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>
110 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
111 cl::desc("Sets the SIMD width. Zero is autoselect."));
113 static cl::opt<unsigned>
114 VectorizationInterleave("force-vector-interleave", cl::init(0), cl::Hidden,
115 cl::desc("Sets the vectorization interleave count. "
116 "Zero is autoselect."));
119 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
120 cl::desc("Enable if-conversion during vectorization."));
122 /// We don't vectorize loops with a known constant trip count below this number.
123 static cl::opt<unsigned>
124 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
126 cl::desc("Don't vectorize loops with a constant "
127 "trip count that is smaller than this "
130 /// This enables versioning on the strides of symbolically striding memory
131 /// accesses in code like the following.
132 /// for (i = 0; i < N; ++i)
133 /// A[i * Stride1] += B[i * Stride2] ...
135 /// Will be roughly translated to
136 /// if (Stride1 == 1 && Stride2 == 1) {
137 /// for (i = 0; i < N; i+=4)
141 static cl::opt<bool> EnableMemAccessVersioning(
142 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
143 cl::desc("Enable symblic stride memory access versioning"));
145 /// We don't unroll loops with a known constant trip count below this number.
146 static const unsigned TinyTripCountUnrollThreshold = 128;
148 /// When performing memory disambiguation checks at runtime do not make more
149 /// than this number of comparisons.
150 static const unsigned RuntimeMemoryCheckThreshold = 8;
152 /// Maximum simd width.
153 static const unsigned MaxVectorWidth = 64;
155 static cl::opt<unsigned> ForceTargetNumScalarRegs(
156 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
157 cl::desc("A flag that overrides the target's number of scalar registers."));
159 static cl::opt<unsigned> ForceTargetNumVectorRegs(
160 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
161 cl::desc("A flag that overrides the target's number of vector registers."));
163 /// Maximum vectorization interleave count.
164 static const unsigned MaxInterleaveFactor = 16;
166 static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
167 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
168 cl::desc("A flag that overrides the target's max interleave factor for "
171 static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
172 "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
173 cl::desc("A flag that overrides the target's max interleave factor for "
174 "vectorized loops."));
176 static cl::opt<unsigned> ForceTargetInstructionCost(
177 "force-target-instruction-cost", cl::init(0), cl::Hidden,
178 cl::desc("A flag that overrides the target's expected cost for "
179 "an instruction to a single constant value. Mostly "
180 "useful for getting consistent testing."));
182 static cl::opt<unsigned> SmallLoopCost(
183 "small-loop-cost", cl::init(20), cl::Hidden,
184 cl::desc("The cost of a loop that is considered 'small' by the unroller."));
186 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
187 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
188 cl::desc("Enable the use of the block frequency analysis to access PGO "
189 "heuristics minimizing code growth in cold regions and being more "
190 "aggressive in hot regions."));
192 // Runtime unroll loops for load/store throughput.
193 static cl::opt<bool> EnableLoadStoreRuntimeUnroll(
194 "enable-loadstore-runtime-unroll", cl::init(true), cl::Hidden,
195 cl::desc("Enable runtime unrolling until load/store ports are saturated"));
197 /// The number of stores in a loop that are allowed to need predication.
198 static cl::opt<unsigned> NumberOfStoresToPredicate(
199 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
200 cl::desc("Max number of stores to be predicated behind an if."));
202 static cl::opt<bool> EnableIndVarRegisterHeur(
203 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
204 cl::desc("Count the induction variable only once when unrolling"));
206 static cl::opt<bool> EnableCondStoresVectorization(
207 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
208 cl::desc("Enable if predication of stores during vectorization."));
210 static cl::opt<unsigned> MaxNestedScalarReductionUF(
211 "max-nested-scalar-reduction-unroll", cl::init(2), cl::Hidden,
212 cl::desc("The maximum unroll factor to use when unrolling a scalar "
213 "reduction in a nested loop."));
217 // Forward declarations.
218 class LoopVectorizationLegality;
219 class LoopVectorizationCostModel;
220 class LoopVectorizeHints;
222 /// InnerLoopVectorizer vectorizes loops which contain only one basic
223 /// block to a specified vectorization factor (VF).
224 /// This class performs the widening of scalars into vectors, or multiple
225 /// scalars. This class also implements the following features:
226 /// * It inserts an epilogue loop for handling loops that don't have iteration
227 /// counts that are known to be a multiple of the vectorization factor.
228 /// * It handles the code generation for reduction variables.
