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");
109 static cl::opt<unsigned, true>
110 VectorizationFactor("force-vector-width", cl::Hidden,
111 cl::desc("Sets the SIMD width. Zero is autoselect."),
112 cl::location(VectorizerParams::VectorizationFactor));
113 unsigned VectorizerParams::VectorizationFactor = 0;
115 static cl::opt<unsigned, true>
116 VectorizationInterleave("force-vector-interleave", cl::Hidden,
117 cl::desc("Sets the vectorization interleave count. "
118 "Zero is autoselect."),
120 VectorizerParams::VectorizationInterleave));
121 unsigned VectorizerParams::VectorizationInterleave = 0;
124 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
125 cl::desc("Enable if-conversion during vectorization."));
127 /// We don't vectorize loops with a known constant trip count below this number.
128 static cl::opt<unsigned>
129 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
131 cl::desc("Don't vectorize loops with a constant "
132 "trip count that is smaller than this "
135 /// This enables versioning on the strides of symbolically striding memory
136 /// accesses in code like the following.
137 /// for (i = 0; i < N; ++i)
138 /// A[i * Stride1] += B[i * Stride2] ...
140 /// Will be roughly translated to
141 /// if (Stride1 == 1 && Stride2 == 1) {
142 /// for (i = 0; i < N; i+=4)
146 static cl::opt<bool> EnableMemAccessVersioning(
147 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
148 cl::desc("Enable symblic stride memory access versioning"));
150 /// We don't unroll loops with a known constant trip count below this number.
151 static const unsigned TinyTripCountUnrollThreshold = 128;
153 /// When performing memory disambiguation checks at runtime do not make more
154 /// than this number of comparisons.
155 const unsigned VectorizerParams::RuntimeMemoryCheckThreshold = 8;
157 /// Maximum simd width.
158 const unsigned VectorizerParams::MaxVectorWidth = 64;
160 static cl::opt<unsigned> ForceTargetNumScalarRegs(
161 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
162 cl::desc("A flag that overrides the target's number of scalar registers."));
164 static cl::opt<unsigned> ForceTargetNumVectorRegs(
165 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
166 cl::desc("A flag that overrides the target's number of vector registers."));
168 /// Maximum vectorization interleave count.
169 static const unsigned MaxInterleaveFactor = 16;
171 static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
172 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
173 cl::desc("A flag that overrides the target's max interleave factor for "
176 static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
177 "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
178 cl::desc("A flag that overrides the target's max interleave factor for "
179 "vectorized loops."));
181 static cl::opt<unsigned> ForceTargetInstructionCost(
182 "force-target-instruction-cost", cl::init(0), cl::Hidden,
183 cl::desc("A flag that overrides the target's expected cost for "
184 "an instruction to a single constant value. Mostly "
185 "useful for getting consistent testing."));
187 static cl::opt<unsigned> SmallLoopCost(
188 "small-loop-cost", cl::init(20), cl::Hidden,
189 cl::desc("The cost of a loop that is considered 'small' by the unroller."));
191 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
192 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
193 cl::desc("Enable the use of the block frequency analysis to access PGO "
194 "heuristics minimizing code growth in cold regions and being more "
195 "aggressive in hot regions."));
197 // Runtime unroll loops for load/store throughput.
198 static cl::opt<bool> EnableLoadStoreRuntimeUnroll(
199 "enable-loadstore-runtime-unroll", cl::init(true), cl::Hidden,
200 cl::desc("Enable runtime unrolling until load/store ports are saturated"));
202 /// The number of stores in a loop that are allowed to need predication.
203 static cl::opt<unsigned> NumberOfStoresToPredicate(
204 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
205 cl::desc("Max number of stores to be predicated behind an if."));
207 static cl::opt<bool> EnableIndVarRegisterHeur(
208 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
209 cl::desc("Count the induction variable only once when unrolling"));
211 static cl::opt<bool> EnableCondStoresVectorization(
212 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
213 cl::desc("Enable if predication of stores during vectorization."));
215 static cl::opt<unsigned> MaxNestedScalarReductionUF(
216 "max-nested-scalar-reduction-unroll", cl::init(2), cl::Hidden,
217 cl::desc("The maximum unroll factor to use when unrolling a scalar "
218 "reduction in a nested loop."));
222 // Forward declarations.
223 class LoopVectorizationLegality;
224 class LoopVectorizationCostModel;
225 class LoopVectorizeHints;
227 /// \brief This modifies LoopAccessReport to initialize message with
228 /// loop-vectorizer-specific part.
229 class VectorizationReport : public LoopAccessReport {
231 VectorizationReport(Instruction *I = nullptr)
232 : LoopAccessReport("loop not vectorized: ", I) {}
234 /// \brief This allows promotion of the loop-access analysis report into the
235 /// loop-vectorizer report. It modifies the message to add the
236 /// loop-vectorizer-specific part of the message.
237 explicit VectorizationReport(const LoopAccessReport &R)
238 : LoopAccessReport(Twine("loop not vectorized: ") + R.str(),
242 /// InnerLoopVectorizer vectorizes loops which contain only one basic
243 /// block to a specified vectorization factor (VF).
244 /// This class performs the widening of scalars into vectors, or multiple
245 /// scalars. This class also implements the following features:
246 /// * It inserts an epilogue loop for handling loops that don't have iteration
247 /// counts that are known to be a multiple of the vectorization factor.
248 /// * It handles the code generation for reduction variables.
249 /// * Scalarization (implementation using scalars) of un-vectorizable
251 /// InnerLoopVectorizer does not perform any vectorization-legality
252 /// checks, and relies on the caller to check for the different legality
253 /// aspects. The InnerLoopVectorizer relies on the
254 /// LoopVectorizationLegality class to provide information about the induction
255 /// and reduction variables that were found to a given vectorization factor.
256 class InnerLoopVectorizer {
258 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
259 DominatorTree *DT, const DataLayout *DL,
260 const TargetLibraryInfo *TLI, unsigned VecWidth,
261 unsigned UnrollFactor)
262 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
263 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
264 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
267 // Perform the actual loop widening (vectorization).
268 void vectorize(LoopVectorizationLegality *L) {
270 // Create a new empty loop. Unlink the old loop and connect the new one.
272 // Widen each instruction in the old loop to a new one in the new loop.
273 // Use the Legality module to find the induction and reduction variables.
275 // Register the new loop and update the analysis passes.
279 virtual ~InnerLoopVectorizer() {}
282 /// A small list of PHINodes.
283 typedef SmallVector<PHINode*, 4> PhiVector;
284 /// When we unroll loops we have multiple vector values for each scalar.
285 /// This data structure holds the unrolled and vectorized values that
286 /// originated from one scalar instruction.
287 typedef SmallVector<Value*, 2> VectorParts;
289 // When we if-convert we need create edge masks. We have to cache values so
290 // that we don't end up with exponential recursion/IR.
291 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
292 VectorParts> EdgeMaskCache;
294 /// \brief Add checks for strides that where assumed to be 1.
296 /// Returns the last check instruction and the first check instruction in the
297 /// pair as (first, last).
298 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
300 /// Create an empty loop, based on the loop ranges of the old loop.
301 void createEmptyLoop();
302 /// Copy and widen the instructions from the old loop.
303 virtual void vectorizeLoop();
305 /// \brief The Loop exit block may have single value PHI nodes where the
306 /// incoming value is 'Undef'. While vectorizing we only handled real values
307 /// that were defined inside the loop. Here we fix the 'undef case'.
311 /// A helper function that computes the predicate of the block BB, assuming
312 /// that the header block of the loop is set to True. It returns the *entry*
313 /// mask for the block BB.
314 VectorParts createBlockInMask(BasicBlock *BB);
315 /// A helper function that computes the predicate of the edge between SRC
317 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
319 /// A helper function to vectorize a single BB within the innermost loop.
320 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
322 /// Vectorize a single PHINode in a block. This method handles the induction
323 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
324 /// arbitrary length vectors.
325 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
326 unsigned UF, unsigned VF, PhiVector *PV);
328 /// Insert the new loop to the loop hierarchy and pass manager
329 /// and update the analysis passes.
330 void updateAnalysis();
332 /// This instruction is un-vectorizable. Implement it as a sequence
333 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
334 /// scalarized instruction behind an if block predicated on the control
335 /// dependence of the instruction.
336 virtual void scalarizeInstruction(Instruction *Instr,
337 bool IfPredicateStore=false);
339 /// Vectorize Load and Store instructions,
340 virtual void vectorizeMemoryInstruction(Instruction *Instr);
342 /// Create a broadcast instruction. This method generates a broadcast
343 /// instruction (shuffle) for loop invariant values and for the induction
344 /// value. If this is the induction variable then we extend it to N, N+1, ...
345 /// this is needed because each iteration in the loop corresponds to a SIMD
347 virtual Value *getBroadcastInstrs(Value *V);
349 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
350 /// to each vector element of Val. The sequence starts at StartIndex.
351 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step);
353 /// When we go over instructions in the basic block we rely on previous
354 /// values within the current basic block or on loop invariant values.
355 /// When we widen (vectorize) values we place them in the map. If the values
356 /// are not within the map, they have to be loop invariant, so we simply
357 /// broadcast them into a vector.
358 VectorParts &getVectorValue(Value *V);
360 /// Generate a shuffle sequence that will reverse the vector Vec.
361 virtual Value *reverseVector(Value *Vec);
363 /// This is a helper class that holds the vectorizer state. It maps scalar
364 /// instructions to vector instructions. When the code is 'unrolled' then
365 /// then a single scalar value is mapped to multiple vector parts. The parts
366 /// are stored in the VectorPart type.
368 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
370 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
372 /// \return True if 'Key' is saved in the Value Map.
373 bool has(Value *Key) const { return MapStorage.count(Key); }
375 /// Initializes a new entry in the map. Sets all of the vector parts to the
376 /// save value in 'Val'.
377 /// \return A reference to a vector with splat values.
378 VectorParts &splat(Value *Key, Value *Val) {
379 VectorParts &Entry = MapStorage[Key];
380 Entry.assign(UF, Val);
384 ///\return A reference to the value that is stored at 'Key'.
385 VectorParts &get(Value *Key) {
386 VectorParts &Entry = MapStorage[Key];
389 assert(Entry.size() == UF);
394 /// The unroll factor. Each entry in the map stores this number of vector
398 /// Map storage. We use std::map and not DenseMap because insertions to a
399 /// dense map invalidates its iterators.
400 std::map<Value *, VectorParts> MapStorage;
403 /// The original loop.
405 /// Scev analysis to use.
414 const DataLayout *DL;
415 /// Target Library Info.
416 const TargetLibraryInfo *TLI;
418 /// The vectorization SIMD factor to use. Each vector will have this many
423 /// The vectorization unroll factor to use. Each scalar is vectorized to this
424 /// many different vector instructions.
427 /// The builder that we use
430 // --- Vectorization state ---
432 /// The vector-loop preheader.
433 BasicBlock *LoopVectorPreHeader;
434 /// The scalar-loop preheader.
435 BasicBlock *LoopScalarPreHeader;
436 /// Middle Block between the vector and the scalar.
437 BasicBlock *LoopMiddleBlock;
438 ///The ExitBlock of the scalar loop.
439 BasicBlock *LoopExitBlock;
440 ///The vector loop body.
441 SmallVector<BasicBlock *, 4> LoopVectorBody;
442 ///The scalar loop body.
443 BasicBlock *LoopScalarBody;
444 /// A list of all bypass blocks. The first block is the entry of the loop.
445 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
447 /// The new Induction variable which was added to the new block.
449 /// The induction variable of the old basic block.
450 PHINode *OldInduction;
451 /// Holds the extended (to the widest induction type) start index.
453 /// Maps scalars to widened vectors.
455 EdgeMaskCache MaskCache;
457 LoopVectorizationLegality *Legal;
460 class InnerLoopUnroller : public InnerLoopVectorizer {
462 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
463 DominatorTree *DT, const DataLayout *DL,
464 const TargetLibraryInfo *TLI, unsigned UnrollFactor) :
465 InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { }
468 void scalarizeInstruction(Instruction *Instr,
469 bool IfPredicateStore = false) override;
470 void vectorizeMemoryInstruction(Instruction *Instr) override;
471 Value *getBroadcastInstrs(Value *V) override;
472 Value *getStepVector(Value *Val, int StartIdx, Value *Step) override;
473 Value *reverseVector(Value *Vec) override;
476 /// \brief Look for a meaningful debug location on the instruction or it's
478 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
483 if (I->getDebugLoc() != Empty)
486 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
487 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
488 if (OpInst->getDebugLoc() != Empty)
495 /// \brief Set the debug location in the builder using the debug location in the
497 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
498 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
499 B.SetCurrentDebugLocation(Inst->getDebugLoc());
501 B.SetCurrentDebugLocation(DebugLoc());
505 /// \return string containing a file name and a line # for the given loop.
506 static std::string getDebugLocString(const Loop *L) {
509 raw_string_ostream OS(Result);
510 const DebugLoc LoopDbgLoc = L->getStartLoc();
511 if (!LoopDbgLoc.isUnknown())
512 LoopDbgLoc.print(L->getHeader()->getContext(), OS);
514 // Just print the module name.
515 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
522 /// \brief Propagate known metadata from one instruction to another.
523 static void propagateMetadata(Instruction *To, const Instruction *From) {
524 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
525 From->getAllMetadataOtherThanDebugLoc(Metadata);
527 for (auto M : Metadata) {
528 unsigned Kind = M.first;
530 // These are safe to transfer (this is safe for TBAA, even when we
531 // if-convert, because should that metadata have had a control dependency
532 // on the condition, and thus actually aliased with some other
533 // non-speculated memory access when the condition was false, this would be
534 // caught by the runtime overlap checks).
535 if (Kind != LLVMContext::MD_tbaa &&
536 Kind != LLVMContext::MD_alias_scope &&
537 Kind != LLVMContext::MD_noalias &&
538 Kind != LLVMContext::MD_fpmath)
541 To->setMetadata(Kind, M.second);
545 /// \brief Propagate known metadata from one instruction to a vector of others.
546 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
548 if (Instruction *I = dyn_cast<Instruction>(V))
549 propagateMetadata(I, From);
552 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
553 /// to what vectorization factor.
554 /// This class does not look at the profitability of vectorization, only the
555 /// legality. This class has two main kinds of checks:
556 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
557 /// will change the order of memory accesses in a way that will change the
558 /// correctness of the program.
559 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
560 /// checks for a number of different conditions, such as the availability of a
561 /// single induction variable, that all types are supported and vectorize-able,
562 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
563 /// This class is also used by InnerLoopVectorizer for identifying
564 /// induction variable and the different reduction variables.
565 class LoopVectorizationLegality {
567 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, const DataLayout *DL,
568 DominatorTree *DT, TargetLibraryInfo *TLI,
569 AliasAnalysis *AA, Function *F,
570 const TargetTransformInfo *TTI,
571 LoopAccessAnalysis *LAA)
572 : NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
573 TLI(TLI), TheFunction(F), TTI(TTI), DT(DT), LAA(LAA), LAI(nullptr),
574 Induction(nullptr), WidestIndTy(nullptr), HasFunNoNaNAttr(false) {}
576 /// This enum represents the kinds of reductions that we support.
578 RK_NoReduction, ///< Not a reduction.
579 RK_IntegerAdd, ///< Sum of integers.
580 RK_IntegerMult, ///< Product of integers.
581 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
582 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
583 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
584 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
585 RK_FloatAdd, ///< Sum of floats.
586 RK_FloatMult, ///< Product of floats.
587 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
590 /// This enum represents the kinds of inductions that we support.
592 IK_NoInduction, ///< Not an induction variable.
593 IK_IntInduction, ///< Integer induction variable. Step = C.
594 IK_PtrInduction ///< Pointer induction var. Step = C / sizeof(elem).
597 // This enum represents the kind of minmax reduction.
598 enum MinMaxReductionKind {
608 /// This struct holds information about reduction variables.
609 struct ReductionDescriptor {
610 ReductionDescriptor() : StartValue(nullptr), LoopExitInstr(nullptr),
611 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
613 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
614 MinMaxReductionKind MK)
615 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
617 // The starting value of the reduction.
618 // It does not have to be zero!
619 TrackingVH<Value> StartValue;
620 // The instruction who's value is used outside the loop.
621 Instruction *LoopExitInstr;
622 // The kind of the reduction.
624 // If this a min/max reduction the kind of reduction.
625 MinMaxReductionKind MinMaxKind;
628 /// This POD struct holds information about a potential reduction operation.
629 struct ReductionInstDesc {
630 ReductionInstDesc(bool IsRedux, Instruction *I) :
631 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
633 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
634 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
636 // Is this instruction a reduction candidate.
638 // The last instruction in a min/max pattern (select of the select(icmp())
639 // pattern), or the current reduction instruction otherwise.
640 Instruction *PatternLastInst;
641 // If this is a min/max pattern the comparison predicate.
642 MinMaxReductionKind MinMaxKind;
645 /// A struct for saving information about induction variables.
646 struct InductionInfo {
647 InductionInfo(Value *Start, InductionKind K, ConstantInt *Step)
648 : StartValue(Start), IK(K), StepValue(Step) {
649 assert(IK != IK_NoInduction && "Not an induction");
650 assert(StartValue && "StartValue is null");
651 assert(StepValue && !StepValue->isZero() && "StepValue is zero");
652 assert((IK != IK_PtrInduction || StartValue->getType()->isPointerTy()) &&
653 "StartValue is not a pointer for pointer induction");
654 assert((IK != IK_IntInduction || StartValue->getType()->isIntegerTy()) &&
655 "StartValue is not an integer for integer induction");
656 assert(StepValue->getType()->isIntegerTy() &&
657 "StepValue is not an integer");
660 : StartValue(nullptr), IK(IK_NoInduction), StepValue(nullptr) {}
662 /// Get the consecutive direction. Returns:
663 /// 0 - unknown or non-consecutive.
664 /// 1 - consecutive and increasing.
665 /// -1 - consecutive and decreasing.
