))
4704 unsigned C = getInstructionCost(it, VF);
4706 // Check if we should override the cost.
4707 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
4708 C = ForceTargetInstructionCost;
4711 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
4712 VF << " For instruction: " << *it << '\n');
4715 // We assume that if-converted blocks have a 50% chance of being executed.
4716 // When the code is scalar then some of the blocks are avoided due to CF.
4717 // When the code is vectorized we execute all code paths.
4718 if (VF == 1 && Legal->blockNeedsPredication(*bb))
4727 /// \brief Check whether the address computation for a non-consecutive memory
4728 /// access looks like an unlikely candidate for being merged into the indexing
4731 /// We look for a GEP which has one index that is an induction variable and all
4732 /// other indices are loop invariant. If the stride of this access is also
4733 /// within a small bound we decide that this address computation can likely be
4734 /// merged into the addressing mode.
4735 /// In all other cases, we identify the address computation as complex.
4736 static bool isLikelyComplexAddressComputation(Value *Ptr,
4737 LoopVectorizationLegality *Legal,
4738 ScalarEvolution *SE,
4739 const Loop *TheLoop) {
4740 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
4744 // We are looking for a gep with all loop invariant indices except for one
4745 // which should be an induction variable.
4746 unsigned NumOperands = Gep->getNumOperands();
4747 for (unsigned i = 1; i < NumOperands; ++i) {
4748 Value *Opd = Gep->getOperand(i);
4749 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
4750 !Legal->isInductionVariable(Opd))
4754 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
4755 // can likely be merged into the address computation.
4756 unsigned MaxMergeDistance = 64;
4758 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
4762 // Check the step is constant.
4763 const SCEV *Step = AddRec->getStepRecurrence(*SE);
4764 // Calculate the pointer stride and check if it is consecutive.
4765 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4769 const APInt &APStepVal = C->getValue()->getValue();
4771 // Huge step value - give up.
4772 if (APStepVal.getBitWidth() > 64)
4775 int64_t StepVal = APStepVal.getSExtValue();
4777 return StepVal > MaxMergeDistance;
4780 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
4781 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
4787 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
4788 // If we know that this instruction will remain uniform, check the cost of
4789 // the scalar version.
4790 if (Legal->isUniformAfterVectorization(I))
4793 Type *RetTy = I->getType();
4794 Type *VectorTy = ToVectorTy(RetTy, VF);
4796 // TODO: We need to estimate the cost of intrinsic calls.
4797 switch (I->getOpcode()) {
4798 case Instruction::GetElementPtr:
4799 // We mark this instruction as zero-cost because the cost of GEPs in
4800 // vectorized code depends on whether the corresponding memory instruction
4801 // is scalarized or not. Therefore, we handle GEPs with the memory
4802 // instruction cost.
4804 case Instruction::Br: {
4805 return TTI.getCFInstrCost(I->getOpcode());
4807 case Instruction::PHI:
4808 //TODO: IF-converted IFs become selects.
4810 case Instruction::Add:
4811 case Instruction::FAdd:
4812 case Instruction::Sub:
4813 case Instruction::FSub:
4814 case Instruction::Mul:
4815 case Instruction::FMul:
4816 case Instruction::UDiv:
4817 case Instruction::SDiv:
4818 case Instruction::FDiv:
4819 case Instruction::URem:
4820 case Instruction::SRem:
4821 case Instruction::FRem:
4822 case Instruction::Shl:
4823 case Instruction::LShr:
4824 case Instruction::AShr:
4825 case Instruction::And:
4826 case Instruction::Or:
4827 case Instruction::Xor: {
4828 // Since we will replace the stride by 1 the multiplication should go away.
4829 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
4831 // Certain instructions can be cheaper to vectorize if they have a constant
4832 // second vector operand. One example of this are shifts on x86.
4833 TargetTransformInfo::OperandValueKind Op1VK =
4834 TargetTransformInfo::OK_AnyValue;
4835 TargetTransformInfo::OperandValueKind Op2VK =
4836 TargetTransformInfo::OK_AnyValue;
4837 TargetTransformInfo::OperandValueProperties Op1VP =
4838 TargetTransformInfo::OP_None;
4839 TargetTransformInfo::OperandValueProperties Op2VP =
4840 TargetTransformInfo::OP_None;
4841 Value *Op2 = I->getOperand(1);
4843 // Check for a splat of a constant or for a non uniform vector of constants.