229 /// * Scalarization (implementation using scalars) of un-vectorizable
231 /// InnerLoopVectorizer does not perform any vectorization-legality
232 /// checks, and relies on the caller to check for the different legality
233 /// aspects. The InnerLoopVectorizer relies on the
234 /// LoopVectorizationLegality class to provide information about the induction
235 /// and reduction variables that were found to a given vectorization factor.
236 class InnerLoopVectorizer {
238 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
239 DominatorTree *DT, const DataLayout *DL,
240 const TargetLibraryInfo *TLI, unsigned VecWidth,
241 unsigned UnrollFactor)
242 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
243 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
244 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
247 // Perform the actual loop widening (vectorization).
248 void vectorize(LoopVectorizationLegality *L) {
250 // Create a new empty loop. Unlink the old loop and connect the new one.
252 // Widen each instruction in the old loop to a new one in the new loop.
253 // Use the Legality module to find the induction and reduction variables.
255 // Register the new loop and update the analysis passes.
259 virtual ~InnerLoopVectorizer() {}
262 /// A small list of PHINodes.
263 typedef SmallVector<PHINode*, 4> PhiVector;
264 /// When we unroll loops we have multiple vector values for each scalar.
265 /// This data structure holds the unrolled and vectorized values that
266 /// originated from one scalar instruction.
267 typedef SmallVector<Value*, 2> VectorParts;
269 // When we if-convert we need create edge masks. We have to cache values so
270 // that we don't end up with exponential recursion/IR.
271 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
272 VectorParts> EdgeMaskCache;
274 /// \brief Add checks for strides that where assumed to be 1.
276 /// Returns the last check instruction and the first check instruction in the
277 /// pair as (first, last).
278 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
280 /// Create an empty loop, based on the loop ranges of the old loop.
281 void createEmptyLoop();
282 /// Copy and widen the instructions from the old loop.
283 virtual void vectorizeLoop();
285 /// \brief The Loop exit block may have single value PHI nodes where the
286 /// incoming value is 'Undef'. While vectorizing we only handled real values
287 /// that were defined inside the loop. Here we fix the 'undef case'.
291 /// A helper function that computes the predicate of the block BB, assuming
292 /// that the header block of the loop is set to True. It returns the *entry*
293 /// mask for the block BB.
294 VectorParts createBlockInMask(BasicBlock *BB);
295 /// A helper function that computes the predicate of the edge between SRC
297 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
299 /// A helper function to vectorize a single BB within the innermost loop.
300 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
302 /// Vectorize a single PHINode in a block. This method handles the induction
303 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
304 /// arbitrary length vectors.
305 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
306 unsigned UF, unsigned VF, PhiVector *PV);
308 /// Insert the new loop to the loop hierarchy and pass manager
309 /// and update the analysis passes.
310 void updateAnalysis();
312 /// This instruction is un-vectorizable. Implement it as a sequence
313 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
314 /// scalarized instruction behind an if block predicated on the control
315 /// dependence of the instruction.
316 virtual void scalarizeInstruction(Instruction *Instr,
317 bool IfPredicateStore=false);
319 /// Vectorize Load and Store instructions,
320 virtual void vectorizeMemoryInstruction(Instruction *Instr);
322 /// Create a broadcast instruction. This method generates a broadcast
323 /// instruction (shuffle) for loop invariant values and for the induction
324 /// value. If this is the induction variable then we extend it to N, N+1, ...
325 /// this is needed because each iteration in the loop corresponds to a SIMD
327 virtual Value *getBroadcastInstrs(Value *V);
329 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
330 /// to each vector element of Val. The sequence starts at StartIndex.
331 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step);
333 /// When we go over instructions in the basic block we rely on previous
334 /// values within the current basic block or on loop invariant values.
335 /// When we widen (vectorize) values we place them in the map. If the values
336 /// are not within the map, they have to be loop invariant, so we simply
337 /// broadcast them into a vector.
338 VectorParts &getVectorValue(Value *V);
340 /// Generate a shuffle sequence that will reverse the vector Vec.
341 virtual Value *reverseVector(Value *Vec);
343 /// This is a helper class that holds the vectorizer state. It maps scalar
344 /// instructions to vector instructions. When the code is 'unrolled' then
345 /// then a single scalar value is mapped to multiple vector parts. The parts
346 /// are stored in the VectorPart type.
348 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
350 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
352 /// \return True if 'Key' is saved in the Value Map.