666 int getConsecutiveDirection() const {
667 if (StepValue && (StepValue->isOne() || StepValue->isMinusOne()))
668 return StepValue->getSExtValue();
672 /// Compute the transformed value of Index at offset StartValue using step
674 /// For integer induction, returns StartValue + Index * StepValue.
675 /// For pointer induction, returns StartValue[Index * StepValue].
676 /// FIXME: The newly created binary instructions should contain nsw/nuw
677 /// flags, which can be found from the original scalar operations.
678 Value *transform(IRBuilder<> &B, Value *Index) const {
680 case IK_IntInduction:
681 assert(Index->getType() == StartValue->getType() &&
682 "Index type does not match StartValue type");
683 if (StepValue->isMinusOne())
684 return B.CreateSub(StartValue, Index);
685 if (!StepValue->isOne())
686 Index = B.CreateMul(Index, StepValue);
687 return B.CreateAdd(StartValue, Index);
689 case IK_PtrInduction:
690 if (StepValue->isMinusOne())
691 Index = B.CreateNeg(Index);
692 else if (!StepValue->isOne())
693 Index = B.CreateMul(Index, StepValue);
694 return B.CreateGEP(StartValue, Index);
699 llvm_unreachable("invalid enum");
703 TrackingVH<Value> StartValue;
707 ConstantInt *StepValue;
710 /// ReductionList contains the reduction descriptors for all
711 /// of the reductions that were found in the loop.
712 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
714 /// InductionList saves induction variables and maps them to the
715 /// induction descriptor.
716 typedef MapVector<PHINode*, InductionInfo> InductionList;
718 /// Returns true if it is legal to vectorize this loop.
719 /// This does not mean that it is profitable to vectorize this
720 /// loop, only that it is legal to do so.
723 /// Returns the Induction variable.
724 PHINode *getInduction() { return Induction; }
726 /// Returns the reduction variables found in the loop.
727 ReductionList *getReductionVars() { return &Reductions; }
729 /// Returns the induction variables found in the loop.
730 InductionList *getInductionVars() { return &Inductions; }
732 /// Returns the widest induction type.
733 Type *getWidestInductionType() { return WidestIndTy; }
735 /// Returns True if V is an induction variable in this loop.
736 bool isInductionVariable(const Value *V);
738 /// Return true if the block BB needs to be predicated in order for the loop
739 /// to be vectorized.
740 bool blockNeedsPredication(BasicBlock *BB);
742 /// Check if this pointer is consecutive when vectorizing. This happens
743 /// when the last index of the GEP is the induction variable, or that the
744 /// pointer itself is an induction variable.
745 /// This check allows us to vectorize A[idx] into a wide load/store.
747 /// 0 - Stride is unknown or non-consecutive.
748 /// 1 - Address is consecutive.
749 /// -1 - Address is consecutive, and decreasing.
750 int isConsecutivePtr(Value *Ptr);
752 /// Returns true if the value V is uniform within the loop.
753 bool isUniform(Value *V);
755 /// Returns true if this instruction will remain scalar after vectorization.
756 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
758 /// Returns the information that we collected about runtime memory check.
759 const LoopAccessInfo::RuntimePointerCheck *getRuntimePointerCheck() const {
760 return LAI->getRuntimePointerCheck();
763 const LoopAccessInfo *getLAI() const {
767 /// This function returns the identity element (or neutral element) for
769 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
771 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
773 bool hasStride(Value *V) { return StrideSet.count(V); }
774 bool mustCheckStrides() { return !StrideSet.empty(); }
775 SmallPtrSet<Value *, 8>::iterator strides_begin() {
776 return StrideSet.begin();
778 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
780 /// Returns true if the target machine supports masked store operation
781 /// for the given \p DataType and kind of access to \p Ptr.
782 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
783 return TTI->isLegalMaskedStore(DataType, isConsecutivePtr(Ptr));
785 /// Returns true if the target machine supports masked load operation
786 /// for the given \p DataType and kind of access to \p Ptr.
787 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
788 return TTI->isLegalMaskedLoad(DataType, isConsecutivePtr(Ptr));
790 /// Returns true if vector representation of the instruction \p I
792 bool isMaskRequired(const Instruction* I) {
793 return (MaskedOp.count(I) != 0);
795 unsigned getNumStores() const {
796 return LAI->getNumStores();
798 unsigned getNumLoads() const {
799 return LAI->getNumLoads();
801 unsigned getNumPredStores() const {
802 return NumPredStores;
805 /// Check if a single basic block loop is vectorizable.
806 /// At this point we know that this is a loop with a constant trip count
807 /// and we only need to check individual instructions.
808 bool canVectorizeInstrs();
810 /// When we vectorize loops we may change the order in which
811 /// we read and write from memory. This method checks if it is
812 /// legal to vectorize the code, considering only memory constrains.
813 /// Returns true if the loop is vectorizable
814 bool canVectorizeMemory();
816 /// Return true if we can vectorize this loop using the IF-conversion
818 bool canVectorizeWithIfConvert();
820 /// Collect the variables that need to stay uniform after vectorization.
821 void collectLoopUniforms();
823 /// Return true if all of the instructions in the block can be speculatively
824 /// executed. \p SafePtrs is a list of addresses that are known to be legal
825 /// and we know that we can read from them without segfault.
826 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
828 /// Returns True, if 'Phi' is the kind of reduction variable for type
829 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
830 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
831 /// Returns a struct describing if the instruction 'I' can be a reduction
832 /// variable of type 'Kind'. If the reduction is a min/max pattern of
833 /// select(icmp()) this function advances the instruction pointer 'I' from the
834 /// compare instruction to the select instruction and stores this pointer in
835 /// 'PatternLastInst' member of the returned struct.
836 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
837 ReductionInstDesc &Desc);
838 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
839 /// pattern corresponding to a min(X, Y) or max(X, Y).
840 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
841 ReductionInstDesc &Prev);
842 /// Returns the induction kind of Phi and record the step. This function may
843 /// return NoInduction if the PHI is not an induction variable.
844 InductionKind isInductionVariable(PHINode *Phi, ConstantInt *&StepValue);
846 /// \brief Collect memory access with loop invariant strides.
848 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
850 void collectStridedAccess(Value *LoadOrStoreInst);
852 /// Report an analysis message to assist the user in diagnosing loops that are
853 /// not vectorized. These are handled as LoopAccessReport rather than
854 /// VectorizationReport because the << operator of VectorizationReport returns
855 /// LoopAccessReport.
856 void emitAnalysis(const LoopAccessReport &Message) {
857 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, LV_NAME);
860 unsigned NumPredStores;
862 /// The loop that we evaluate.
866 /// DataLayout analysis.
867 const DataLayout *DL;
868 /// Target Library Info.
869 TargetLibraryInfo *TLI;
871 Function *TheFunction;
872 /// Target Transform Info
873 const TargetTransformInfo *TTI;
876 // LoopAccess analysis.
877 LoopAccessAnalysis *LAA;
878 // And the loop-accesses info corresponding to this loop. This pointer is
879 // null until canVectorizeMemory sets it up.
880 const LoopAccessInfo *LAI;
882 // --- vectorization state --- //
884 /// Holds the integer induction variable. This is the counter of the
887 /// Holds the reduction variables.
888 ReductionList Reductions;
889 /// Holds all of the induction variables that we found in the loop.
890 /// Notice that inductions don't need to start at zero and that induction
891 /// variables can be pointers.
892 InductionList Inductions;
893 /// Holds the widest induction type encountered.
896 /// Allowed outside users. This holds the reduction
897 /// vars which can be accessed from outside the loop.
898 SmallPtrSet<Value*, 4> AllowedExit;
899 /// This set holds the variables which are known to be uniform after
901 SmallPtrSet<Instruction*, 4> Uniforms;
903 /// Can we assume the absence of NaNs.
904 bool HasFunNoNaNAttr;
906 ValueToValueMap Strides;
907 SmallPtrSet<Value *, 8> StrideSet;
909 /// While vectorizing these instructions we have to generate a
910 /// call to the appropriate masked intrinsic
911 SmallPtrSet<const Instruction*, 8> MaskedOp;
914 /// LoopVectorizationCostModel - estimates the expected speedups due to
916 /// In many cases vectorization is not profitable. This can happen because of
917 /// a number of reasons. In this class we mainly attempt to predict the
918 /// expected speedup/slowdowns due to the supported instruction set. We use the
919 /// TargetTransformInfo to query the different backends for the cost of
920 /// different operations.
921 class LoopVectorizationCostModel {
923 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
924 LoopVectorizationLegality *Legal,
925 const TargetTransformInfo &TTI,
926 const DataLayout *DL, const TargetLibraryInfo *TLI,
927 AssumptionCache *AC, const Function *F,
928 const LoopVectorizeHints *Hints)
929 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI),
930 TheFunction(F), Hints(Hints) {
931 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
934 /// Information about vectorization costs
935 struct VectorizationFactor {
936 unsigned Width; // Vector width with best cost
937 unsigned Cost; // Cost of the loop with that width
939 /// \return The most profitable vectorization factor and the cost of that VF.
940 /// This method checks every power of two up to VF. If UserVF is not ZERO
941 /// then this vectorization factor will be selected if vectorization is
943 VectorizationFactor selectVectorizationFactor(bool OptForSize);
945 /// \return The size (in bits) of the widest type in the code that
946 /// needs to be vectorized. We ignore values that remain scalar such as
947 /// 64 bit loop indices.
948 unsigned getWidestType();
950 /// \return The most profitable unroll factor.
951 /// If UserUF is non-zero then this method finds the best unroll-factor
952 /// based on register pressure and other parameters.
953 /// VF and LoopCost are the selected vectorization factor and the cost of the
955 unsigned selectUnrollFactor(bool OptForSize, unsigned VF, unsigned LoopCost);
957 /// \brief A struct that represents some properties of the register usage
959 struct RegisterUsage {
960 /// Holds the number of loop invariant values that are used in the loop.
961 unsigned LoopInvariantRegs;
962 /// Holds the maximum number of concurrent live intervals in the loop.
963 unsigned MaxLocalUsers;
964 /// Holds the number of instructions in the loop.
965 unsigned NumInstructions;
968 /// \return information about the register usage of the loop.
969 RegisterUsage calculateRegisterUsage();
972 /// Returns the expected execution cost. The unit of the cost does
973 /// not matter because we use the 'cost' units to compare different
974 /// vector widths. The cost that is returned is *not* normalized by
975 /// the factor width.
976 unsigned expectedCost(unsigned VF);
978 /// Returns the execution time cost of an instruction for a given vector
979 /// width. Vector width of one means scalar.
980 unsigned getInstructionCost(Instruction *I, unsigned VF);
982 /// A helper function for converting Scalar types to vector types.
983 /// If the incoming type is void, we return void. If the VF is 1, we return
985 static Type* ToVectorTy(Type *Scalar, unsigned VF);
987 /// Returns whether the instruction is a load or store and will be a emitted
988 /// as a vector operation.
989 bool isConsecutiveLoadOrStore(Instruction *I);
991 /// Report an analysis message to assist the user in diagnosing loops that are
992 /// not vectorized. These are handled as LoopAccessReport rather than
993 /// VectorizationReport because the << operator of VectorizationReport returns
994 /// LoopAccessReport.
995 void emitAnalysis(const LoopAccessReport &Message) {
996 LoopAccessReport::emitAnalysis(Message, TheFunction, TheLoop, LV_NAME);
999 /// Values used only by @llvm.assume calls.
1000 SmallPtrSet<const Value *, 32> EphValues;
1002 /// The loop that we evaluate.
1005 ScalarEvolution *SE;
1006 /// Loop Info analysis.
1008 /// Vectorization legality.
1009 LoopVectorizationLegality *Legal;
1010 /// Vector target information.
1011 const TargetTransformInfo &TTI;
1012 /// Target data layout information.
1013 const DataLayout *DL;
1014 /// Target Library Info.
1015 const TargetLibraryInfo *TLI;
1016 const Function *TheFunction;
1017 // Loop Vectorize Hint.
1018 const LoopVectorizeHints *Hints;
1021 /// Utility class for getting and setting loop vectorizer hints in the form
1022 /// of loop metadata.
1023 /// This class keeps a number of loop annotations locally (as member variables)
1024 /// and can, upon request, write them back as metadata on the loop. It will
1025 /// initially scan the loop for existing metadata, and will update the local
1026 /// values based on information in the loop.
1027 /// We cannot write all values to metadata, as the mere presence of some info,
1028 /// for example 'force', means a decision has been made. So, we need to be
1029 /// careful NOT to add them if the user hasn't specifically asked so.
1030 class LoopVectorizeHints {
1037 /// Hint - associates name and validation with the hint value.
1040 unsigned Value; // This may have to change for non-numeric values.
1043 Hint(const char * Name, unsigned Value, HintKind Kind)
1044 : Name(Name), Value(Value), Kind(Kind) { }
1046 bool validate(unsigned Val) {
1049 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
1051 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1059 /// Vectorization width.
1061 /// Vectorization interleave factor.
1063 /// Vectorization forced
1066 /// Return the loop metadata prefix.
1067 static StringRef Prefix() { return "llvm.loop."; }
1071 FK_Undefined = -1, ///< Not selected.
1072 FK_Disabled = 0, ///< Forcing disabled.
1073 FK_Enabled = 1, ///< Forcing enabled.
1076 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
1077 : Width("vectorize.width", VectorizationFactor, HK_WIDTH),
1078 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1079 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1081 // Populate values with existing loop metadata.
1082 getHintsFromMetadata();
1084 // force-vector-interleave overrides DisableInterleaving.
1085 if (VectorizationInterleave.getNumOccurrences() > 0)
1086 Interleave.Value = VectorizationInterleave;
1088 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
1089 << "LV: Interleaving disabled by the pass manager\n");
1092 /// Mark the loop L as already vectorized by setting the width to 1.
1093 void setAlreadyVectorized() {
1094 Width.Value = Interleave.Value = 1;
1095 Hint Hints[] = {Width, Interleave};
1096 writeHintsToMetadata(Hints);
1099 /// Dumps all the hint information.
1100 std::string emitRemark() const {
1101 VectorizationReport R;
1102 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1103 R << "vectorization is explicitly disabled";
1105 R << "use -Rpass-analysis=loop-vectorize for more info";
1106 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1107 R << " (Force=true";
1108 if (Width.Value != 0)
1109 R << ", Vector Width=" << Width.Value;
1110 if (Interleave.Value != 0)
1111 R << ", Interleave Count=" << Interleave.Value;
1119 unsigned getWidth() const { return Width.Value; }
1120 unsigned getInterleave() const { return Interleave.Value; }
1121 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1124 /// Find hints specified in the loop metadata and update local values.
1125 void getHintsFromMetadata() {
1126 MDNode *LoopID = TheLoop->getLoopID();
1130 // First operand should refer to the loop id itself.
1131 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1132 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1134 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1135 const MDString *S = nullptr;
1136 SmallVector<Metadata *, 4> Args;
1138 // The expected hint is either a MDString or a MDNode with the first
1139 // operand a MDString.
1140 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1141 if (!MD || MD->getNumOperands() == 0)
1143 S = dyn_cast<MDString>(MD->getOperand(0));
1144 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1145 Args.push_back(MD->getOperand(i));
1147 S = dyn_cast<MDString>(LoopID->getOperand(i));
1148 assert(Args.size() == 0 && "too many arguments for MDString");
1154 // Check if the hint starts with the loop metadata prefix.
1155 StringRef Name = S->getString();
1156 if (Args.size() == 1)
1157 setHint(Name, Args[0]);
1161 /// Checks string hint with one operand and set value if valid.
1162 void setHint(StringRef Name, Metadata *Arg) {
1163 if (!Name.startswith(Prefix()))
1165 Name = Name.substr(Prefix().size(), StringRef::npos);
1167 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1169 unsigned Val = C->getZExtValue();
1171 Hint *Hints[] = {&Width, &Interleave, &Force};
1172 for (auto H : Hints) {
1173 if (Name == H->Name) {
1174 if (H->validate(Val))
1177 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1183 /// Create a new hint from name / value pair.
1184 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1185 LLVMContext &Context = TheLoop->getHeader()->getContext();
1186 Metadata *MDs[] = {MDString::get(Context, Name),
1187 ConstantAsMetadata::get(
1188 ConstantInt::get(Type::getInt32Ty(Context), V))};
1189 return MDNode::get(Context, MDs);
1192 /// Matches metadata with hint name.
1193 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1194 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1198 for (auto H : HintTypes)
1199 if (Name->getString().endswith(H.Name))
1204 /// Sets current hints into loop metadata, keeping other values intact.
1205 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1206 if (HintTypes.size() == 0)
1209 // Reserve the first element to LoopID (see below).
1210 SmallVector<Metadata *, 4> MDs(1);
1211 // If the loop already has metadata, then ignore the existing operands.
1212 MDNode *LoopID = TheLoop->getLoopID();
1214 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1215 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1216 // If node in update list, ignore old value.
1217 if (!matchesHintMetadataName(Node, HintTypes))
1218 MDs.push_back(Node);
1222 // Now, add the missing hints.
1223 for (auto H : HintTypes)
1224 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1226 // Replace current metadata node with new one.
1227 LLVMContext &Context = TheLoop->getHeader()->getContext();
1228 MDNode *NewLoopID = MDNode::get(Context, MDs);
1229 // Set operand 0 to refer to the loop id itself.
1230 NewLoopID->replaceOperandWith(0, NewLoopID);
1232 TheLoop->setLoopID(NewLoopID);
1235 /// The loop these hints belong to.
1236 const Loop *TheLoop;
1239 static void emitMissedWarning(Function *F, Loop *L,
1240 const LoopVectorizeHints &LH) {
1241 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1242 L->getStartLoc(), LH.emitRemark());
1244 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1245 if (LH.getWidth() != 1)
1246 emitLoopVectorizeWarning(
1247 F->getContext(), *F, L->getStartLoc(),
1248 "failed explicitly specified loop vectorization");
1249 else if (LH.getInterleave() != 1)
1250 emitLoopInterleaveWarning(
1251 F->getContext(), *F, L->getStartLoc(),
1252 "failed explicitly specified loop interleaving");
1256 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1258 return V.push_back(&L);
1260 for (Loop *InnerL : L)
1261 addInnerLoop(*InnerL, V);
1264 /// The LoopVectorize Pass.