4844 if (isa<ConstantInt>(Op2)) {
4845 ConstantInt *CInt = cast<ConstantInt>(Op2);
4846 if (CInt && CInt->getValue().isPowerOf2())
4847 Op2VP = TargetTransformInfo::OP_PowerOf2;
4848 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4849 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
4850 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
4851 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
4853 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
4854 if (CInt && CInt->getValue().isPowerOf2())
4855 Op2VP = TargetTransformInfo::OP_PowerOf2;
4856 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4860 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
4863 case Instruction::Select: {
4864 SelectInst *SI = cast<SelectInst>(I);
4865 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
4866 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
4867 Type *CondTy = SI->getCondition()->getType();
4869 CondTy = VectorType::get(CondTy, VF);
4871 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
4873 case Instruction::ICmp:
4874 case Instruction::FCmp: {
4875 Type *ValTy = I->getOperand(0)->getType();
4876 VectorTy = ToVectorTy(ValTy, VF);
4877 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
4879 case Instruction::Store:
4880 case Instruction::Load: {
4881 StoreInst *SI = dyn_cast<StoreInst>(I);
4882 LoadInst *LI = dyn_cast<LoadInst>(I);
4883 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
4885 VectorTy = ToVectorTy(ValTy, VF);
4887 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
4888 unsigned AS = SI ? SI->getPointerAddressSpace() :
4889 LI->getPointerAddressSpace();
4890 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
4891 // We add the cost of address computation here instead of with the gep
4892 // instruction because only here we know whether the operation is
4895 return TTI.getAddressComputationCost(VectorTy) +
4896 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4898 // Scalarized loads/stores.
4899 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
4900 bool Reverse = ConsecutiveStride < 0;
4901 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
4902 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
4903 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
4904 bool IsComplexComputation =
4905 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
4907 // The cost of extracting from the value vector and pointer vector.
4908 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
4909 for (unsigned i = 0; i < VF; ++i) {
4910 // The cost of extracting the pointer operand.
4911 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
4912 // In case of STORE, the cost of ExtractElement from the vector.
4913 // In case of LOAD, the cost of InsertElement into the returned
4915 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
4916 Instruction::InsertElement,
4920 // The cost of the scalar loads/stores.
4921 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
4922 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
4927 // Wide load/stores.
4928 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
4929 if (Legal->isMaskRequired(I))
4930 Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment,
4933 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4936 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
4940 case Instruction::ZExt:
4941 case Instruction::SExt:
4942 case Instruction::FPToUI:
4943 case Instruction::FPToSI:
4944 case Instruction::FPExt:
4945 case Instruction::PtrToInt:
4946 case Instruction::IntToPtr:
4947 case Instruction::SIToFP:
4948 case Instruction::UIToFP:
4949 case Instruction::Trunc:
4950 case Instruction::FPTrunc:
4951 case Instruction::BitCast: {
4952 // We optimize the truncation of induction variable.
4953 // The cost of these is the same as the scalar operation.
4954 if (I->getOpcode() == Instruction::Trunc &&
4955 Legal->isInductionVariable(I->getOperand(0)))
4956 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
4957 I->getOperand(0)->getType());
4959 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
4960 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
4962 case Instruction::Call: {
4963 CallInst *CI = cast<CallInst>(I);
4964 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
4965 assert(ID && "Not an intrinsic call!");
4966 Type *RetTy = ToVectorTy(CI->getType(), VF);
4967 SmallVector<Type*, 4> Tys;
4968 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
4969 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
4970 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
4973 // We are scalarizing the instruction. Return the cost of the scalar
4974 // instruction, plus the cost of insert and extract into vector
4975 // elements, times the vector width.
4978 if (!RetTy->isVoidTy() && VF != 1) {
4979 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
4981 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
4984 // The cost of inserting the results plus extracting each one of the
4986 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
4989 // The cost of executing VF copies of the scalar instruction. This opcode
4990 // is unknown. Assume that it is the same as 'mul'.