353 bool has(Value *Key) const { return MapStorage.count(Key); }
355 /// Initializes a new entry in the map. Sets all of the vector parts to the
356 /// save value in 'Val'.
357 /// \return A reference to a vector with splat values.
358 VectorParts &splat(Value *Key, Value *Val) {
359 VectorParts &Entry = MapStorage[Key];
360 Entry.assign(UF, Val);
364 ///\return A reference to the value that is stored at 'Key'.
365 VectorParts &get(Value *Key) {
366 VectorParts &Entry = MapStorage[Key];
369 assert(Entry.size() == UF);
374 /// The unroll factor. Each entry in the map stores this number of vector
378 /// Map storage. We use std::map and not DenseMap because insertions to a
379 /// dense map invalidates its iterators.
380 std::map<Value *, VectorParts> MapStorage;
383 /// The original loop.
385 /// Scev analysis to use.
394 const DataLayout *DL;
395 /// Target Library Info.
396 const TargetLibraryInfo *TLI;
398 /// The vectorization SIMD factor to use. Each vector will have this many
403 /// The vectorization unroll factor to use. Each scalar is vectorized to this
404 /// many different vector instructions.
407 /// The builder that we use
410 // --- Vectorization state ---
412 /// The vector-loop preheader.
413 BasicBlock *LoopVectorPreHeader;
414 /// The scalar-loop preheader.
415 BasicBlock *LoopScalarPreHeader;
416 /// Middle Block between the vector and the scalar.
417 BasicBlock *LoopMiddleBlock;
418 ///The ExitBlock of the scalar loop.
419 BasicBlock *LoopExitBlock;
420 ///The vector loop body.
421 SmallVector<BasicBlock *, 4> LoopVectorBody;
422 ///The scalar loop body.
423 BasicBlock *LoopScalarBody;
424 /// A list of all bypass blocks. The first block is the entry of the loop.
425 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
427 /// The new Induction variable which was added to the new block.
429 /// The induction variable of the old basic block.
430 PHINode *OldInduction;
431 /// Holds the extended (to the widest induction type) start index.
433 /// Maps scalars to widened vectors.
435 EdgeMaskCache MaskCache;
437 LoopVectorizationLegality *Legal;
440 class InnerLoopUnroller : public InnerLoopVectorizer {
442 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
443 DominatorTree *DT, const DataLayout *DL,
444 const TargetLibraryInfo *TLI, unsigned UnrollFactor) :
445 InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { }
448 void scalarizeInstruction(Instruction *Instr,
449 bool IfPredicateStore = false) override;
450 void vectorizeMemoryInstruction(Instruction *Instr) override;
451 Value *getBroadcastInstrs(Value *V) override;
452 Value *getStepVector(Value *Val, int StartIdx, Value *Step) override;
453 Value *reverseVector(Value *Vec) override;
456 /// \brief Look for a meaningful debug location on the instruction or it's
458 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
463 if (I->getDebugLoc() != Empty)
466 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
467 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
468 if (OpInst->getDebugLoc() != Empty)
475 /// \brief Set the debug location in the builder using the debug location in the
477 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
478 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
479 B.SetCurrentDebugLocation(Inst->getDebugLoc());
481 B.SetCurrentDebugLocation(DebugLoc());
485 /// \return string containing a file name and a line # for the given loop.
486 static std::string getDebugLocString(const Loop *L) {
489 raw_string_ostream OS(Result);
490 const DebugLoc LoopDbgLoc = L->getStartLoc();
491 if (!LoopDbgLoc.isUnknown())
492 LoopDbgLoc.print(L->getHeader()->getContext(), OS);
494 // Just print the module name.
495 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
502 /// \brief Propagate known metadata from one instruction to another.
503 static void propagateMetadata(Instruction *To, const Instruction *From) {
504 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
505 From->getAllMetadataOtherThanDebugLoc(Metadata);
507 for (auto M : Metadata) {
508 unsigned Kind = M.first;
510 // These are safe to transfer (this is safe for TBAA, even when we
511 // if-convert, because should that metadata have had a control dependency
512 // on the condition, and thus actually aliased with some other
513 // non-speculated memory access when the condition was false, this would be
514 // caught by the runtime overlap checks).
515 if (Kind != LLVMContext::MD_tbaa &&
516 Kind != LLVMContext::MD_alias_scope &&
517 Kind != LLVMContext::MD_noalias &&
518 Kind != LLVMContext::MD_fpmath)
521 To->setMetadata(Kind, M.second);
525 /// \brief Propagate known metadata from one instruction to a vector of others.