1265 struct LoopVectorize : public FunctionPass {
1266 /// Pass identification, replacement for typeid
1269 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1271 DisableUnrolling(NoUnrolling),
1272 AlwaysVectorize(AlwaysVectorize) {
1273 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1276 ScalarEvolution *SE;
1277 const DataLayout *DL;
1279 TargetTransformInfo *TTI;
1281 BlockFrequencyInfo *BFI;
1282 TargetLibraryInfo *TLI;
1284 AssumptionCache *AC;
1285 LoopAccessAnalysis *LAA;
1286 bool DisableUnrolling;
1287 bool AlwaysVectorize;
1289 BlockFrequency ColdEntryFreq;
1291 bool runOnFunction(Function &F) override {
1292 SE = &getAnalysis<ScalarEvolution>();
1293 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1294 DL = DLP ? &DLP->getDataLayout() : nullptr;
1295 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1296 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1297 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1298 BFI = &getAnalysis<BlockFrequencyInfo>();
1299 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1300 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1301 AA = &getAnalysis<AliasAnalysis>();
1302 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1303 LAA = &getAnalysis<LoopAccessAnalysis>();
1305 // Compute some weights outside of the loop over the loops. Compute this
1306 // using a BranchProbability to re-use its scaling math.
1307 const BranchProbability ColdProb(1, 5); // 20%
1308 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1310 // If the target claims to have no vector registers don't attempt
1312 if (!TTI->getNumberOfRegisters(true))
1316 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1317 << ": Missing data layout\n");
1321 // Build up a worklist of inner-loops to vectorize. This is necessary as
1322 // the act of vectorizing or partially unrolling a loop creates new loops
1323 // and can invalidate iterators across the loops.
1324 SmallVector<Loop *, 8> Worklist;
1327 addInnerLoop(*L, Worklist);
1329 LoopsAnalyzed += Worklist.size();
1331 // Now walk the identified inner loops.
1332 bool Changed = false;
1333 while (!Worklist.empty())
1334 Changed |= processLoop(Worklist.pop_back_val());
1336 // Process each loop nest in the function.
1340 bool processLoop(Loop *L) {
1341 assert(L->empty() && "Only process inner loops.");
1344 const std::string DebugLocStr = getDebugLocString(L);
1347 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1348 << L->getHeader()->getParent()->getName() << "\" from "
1349 << DebugLocStr << "\n");
1351 LoopVectorizeHints Hints(L, DisableUnrolling);
1353 DEBUG(dbgs() << "LV: Loop hints:"
1355 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1357 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1359 : "?")) << " width=" << Hints.getWidth()
1360 << " unroll=" << Hints.getInterleave() << "\n");
1362 // Function containing loop
1363 Function *F = L->getHeader()->getParent();
1365 // Looking at the diagnostic output is the only way to determine if a loop
1366 // was vectorized (other than looking at the IR or machine code), so it
1367 // is important to generate an optimization remark for each loop. Most of
1368 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1369 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1370 // less verbose reporting vectorized loops and unvectorized loops that may
1371 // benefit from vectorization, respectively.
1373 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1374 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1375 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1376 L->getStartLoc(), Hints.emitRemark());
1380 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1381 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1382 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1383 L->getStartLoc(), Hints.emitRemark());
1387 if (Hints.getWidth() == 1 && Hints.getInterleave() == 1) {
1388 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1389 emitOptimizationRemarkAnalysis(
1390 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1391 "loop not vectorized: vector width and interleave count are "
1392 "explicitly set to 1");
1396 // Check the loop for a trip count threshold:
1397 // do not vectorize loops with a tiny trip count.
1398 const unsigned TC = SE->getSmallConstantTripCount(L);
1399 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1400 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1401 << "This loop is not worth vectorizing.");
1402 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1403 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1405 DEBUG(dbgs() << "\n");
1406 emitOptimizationRemarkAnalysis(
1407 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1408 "vectorization is not beneficial and is not explicitly forced");
1413 // Check if it is legal to vectorize the loop.
1414 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, AA, F, TTI, LAA);
1415 if (!LVL.canVectorize()) {
1416 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1417 emitMissedWarning(F, L, Hints);
1421 // Use the cost model.
1422 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI, AC, F,
1425 // Check the function attributes to find out if this function should be
1426 // optimized for size.
1427 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1428 F->hasFnAttribute(Attribute::OptimizeForSize);
1430 // Compute the weighted frequency of this loop being executed and see if it
1431 // is less than 20% of the function entry baseline frequency. Note that we
1432 // always have a canonical loop here because we think we *can* vectoriez.
1433 // FIXME: This is hidden behind a flag due to pervasive problems with
1434 // exactly what block frequency models.
1435 if (LoopVectorizeWithBlockFrequency) {
1436 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1437 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1438 LoopEntryFreq < ColdEntryFreq)
1442 // Check the function attributes to see if implicit floats are allowed.a
1443 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1444 // an integer loop and the vector instructions selected are purely integer
1445 // vector instructions?
1446 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1447 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1448 "attribute is used.\n");
1449 emitOptimizationRemarkAnalysis(
1450 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1451 "loop not vectorized due to NoImplicitFloat attribute");
1452 emitMissedWarning(F, L, Hints);
1456 // Select the optimal vectorization factor.
1457 const LoopVectorizationCostModel::VectorizationFactor VF =
1458 CM.selectVectorizationFactor(OptForSize);
1460 // Select the unroll factor.
1462 CM.selectUnrollFactor(OptForSize, VF.Width, VF.Cost);
1464 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1465 << DebugLocStr << '\n');
1466 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1468 if (VF.Width == 1) {
1469 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
1472 emitOptimizationRemarkAnalysis(
1473 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1474 "not beneficial to vectorize and user disabled interleaving");
1477 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1479 // Report the unrolling decision.
1480 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1481 Twine("unrolled with interleaving factor " +
1483 " (vectorization not beneficial)"));
1485 // We decided not to vectorize, but we may want to unroll.
1487 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1488 Unroller.vectorize(&LVL);
1490 // If we decided that it is *legal* to vectorize the loop then do it.
1491 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1495 // Report the vectorization decision.
1496 emitOptimizationRemark(
1497 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1498 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1499 ", unrolling interleave factor: " + Twine(UF) + ")");
1502 // Mark the loop as already vectorized to avoid vectorizing again.
1503 Hints.setAlreadyVectorized();
1505 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1509 void getAnalysisUsage(AnalysisUsage &AU) const override {
1510 AU.addRequired<AssumptionCacheTracker>();
1511 AU.addRequiredID(LoopSimplifyID);
1512 AU.addRequiredID(LCSSAID);
1513 AU.addRequired<BlockFrequencyInfo>();
1514 AU.addRequired<DominatorTreeWrapperPass>();
1515 AU.addRequired<LoopInfoWrapperPass>();
1516 AU.addRequired<ScalarEvolution>();
1517 AU.addRequired<TargetTransformInfoWrapperPass>();
1518 AU.addRequired<AliasAnalysis>();
1519 AU.addRequired<LoopAccessAnalysis>();
1520 AU.addPreserved<LoopInfoWrapperPass>();
1521 AU.addPreserved<DominatorTreeWrapperPass>();
1522 AU.addPreserved<AliasAnalysis>();
1527 } // end anonymous namespace
1529 //===----------------------------------------------------------------------===//
1530 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1531 // LoopVectorizationCostModel.
1532 //===----------------------------------------------------------------------===//
1534 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1535 // We need to place the broadcast of invariant variables outside the loop.
1536 Instruction *Instr = dyn_cast<Instruction>(V);
1538 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1539 Instr->getParent()) != LoopVectorBody.end());
1540 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1542 // Place the code for broadcasting invariant variables in the new preheader.
1543 IRBuilder<>::InsertPointGuard Guard(Builder);
1545 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1547 // Broadcast the scalar into all locations in the vector.
1548 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1553 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
1555 assert(Val->getType()->isVectorTy() && "Must be a vector");
1556 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1557 "Elem must be an integer");
1558 assert(Step->getType() == Val->getType()->getScalarType() &&
1559 "Step has wrong type");
1560 // Create the types.
1561 Type *ITy = Val->getType()->getScalarType();
1562 VectorType *Ty = cast<VectorType>(Val->getType());
1563 int VLen = Ty->getNumElements();
1564 SmallVector<Constant*, 8> Indices;
1566 // Create a vector of consecutive numbers from zero to VF.
1567 for (int i = 0; i < VLen; ++i)
1568 Indices.push_back(ConstantInt::get(ITy, StartIdx + i));
1570 // Add the consecutive indices to the vector value.
1571 Constant *Cv = ConstantVector::get(Indices);
1572 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1573 Step = Builder.CreateVectorSplat(VLen, Step);
1574 assert(Step->getType() == Val->getType() && "Invalid step vec");
1575 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
1576 // which can be found from the original scalar operations.
1577 Step = Builder.CreateMul(Cv, Step);
1578 return Builder.CreateAdd(Val, Step, "induction");
1581 /// \brief Find the operand of the GEP that should be checked for consecutive
1582 /// stores. This ignores trailing indices that have no effect on the final
1584 static unsigned getGEPInductionOperand(const DataLayout *DL,
1585 const GetElementPtrInst *Gep) {
1586 unsigned LastOperand = Gep->getNumOperands() - 1;
1587 unsigned GEPAllocSize = DL->getTypeAllocSize(
1588 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1590 // Walk backwards and try to peel off zeros.
1591 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1592 // Find the type we're currently indexing into.
1593 gep_type_iterator GEPTI = gep_type_begin(Gep);
1594 std::advance(GEPTI, LastOperand - 1);
1596 // If it's a type with the same allocation size as the result of the GEP we
1597 // can peel off the zero index.
1598 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1606 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1607 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1608 // Make sure that the pointer does not point to structs.
1609 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1612 // If this value is a pointer induction variable we know it is consecutive.
1613 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1614 if (Phi && Inductions.count(Phi)) {
1615 InductionInfo II = Inductions[Phi];
1616 return II.getConsecutiveDirection();
1619 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1623 unsigned NumOperands = Gep->getNumOperands();
1624 Value *GpPtr = Gep->getPointerOperand();
1625 // If this GEP value is a consecutive pointer induction variable and all of
1626 // the indices are constant then we know it is consecutive. We can
1627 Phi = dyn_cast<PHINode>(GpPtr);
1628 if (Phi && Inductions.count(Phi)) {
1630 // Make sure that the pointer does not point to structs.
1631 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1632 if (GepPtrType->getElementType()->isAggregateType())
1635 // Make sure that all of the index operands are loop invariant.
1636 for (unsigned i = 1; i < NumOperands; ++i)
1637 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1640 InductionInfo II = Inductions[Phi];
1641 return II.getConsecutiveDirection();
1644 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1646 // Check that all of the gep indices are uniform except for our induction
1648 for (unsigned i = 0; i != NumOperands; ++i)
1649 if (i != InductionOperand &&
1650 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1653 // We can emit wide load/stores only if the last non-zero index is the
1654 // induction variable.
1655 const SCEV *Last = nullptr;
1656 if (!Strides.count(Gep))
1657 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1659 // Because of the multiplication by a stride we can have a s/zext cast.
1660 // We are going to replace this stride by 1 so the cast is safe to ignore.
1662 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1663 // %0 = trunc i64 %indvars.iv to i32
1664 // %mul = mul i32 %0, %Stride1
1665 // %idxprom = zext i32 %mul to i64 << Safe cast.
1666 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1668 Last = replaceSymbolicStrideSCEV(SE, Strides,
1669 Gep->getOperand(InductionOperand), Gep);
1670 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1672 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1676 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1677 const SCEV *Step = AR->getStepRecurrence(*SE);
1679 // The memory is consecutive because the last index is consecutive
1680 // and all other indices are loop invariant.
1683 if (Step->isAllOnesValue())
1690 bool LoopVectorizationLegality::isUniform(Value *V) {
1691 return LAI->isUniform(V);
1694 InnerLoopVectorizer::VectorParts&
1695 InnerLoopVectorizer::getVectorValue(Value *V) {
1696 assert(V != Induction && "The new induction variable should not be used.");
1697 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1699 // If we have a stride that is replaced by one, do it here.
1700 if (Legal->hasStride(V))
1701 V = ConstantInt::get(V->getType(), 1);
1703 // If we have this scalar in the map, return it.
1704 if (WidenMap.has(V))
1705 return WidenMap.get(V);
1707 // If this scalar is unknown, assume that it is a constant or that it is
1708 // loop invariant. Broadcast V and save the value for future uses.
1709 Value *B = getBroadcastInstrs(V);
1710 return WidenMap.splat(V, B);
1713 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1714 assert(Vec->getType()->isVectorTy() && "Invalid type");
1715 SmallVector<Constant*, 8> ShuffleMask;
1716 for (unsigned i = 0; i < VF; ++i)
1717 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1719 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1720 ConstantVector::get(ShuffleMask),
1724 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1725 // Attempt to issue a wide load.
1726 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1727 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1729 assert((LI || SI) && "Invalid Load/Store instruction");
1731 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1732 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1733 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1734 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1735 // An alignment of 0 means target abi alignment. We need to use the scalar's
1736 // target abi alignment in such a case.
1738 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1739 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1740 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1741 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1743 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
1744 !Legal->isMaskRequired(SI))
1745 return scalarizeInstruction(Instr, true);
1747 if (ScalarAllocatedSize != VectorElementSize)
1748 return scalarizeInstruction(Instr);
1750 // If the pointer is loop invariant or if it is non-consecutive,
1751 // scalarize the load.
1752 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1753 bool Reverse = ConsecutiveStride < 0;
1754 bool UniformLoad = LI && Legal->isUniform(Ptr);
1755 if (!ConsecutiveStride || UniformLoad)
1756 return scalarizeInstruction(Instr);
1758 Constant *Zero = Builder.getInt32(0);
1759 VectorParts &Entry = WidenMap.get(Instr);
1761 // Handle consecutive loads/stores.
1762 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1763 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1764 setDebugLocFromInst(Builder, Gep);
1765 Value *PtrOperand = Gep->getPointerOperand();
1766 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1767 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1769 // Create the new GEP with the new induction variable.
1770 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1771 Gep2->setOperand(0, FirstBasePtr);
1772 Gep2->setName("gep.indvar.base");
1773 Ptr = Builder.Insert(Gep2);
1775 setDebugLocFromInst(Builder, Gep);
1776 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1777 OrigLoop) && "Base ptr must be invariant");
1779 // The last index does not have to be the induction. It can be
1780 // consecutive and be a function of the index. For example A[I+1];
1781 unsigned NumOperands = Gep->getNumOperands();
1782 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1783 // Create the new GEP with the new induction variable.
1784 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1786 for (unsigned i = 0; i < NumOperands; ++i) {
1787 Value *GepOperand = Gep->getOperand(i);
1788 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1790 // Update last index or loop invariant instruction anchored in loop.
1791 if (i == InductionOperand ||
1792 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1793 assert((i == InductionOperand ||
1794 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1795 "Must be last index or loop invariant");
1797 VectorParts &GEPParts = getVectorValue(GepOperand);
1798 Value *Index = GEPParts[0];
1799 Index = Builder.CreateExtractElement(Index, Zero);
1800 Gep2->setOperand(i, Index);
1801 Gep2->setName("gep.indvar.idx");
1804 Ptr = Builder.Insert(Gep2);
1806 // Use the induction element ptr.
1807 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1808 setDebugLocFromInst(Builder, Ptr);
1809 VectorParts &PtrVal = getVectorValue(Ptr);
1810 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1813 VectorParts Mask = createBlockInMask(Instr->getParent());
1816 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1817 "We do not allow storing to uniform addresses");
1818 setDebugLocFromInst(Builder, SI);
1819 // We don't want to update the value in the map as it might be used in
1820 // another expression. So don't use a reference type for "StoredVal".
1821 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1823 for (unsigned Part = 0; Part < UF; ++Part) {
1824 // Calculate the pointer for the specific unroll-part.
1825 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1828 // If we store to reverse consecutive memory locations then we need
1829 // to reverse the order of elements in the stored value.
1830 StoredVal[Part] = reverseVector(StoredVal[Part]);
1831 // If the address is consecutive but reversed, then the
1832 // wide store needs to start at the last vector element.
1833 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1834 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1835 Mask[Part] = reverseVector(Mask[Part]);
1838 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1839 DataTy->getPointerTo(AddressSpace));
1842 if (Legal->isMaskRequired(SI))
1843 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
1846 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
1847 propagateMetadata(NewSI, SI);
1853 assert(LI && "Must have a load instruction");
1854 setDebugLocFromInst(Builder, LI);
1855 for (unsigned Part = 0; Part < UF; ++Part) {
1856 // Calculate the pointer for the specific unroll-part.
1857 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1860 // If the address is consecutive but reversed, then the
1861 // wide load needs to start at the last vector element.
1862 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1863 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1864 Mask[Part] = reverseVector(Mask[Part]);
1868 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1869 DataTy->getPointerTo(AddressSpace));
1870 if (Legal->isMaskRequired(LI))
1871 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
1872 UndefValue::get(DataTy),
1873 "wide.masked.load");
1875 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
1876 propagateMetadata(NewLI, LI);
1877 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
1881 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1882 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1883 // Holds vector parameters or scalars, in case of uniform vals.
1884 SmallVector<VectorParts, 4> Params;
1886 setDebugLocFromInst(Builder, Instr);
1888 // Find all of the vectorized parameters.
1889 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1890 Value *SrcOp = Instr->getOperand(op);
1892 // If we are accessing the old induction variable, use the new one.
1893 if (SrcOp == OldInduction) {
1894 Params.push_back(getVectorValue(SrcOp));
1898 // Try using previously calculated values.
1899 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1901 // If the src is an instruction that appeared earlier in the basic block
1902 // then it should already be vectorized.