4991 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
4997 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
4998 if (Scalar->isVoidTy() || VF == 1)
5000 return VectorType::get(Scalar, VF);
5003 char LoopVectorize::ID = 0;
5004 static const char lv_name[] = "Loop Vectorization";
5005 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5006 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
5007 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
5008 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
5009 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
5010 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5011 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5012 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5013 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
5014 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5015 INITIALIZE_PASS_DEPENDENCY(LoopAccessAnalysis)
5016 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5019 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5020 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5024 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5025 // Check for a store.
5026 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5027 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5029 // Check for a load.
5030 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5031 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5037 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5038 bool IfPredicateStore) {
5039 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5040 // Holds vector parameters or scalars, in case of uniform vals.
5041 SmallVector<VectorParts, 4> Params;
5043 setDebugLocFromInst(Builder, Instr);
5045 // Find all of the vectorized parameters.
5046 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5047 Value *SrcOp = Instr->getOperand(op);
5049 // If we are accessing the old induction variable, use the new one.
5050 if (SrcOp == OldInduction) {
5051 Params.push_back(getVectorValue(SrcOp));
5055 // Try using previously calculated values.
5056 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5058 // If the src is an instruction that appeared earlier in the basic block
5059 // then it should already be vectorized.
5060 if (SrcInst && OrigLoop->contains(SrcInst)) {
5061 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5062 // The parameter is a vector value from earlier.
5063 Params.push_back(WidenMap.get(SrcInst));
5065 // The parameter is a scalar from outside the loop. Maybe even a constant.
5066 VectorParts Scalars;
5067 Scalars.append(UF, SrcOp);
5068 Params.push_back(Scalars);
5072 assert(Params.size() == Instr->getNumOperands() &&
5073 "Invalid number of operands");
5075 // Does this instruction return a value ?
5076 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5078 Value *UndefVec = IsVoidRetTy ? nullptr :
5079 UndefValue::get(Instr->getType());
5080 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5081 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5083 Instruction *InsertPt = Builder.GetInsertPoint();
5084 BasicBlock *IfBlock = Builder.GetInsertBlock();
5085 BasicBlock *CondBlock = nullptr;
5088 Loop *VectorLp = nullptr;
5089 if (IfPredicateStore) {
5090 assert(Instr->getParent()->getSinglePredecessor() &&
5091 "Only support single predecessor blocks");
5092 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5093 Instr->getParent());
5094 VectorLp = LI->getLoopFor(IfBlock);
5095 assert(VectorLp && "Must have a loop for this block");
5098 // For each vector unroll 'part':
5099 for (unsigned Part = 0; Part < UF; ++Part) {
5100 // For each scalar that we create:
5102 // Start an "if (pred) a[i] = ..." block.
5103 Value *Cmp = nullptr;
5104 if (IfPredicateStore) {
5105 if (Cond[Part]->getType()->isVectorTy())
5107 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5108 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5109 ConstantInt::get(Cond[Part]->getType(), 1));
5110 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5111 LoopVectorBody.push_back(CondBlock);
5112 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
5113 // Update Builder with newly created basic block.
5114 Builder.SetInsertPoint(InsertPt);
5117 Instruction *Cloned = Instr->clone();
5119 Cloned->setName(Instr->getName() + ".cloned");
5120 // Replace the operands of the cloned instructions with extracted scalars.
5121 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5122 Value *Op = Params[op][Part];
5123 Cloned->setOperand(op, Op);
5126 // Place the cloned scalar in the new loop.
5127 Builder.Insert(Cloned);
5129 // If the original scalar returns a value we need to place it in a vector
5130 // so that future users will be able to use it.
5132 VecResults[Part] = Cloned;
5135 if (IfPredicateStore) {
5136 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
5137 LoopVectorBody.push_back(NewIfBlock);
5138 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
5139 Builder.SetInsertPoint(InsertPt);
5140 Instruction *OldBr = IfBlock->getTerminator();
5141 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
5142 OldBr->eraseFromParent();
5143 IfBlock = NewIfBlock;
5148 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5149 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5150 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5152 return scalarizeInstruction(Instr, IfPredicateStore);
5155 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5159 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5163 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) {
5164 // When unrolling and the VF is 1, we only need to add a simple scalar.
5165 Type *ITy = Val->getType();
5166 assert(!ITy->isVectorTy() && "Val must be a scalar");
5167 Constant *C = ConstantInt::get(ITy, StartIdx);
5168 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");