526 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
528 if (Instruction *I = dyn_cast<Instruction>(V))
529 propagateMetadata(I, From);
532 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
533 /// to what vectorization factor.
534 /// This class does not look at the profitability of vectorization, only the
535 /// legality. This class has two main kinds of checks:
536 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
537 /// will change the order of memory accesses in a way that will change the
538 /// correctness of the program.
539 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
540 /// checks for a number of different conditions, such as the availability of a
541 /// single induction variable, that all types are supported and vectorize-able,
542 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
543 /// This class is also used by InnerLoopVectorizer for identifying
544 /// induction variable and the different reduction variables.
545 class LoopVectorizationLegality {
547 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, const DataLayout *DL,
548 DominatorTree *DT, TargetLibraryInfo *TLI,
549 AliasAnalysis *AA, Function *F,
550 const TargetTransformInfo *TTI)
551 : NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
552 TLI(TLI), TheFunction(F), TTI(TTI), Induction(nullptr),
553 WidestIndTy(nullptr),
554 LAA(F, L, SE, DL, TLI, AA, DT,
555 LoopAccessAnalysis::VectorizerParams(
556 MaxVectorWidth, VectorizationFactor, VectorizationInterleave,
557 RuntimeMemoryCheckThreshold)),
558 HasFunNoNaNAttr(false) {}
560 /// This enum represents the kinds of reductions that we support.
562 RK_NoReduction, ///< Not a reduction.
563 RK_IntegerAdd, ///< Sum of integers.
564 RK_IntegerMult, ///< Product of integers.
565 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
566 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
567 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
568 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
569 RK_FloatAdd, ///< Sum of floats.
570 RK_FloatMult, ///< Product of floats.
571 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
574 /// This enum represents the kinds of inductions that we support.
576 IK_NoInduction, ///< Not an induction variable.
577 IK_IntInduction, ///< Integer induction variable. Step = C.
578 IK_PtrInduction ///< Pointer induction var. Step = C / sizeof(elem).
581 // This enum represents the kind of minmax reduction.
582 enum MinMaxReductionKind {
592 /// This struct holds information about reduction variables.
593 struct ReductionDescriptor {
594 ReductionDescriptor() : StartValue(nullptr), LoopExitInstr(nullptr),
595 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
597 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
598 MinMaxReductionKind MK)
599 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
601 // The starting value of the reduction.
602 // It does not have to be zero!
603 TrackingVH<Value> StartValue;
604 // The instruction who's value is used outside the loop.
605 Instruction *LoopExitInstr;
606 // The kind of the reduction.
608 // If this a min/max reduction the kind of reduction.
609 MinMaxReductionKind MinMaxKind;
612 /// This POD struct holds information about a potential reduction operation.
613 struct ReductionInstDesc {
614 ReductionInstDesc(bool IsRedux, Instruction *I) :
615 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
617 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
618 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
620 // Is this instruction a reduction candidate.
622 // The last instruction in a min/max pattern (select of the select(icmp())
623 // pattern), or the current reduction instruction otherwise.
624 Instruction *PatternLastInst;
625 // If this is a min/max pattern the comparison predicate.
626 MinMaxReductionKind MinMaxKind;
629 /// A struct for saving information about induction variables.
630 struct InductionInfo {
631 InductionInfo(Value *Start, InductionKind K, ConstantInt *Step)
632 : StartValue(Start), IK(K), StepValue(Step) {
633 assert(IK != IK_NoInduction && "Not an induction");
634 assert(StartValue && "StartValue is null");
635 assert(StepValue && !StepValue->isZero() && "StepValue is zero");
636 assert((IK != IK_PtrInduction || StartValue->getType()->isPointerTy()) &&
637 "StartValue is not a pointer for pointer induction");
638 assert((IK != IK_IntInduction || StartValue->getType()->isIntegerTy()) &&
639 "StartValue is not an integer for integer induction");
640 assert(StepValue->getType()->isIntegerTy() &&
641 "StepValue is not an integer");
644 : StartValue(nullptr), IK(IK_NoInduction), StepValue(nullptr) {}
646 /// Get the consecutive direction. Returns:
647 /// 0 - unknown or non-consecutive.
648 /// 1 - consecutive and increasing.
649 /// -1 - consecutive and decreasing.