1903 if (SrcInst && OrigLoop->contains(SrcInst)) {
1904 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1905 // The parameter is a vector value from earlier.
1906 Params.push_back(WidenMap.get(SrcInst));
1908 // The parameter is a scalar from outside the loop. Maybe even a constant.
1909 VectorParts Scalars;
1910 Scalars.append(UF, SrcOp);
1911 Params.push_back(Scalars);
1915 assert(Params.size() == Instr->getNumOperands() &&
1916 "Invalid number of operands");
1918 // Does this instruction return a value ?
1919 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1921 Value *UndefVec = IsVoidRetTy ? nullptr :
1922 UndefValue::get(VectorType::get(Instr->getType(), VF));
1923 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1924 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1926 Instruction *InsertPt = Builder.GetInsertPoint();
1927 BasicBlock *IfBlock = Builder.GetInsertBlock();
1928 BasicBlock *CondBlock = nullptr;
1931 Loop *VectorLp = nullptr;
1932 if (IfPredicateStore) {
1933 assert(Instr->getParent()->getSinglePredecessor() &&
1934 "Only support single predecessor blocks");
1935 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1936 Instr->getParent());
1937 VectorLp = LI->getLoopFor(IfBlock);
1938 assert(VectorLp && "Must have a loop for this block");
1941 // For each vector unroll 'part':
1942 for (unsigned Part = 0; Part < UF; ++Part) {
1943 // For each scalar that we create:
1944 for (unsigned Width = 0; Width < VF; ++Width) {
1947 Value *Cmp = nullptr;
1948 if (IfPredicateStore) {
1949 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1950 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1951 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1952 LoopVectorBody.push_back(CondBlock);
1953 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
1954 // Update Builder with newly created basic block.
1955 Builder.SetInsertPoint(InsertPt);
1958 Instruction *Cloned = Instr->clone();
1960 Cloned->setName(Instr->getName() + ".cloned");
1961 // Replace the operands of the cloned instructions with extracted scalars.
1962 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1963 Value *Op = Params[op][Part];
1964 // Param is a vector. Need to extract the right lane.
1965 if (Op->getType()->isVectorTy())
1966 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1967 Cloned->setOperand(op, Op);
1970 // Place the cloned scalar in the new loop.
1971 Builder.Insert(Cloned);
1973 // If the original scalar returns a value we need to place it in a vector
1974 // so that future users will be able to use it.
1976 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1977 Builder.getInt32(Width));
1979 if (IfPredicateStore) {
1980 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1981 LoopVectorBody.push_back(NewIfBlock);
1982 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
1983 Builder.SetInsertPoint(InsertPt);
1984 Instruction *OldBr = IfBlock->getTerminator();
1985 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1986 OldBr->eraseFromParent();
1987 IfBlock = NewIfBlock;
1993 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1997 if (Instruction *I = dyn_cast<Instruction>(V))
1998 return I->getParent() == Loc->getParent() ? I : nullptr;
2002 std::pair<Instruction *, Instruction *>
2003 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
2004 Instruction *tnullptr = nullptr;
2005 if (!Legal->mustCheckStrides())
2006 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2008 IRBuilder<> ChkBuilder(Loc);
2011 Value *Check = nullptr;
2012 Instruction *FirstInst = nullptr;
2013 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
2014 SE = Legal->strides_end();
2016 Value *Ptr = stripIntegerCast(*SI);
2017 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
2019 // Store the first instruction we create.
2020 FirstInst = getFirstInst(FirstInst, C, Loc);
2022 Check = ChkBuilder.CreateOr(Check, C);
2027 // We have to do this trickery because the IRBuilder might fold the check to a
2028 // constant expression in which case there is no Instruction anchored in a
2030 LLVMContext &Ctx = Loc->getContext();
2031 Instruction *TheCheck =
2032 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2033 ChkBuilder.Insert(TheCheck, "stride.not.one");
2034 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2036 return std::make_pair(FirstInst, TheCheck);
2039 void InnerLoopVectorizer::createEmptyLoop() {
2041 In this function we generate a new loop. The new loop will contain
2042 the vectorized instructions while the old loop will continue to run the
2045 [ ] <-- Back-edge taken count overflow check.
2048 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2051 || [ ] <-- vector pre header.
2055 || [ ]_| <-- vector loop.
2058 | >[ ] <--- middle-block.
2061 -|- >[ ] <--- new preheader.
2065 | [ ]_| <-- old scalar loop to handle remainder.
2068 >[ ] <-- exit block.
2072 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2073 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2074 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2075 assert(BypassBlock && "Invalid loop structure");
2076 assert(ExitBlock && "Must have an exit block");
2078 // Some loops have a single integer induction variable, while other loops
2079 // don't. One example is c++ iterators that often have multiple pointer
2080 // induction variables. In the code below we also support a case where we
2081 // don't have a single induction variable.
2082 OldInduction = Legal->getInduction();
2083 Type *IdxTy = Legal->getWidestInductionType();
2085 // Find the loop boundaries.
2086 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2087 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2089 // The exit count might have the type of i64 while the phi is i32. This can
2090 // happen if we have an induction variable that is sign extended before the
2091 // compare. The only way that we get a backedge taken count is that the
2092 // induction variable was signed and as such will not overflow. In such a case
2093 // truncation is legal.
2094 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2095 IdxTy->getPrimitiveSizeInBits())
2096 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2098 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2099 // Get the total trip count from the count by adding 1.
2100 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2101 SE->getConstant(BackedgeTakeCount->getType(), 1));
2103 // Expand the trip count and place the new instructions in the preheader.
2104 // Notice that the pre-header does not change, only the loop body.
2105 SCEVExpander Exp(*SE, "induction");
2107 // We need to test whether the backedge-taken count is uint##_max. Adding one
2108 // to it will cause overflow and an incorrect loop trip count in the vector
2109 // body. In case of overflow we want to directly jump to the scalar remainder
2111 Value *BackedgeCount =
2112 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2113 BypassBlock->getTerminator());
2114 if (BackedgeCount->getType()->isPointerTy())
2115 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2116 "backedge.ptrcnt.to.int",
2117 BypassBlock->getTerminator());
2118 Instruction *CheckBCOverflow =
2119 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2120 Constant::getAllOnesValue(BackedgeCount->getType()),
2121 "backedge.overflow", BypassBlock->getTerminator());
2123 // The loop index does not have to start at Zero. Find the original start
2124 // value from the induction PHI node. If we don't have an induction variable
2125 // then we know that it starts at zero.
2126 Builder.SetInsertPoint(BypassBlock->getTerminator());
2127 Value *StartIdx = ExtendedIdx = OldInduction ?
2128 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2130 ConstantInt::get(IdxTy, 0);
2132 // We need an instruction to anchor the overflow check on. StartIdx needs to
2133 // be defined before the overflow check branch. Because the scalar preheader
2134 // is going to merge the start index and so the overflow branch block needs to
2135 // contain a definition of the start index.
2136 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2137 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2138 BypassBlock->getTerminator());
2140 // Count holds the overall loop count (N).
2141 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2142 BypassBlock->getTerminator());
2144 LoopBypassBlocks.push_back(BypassBlock);
2146 // Split the single block loop into the two loop structure described above.
2147 BasicBlock *VectorPH =
2148 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2149 BasicBlock *VecBody =
2150 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2151 BasicBlock *MiddleBlock =
2152 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2153 BasicBlock *ScalarPH =
2154 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2156 // Create and register the new vector loop.
2157 Loop* Lp = new Loop();
2158 Loop *ParentLoop = OrigLoop->getParentLoop();
2160 // Insert the new loop into the loop nest and register the new basic blocks
2161 // before calling any utilities such as SCEV that require valid LoopInfo.
2163 ParentLoop->addChildLoop(Lp);
2164 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
2165 ParentLoop->addBasicBlockToLoop(VectorPH, *LI);
2166 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
2168 LI->addTopLevelLoop(Lp);
2170 Lp->addBasicBlockToLoop(VecBody, *LI);
2172 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2174 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2176 // Generate the induction variable.
2177 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2178 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2179 // The loop step is equal to the vectorization factor (num of SIMD elements)
2180 // times the unroll factor (num of SIMD instructions).
2181 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2183 // This is the IR builder that we use to add all of the logic for bypassing
2184 // the new vector loop.
2185 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2186 setDebugLocFromInst(BypassBuilder,
2187 getDebugLocFromInstOrOperands(OldInduction));
2189 // We may need to extend the index in case there is a type mismatch.
2190 // We know that the count starts at zero and does not overflow.
2191 if (Count->getType() != IdxTy) {
2192 // The exit count can be of pointer type. Convert it to the correct
2194 if (ExitCount->getType()->isPointerTy())
2195 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2197 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2200 // Add the start index to the loop count to get the new end index.
2201 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2203 // Now we need to generate the expression for N - (N % VF), which is
2204 // the part that the vectorized body will execute.
2205 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2206 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2207 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2208 "end.idx.rnd.down");
2210 // Now, compare the new count to zero. If it is zero skip the vector loop and
2211 // jump to the scalar loop.
2213 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2215 BasicBlock *LastBypassBlock = BypassBlock;
2217 // Generate code to check that the loops trip count that we computed by adding
2218 // one to the backedge-taken count will not overflow.
2220 auto PastOverflowCheck =
2221 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2222 BasicBlock *CheckBlock =
2223 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2225 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2226 LoopBypassBlocks.push_back(CheckBlock);
2227 Instruction *OldTerm = LastBypassBlock->getTerminator();
2228 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2229 OldTerm->eraseFromParent();
2230 LastBypassBlock = CheckBlock;
2233 // Generate the code to check that the strides we assumed to be one are really
2234 // one. We want the new basic block to start at the first instruction in a
2235 // sequence of instructions that form a check.
2236 Instruction *StrideCheck;
2237 Instruction *FirstCheckInst;
2238 std::tie(FirstCheckInst, StrideCheck) =
2239 addStrideCheck(LastBypassBlock->getTerminator());
2241 // Create a new block containing the stride check.
2242 BasicBlock *CheckBlock =
2243 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2245 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2246 LoopBypassBlocks.push_back(CheckBlock);
2248 // Replace the branch into the memory check block with a conditional branch
2249 // for the "few elements case".
2250 Instruction *OldTerm = LastBypassBlock->getTerminator();
2251 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2252 OldTerm->eraseFromParent();
2255 LastBypassBlock = CheckBlock;
2258 // Generate the code that checks in runtime if arrays overlap. We put the
2259 // checks into a separate block to make the more common case of few elements
2261 Instruction *MemRuntimeCheck;
2262 std::tie(FirstCheckInst, MemRuntimeCheck) =
2263 Legal->getLAI()->addRuntimeCheck(LastBypassBlock->getTerminator());
2264 if (MemRuntimeCheck) {
2265 // Create a new block containing the memory check.
2266 BasicBlock *CheckBlock =
2267 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.memcheck");
2269 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2270 LoopBypassBlocks.push_back(CheckBlock);
2272 // Replace the branch into the memory check block with a conditional branch
2273 // for the "few elements case".
2274 Instruction *OldTerm = LastBypassBlock->getTerminator();
2275 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2276 OldTerm->eraseFromParent();
2278 Cmp = MemRuntimeCheck;
2279 LastBypassBlock = CheckBlock;
2282 LastBypassBlock->getTerminator()->eraseFromParent();
2283 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2286 // We are going to resume the execution of the scalar loop.
2287 // Go over all of the induction variables that we found and fix the
2288 // PHIs that are left in the scalar version of the loop.
2289 // The starting values of PHI nodes depend on the counter of the last
2290 // iteration in the vectorized loop.
2291 // If we come from a bypass edge then we need to start from the original
2294 // This variable saves the new starting index for the scalar loop.
2295 PHINode *ResumeIndex = nullptr;
2296 LoopVectorizationLegality::InductionList::iterator I, E;
2297 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2298 // Set builder to point to last bypass block.
2299 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2300 for (I = List->begin(), E = List->end(); I != E; ++I) {
2301 PHINode *OrigPhi = I->first;
2302 LoopVectorizationLegality::InductionInfo II = I->second;
2304 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2305 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2306 MiddleBlock->getTerminator());
2307 // We might have extended the type of the induction variable but we need a
2308 // truncated version for the scalar loop.
2309 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2310 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2311 MiddleBlock->getTerminator()) : nullptr;
2313 // Create phi nodes to merge from the backedge-taken check block.
2314 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2315 ScalarPH->getTerminator());
2316 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2318 PHINode *BCTruncResumeVal = nullptr;
2319 if (OrigPhi == OldInduction) {
2321 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2322 ScalarPH->getTerminator());
2323 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2326 Value *EndValue = nullptr;
2328 case LoopVectorizationLegality::IK_NoInduction:
2329 llvm_unreachable("Unknown induction");
2330 case LoopVectorizationLegality::IK_IntInduction: {
2331 // Handle the integer induction counter.
2332 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2334 // We have the canonical induction variable.
2335 if (OrigPhi == OldInduction) {
2336 // Create a truncated version of the resume value for the scalar loop,
2337 // we might have promoted the type to a larger width.
2339 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2340 // The new PHI merges the original incoming value, in case of a bypass,
2341 // or the value at the end of the vectorized loop.
2342 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2343 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2344 TruncResumeVal->addIncoming(EndValue, VecBody);
2346 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2348 // We know what the end value is.
2349 EndValue = IdxEndRoundDown;
2350 // We also know which PHI node holds it.
2351 ResumeIndex = ResumeVal;
2355 // Not the canonical induction variable - add the vector loop count to the
2357 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2358 II.StartValue->getType(),
2360 EndValue = II.transform(BypassBuilder, CRD);
2361 EndValue->setName("ind.end");
2364 case LoopVectorizationLegality::IK_PtrInduction: {
2365 EndValue = II.transform(BypassBuilder, CountRoundDown);
2366 EndValue->setName("ptr.ind.end");
2371 // The new PHI merges the original incoming value, in case of a bypass,
2372 // or the value at the end of the vectorized loop.
2373 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2374 if (OrigPhi == OldInduction)
2375 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2377 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2379 ResumeVal->addIncoming(EndValue, VecBody);
2381 // Fix the scalar body counter (PHI node).
2382 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2384 // The old induction's phi node in the scalar body needs the truncated
2386 if (OrigPhi == OldInduction) {
2387 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2388 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2390 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2391 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2395 // If we are generating a new induction variable then we also need to
2396 // generate the code that calculates the exit value. This value is not
2397 // simply the end of the counter because we may skip the vectorized body
2398 // in case of a runtime check.
2400 assert(!ResumeIndex && "Unexpected resume value found");
2401 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2402 MiddleBlock->getTerminator());
2403 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2404 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2405 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2408 // Make sure that we found the index where scalar loop needs to continue.
2409 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2410 "Invalid resume Index");
2412 // Add a check in the middle block to see if we have completed
2413 // all of the iterations in the first vector loop.
2414 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2415 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2416 ResumeIndex, "cmp.n",
2417 MiddleBlock->getTerminator());
2419 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2420 // Remove the old terminator.
2421 MiddleBlock->getTerminator()->eraseFromParent();
2423 // Create i+1 and fill the PHINode.
2424 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2425 Induction->addIncoming(StartIdx, VectorPH);
2426 Induction->addIncoming(NextIdx, VecBody);
2427 // Create the compare.
2428 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2429 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2431 // Now we have two terminators. Remove the old one from the block.
2432 VecBody->getTerminator()->eraseFromParent();
2434 // Get ready to start creating new instructions into the vectorized body.
2435 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2438 LoopVectorPreHeader = VectorPH;
2439 LoopScalarPreHeader = ScalarPH;
2440 LoopMiddleBlock = MiddleBlock;
2441 LoopExitBlock = ExitBlock;
2442 LoopVectorBody.push_back(VecBody);
2443 LoopScalarBody = OldBasicBlock;
2445 LoopVectorizeHints Hints(Lp, true);
2446 Hints.setAlreadyVectorized();
2449 /// This function returns the identity element (or neutral element) for
2450 /// the operation K.
2452 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2457 // Adding, Xoring, Oring zero to a number does not change it.
2458 return ConstantInt::get(Tp, 0);
2459 case RK_IntegerMult:
2460 // Multiplying a number by 1 does not change it.
2461 return ConstantInt::get(Tp, 1);
2463 // AND-ing a number with an all-1 value does not change it.
2464 return ConstantInt::get(Tp, -1, true);
2466 // Multiplying a number by 1 does not change it.
2467 return ConstantFP::get(Tp, 1.0L);
2469 // Adding zero to a number does not change it.
2470 return ConstantFP::get(Tp, 0.0L);
2472 llvm_unreachable("Unknown reduction kind");
2476 /// This function translates the reduction kind to an LLVM binary operator.
2478 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2480 case LoopVectorizationLegality::RK_IntegerAdd:
2481 return Instruction::Add;
2482 case LoopVectorizationLegality::RK_IntegerMult:
2483 return Instruction::Mul;
2484 case LoopVectorizationLegality::RK_IntegerOr:
2485 return Instruction::Or;
2486 case LoopVectorizationLegality::RK_IntegerAnd:
2487 return Instruction::And;
2488 case LoopVectorizationLegality::RK_IntegerXor:
2489 return Instruction::Xor;
2490 case LoopVectorizationLegality::RK_FloatMult:
2491 return Instruction::FMul;
2492 case LoopVectorizationLegality::RK_FloatAdd:
2493 return Instruction::FAdd;
2494 case LoopVectorizationLegality::RK_IntegerMinMax:
2495 return Instruction::ICmp;
2496 case LoopVectorizationLegality::RK_FloatMinMax:
2497 return Instruction::FCmp;
2499 llvm_unreachable("Unknown reduction operation");
2503 Value *createMinMaxOp(IRBuilder<> &Builder,
2504 LoopVectorizationLegality::MinMaxReductionKind RK,
2507 CmpInst::Predicate P = CmpInst::ICMP_NE;
2510 llvm_unreachable("Unknown min/max reduction kind");
2511 case LoopVectorizationLegality::MRK_UIntMin:
2512 P = CmpInst::ICMP_ULT;
2514 case LoopVectorizationLegality::MRK_UIntMax:
2515 P = CmpInst::ICMP_UGT;
2517 case LoopVectorizationLegality::MRK_SIntMin:
2518 P = CmpInst::ICMP_SLT;
2520 case LoopVectorizationLegality::MRK_SIntMax:
2521 P = CmpInst::ICMP_SGT;
2523 case LoopVectorizationLegality::MRK_FloatMin:
2524 P = CmpInst::FCMP_OLT;
2526 case LoopVectorizationLegality::MRK_FloatMax:
2527 P = CmpInst::FCMP_OGT;
2532 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2533 RK == LoopVectorizationLegality::MRK_FloatMax)
2534 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2536 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2538 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2543 struct CSEDenseMapInfo {
2544 static bool canHandle(Instruction *I) {
2545 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2546 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2548 static inline Instruction *getEmptyKey() {
2549 return DenseMapInfo<Instruction *>::getEmptyKey();
2551 static inline Instruction *getTombstoneKey() {
2552 return DenseMapInfo<Instruction *>::getTombstoneKey();
2554 static unsigned getHashValue(Instruction *I) {
2555 assert(canHandle(I) && "Unknown instruction!");
2556 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2557 I->value_op_end()));
2559 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2560 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2561 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2563 return LHS->isIdenticalTo(RHS);
2568 /// \brief Check whether this block is a predicated block.
2569 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2570 /// = ...; " blocks. We start with one vectorized basic block. For every
2571 /// conditional block we split this vectorized block. Therefore, every second
2572 /// block will be a predicated one.
2573 static bool isPredicatedBlock(unsigned BlockNum) {
2574 return BlockNum % 2;
2577 ///\brief Perform cse of induction variable instructions.
2578 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2579 // Perform simple cse.
2580 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2581 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2582 BasicBlock *BB = BBs[i];
2583 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2584 Instruction *In = I++;
2586 if (!CSEDenseMapInfo::canHandle(In))
2589 // Check if we can replace this instruction with any of the
2590 // visited instructions.
2591 if (Instruction *V = CSEMap.lookup(In)) {
2592 In->replaceAllUsesWith(V);
2593 In->eraseFromParent();
2596 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2597 // ...;" blocks for predicated stores. Every second block is a predicated
2599 if (isPredicatedBlock(i))
2607 /// \brief Adds a 'fast' flag to floating point operations.
2608 static Value *addFastMathFlag(Value *V) {
2609 if (isa<FPMathOperator>(V)){
2610 FastMathFlags Flags;
2611 Flags.setUnsafeAlgebra();
2612 cast<Instruction>(V)->setFastMathFlags(Flags);
2617 void InnerLoopVectorizer::vectorizeLoop() {
2618 //===------------------------------------------------===//
2620 // Notice: any optimization or new instruction that go
2621 // into the code below should be also be implemented in
2624 //===------------------------------------------------===//
2625 Constant *Zero = Builder.getInt32(0);
2627 // In order to support reduction variables we need to be able to vectorize
2628 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2629 // stages. First, we create a new vector PHI node with no incoming edges.
2630 // We use this value when we vectorize all of the instructions that use the
2631 // PHI. Next, after all of the instructions in the block are complete we
2632 // add the new incoming edges to the PHI. At this point all of the
2633 // instructions in the basic block are vectorized, so we can use them to
2634 // construct the PHI.
2635 PhiVector RdxPHIsToFix;
2637 // Scan the loop in a topological order to ensure that defs are vectorized
2639 LoopBlocksDFS DFS(OrigLoop);
2642 // Vectorize all of the blocks in the original loop.
2643 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2644 be = DFS.endRPO(); bb != be; ++bb)
2645 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2647 // At this point every instruction in the original loop is widened to
2648 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2649 // that we vectorized. The PHI nodes are currently empty because we did
2650 // not want to introduce cycles. Notice that the remaining PHI nodes
2651 // that we need to fix are reduction variables.
2653 // Create the 'reduced' values for each of the induction vars.
2654 // The reduced values are the vector values that we scalarize and combine
2655 // after the loop is finished.
2656 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2658 PHINode *RdxPhi = *it;
2659 assert(RdxPhi && "Unable to recover vectorized PHI");
2661 // Find the reduction variable descriptor.
2662 assert(Legal->getReductionVars()->count(RdxPhi) &&
2663 "Unable to find the reduction variable");
2664 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2665 (*Legal->getReductionVars())[RdxPhi];
2667 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2669 // We need to generate a reduction vector from the incoming scalar.
2670 // To do so, we need to generate the 'identity' vector and override
2671 // one of the elements with the incoming scalar reduction. We need
2672 // to do it in the vector-loop preheader.
2673 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
2675 // This is the vector-clone of the value that leaves the loop.
2676 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2677 Type *VecTy = VectorExit[0]->getType();
2679 // Find the reduction identity variable. Zero for addition, or, xor,
2680 // one for multiplication, -1 for And.
2683 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2684 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2685 // MinMax reduction have the start value as their identify.
2687 VectorStart = Identity = RdxDesc.StartValue;
2689 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2694 // Handle other reduction kinds:
2696 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2697 VecTy->getScalarType());
2700 // This vector is the Identity vector where the first element is the
2701 // incoming scalar reduction.
2702 VectorStart = RdxDesc.StartValue;
2704 Identity = ConstantVector::getSplat(VF, Iden);
2706 // This vector is the Identity vector where the first element is the
2707 // incoming scalar reduction.
2708 VectorStart = Builder.CreateInsertElement(Identity,
2709 RdxDesc.StartValue, Zero);
2713 // Fix the vector-loop phi.
2715 // Reductions do not have to start at zero. They can start with
2716 // any loop invariant values.
2717 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2718 BasicBlock *Latch = OrigLoop->getLoopLatch();
2719 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2720 VectorParts &Val = getVectorValue(LoopVal);
2721 for (unsigned part = 0; part < UF; ++part) {
2722 // Make sure to add the reduction stat value only to the
2723 // first unroll part.
2724 Value *StartVal = (part == 0) ? VectorStart : Identity;
2725 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
2726 LoopVectorPreHeader);
2727 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2728 LoopVectorBody.back());
2731 // Before each round, move the insertion point right between
2732 // the PHIs and the values we are going to write.
2733 // This allows us to write both PHINodes and the extractelement
2735 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2737 VectorParts RdxParts;
2738 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2739 for (unsigned part = 0; part < UF; ++part) {
2740 // This PHINode contains the vectorized reduction variable, or
2741 // the initial value vector, if we bypass the vector loop.
2742 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2743 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2744 Value *StartVal = (part == 0) ? VectorStart : Identity;
2745 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2746 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2747 NewPhi->addIncoming(RdxExitVal[part],
2748 LoopVectorBody.back());
2749 RdxParts.push_back(NewPhi);
2752 // Reduce all of the unrolled parts into a single vector.
2753 Value *ReducedPartRdx = RdxParts[0];
2754 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2755 setDebugLocFromInst(Builder, ReducedPartRdx);
2756 for (unsigned part = 1; part < UF; ++part) {
2757 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2758 // Floating point operations had to be 'fast' to enable the reduction.
2759 ReducedPartRdx = addFastMathFlag(
2760 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2761 ReducedPartRdx, "bin.rdx"));
2763 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2764 ReducedPartRdx, RdxParts[part]);
2768 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2769 // and vector ops, reducing the set of values being computed by half each
2771 assert(isPowerOf2_32(VF) &&
2772 "Reduction emission only supported for pow2 vectors!");
2773 Value *TmpVec = ReducedPartRdx;
2774 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2775 for (unsigned i = VF; i != 1; i >>= 1) {
2776 // Move the upper half of the vector to the lower half.
2777 for (unsigned j = 0; j != i/2; ++j)
2778 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2780 // Fill the rest of the mask with undef.
2781 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2782 UndefValue::get(Builder.getInt32Ty()));
2785 Builder.CreateShuffleVector(TmpVec,
2786 UndefValue::get(TmpVec->getType()),
2787 ConstantVector::get(ShuffleMask),
2790 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2791 // Floating point operations had to be 'fast' to enable the reduction.
2792 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2793 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2795 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2798 // The result is in the first element of the vector.
2799 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2800 Builder.getInt32(0));
2803 // Create a phi node that merges control-flow from the backedge-taken check
2804 // block and the middle block.
2805 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
2806 LoopScalarPreHeader->getTerminator());
2807 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
2808 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2810 // Now, we need to fix the users of the reduction variable
2811 // inside and outside of the scalar remainder loop.
2812 // We know that the loop is in LCSSA form. We need to update the
2813 // PHI nodes in the exit blocks.
2814 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2815 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2816 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2817 if (!LCSSAPhi) break;
2819 // All PHINodes need to have a single entry edge, or two if
2820 // we already fixed them.
2821 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2823 // We found our reduction value exit-PHI. Update it with the
2824 // incoming bypass edge.
2825 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2826 // Add an edge coming from the bypass.
2827 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2830 }// end of the LCSSA phi scan.
2832 // Fix the scalar loop reduction variable with the incoming reduction sum
2833 // from the vector body and from the backedge value.
2834 int IncomingEdgeBlockIdx =
2835 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2836 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2837 // Pick the other block.
2838 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2839 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
2840 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2841 }// end of for each redux variable.
2845 // Remove redundant induction instructions.
2846 cse(LoopVectorBody);
2849 void InnerLoopVectorizer::fixLCSSAPHIs() {
2850 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2851 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2852 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2853 if (!LCSSAPhi) break;
2854 if (LCSSAPhi->getNumIncomingValues() == 1)
2855 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2860 InnerLoopVectorizer::VectorParts
2861 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2862 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2865 // Look for cached value.
2866 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2867 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2868 if (ECEntryIt != MaskCache.end())
2869 return ECEntryIt->second;
2871 VectorParts SrcMask = createBlockInMask(Src);
2873 // The terminator has to be a branch inst!
2874 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2875 assert(BI && "Unexpected terminator found");
2877 if (BI->isConditional()) {
2878 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2880 if (BI->getSuccessor(0) != Dst)
2881 for (unsigned part = 0; part < UF; ++part)
2882 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2884 for (unsigned part = 0; part < UF; ++part)
2885 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2887 MaskCache[Edge] = EdgeMask;
2891 MaskCache[Edge] = SrcMask;
2895 InnerLoopVectorizer::VectorParts
2896 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2897 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2899 // Loop incoming mask is all-one.
2900 if (OrigLoop->getHeader() == BB) {
2901 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2902 return getVectorValue(C);
2905 // This is the block mask. We OR all incoming edges, and with zero.
2906 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2907 VectorParts BlockMask = getVectorValue(Zero);
2910 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2911 VectorParts EM = createEdgeMask(*it, BB);
2912 for (unsigned part = 0; part < UF; ++part)
2913 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2919 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2920 InnerLoopVectorizer::VectorParts &Entry,
2921 unsigned UF, unsigned VF, PhiVector *PV) {
2922 PHINode* P = cast<PHINode>(PN);
2923 // Handle reduction variables:
2924 if (Legal->getReductionVars()->count(P)) {
2925 for (unsigned part = 0; part < UF; ++part) {
2926 // This is phase one of vectorizing PHIs.
2927 Type *VecTy = (VF == 1) ? PN->getType() :
2928 VectorType::get(PN->getType(), VF);
2929 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2930 LoopVectorBody.back()-> getFirstInsertionPt());
2936 setDebugLocFromInst(Builder, P);
2937 // Check for PHI nodes that are lowered to vector selects.
2938 if (P->getParent() != OrigLoop->getHeader()) {
2939 // We know that all PHIs in non-header blocks are converted into
2940 // selects, so we don't have to worry about the insertion order and we
2941 // can just use the builder.
2942 // At this point we generate the predication tree. There may be
2943 // duplications since this is a simple recursive scan, but future
2944 // optimizations will clean it up.
2946 unsigned NumIncoming = P->getNumIncomingValues();
2948 // Generate a sequence of selects of the form:
2949 // SELECT(Mask3, In3,
2950 // SELECT(Mask2, In2,
2952 for (unsigned In = 0; In < NumIncoming; In++) {
2953 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2955 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2957 for (unsigned part = 0; part < UF; ++part) {
2958 // We might have single edge PHIs (blocks) - use an identity
2959 // 'select' for the first PHI operand.
2961 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2964 // Select between the current value and the previous incoming edge
2965 // based on the incoming mask.
2966 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2967 Entry[part], "predphi");
2973 // This PHINode must be an induction variable.
2974 // Make sure that we know about it.
2975 assert(Legal->getInductionVars()->count(P) &&
2976 "Not an induction variable");
2978 LoopVectorizationLegality::InductionInfo II =
2979 Legal->getInductionVars()->lookup(P);
2981 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
2982 // which can be found from the original scalar operations.
2984 case LoopVectorizationLegality::IK_NoInduction:
2985 llvm_unreachable("Unknown induction");
2986 case LoopVectorizationLegality::IK_IntInduction: {
2987 assert(P->getType() == II.StartValue->getType() && "Types must match");
2988 Type *PhiTy = P->getType();
2990 if (P == OldInduction) {
2991 // Handle the canonical induction variable. We might have had to
2993 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2995 // Handle other induction variables that are now based on the
2997 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2999 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
3000 Broadcasted = II.transform(Builder, NormalizedIdx);
3001 Broadcasted->setName("offset.idx");
3003 Broadcasted = getBroadcastInstrs(Broadcasted);
3004 // After broadcasting the induction variable we need to make the vector
3005 // consecutive by adding 0, 1, 2, etc.
3006 for (unsigned part = 0; part < UF; ++part)
3007 Entry[part] = getStepVector(Broadcasted, VF * part, II.StepValue);
3010 case LoopVectorizationLegality::IK_PtrInduction:
3011 // Handle the pointer induction variable case.
3012 assert(P->getType()->isPointerTy() && "Unexpected type.");
3013 // This is the normalized GEP that starts counting at zero.
3014 Value *NormalizedIdx =
3015 Builder.CreateSub(Induction, ExtendedIdx, "normalized.idx");
3016 // This is the vector of results. Notice that we don't generate
3017 // vector geps because scalar geps result in better code.
3018 for (unsigned part = 0; part < UF; ++part) {
3020 int EltIndex = part;
3021 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3022 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3023 Value *SclrGep = II.transform(Builder, GlobalIdx);
3024 SclrGep->setName("next.gep");
3025 Entry[part] = SclrGep;
3029 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3030 for (unsigned int i = 0; i < VF; ++i) {
3031 int EltIndex = i + part * VF;
3032 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3033 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3034 Value *SclrGep = II.transform(Builder, GlobalIdx);
3035 SclrGep->setName("next.gep");
3036 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3037 Builder.getInt32(i),
3040 Entry[part] = VecVal;
3046 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3047 // For each instruction in the old loop.
3048 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3049 VectorParts &Entry = WidenMap.get(it);
3050 switch (it->getOpcode()) {
3051 case Instruction::Br:
3052 // Nothing to do for PHIs and BR, since we already took care of the
3053 // loop control flow instructions.
3055 case Instruction::PHI: {
3056 // Vectorize PHINodes.
3057 widenPHIInstruction(it, Entry, UF, VF, PV);
3061 case Instruction::Add:
3062 case Instruction::FAdd:
3063 case Instruction::Sub:
3064 case Instruction::FSub:
3065 case Instruction::Mul:
3066 case Instruction::FMul:
3067 case Instruction::UDiv:
3068 case Instruction::SDiv:
3069 case Instruction::FDiv:
3070 case Instruction::URem:
3071 case Instruction::SRem:
3072 case Instruction::FRem:
3073 case Instruction::Shl:
3074 case Instruction::LShr:
3075 case Instruction::AShr:
3076 case Instruction::And:
3077 case Instruction::Or:
3078 case Instruction::Xor: {
3079 // Just widen binops.
3080 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3081 setDebugLocFromInst(Builder, BinOp);
3082 VectorParts &A = getVectorValue(it->getOperand(0));
3083 VectorParts &B = getVectorValue(it->getOperand(1));
3085 // Use this vector value for all users of the original instruction.
3086 for (unsigned Part = 0; Part < UF; ++Part) {
3087 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3089 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3090 VecOp->copyIRFlags(BinOp);
3095 propagateMetadata(Entry, it);
3098 case Instruction::Select: {
3100 // If the selector is loop invariant we can create a select
3101 // instruction with a scalar condition. Otherwise, use vector-select.
3102 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3104 setDebugLocFromInst(Builder, it);
3106 // The condition can be loop invariant but still defined inside the
3107 // loop. This means that we can't just use the original 'cond' value.
3108 // We have to take the 'vectorized' value and pick the first lane.
3109 // Instcombine will make this a no-op.
3110 VectorParts &Cond = getVectorValue(it->getOperand(0));
3111 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3112 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3114 Value *ScalarCond = (VF == 1) ? Cond[0] :
3115 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3117 for (unsigned Part = 0; Part < UF; ++Part) {
3118 Entry[Part] = Builder.CreateSelect(
3119 InvariantCond ? ScalarCond : Cond[Part],
3124 propagateMetadata(Entry, it);
3128 case Instruction::ICmp:
3129 case Instruction::FCmp: {
3130 // Widen compares. Generate vector compares.
3131 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3132 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3133 setDebugLocFromInst(Builder, it);
3134 VectorParts &A = getVectorValue(it->getOperand(0));
3135 VectorParts &B = getVectorValue(it->getOperand(1));
3136 for (unsigned Part = 0; Part < UF; ++Part) {
3139 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3141 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3145 propagateMetadata(Entry, it);
3149 case Instruction::Store:
3150 case Instruction::Load:
3151 vectorizeMemoryInstruction(it);
3153 case Instruction::ZExt:
3154 case Instruction::SExt:
3155 case Instruction::FPToUI:
3156 case Instruction::FPToSI:
3157 case Instruction::FPExt:
3158 case Instruction::PtrToInt:
3159 case Instruction::IntToPtr:
3160 case Instruction::SIToFP:
3161 case Instruction::UIToFP:
3162 case Instruction::Trunc:
3163 case Instruction::FPTrunc:
3164 case Instruction::BitCast: {
3165 CastInst *CI = dyn_cast<CastInst>(it);
3166 setDebugLocFromInst(Builder, it);
3167 /// Optimize the special case where the source is the induction
3168 /// variable. Notice that we can only optimize the 'trunc' case
3169 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3170 /// c. other casts depend on pointer size.