650 int getConsecutiveDirection() const {
651 if (StepValue && (StepValue->isOne() || StepValue->isMinusOne()))
652 return StepValue->getSExtValue();
656 /// Compute the transformed value of Index at offset StartValue using step
658 /// For integer induction, returns StartValue + Index * StepValue.
659 /// For pointer induction, returns StartValue[Index * StepValue].
660 /// FIXME: The newly created binary instructions should contain nsw/nuw
661 /// flags, which can be found from the original scalar operations.
662 Value *transform(IRBuilder<> &B, Value *Index) const {
664 case IK_IntInduction:
665 assert(Index->getType() == StartValue->getType() &&
666 "Index type does not match StartValue type");
667 if (StepValue->isMinusOne())
668 return B.CreateSub(StartValue, Index);
669 if (!StepValue->isOne())
670 Index = B.CreateMul(Index, StepValue);
671 return B.CreateAdd(StartValue, Index);
673 case IK_PtrInduction:
674 if (StepValue->isMinusOne())
675 Index = B.CreateNeg(Index);
676 else if (!StepValue->isOne())
677 Index = B.CreateMul(Index, StepValue);
678 return B.CreateGEP(StartValue, Index);
683 llvm_unreachable("invalid enum");
687 TrackingVH<Value> StartValue;
691 ConstantInt *StepValue;
694 /// ReductionList contains the reduction descriptors for all
695 /// of the reductions that were found in the loop.
696 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
698 /// InductionList saves induction variables and maps them to the
699 /// induction descriptor.
700 typedef MapVector<PHINode*, InductionInfo> InductionList;
702 /// Returns true if it is legal to vectorize this loop.
703 /// This does not mean that it is profitable to vectorize this
704 /// loop, only that it is legal to do so.
707 /// Returns the Induction variable.
708 PHINode *getInduction() { return Induction; }
710 /// Returns the reduction variables found in the loop.
711 ReductionList *getReductionVars() { return &Reductions; }
713 /// Returns the induction variables found in the loop.
714 InductionList *getInductionVars() { return &Inductions; }
716 /// Returns the widest induction type.
717 Type *getWidestInductionType() { return WidestIndTy; }
719 /// Returns True if V is an induction variable in this loop.
720 bool isInductionVariable(const Value *V);
722 /// Return true if the block BB needs to be predicated in order for the loop
723 /// to be vectorized.
724 bool blockNeedsPredication(BasicBlock *BB);
726 /// Check if this pointer is consecutive when vectorizing. This happens
727 /// when the last index of the GEP is the induction variable, or that the
728 /// pointer itself is an induction variable.
729 /// This check allows us to vectorize A[idx] into a wide load/store.
731 /// 0 - Stride is unknown or non-consecutive.
732 /// 1 - Address is consecutive.
733 /// -1 - Address is consecutive, and decreasing.
734 int isConsecutivePtr(Value *Ptr);
736 /// Returns true if the value V is uniform within the loop.
737 bool isUniform(Value *V);
739 /// Returns true if this instruction will remain scalar after vectorization.
740 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
742 /// Returns the information that we collected about runtime memory check.
743 LoopAccessAnalysis::RuntimePointerCheck *getRuntimePointerCheck() {
744 return LAA.getRuntimePointerCheck();
747 LoopAccessAnalysis *getLAA() {
751 /// This function returns the identity element (or neutral element) for
753 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
755 unsigned getMaxSafeDepDistBytes() { return LAA.getMaxSafeDepDistBytes(); }
757 bool hasStride(Value *V) { return StrideSet.count(V); }
758 bool mustCheckStrides() { return !StrideSet.empty(); }
759 SmallPtrSet<Value *, 8>::iterator strides_begin() {
760 return StrideSet.begin();
762 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
764 /// Returns true if the target machine supports masked store operation
765 /// for the given \p DataType and kind of access to \p Ptr.
766 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
767 return TTI->isLegalMaskedStore(DataType, isConsecutivePtr(Ptr));
769 /// Returns true if the target machine supports masked load operation
770 /// for the given \p DataType and kind of access to \p Ptr.
771 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
772 return TTI->isLegalMaskedLoad(DataType, isConsecutivePtr(Ptr));
774 /// Returns true if vector representation of the instruction \p I
776 bool isMaskRequired(const Instruction* I) {
777 return (MaskedOp.count(I) != 0);
779 unsigned getNumStores() const {
780 return LAA.getNumStores();
782 unsigned getNumLoads() const {
783 return LAA.getNumLoads();
785 unsigned getNumPredStores() const {
786 return NumPredStores;
789 /// Check if a single basic block loop is vectorizable.