3171 if (CI->getOperand(0) == OldInduction &&
3172 it->getOpcode() == Instruction::Trunc) {
3173 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3175 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3176 LoopVectorizationLegality::InductionInfo II =
3177 Legal->getInductionVars()->lookup(OldInduction);
3179 ConstantInt::getSigned(CI->getType(), II.StepValue->getSExtValue());
3180 for (unsigned Part = 0; Part < UF; ++Part)
3181 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
3182 propagateMetadata(Entry, it);
3185 /// Vectorize casts.
3186 Type *DestTy = (VF == 1) ? CI->getType() :
3187 VectorType::get(CI->getType(), VF);
3189 VectorParts &A = getVectorValue(it->getOperand(0));
3190 for (unsigned Part = 0; Part < UF; ++Part)
3191 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3192 propagateMetadata(Entry, it);
3196 case Instruction::Call: {
3197 // Ignore dbg intrinsics.
3198 if (isa<DbgInfoIntrinsic>(it))
3200 setDebugLocFromInst(Builder, it);
3202 Module *M = BB->getParent()->getParent();
3203 CallInst *CI = cast<CallInst>(it);
3204 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3205 assert(ID && "Not an intrinsic call!");
3207 case Intrinsic::assume:
3208 case Intrinsic::lifetime_end:
3209 case Intrinsic::lifetime_start:
3210 scalarizeInstruction(it);
3213 bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
3214 for (unsigned Part = 0; Part < UF; ++Part) {
3215 SmallVector<Value *, 4> Args;
3216 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3217 if (HasScalarOpd && i == 1) {
3218 Args.push_back(CI->getArgOperand(i));
3221 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3222 Args.push_back(Arg[Part]);
3224 Type *Tys[] = {CI->getType()};
3226 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3228 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3229 Entry[Part] = Builder.CreateCall(F, Args);
3232 propagateMetadata(Entry, it);
3239 // All other instructions are unsupported. Scalarize them.
3240 scalarizeInstruction(it);
3243 }// end of for_each instr.
3246 void InnerLoopVectorizer::updateAnalysis() {
3247 // Forget the original basic block.
3248 SE->forgetLoop(OrigLoop);
3250 // Update the dominator tree information.
3251 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3252 "Entry does not dominate exit.");
3254 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3255 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3256 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3258 // Due to if predication of stores we might create a sequence of "if(pred)
3259 // a[i] = ...; " blocks.
3260 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3262 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3263 else if (isPredicatedBlock(i)) {
3264 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3266 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3270 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3271 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3272 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3273 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3275 DEBUG(DT->verifyDomTree());
3278 /// \brief Check whether it is safe to if-convert this phi node.
3280 /// Phi nodes with constant expressions that can trap are not safe to if
3282 static bool canIfConvertPHINodes(BasicBlock *BB) {
3283 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3284 PHINode *Phi = dyn_cast<PHINode>(I);
3287 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3288 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3295 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3296 if (!EnableIfConversion) {
3297 emitAnalysis(VectorizationReport() << "if-conversion is disabled");
3301 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3303 // A list of pointers that we can safely read and write to.
3304 SmallPtrSet<Value *, 8> SafePointes;
3306 // Collect safe addresses.
3307 for (Loop::block_iterator BI = TheLoop->block_begin(),
3308 BE = TheLoop->block_end(); BI != BE; ++BI) {
3309 BasicBlock *BB = *BI;
3311 if (blockNeedsPredication(BB))
3314 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3315 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3316 SafePointes.insert(LI->getPointerOperand());
3317 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3318 SafePointes.insert(SI->getPointerOperand());
3322 // Collect the blocks that need predication.
3323 BasicBlock *Header = TheLoop->getHeader();
3324 for (Loop::block_iterator BI = TheLoop->block_begin(),
3325 BE = TheLoop->block_end(); BI != BE; ++BI) {
3326 BasicBlock *BB = *BI;
3328 // We don't support switch statements inside loops.
3329 if (!isa<BranchInst>(BB->getTerminator())) {
3330 emitAnalysis(VectorizationReport(BB->getTerminator())
3331 << "loop contains a switch statement");
3335 // We must be able to predicate all blocks that need to be predicated.
3336 if (blockNeedsPredication(BB)) {
3337 if (!blockCanBePredicated(BB, SafePointes)) {
3338 emitAnalysis(VectorizationReport(BB->getTerminator())
3339 << "control flow cannot be substituted for a select");
3342 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3343 emitAnalysis(VectorizationReport(BB->getTerminator())
3344 << "control flow cannot be substituted for a select");
3349 // We can if-convert this loop.
3353 bool LoopVectorizationLegality::canVectorize() {
3354 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3355 // be canonicalized.
3356 if (!TheLoop->getLoopPreheader()) {
3358 VectorizationReport() <<
3359 "loop control flow is not understood by vectorizer");
3363 // We can only vectorize innermost loops.
3364 if (!TheLoop->getSubLoopsVector().empty()) {
3365 emitAnalysis(VectorizationReport() << "loop is not the innermost loop");
3369 // We must have a single backedge.
3370 if (TheLoop->getNumBackEdges() != 1) {
3372 VectorizationReport() <<
3373 "loop control flow is not understood by vectorizer");
3377 // We must have a single exiting block.
3378 if (!TheLoop->getExitingBlock()) {
3380 VectorizationReport() <<
3381 "loop control flow is not understood by vectorizer");
3385 // We only handle bottom-tested loops, i.e. loop in which the condition is
3386 // checked at the end of each iteration. With that we can assume that all
3387 // instructions in the loop are executed the same number of times.
3388 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
3390 VectorizationReport() <<
3391 "loop control flow is not understood by vectorizer");
3395 // We need to have a loop header.
3396 DEBUG(dbgs() << "LV: Found a loop: " <<
3397 TheLoop->getHeader()->getName() << '\n');
3399 // Check if we can if-convert non-single-bb loops.
3400 unsigned NumBlocks = TheLoop->getNumBlocks();
3401 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3402 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3406 // ScalarEvolution needs to be able to find the exit count.
3407 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3408 if (ExitCount == SE->getCouldNotCompute()) {
3409 emitAnalysis(VectorizationReport() <<
3410 "could not determine number of loop iterations");
3411 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3415 // Check if we can vectorize the instructions and CFG in this loop.
3416 if (!canVectorizeInstrs()) {
3417 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3421 // Go over each instruction and look at memory deps.
3422 if (!canVectorizeMemory()) {
3423 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3427 // Collect all of the variables that remain uniform after vectorization.
3428 collectLoopUniforms();
3430 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3431 (LAI->getRuntimePointerCheck()->Need ? " (with a runtime bound check)" :
3435 // Okay! We can vectorize. At this point we don't have any other mem analysis
3436 // which may limit our maximum vectorization factor, so just return true with
3441 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3442 if (Ty->isPointerTy())
3443 return DL.getIntPtrType(Ty);
3445 // It is possible that char's or short's overflow when we ask for the loop's
3446 // trip count, work around this by changing the type size.
3447 if (Ty->getScalarSizeInBits() < 32)
3448 return Type::getInt32Ty(Ty->getContext());
3453 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3454 Ty0 = convertPointerToIntegerType(DL, Ty0);
3455 Ty1 = convertPointerToIntegerType(DL, Ty1);
3456 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3461 /// \brief Check that the instruction has outside loop users and is not an
3462 /// identified reduction variable.
3463 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3464 SmallPtrSetImpl<Value *> &Reductions) {
3465 // Reduction instructions are allowed to have exit users. All other
3466 // instructions must not have external users.
3467 if (!Reductions.count(Inst))
3468 //Check that all of the users of the loop are inside the BB.
3469 for (User *U : Inst->users()) {
3470 Instruction *UI = cast<Instruction>(U);
3471 // This user may be a reduction exit value.
3472 if (!TheLoop->contains(UI)) {
3473 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3480 bool LoopVectorizationLegality::canVectorizeInstrs() {
3481 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3482 BasicBlock *Header = TheLoop->getHeader();
3484 // Look for the attribute signaling the absence of NaNs.
3485 Function &F = *Header->getParent();
3486 if (F.hasFnAttribute("no-nans-fp-math"))
3488 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
3490 // For each block in the loop.
3491 for (Loop::block_iterator bb = TheLoop->block_begin(),
3492 be = TheLoop->block_end(); bb != be; ++bb) {
3494 // Scan the instructions in the block and look for hazards.
3495 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3498 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3499 Type *PhiTy = Phi->getType();
3500 // Check that this PHI type is allowed.
3501 if (!PhiTy->isIntegerTy() &&
3502 !PhiTy->isFloatingPointTy() &&
3503 !PhiTy->isPointerTy()) {
3504 emitAnalysis(VectorizationReport(it)
3505 << "loop control flow is not understood by vectorizer");
3506 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3510 // If this PHINode is not in the header block, then we know that we
3511 // can convert it to select during if-conversion. No need to check if
3512 // the PHIs in this block are induction or reduction variables.
3513 if (*bb != Header) {
3514 // Check that this instruction has no outside users or is an
3515 // identified reduction value with an outside user.
3516 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3518 emitAnalysis(VectorizationReport(it) <<
3519 "value could not be identified as "
3520 "an induction or reduction variable");
3524 // We only allow if-converted PHIs with exactly two incoming values.
3525 if (Phi->getNumIncomingValues() != 2) {
3526 emitAnalysis(VectorizationReport(it)
3527 << "control flow not understood by vectorizer");
3528 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3532 // This is the value coming from the preheader.
3533 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3534 ConstantInt *StepValue = nullptr;
3535 // Check if this is an induction variable.
3536 InductionKind IK = isInductionVariable(Phi, StepValue);
3538 if (IK_NoInduction != IK) {
3539 // Get the widest type.
3541 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3543 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3545 // Int inductions are special because we only allow one IV.
3546 if (IK == IK_IntInduction && StepValue->isOne()) {
3547 // Use the phi node with the widest type as induction. Use the last
3548 // one if there are multiple (no good reason for doing this other
3549 // than it is expedient).
3550 if (!Induction || PhiTy == WidestIndTy)
3554 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3555 Inductions[Phi] = InductionInfo(StartValue, IK, StepValue);
3557 // Until we explicitly handle the case of an induction variable with
3558 // an outside loop user we have to give up vectorizing this loop.
3559 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3560 emitAnalysis(VectorizationReport(it) <<
3561 "use of induction value outside of the "
3562 "loop is not handled by vectorizer");
3569 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3570 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3573 if (AddReductionVar(Phi, RK_IntegerMult)) {
3574 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3577 if (AddReductionVar(Phi, RK_IntegerOr)) {
3578 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3581 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3582 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3585 if (AddReductionVar(Phi, RK_IntegerXor)) {
3586 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3589 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3590 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3593 if (AddReductionVar(Phi, RK_FloatMult)) {
3594 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3597 if (AddReductionVar(Phi, RK_FloatAdd)) {
3598 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3601 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3602 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3607 emitAnalysis(VectorizationReport(it) <<
3608 "value that could not be identified as "
3609 "reduction is used outside the loop");
3610 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3612 }// end of PHI handling
3614 // We still don't handle functions. However, we can ignore dbg intrinsic
3615 // calls and we do handle certain intrinsic and libm functions.
3616 CallInst *CI = dyn_cast<CallInst>(it);
3617 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3618 emitAnalysis(VectorizationReport(it) <<
3619 "call instruction cannot be vectorized");
3620 DEBUG(dbgs() << "LV: Found a call site.\n");
3624 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
3625 // second argument is the same (i.e. loop invariant)
3627 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
3628 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
3629 emitAnalysis(VectorizationReport(it)
3630 << "intrinsic instruction cannot be vectorized");
3631 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
3636 // Check that the instruction return type is vectorizable.
3637 // Also, we can't vectorize extractelement instructions.
3638 if ((!VectorType::isValidElementType(it->getType()) &&
3639 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3640 emitAnalysis(VectorizationReport(it)
3641 << "instruction return type cannot be vectorized");
3642 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3646 // Check that the stored type is vectorizable.
3647 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3648 Type *T = ST->getValueOperand()->getType();
3649 if (!VectorType::isValidElementType(T)) {
3650 emitAnalysis(VectorizationReport(ST) <<
3651 "store instruction cannot be vectorized");
3654 if (EnableMemAccessVersioning)
3655 collectStridedAccess(ST);
3658 if (EnableMemAccessVersioning)
3659 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3660 collectStridedAccess(LI);
3662 // Reduction instructions are allowed to have exit users.
3663 // All other instructions must not have external users.
3664 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3665 emitAnalysis(VectorizationReport(it) <<
3666 "value cannot be used outside the loop");
3675 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3676 if (Inductions.empty()) {
3677 emitAnalysis(VectorizationReport()
3678 << "loop induction variable could not be identified");
3686 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3687 /// return the induction operand of the gep pointer.
3688 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3689 const DataLayout *DL, Loop *Lp) {
3690 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3694 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3696 // Check that all of the gep indices are uniform except for our induction
3698 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3699 if (i != InductionOperand &&
3700 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3702 return GEP->getOperand(InductionOperand);
3705 ///\brief Look for a cast use of the passed value.
3706 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3707 Value *UniqueCast = nullptr;
3708 for (User *U : Ptr->users()) {
3709 CastInst *CI = dyn_cast<CastInst>(U);
3710 if (CI && CI->getType() == Ty) {
3720 ///\brief Get the stride of a pointer access in a loop.
3721 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3722 /// pointer to the Value, or null otherwise.
3723 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3724 const DataLayout *DL, Loop *Lp) {
3725 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3726 if (!PtrTy || PtrTy->isAggregateType())
3729 // Try to remove a gep instruction to make the pointer (actually index at this
3730 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3731 // pointer, otherwise, we are analyzing the index.
3732 Value *OrigPtr = Ptr;
3734 // The size of the pointer access.
3735 int64_t PtrAccessSize = 1;
3737 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3738 const SCEV *V = SE->getSCEV(Ptr);
3742 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3743 V = C->getOperand();
3745 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3749 V = S->getStepRecurrence(*SE);
3753 // Strip off the size of access multiplication if we are still analyzing the
3755 if (OrigPtr == Ptr) {
3756 DL->getTypeAllocSize(PtrTy->getElementType());
3757 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3758 if (M->getOperand(0)->getSCEVType() != scConstant)
3761 const APInt &APStepVal =
3762 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3764 // Huge step value - give up.
3765 if (APStepVal.getBitWidth() > 64)
3768 int64_t StepVal = APStepVal.getSExtValue();
3769 if (PtrAccessSize != StepVal)
3771 V = M->getOperand(1);
3776 Type *StripedOffRecurrenceCast = nullptr;
3777 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3778 StripedOffRecurrenceCast = C->getType();
3779 V = C->getOperand();
3782 // Look for the loop invariant symbolic value.
3783 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3787 Value *Stride = U->getValue();
3788 if (!Lp->isLoopInvariant(Stride))
3791 // If we have stripped off the recurrence cast we have to make sure that we
3792 // return the value that is used in this loop so that we can replace it later.
3793 if (StripedOffRecurrenceCast)
3794 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3799 void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) {
3800 Value *Ptr = nullptr;
3801 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3802 Ptr = LI->getPointerOperand();
3803 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3804 Ptr = SI->getPointerOperand();
3808 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3812 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3813 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3814 Strides[Ptr] = Stride;
3815 StrideSet.insert(Stride);
3818 void LoopVectorizationLegality::collectLoopUniforms() {
3819 // We now know that the loop is vectorizable!
3820 // Collect variables that will remain uniform after vectorization.
3821 std::vector<Value*> Worklist;
3822 BasicBlock *Latch = TheLoop->getLoopLatch();
3824 // Start with the conditional branch and walk up the block.
3825 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3827 // Also add all consecutive pointer values; these values will be uniform
3828 // after vectorization (and subsequent cleanup) and, until revectorization is
3829 // supported, all dependencies must also be uniform.
3830 for (Loop::block_iterator B = TheLoop->block_begin(),
3831 BE = TheLoop->block_end(); B != BE; ++B)
3832 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
3834 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
3835 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3837 while (!Worklist.empty()) {
3838 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3839 Worklist.pop_back();
3841 // Look at instructions inside this loop.
3842 // Stop when reaching PHI nodes.
3843 // TODO: we need to follow values all over the loop, not only in this block.
3844 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3847 // This is a known uniform.
3850 // Insert all operands.
3851 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3855 bool LoopVectorizationLegality::canVectorizeMemory() {
3856 LAI = &LAA->getInfo(TheLoop, Strides);
3857 auto &OptionalReport = LAI->getReport();
3859 emitAnalysis(VectorizationReport(*OptionalReport));
3860 return LAI->canVectorizeMemory();
3863 static bool hasMultipleUsesOf(Instruction *I,
3864 SmallPtrSetImpl<Instruction *> &Insts) {
3865 unsigned NumUses = 0;
3866 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
3867 if (Insts.count(dyn_cast<Instruction>(*Use)))
3876 static bool areAllUsesIn(Instruction *I, SmallPtrSetImpl<Instruction *> &Set) {
3877 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
3878 if (!Set.count(dyn_cast<Instruction>(*Use)))
3883 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
3884 ReductionKind Kind) {
3885 if (Phi->getNumIncomingValues() != 2)
3888 // Reduction variables are only found in the loop header block.
3889 if (Phi->getParent() != TheLoop->getHeader())
3892 // Obtain the reduction start value from the value that comes from the loop
3894 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
3896 // ExitInstruction is the single value which is used outside the loop.
3897 // We only allow for a single reduction value to be used outside the loop.
3898 // This includes users of the reduction, variables (which form a cycle
3899 // which ends in the phi node).
3900 Instruction *ExitInstruction = nullptr;
3901 // Indicates that we found a reduction operation in our scan.
3902 bool FoundReduxOp = false;
3904 // We start with the PHI node and scan for all of the users of this
3905 // instruction. All users must be instructions that can be used as reduction
3906 // variables (such as ADD). We must have a single out-of-block user. The cycle
3907 // must include the original PHI.
3908 bool FoundStartPHI = false;
3910 // To recognize min/max patterns formed by a icmp select sequence, we store
3911 // the number of instruction we saw from the recognized min/max pattern,
3912 // to make sure we only see exactly the two instructions.
3913 unsigned NumCmpSelectPatternInst = 0;
3914 ReductionInstDesc ReduxDesc(false, nullptr);
3916 SmallPtrSet<Instruction *, 8> VisitedInsts;
3917 SmallVector<Instruction *, 8> Worklist;
3918 Worklist.push_back(Phi);
3919 VisitedInsts.insert(Phi);
3921 // A value in the reduction can be used:
3922 // - By the reduction:
3923 // - Reduction operation:
3924 // - One use of reduction value (safe).
3925 // - Multiple use of reduction value (not safe).
3927 // - All uses of the PHI must be the reduction (safe).
3928 // - Otherwise, not safe.
3929 // - By one instruction outside of the loop (safe).
3930 // - By further instructions outside of the loop (not safe).
3931 // - By an instruction that is not part of the reduction (not safe).
3933 // * An instruction type other than PHI or the reduction operation.
3934 // * A PHI in the header other than the initial PHI.
3935 while (!Worklist.empty()) {
3936 Instruction *Cur = Worklist.back();
3937 Worklist.pop_back();
3940 // If the instruction has no users then this is a broken chain and can't be
3941 // a reduction variable.
3942 if (Cur->use_empty())
3945 bool IsAPhi = isa<PHINode>(Cur);
3947 // A header PHI use other than the original PHI.
3948 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
3951 // Reductions of instructions such as Div, and Sub is only possible if the
3952 // LHS is the reduction variable.
3953 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
3954 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
3955 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
3958 // Any reduction instruction must be of one of the allowed kinds.
3959 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
3960 if (!ReduxDesc.IsReduction)
3963 // A reduction operation must only have one use of the reduction value.
3964 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
3965 hasMultipleUsesOf(Cur, VisitedInsts))
3968 // All inputs to a PHI node must be a reduction value.
3969 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
3972 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
3973 isa<SelectInst>(Cur)))
3974 ++NumCmpSelectPatternInst;
3975 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
3976 isa<SelectInst>(Cur)))
3977 ++NumCmpSelectPatternInst;
3979 // Check whether we found a reduction operator.
3980 FoundReduxOp |= !IsAPhi;
3982 // Process users of current instruction. Push non-PHI nodes after PHI nodes
3983 // onto the stack. This way we are going to have seen all inputs to PHI
3984 // nodes once we get to them.
3985 SmallVector<Instruction *, 8> NonPHIs;
3986 SmallVector<Instruction *, 8> PHIs;
3987 for (User *U : Cur->users()) {
3988 Instruction *UI = cast<Instruction>(U);
3990 // Check if we found the exit user.
3991 BasicBlock *Parent = UI->getParent();
3992 if (!TheLoop->contains(Parent)) {
3993 // Exit if you find multiple outside users or if the header phi node is
3994 // being used. In this case the user uses the value of the previous
3995 // iteration, in which case we would loose "VF-1" iterations of the
3996 // reduction operation if we vectorize.
3997 if (ExitInstruction != nullptr || Cur == Phi)
4000 // The instruction used by an outside user must be the last instruction
4001 // before we feed back to the reduction phi. Otherwise, we loose VF-1
4002 // operations on the value.
4003 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
4006 ExitInstruction = Cur;
4010 // Process instructions only once (termination). Each reduction cycle
4011 // value must only be used once, except by phi nodes and min/max
4012 // reductions which are represented as a cmp followed by a select.
4013 ReductionInstDesc IgnoredVal(false, nullptr);
4014 if (VisitedInsts.insert(UI).second) {
4015 if (isa<PHINode>(UI))
4018 NonPHIs.push_back(UI);
4019 } else if (!isa<PHINode>(UI) &&
4020 ((!isa<FCmpInst>(UI) &&
4021 !isa<ICmpInst>(UI) &&
4022 !isa<SelectInst>(UI)) ||
4023 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
4026 // Remember that we completed the cycle.
4028 FoundStartPHI = true;
4030 Worklist.append(PHIs.begin(), PHIs.end());
4031 Worklist.append(NonPHIs.begin(), NonPHIs.end());
4034 // This means we have seen one but not the other instruction of the
4035 // pattern or more than just a select and cmp.
4036 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
4037 NumCmpSelectPatternInst != 2)
4040 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
4043 // We found a reduction var if we have reached the original phi node and we
4044 // only have a single instruction with out-of-loop users.
4046 // This instruction is allowed to have out-of-loop users.
4047 AllowedExit.insert(ExitInstruction);
4049 // Save the description of this reduction variable.
4050 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
4051 ReduxDesc.MinMaxKind);
4052 Reductions[Phi] = RD;
4053 // We've ended the cycle. This is a reduction variable if we have an
4054 // outside user and it has a binary op.
4059 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
4060 /// pattern corresponding to a min(X, Y) or max(X, Y).
4061 LoopVectorizationLegality::ReductionInstDesc
4062 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
4063 ReductionInstDesc &Prev) {
4065 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
4066 "Expect a select instruction");
4067 Instruction *Cmp = nullptr;
4068 SelectInst *Select = nullptr;
4070 // We must handle the select(cmp()) as a single instruction. Advance to the
4072 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
4073 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
4074 return ReductionInstDesc(false, I);
4075 return ReductionInstDesc(Select, Prev.MinMaxKind);
4078 // Only handle single use cases for now.
4079 if (!(Select = dyn_cast<SelectInst>(I)))
4080 return ReductionInstDesc(false, I);
4081 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
4082 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
4083 return ReductionInstDesc(false, I);
4084 if (!Cmp->hasOneUse())
4085 return ReductionInstDesc(false, I);
4090 // Look for a min/max pattern.
4091 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4092 return ReductionInstDesc(Select, MRK_UIntMin);
4093 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4094 return ReductionInstDesc(Select, MRK_UIntMax);
4095 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4096 return ReductionInstDesc(Select, MRK_SIntMax);
4097 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4098 return ReductionInstDesc(Select, MRK_SIntMin);
4099 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4100 return ReductionInstDesc(Select, MRK_FloatMin);
4101 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4102 return ReductionInstDesc(Select, MRK_FloatMax);
4103 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4104 return ReductionInstDesc(Select, MRK_FloatMin);
4105 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4106 return ReductionInstDesc(Select, MRK_FloatMax);
4108 return ReductionInstDesc(false, I);
4111 LoopVectorizationLegality::ReductionInstDesc
4112 LoopVectorizationLegality::isReductionInstr(Instruction *I,
4114 ReductionInstDesc &Prev) {
4115 bool FP = I->getType()->isFloatingPointTy();
4116 bool FastMath = FP && I->hasUnsafeAlgebra();
4117 switch (I->getOpcode()) {
4119 return ReductionInstDesc(false, I);
4120 case Instruction::PHI:
4121 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
4122 Kind != RK_FloatMinMax))
4123 return ReductionInstDesc(false, I);
4124 return ReductionInstDesc(I, Prev.MinMaxKind);
4125 case Instruction::Sub:
4126 case Instruction::Add:
4127 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4128 case Instruction::Mul:
4129 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4130 case Instruction::And:
4131 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4132 case Instruction::Or:
4133 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4134 case Instruction::Xor:
4135 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4136 case Instruction::FMul:
4137 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4138 case Instruction::FSub:
4139 case Instruction::FAdd:
4140 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4141 case Instruction::FCmp:
4142 case Instruction::ICmp:
4143 case Instruction::Select:
4144 if (Kind != RK_IntegerMinMax &&
4145 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4146 return ReductionInstDesc(false, I);
4147 return isMinMaxSelectCmpPattern(I, Prev);
4151 LoopVectorizationLegality::InductionKind
4152 LoopVectorizationLegality::isInductionVariable(PHINode *Phi,
4153 ConstantInt *&StepValue) {
4154 Type *PhiTy = Phi->getType();
4155 // We only handle integer and pointer inductions variables.
4156 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4157 return IK_NoInduction;
4159 // Check that the PHI is consecutive.
4160 const SCEV *PhiScev = SE->getSCEV(Phi);
4161 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4163 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4164 return IK_NoInduction;
4167 const SCEV *Step = AR->getStepRecurrence(*SE);
4168 // Calculate the pointer stride and check if it is consecutive.
4169 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4171 return IK_NoInduction;
4173 ConstantInt *CV = C->getValue();
4174 if (PhiTy->isIntegerTy()) {
4176 return IK_IntInduction;
4179 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4180 Type *PointerElementType = PhiTy->getPointerElementType();
4181 // The pointer stride cannot be determined if the pointer element type is not
4183 if (!PointerElementType->isSized())
4184 return IK_NoInduction;
4186 int64_t Size = static_cast<int64_t>(DL->getTypeAllocSize(PointerElementType));
4187 int64_t CVSize = CV->getSExtValue();
4189 return IK_NoInduction;
4190 StepValue = ConstantInt::getSigned(CV->getType(), CVSize / Size);
4191 return IK_PtrInduction;
4194 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4195 Value *In0 = const_cast<Value*>(V);
4196 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4200 return Inductions.count(PN);
4203 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4204 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4207 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4208 SmallPtrSetImpl<Value *> &SafePtrs) {
4210 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4211 // Check that we don't have a constant expression that can trap as operand.
4212 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4214 if (Constant *C = dyn_cast<Constant>(*OI))
4218 // We might be able to hoist the load.
4219 if (it->mayReadFromMemory()) {
4220 LoadInst *LI = dyn_cast<LoadInst>(it);
4223 if (!SafePtrs.count(LI->getPointerOperand())) {
4224 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
4225 MaskedOp.insert(LI);
4232 // We don't predicate stores at the moment.
4233 if (it->mayWriteToMemory()) {
4234 StoreInst *SI = dyn_cast<StoreInst>(it);
4235 // We only support predication of stores in basic blocks with one
4240 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
4241 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
4243 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
4244 !isSinglePredecessor) {
4245 // Build a masked store if it is legal for the target, otherwise scalarize
4247 bool isLegalMaskedOp =
4248 isLegalMaskedStore(SI->getValueOperand()->getType(),
4249 SI->getPointerOperand());
4250 if (isLegalMaskedOp) {
4252 MaskedOp.insert(SI);
4261 // The instructions below can trap.
4262 switch (it->getOpcode()) {
4264 case Instruction::UDiv:
4265 case Instruction::SDiv:
4266 case Instruction::URem:
4267 case Instruction::SRem:
4275 LoopVectorizationCostModel::VectorizationFactor
4276 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
4277 // Width 1 means no vectorize
4278 VectorizationFactor Factor = { 1U, 0U };
4279 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4280 emitAnalysis(VectorizationReport() <<
4281 "runtime pointer checks needed. Enable vectorization of this "
4282 "loop with '#pragma clang loop vectorize(enable)' when "
4283 "compiling with -Os");
4284 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4288 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
4289 emitAnalysis(VectorizationReport() <<
4290 "store that is conditionally executed prevents vectorization");
4291 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
4295 // Find the trip count.
4296 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4297 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4299 unsigned WidestType = getWidestType();
4300 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4301 unsigned MaxSafeDepDist = -1U;
4302 if (Legal->getMaxSafeDepDistBytes() != -1U)
4303 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4304 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4305 WidestRegister : MaxSafeDepDist);
4306 unsigned MaxVectorSize = WidestRegister / WidestType;
4307 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4308 DEBUG(dbgs() << "LV: The Widest register is: "
4309 << WidestRegister << " bits.\n");
4311 if (MaxVectorSize == 0) {
4312 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4316 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
4317 " into one vector!");
4319 unsigned VF = MaxVectorSize;
4321 // If we optimize the program for size, avoid creating the tail loop.
4323 // If we are unable to calculate the trip count then don't try to vectorize.
4326 (VectorizationReport() <<
4327 "unable to calculate the loop count due to complex control flow");
4328 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4332 // Find the maximum SIMD width that can fit within the trip count.
4333 VF = TC % MaxVectorSize;
4338 // If the trip count that we found modulo the vectorization factor is not
4339 // zero then we require a tail.
4341 emitAnalysis(VectorizationReport() <<
4342 "cannot optimize for size and vectorize at the "
4343 "same time. Enable vectorization of this loop "
4344 "with '#pragma clang loop vectorize(enable)' "
4345 "when compiling with -Os");
4346 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4351 int UserVF = Hints->getWidth();
4353 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4354 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4356 Factor.Width = UserVF;
4360 float Cost = expectedCost(1);
4362 const float ScalarCost = Cost;
4365 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
4367 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
4368 // Ignore scalar width, because the user explicitly wants vectorization.
4369 if (ForceVectorization && VF > 1) {
4371 Cost = expectedCost(Width) / (float)Width;
4374 for (unsigned i=2; i <= VF; i*=2) {
4375 // Notice that the vector loop needs to be executed less times, so
4376 // we need to divide the cost of the vector loops by the width of
4377 // the vector elements.
4378 float VectorCost = expectedCost(i) / (float)i;
4379 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
4380 (int)VectorCost << ".\n");
4381 if (VectorCost < Cost) {
4387 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
4388 << "LV: Vectorization seems to be not beneficial, "
4389 << "but was forced by a user.\n");
4390 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
4391 Factor.Width = Width;
4392 Factor.Cost = Width * Cost;
4396 unsigned LoopVectorizationCostModel::getWidestType() {
4397 unsigned MaxWidth = 8;
4400 for (Loop::block_iterator bb = TheLoop->block_begin(),
4401 be = TheLoop->block_end(); bb != be; ++bb) {
4402 BasicBlock *BB = *bb;
4404 // For each instruction in the loop.
4405 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4406 Type *T = it->getType();
4408 // Ignore ephemeral values.
4409 if (EphValues.count(it))
4412 // Only examine Loads, Stores and PHINodes.
4413 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4416 // Examine PHI nodes that are reduction variables.
4417 if (PHINode *PN = dyn_cast<PHINode>(it))
4418 if (!Legal->getReductionVars()->count(PN))
4421 // Examine the stored values.
4422 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4423 T = ST->getValueOperand()->getType();
4425 // Ignore loaded pointer types and stored pointer types that are not
4426 // consecutive. However, we do want to take consecutive stores/loads of
4427 // pointer vectors into account.
4428 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4431 MaxWidth = std::max(MaxWidth,
4432 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
4440 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
4442 unsigned LoopCost) {
4444 // -- The unroll heuristics --
4445 // We unroll the loop in order to expose ILP and reduce the loop overhead.
4446 // There are many micro-architectural considerations that we can't predict
4447 // at this level. For example, frontend pressure (on decode or fetch) due to
4448 // code size, or the number and capabilities of the execution ports.
4450 // We use the following heuristics to select the unroll factor:
4451 // 1. If the code has reductions, then we unroll in order to break the cross
4452 // iteration dependency.
4453 // 2. If the loop is really small, then we unroll in order to reduce the loop
4455 // 3. We don't unroll if we think that we will spill registers to memory due
4456 // to the increased register pressure.
4458 // Use the user preference, unless 'auto' is selected.
4459 int UserUF = Hints->getInterleave();
4463 // When we optimize for size, we don't unroll.
4467 // We used the distance for the unroll factor.
4468 if (Legal->getMaxSafeDepDistBytes() != -1U)
4471 // Do not unroll loops with a relatively small trip count.
4472 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4473 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
4476 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
4477 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
4481 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
4482 TargetNumRegisters = ForceTargetNumScalarRegs;
4484 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
4485 TargetNumRegisters = ForceTargetNumVectorRegs;
4488 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4489 // We divide by these constants so assume that we have at least one
4490 // instruction that uses at least one register.
4491 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4492 R.NumInstructions = std::max(R.NumInstructions, 1U);
4494 // We calculate the unroll factor using the following formula.
4495 // Subtract the number of loop invariants from the number of available
4496 // registers. These registers are used by all of the unrolled instances.
4497 // Next, divide the remaining registers by the number of registers that is
4498 // required by the loop, in order to estimate how many parallel instances
4499 // fit without causing spills. All of this is rounded down if necessary to be
4500 // a power of two. We want power of two unroll factors to simplify any
4501 // addressing operations or alignment considerations.
4502 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
4505 // Don't count the induction variable as unrolled.
4506 if (EnableIndVarRegisterHeur)
4507 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
4508 std::max(1U, (R.MaxLocalUsers - 1)));
4510 // Clamp the unroll factor ranges to reasonable factors.
4511 unsigned MaxInterleaveSize = TTI.getMaxInterleaveFactor();
4513 // Check if the user has overridden the unroll max.
4515 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
4516 MaxInterleaveSize = ForceTargetMaxScalarInterleaveFactor;
4518 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
4519 MaxInterleaveSize = ForceTargetMaxVectorInterleaveFactor;
4522 // If we did not calculate the cost for VF (because the user selected the VF)
4523 // then we calculate the cost of VF here.
4525 LoopCost = expectedCost(VF);
4527 // Clamp the calculated UF to be between the 1 and the max unroll factor
4528 // that the target allows.
4529 if (UF > MaxInterleaveSize)
4530 UF = MaxInterleaveSize;
4534 // Unroll if we vectorized this loop and there is a reduction that could
4535 // benefit from unrolling.
4536 if (VF > 1 && Legal->getReductionVars()->size()) {
4537 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
4541 // Note that if we've already vectorized the loop we will have done the
4542 // runtime check and so unrolling won't require further checks.
4543 bool UnrollingRequiresRuntimePointerCheck =
4544 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
4546 // We want to unroll small loops in order to reduce the loop overhead and
4547 // potentially expose ILP opportunities.