790 /// At this point we know that this is a loop with a constant trip count
791 /// and we only need to check individual instructions.
792 bool canVectorizeInstrs();
794 /// When we vectorize loops we may change the order in which
795 /// we read and write from memory. This method checks if it is
796 /// legal to vectorize the code, considering only memory constrains.
797 /// Returns true if the loop is vectorizable
798 bool canVectorizeMemory();
800 /// Return true if we can vectorize this loop using the IF-conversion
802 bool canVectorizeWithIfConvert();
804 /// Collect the variables that need to stay uniform after vectorization.
805 void collectLoopUniforms();
807 /// Return true if all of the instructions in the block can be speculatively
808 /// executed. \p SafePtrs is a list of addresses that are known to be legal
809 /// and we know that we can read from them without segfault.
810 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
812 /// Returns True, if 'Phi' is the kind of reduction variable for type
813 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
814 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
815 /// Returns a struct describing if the instruction 'I' can be a reduction
816 /// variable of type 'Kind'. If the reduction is a min/max pattern of
817 /// select(icmp()) this function advances the instruction pointer 'I' from the
818 /// compare instruction to the select instruction and stores this pointer in
819 /// 'PatternLastInst' member of the returned struct.
820 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
821 ReductionInstDesc &Desc);
822 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
823 /// pattern corresponding to a min(X, Y) or max(X, Y).
824 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
825 ReductionInstDesc &Prev);
826 /// Returns the induction kind of Phi and record the step. This function may
827 /// return NoInduction if the PHI is not an induction variable.
828 InductionKind isInductionVariable(PHINode *Phi, ConstantInt *&StepValue);
830 /// \brief Collect memory access with loop invariant strides.
832 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
834 void collectStridedAccess(Value *LoadOrStoreInst);
836 /// Report an analysis message to assist the user in diagnosing loops that are
838 void emitAnalysis(VectorizationReport &Message) {
839 VectorizationReport::emitAnalysis(Message, TheFunction, TheLoop);
842 unsigned NumPredStores;
844 /// The loop that we evaluate.
848 /// DataLayout analysis.
849 const DataLayout *DL;
850 /// Target Library Info.
851 TargetLibraryInfo *TLI;
853 Function *TheFunction;
854 /// Target Transform Info
855 const TargetTransformInfo *TTI;
857 // --- vectorization state --- //
859 /// Holds the integer induction variable. This is the counter of the
862 /// Holds the reduction variables.
863 ReductionList Reductions;
864 /// Holds all of the induction variables that we found in the loop.
865 /// Notice that inductions don't need to start at zero and that induction
866 /// variables can be pointers.
867 InductionList Inductions;
868 /// Holds the widest induction type encountered.
871 /// Allowed outside users. This holds the reduction
872 /// vars which can be accessed from outside the loop.
873 SmallPtrSet<Value*, 4> AllowedExit;
874 /// This set holds the variables which are known to be uniform after
876 SmallPtrSet<Instruction*, 4> Uniforms;
877 LoopAccessAnalysis LAA;
878 /// Can we assume the absence of NaNs.
879 bool HasFunNoNaNAttr;
881 ValueToValueMap Strides;
882 SmallPtrSet<Value *, 8> StrideSet;
884 /// While vectorizing these instructions we have to generate a
885 /// call to the appropriate masked intrinsic
886 SmallPtrSet<const Instruction*, 8> MaskedOp;
889 /// LoopVectorizationCostModel - estimates the expected speedups due to
891 /// In many cases vectorization is not profitable. This can happen because of
892 /// a number of reasons. In this class we mainly attempt to predict the
893 /// expected speedup/slowdowns due to the supported instruction set. We use the
894 /// TargetTransformInfo to query the different backends for the cost of
895 /// different operations.
896 class LoopVectorizationCostModel {
898 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
899 LoopVectorizationLegality *Legal,
900 const TargetTransformInfo &TTI,
901 const DataLayout *DL, const TargetLibraryInfo *TLI,
902 AssumptionCache *AC, const Function *F,
903 const LoopVectorizeHints *Hints)
904 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI),
905 TheFunction(F), Hints(Hints) {
906 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
909 /// Information about vectorization costs
910 struct VectorizationFactor {
911 unsigned Width; // Vector width with best cost
912 unsigned Cost; // Cost of the loop with that width
914 /// \return The most profitable vectorization factor and the cost of that VF.
915 /// This method checks every power of two up to VF. If UserVF is not ZERO
916 /// then this vectorization factor will be selected if vectorization is
918 VectorizationFactor selectVectorizationFactor(bool OptForSize);
920 /// \return The size (in bits) of the widest type in the code that
921 /// needs to be vectorized. We ignore values that remain scalar such as
922 /// 64 bit loop indices.
923 unsigned getWidestType();
925 /// \return The most profitable unroll factor.
926 /// If UserUF is non-zero then this method finds the best unroll-factor
927 /// based on register pressure and other parameters.
928 /// VF and LoopCost are the selected vectorization factor and the cost of the
930 unsigned selectUnrollFactor(bool OptForSize, unsigned VF, unsigned LoopCost);
932 /// \brief A struct that represents some properties of the register usage
934 struct RegisterUsage {
935 /// Holds the number of loop invariant values that are used in the loop.
936 unsigned LoopInvariantRegs;
937 /// Holds the maximum number of concurrent live intervals in the loop.
938 unsigned MaxLocalUsers;
939 /// Holds the number of instructions in the loop.
940 unsigned NumInstructions;
943 /// \return information about the register usage of the loop.
944 RegisterUsage calculateRegisterUsage();
947 /// Returns the expected execution cost. The unit of the cost does
948 /// not matter because we use the 'cost' units to compare different
949 /// vector widths. The cost that is returned is *not* normalized by
950 /// the factor width.
951 unsigned expectedCost(unsigned VF);
953 /// Returns the execution time cost of an instruction for a given vector
954 /// width. Vector width of one means scalar.
955 unsigned getInstructionCost(Instruction *I, unsigned VF);
957 /// A helper function for converting Scalar types to vector types.
958 /// If the incoming type is void, we return void. If the VF is 1, we return
960 static Type* ToVectorTy(Type *Scalar, unsigned VF);
962 /// Returns whether the instruction is a load or store and will be a emitted
963 /// as a vector operation.
964 bool isConsecutiveLoadOrStore(Instruction *I);
966 /// Report an analysis message to assist the user in diagnosing loops that are
968 void emitAnalysis(VectorizationReport &Message) {
969 VectorizationReport::emitAnalysis(Message, TheFunction, TheLoop);
972 /// Values used only by @llvm.assume calls.
973 SmallPtrSet<const Value *, 32> EphValues;
975 /// The loop that we evaluate.
979 /// Loop Info analysis.
981 /// Vectorization legality.
982 LoopVectorizationLegality *Legal;
983 /// Vector target information.
984 const TargetTransformInfo &TTI;
985 /// Target data layout information.
986 const DataLayout *DL;
987 /// Target Library Info.
988 const TargetLibraryInfo *TLI;
989 const Function *TheFunction;
990 // Loop Vectorize Hint.
991 const LoopVectorizeHints *Hints;
994 /// Utility class for getting and setting loop vectorizer hints in the form
995 /// of loop metadata.
996 /// This class keeps a number of loop annotations locally (as member variables)
997 /// and can, upon request, write them back as metadata on the loop. It will
998 /// initially scan the loop for existing metadata, and will update the local
999 /// values based on information in the loop.
1000 /// We cannot write all values to metadata, as the mere presence of some info,
1001 /// for example 'force', means a decision has been made. So, we need to be
1002 /// careful NOT to add them if the user hasn't specifically asked so.
1003 class LoopVectorizeHints {
1010 /// Hint - associates name and validation with the hint value.
1013 unsigned Value; // This may have to change for non-numeric values.
1016 Hint(const char * Name, unsigned Value, HintKind Kind)
1017 : Name(Name), Value(Value), Kind(Kind) { }
1019 bool validate(unsigned Val) {
1022 return isPowerOf2_32(Val) && Val <= MaxVectorWidth;
1024 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1032 /// Vectorization width.
1034 /// Vectorization interleave factor.
1036 /// Vectorization forced
1039 /// Return the loop metadata prefix.
1040 static StringRef Prefix() { return "llvm.loop."; }
1044 FK_Undefined = -1, ///< Not selected.
1045 FK_Disabled = 0, ///< Forcing disabled.
1046 FK_Enabled = 1, ///< Forcing enabled.
1049 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
1050 : Width("vectorize.width", VectorizationFactor, HK_WIDTH),
1051 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1052 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1054 // Populate values with existing loop metadata.