4548 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
4549 if (!UnrollingRequiresRuntimePointerCheck &&
4550 LoopCost < SmallLoopCost) {
4551 // We assume that the cost overhead is 1 and we use the cost model
4552 // to estimate the cost of the loop and unroll until the cost of the
4553 // loop overhead is about 5% of the cost of the loop.
4554 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
4556 // Unroll until store/load ports (estimated by max unroll factor) are
4558 unsigned NumStores = Legal->getNumStores();
4559 unsigned NumLoads = Legal->getNumLoads();
4560 unsigned StoresUF = UF / (NumStores ? NumStores : 1);
4561 unsigned LoadsUF = UF / (NumLoads ? NumLoads : 1);
4563 // If we have a scalar reduction (vector reductions are already dealt with
4564 // by this point), we can increase the critical path length if the loop
4565 // we're unrolling is inside another loop. Limit, by default to 2, so the
4566 // critical path only gets increased by one reduction operation.
4567 if (Legal->getReductionVars()->size() &&
4568 TheLoop->getLoopDepth() > 1) {
4569 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionUF);
4570 SmallUF = std::min(SmallUF, F);
4571 StoresUF = std::min(StoresUF, F);
4572 LoadsUF = std::min(LoadsUF, F);
4575 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
4576 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
4577 return std::max(StoresUF, LoadsUF);
4580 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
4584 DEBUG(dbgs() << "LV: Not Unrolling.\n");
4588 LoopVectorizationCostModel::RegisterUsage
4589 LoopVectorizationCostModel::calculateRegisterUsage() {
4590 // This function calculates the register usage by measuring the highest number
4591 // of values that are alive at a single location. Obviously, this is a very
4592 // rough estimation. We scan the loop in a topological order in order and
4593 // assign a number to each instruction. We use RPO to ensure that defs are
4594 // met before their users. We assume that each instruction that has in-loop
4595 // users starts an interval. We record every time that an in-loop value is
4596 // used, so we have a list of the first and last occurrences of each
4597 // instruction. Next, we transpose this data structure into a multi map that
4598 // holds the list of intervals that *end* at a specific location. This multi
4599 // map allows us to perform a linear search. We scan the instructions linearly
4600 // and record each time that a new interval starts, by placing it in a set.
4601 // If we find this value in the multi-map then we remove it from the set.
4602 // The max register usage is the maximum size of the set.
4603 // We also search for instructions that are defined outside the loop, but are
4604 // used inside the loop. We need this number separately from the max-interval
4605 // usage number because when we unroll, loop-invariant values do not take
4607 LoopBlocksDFS DFS(TheLoop);
4611 R.NumInstructions = 0;
4613 // Each 'key' in the map opens a new interval. The values
4614 // of the map are the index of the 'last seen' usage of the
4615 // instruction that is the key.
4616 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4617 // Maps instruction to its index.
4618 DenseMap<unsigned, Instruction*> IdxToInstr;
4619 // Marks the end of each interval.
4620 IntervalMap EndPoint;
4621 // Saves the list of instruction indices that are used in the loop.
4622 SmallSet<Instruction*, 8> Ends;
4623 // Saves the list of values that are used in the loop but are
4624 // defined outside the loop, such as arguments and constants.
4625 SmallPtrSet<Value*, 8> LoopInvariants;
4628 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4629 be = DFS.endRPO(); bb != be; ++bb) {
4630 R.NumInstructions += (*bb)->size();
4631 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4633 Instruction *I = it;
4634 IdxToInstr[Index++] = I;
4636 // Save the end location of each USE.
4637 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4638 Value *U = I->getOperand(i);
4639 Instruction *Instr = dyn_cast<Instruction>(U);
4641 // Ignore non-instruction values such as arguments, constants, etc.
4642 if (!Instr) continue;
4644 // If this instruction is outside the loop then record it and continue.
4645 if (!TheLoop->contains(Instr)) {
4646 LoopInvariants.insert(Instr);
4650 // Overwrite previous end points.
4651 EndPoint[Instr] = Index;
4657 // Saves the list of intervals that end with the index in 'key'.
4658 typedef SmallVector<Instruction*, 2> InstrList;
4659 DenseMap<unsigned, InstrList> TransposeEnds;
4661 // Transpose the EndPoints to a list of values that end at each index.
4662 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4664 TransposeEnds[it->second].push_back(it->first);
4666 SmallSet<Instruction*, 8> OpenIntervals;
4667 unsigned MaxUsage = 0;
4670 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4671 for (unsigned int i = 0; i < Index; ++i) {
4672 Instruction *I = IdxToInstr[i];
4673 // Ignore instructions that are never used within the loop.
4674 if (!Ends.count(I)) continue;
4676 // Ignore ephemeral values.
4677 if (EphValues.count(I))
4680 // Remove all of the instructions that end at this location.
4681 InstrList &List = TransposeEnds[i];
4682 for (unsigned int j=0, e = List.size(); j < e; ++j)
4683 OpenIntervals.erase(List[j]);
4685 // Count the number of live interals.
4686 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4688 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4689 OpenIntervals.size() << '\n');
4691 // Add the current instruction to the list of open intervals.
4692 OpenIntervals.insert(I);
4695 unsigned Invariant = LoopInvariants.size();
4696 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
4697 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
4698 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
4700 R.LoopInvariantRegs = Invariant;
4701 R.MaxLocalUsers = MaxUsage;
4705 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4709 for (Loop::block_iterator bb = TheLoop->block_begin(),
4710 be = TheLoop->block_end(); bb != be; ++bb) {
4711 unsigned BlockCost = 0;
4712 BasicBlock *BB = *bb;
4714 // For each instruction in the old loop.
4715 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4716 // Skip dbg intrinsics.
4717 if (isa<DbgInfoIntrinsic>(it))
4720 // Ignore ephemeral values.
4721 if (EphValues.count(it))
4724 unsigned C = getInstructionCost(it, VF);
4726 // Check if we should override the cost.
4727 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
4728 C = ForceTargetInstructionCost;
4731 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
4732 VF << " For instruction: " << *it << '\n');
4735 // We assume that if-converted blocks have a 50% chance of being executed.
4736 // When the code is scalar then some of the blocks are avoided due to CF.
4737 // When the code is vectorized we execute all code paths.
4738 if (VF == 1 && Legal->blockNeedsPredication(*bb))
4747 /// \brief Check whether the address computation for a non-consecutive memory
4748 /// access looks like an unlikely candidate for being merged into the indexing
4751 /// We look for a GEP which has one index that is an induction variable and all
4752 /// other indices are loop invariant. If the stride of this access is also
4753 /// within a small bound we decide that this address computation can likely be
4754 /// merged into the addressing mode.
4755 /// In all other cases, we identify the address computation as complex.
4756 static bool isLikelyComplexAddressComputation(Value *Ptr,
4757 LoopVectorizationLegality *Legal,
4758 ScalarEvolution *SE,
4759 const Loop *TheLoop) {
4760 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
4764 // We are looking for a gep with all loop invariant indices except for one
4765 // which should be an induction variable.
4766 unsigned NumOperands = Gep->getNumOperands();
4767 for (unsigned i = 1; i < NumOperands; ++i) {
4768 Value *Opd = Gep->getOperand(i);
4769 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
4770 !Legal->isInductionVariable(Opd))
4774 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
4775 // can likely be merged into the address computation.
4776 unsigned MaxMergeDistance = 64;
4778 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
4782 // Check the step is constant.
4783 const SCEV *Step = AddRec->getStepRecurrence(*SE);
4784 // Calculate the pointer stride and check if it is consecutive.
4785 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4789 const APInt &APStepVal = C->getValue()->getValue();
4791 // Huge step value - give up.
4792 if (APStepVal.getBitWidth() > 64)
4795 int64_t StepVal = APStepVal.getSExtValue();
4797 return StepVal > MaxMergeDistance;
4800 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
4801 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
4807 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
4808 // If we know that this instruction will remain uniform, check the cost of
4809 // the scalar version.
4810 if (Legal->isUniformAfterVectorization(I))
4813 Type *RetTy = I->getType();
4814 Type *VectorTy = ToVectorTy(RetTy, VF);
4816 // TODO: We need to estimate the cost of intrinsic calls.
4817 switch (I->getOpcode()) {
4818 case Instruction::GetElementPtr:
4819 // We mark this instruction as zero-cost because the cost of GEPs in
4820 // vectorized code depends on whether the corresponding memory instruction
4821 // is scalarized or not. Therefore, we handle GEPs with the memory
4822 // instruction cost.
4824 case Instruction::Br: {
4825 return TTI.getCFInstrCost(I->getOpcode());
4827 case Instruction::PHI:
4828 //TODO: IF-converted IFs become selects.
4830 case Instruction::Add:
4831 case Instruction::FAdd:
4832 case Instruction::Sub:
4833 case Instruction::FSub:
4834 case Instruction::Mul:
4835 case Instruction::FMul:
4836 case Instruction::UDiv:
4837 case Instruction::SDiv:
4838 case Instruction::FDiv:
4839 case Instruction::URem:
4840 case Instruction::SRem:
4841 case Instruction::FRem:
4842 case Instruction::Shl:
4843 case Instruction::LShr:
4844 case Instruction::AShr:
4845 case Instruction::And:
4846 case Instruction::Or:
4847 case Instruction::Xor: {
4848 // Since we will replace the stride by 1 the multiplication should go away.
4849 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
4851 // Certain instructions can be cheaper to vectorize if they have a constant
4852 // second vector operand. One example of this are shifts on x86.
4853 TargetTransformInfo::OperandValueKind Op1VK =
4854 TargetTransformInfo::OK_AnyValue;
4855 TargetTransformInfo::OperandValueKind Op2VK =
4856 TargetTransformInfo::OK_AnyValue;
4857 TargetTransformInfo::OperandValueProperties Op1VP =
4858 TargetTransformInfo::OP_None;
4859 TargetTransformInfo::OperandValueProperties Op2VP =
4860 TargetTransformInfo::OP_None;
4861 Value *Op2 = I->getOperand(1);
4863 // Check for a splat of a constant or for a non uniform vector of constants.
4864 if (isa<ConstantInt>(Op2)) {
4865 ConstantInt *CInt = cast<ConstantInt>(Op2);
4866 if (CInt && CInt->getValue().isPowerOf2())
4867 Op2VP = TargetTransformInfo::OP_PowerOf2;
4868 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4869 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
4870 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
4871 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
4873 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
4874 if (CInt && CInt->getValue().isPowerOf2())
4875 Op2VP = TargetTransformInfo::OP_PowerOf2;
4876 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4880 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
4883 case Instruction::Select: {
4884 SelectInst *SI = cast<SelectInst>(I);
4885 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
4886 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
4887 Type *CondTy = SI->getCondition()->getType();
4889 CondTy = VectorType::get(CondTy, VF);
4891 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
4893 case Instruction::ICmp:
4894 case Instruction::FCmp: {
4895 Type *ValTy = I->getOperand(0)->getType();
4896 VectorTy = ToVectorTy(ValTy, VF);
4897 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
4899 case Instruction::Store:
4900 case Instruction::Load: {
4901 StoreInst *SI = dyn_cast<StoreInst>(I);
4902 LoadInst *LI = dyn_cast<LoadInst>(I);
4903 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
4905 VectorTy = ToVectorTy(ValTy, VF);
4907 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
4908 unsigned AS = SI ? SI->getPointerAddressSpace() :
4909 LI->getPointerAddressSpace();
4910 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
4911 // We add the cost of address computation here instead of with the gep
4912 // instruction because only here we know whether the operation is
4915 return TTI.getAddressComputationCost(VectorTy) +
4916 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4918 // Scalarized loads/stores.
4919 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
4920 bool Reverse = ConsecutiveStride < 0;
4921 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
4922 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
4923 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
4924 bool IsComplexComputation =
4925 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
4927 // The cost of extracting from the value vector and pointer vector.
4928 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
4929 for (unsigned i = 0; i < VF; ++i) {
4930 // The cost of extracting the pointer operand.
4931 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
4932 // In case of STORE, the cost of ExtractElement from the vector.
4933 // In case of LOAD, the cost of InsertElement into the returned
4935 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
4936 Instruction::InsertElement,
4940 // The cost of the scalar loads/stores.
4941 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
4942 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
4947 // Wide load/stores.
4948 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
4949 if (Legal->isMaskRequired(I))
4950 Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment,
4953 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4956 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
4960 case Instruction::ZExt:
4961 case Instruction::SExt:
4962 case Instruction::FPToUI:
4963 case Instruction::FPToSI:
4964 case Instruction::FPExt:
4965 case Instruction::PtrToInt:
4966 case Instruction::IntToPtr:
4967 case Instruction::SIToFP:
4968 case Instruction::UIToFP:
4969 case Instruction::Trunc:
4970 case Instruction::FPTrunc:
4971 case Instruction::BitCast: {
4972 // We optimize the truncation of induction variable.
4973 // The cost of these is the same as the scalar operation.
4974 if (I->getOpcode() == Instruction::Trunc &&
4975 Legal->isInductionVariable(I->getOperand(0)))
4976 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
4977 I->getOperand(0)->getType());
4979 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
4980 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
4982 case Instruction::Call: {
4983 CallInst *CI = cast<CallInst>(I);
4984 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
4985 assert(ID && "Not an intrinsic call!");
4986 Type *RetTy = ToVectorTy(CI->getType(), VF);
4987 SmallVector<Type*, 4> Tys;
4988 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
4989 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
4990 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
4993 // We are scalarizing the instruction. Return the cost of the scalar
4994 // instruction, plus the cost of insert and extract into vector
4995 // elements, times the vector width.
4998 if (!RetTy->isVoidTy() && VF != 1) {
4999 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5001 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5004 // The cost of inserting the results plus extracting each one of the
5006 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5009 // The cost of executing VF copies of the scalar instruction. This opcode
5010 // is unknown. Assume that it is the same as 'mul'.
5011 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5017 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
5018 if (Scalar->isVoidTy() || VF == 1)
5020 return VectorType::get(Scalar, VF);
5023 char LoopVectorize::ID = 0;
5024 static const char lv_name[] = "Loop Vectorization";
5025 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5026 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
5027 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
5028 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
5029 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
5030 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5031 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5032 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5033 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
5034 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5035 INITIALIZE_PASS_DEPENDENCY(LoopAccessAnalysis)
5036 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5039 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5040 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5044 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5045 // Check for a store.
5046 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5047 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5049 // Check for a load.
5050 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5051 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5057 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5058 bool IfPredicateStore) {
5059 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5060 // Holds vector parameters or scalars, in case of uniform vals.
5061 SmallVector<VectorParts, 4> Params;
5063 setDebugLocFromInst(Builder, Instr);
5065 // Find all of the vectorized parameters.
5066 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5067 Value *SrcOp = Instr->getOperand(op);
5069 // If we are accessing the old induction variable, use the new one.
5070 if (SrcOp == OldInduction) {
5071 Params.push_back(getVectorValue(SrcOp));
5075 // Try using previously calculated values.
5076 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5078 // If the src is an instruction that appeared earlier in the basic block
5079 // then it should already be vectorized.
5080 if (SrcInst && OrigLoop->contains(SrcInst)) {
5081 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5082 // The parameter is a vector value from earlier.
5083 Params.push_back(WidenMap.get(SrcInst));
5085 // The parameter is a scalar from outside the loop. Maybe even a constant.
5086 VectorParts Scalars;
5087 Scalars.append(UF, SrcOp);
5088 Params.push_back(Scalars);
5092 assert(Params.size() == Instr->getNumOperands() &&
5093 "Invalid number of operands");
5095 // Does this instruction return a value ?
5096 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5098 Value *UndefVec = IsVoidRetTy ? nullptr :
5099 UndefValue::get(Instr->getType());
5100 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5101 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5103 Instruction *InsertPt = Builder.GetInsertPoint();
5104 BasicBlock *IfBlock = Builder.GetInsertBlock();
5105 BasicBlock *CondBlock = nullptr;
5108 Loop *VectorLp = nullptr;
5109 if (IfPredicateStore) {
5110 assert(Instr->getParent()->getSinglePredecessor() &&
5111 "Only support single predecessor blocks");
5112 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5113 Instr->getParent());
5114 VectorLp = LI->getLoopFor(IfBlock);
5115 assert(VectorLp && "Must have a loop for this block");
5118 // For each vector unroll 'part':
5119 for (unsigned Part = 0; Part < UF; ++Part) {
5120 // For each scalar that we create:
5122 // Start an "if (pred) a[i] = ..." block.
5123 Value *Cmp = nullptr;
5124 if (IfPredicateStore) {
5125 if (Cond[Part]->getType()->isVectorTy())
5127 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5128 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5129 ConstantInt::get(Cond[Part]->getType(), 1));
5130 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5131 LoopVectorBody.push_back(CondBlock);
5132 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
5133 // Update Builder with newly created basic block.
5134 Builder.SetInsertPoint(InsertPt);
5137 Instruction *Cloned = Instr->clone();
5139 Cloned->setName(Instr->getName() + ".cloned");
5140 // Replace the operands of the cloned instructions with extracted scalars.
5141 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5142 Value *Op = Params[op][Part];
5143 Cloned->setOperand(op, Op);
5146 // Place the cloned scalar in the new loop.
5147 Builder.Insert(Cloned);
5149 // If the original scalar returns a value we need to place it in a vector
5150 // so that future users will be able to use it.
5152 VecResults[Part] = Cloned;
5155 if (IfPredicateStore) {
5156 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
5157 LoopVectorBody.push_back(NewIfBlock);
5158 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
5159 Builder.SetInsertPoint(InsertPt);
5160 Instruction *OldBr = IfBlock->getTerminator();
5161 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
5162 OldBr->eraseFromParent();
5163 IfBlock = NewIfBlock;
5168 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5169 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5170 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5172 return scalarizeInstruction(Instr, IfPredicateStore);
5175 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5179 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5183 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) {
5184 // When unrolling and the VF is 1, we only need to add a simple scalar.
5185 Type *ITy = Val->getType();
5186 assert(!ITy->isVectorTy() && "Val must be a scalar");
5187 Constant *C = ConstantInt::get(ITy, StartIdx);
5188 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");