1055 getHintsFromMetadata();
1057 // force-vector-interleave overrides DisableInterleaving.
1058 if (VectorizationInterleave.getNumOccurrences() > 0)
1059 Interleave.Value = VectorizationInterleave;
1061 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
1062 << "LV: Interleaving disabled by the pass manager\n");
1065 /// Mark the loop L as already vectorized by setting the width to 1.
1066 void setAlreadyVectorized() {
1067 Width.Value = Interleave.Value = 1;
1068 Hint Hints[] = {Width, Interleave};
1069 writeHintsToMetadata(Hints);
1072 /// Dumps all the hint information.
1073 std::string emitRemark() const {
1074 VectorizationReport R;
1075 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1076 R << "vectorization is explicitly disabled";
1078 R << "use -Rpass-analysis=loop-vectorize for more info";
1079 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1080 R << " (Force=true";
1081 if (Width.Value != 0)
1082 R << ", Vector Width=" << Width.Value;
1083 if (Interleave.Value != 0)
1084 R << ", Interleave Count=" << Interleave.Value;
1092 unsigned getWidth() const { return Width.Value; }
1093 unsigned getInterleave() const { return Interleave.Value; }
1094 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1097 /// Find hints specified in the loop metadata and update local values.
1098 void getHintsFromMetadata() {
1099 MDNode *LoopID = TheLoop->getLoopID();
1103 // First operand should refer to the loop id itself.
1104 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1105 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1107 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1108 const MDString *S = nullptr;
1109 SmallVector<Metadata *, 4> Args;
1111 // The expected hint is either a MDString or a MDNode with the first
1112 // operand a MDString.
1113 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1114 if (!MD || MD->getNumOperands() == 0)
1116 S = dyn_cast<MDString>(MD->getOperand(0));
1117 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1118 Args.push_back(MD->getOperand(i));
1120 S = dyn_cast<MDString>(LoopID->getOperand(i));
1121 assert(Args.size() == 0 && "too many arguments for MDString");
1127 // Check if the hint starts with the loop metadata prefix.
1128 StringRef Name = S->getString();
1129 if (Args.size() == 1)
1130 setHint(Name, Args[0]);
1134 /// Checks string hint with one operand and set value if valid.
1135 void setHint(StringRef Name, Metadata *Arg) {
1136 if (!Name.startswith(Prefix()))
1138 Name = Name.substr(Prefix().size(), StringRef::npos);
1140 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1142 unsigned Val = C->getZExtValue();
1144 Hint *Hints[] = {&Width, &Interleave, &Force};
1145 for (auto H : Hints) {
1146 if (Name == H->Name) {
1147 if (H->validate(Val))
1150 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1156 /// Create a new hint from name / value pair.
1157 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1158 LLVMContext &Context = TheLoop->getHeader()->getContext();
1159 Metadata *MDs[] = {MDString::get(Context, Name),
1160 ConstantAsMetadata::get(
1161 ConstantInt::get(Type::getInt32Ty(Context), V))};
1162 return MDNode::get(Context, MDs);
1165 /// Matches metadata with hint name.
1166 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1167 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1171 for (auto H : HintTypes)
1172 if (Name->getString().endswith(H.Name))
1177 /// Sets current hints into loop metadata, keeping other values intact.
1178 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1179 if (HintTypes.size() == 0)
1182 // Reserve the first element to LoopID (see below).
1183 SmallVector<Metadata *, 4> MDs(1);
1184 // If the loop already has metadata, then ignore the existing operands.
1185 MDNode *LoopID = TheLoop->getLoopID();
1187 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1188 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1189 // If node in update list, ignore old value.
1190 if (!matchesHintMetadataName(Node, HintTypes))
1191 MDs.push_back(Node);
1195 // Now, add the missing hints.
1196 for (auto H : HintTypes)
1197 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1199 // Replace current metadata node with new one.
1200 LLVMContext &Context = TheLoop->getHeader()->getContext();
1201 MDNode *NewLoopID = MDNode::get(Context, MDs);
1202 // Set operand 0 to refer to the loop id itself.
1203 NewLoopID->replaceOperandWith(0, NewLoopID);
1205 TheLoop->setLoopID(NewLoopID);
1208 /// The loop these hints belong to.
1209 const Loop *TheLoop;
1212 static void emitMissedWarning(Function *F, Loop *L,
1213 const LoopVectorizeHints &LH) {
1214 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1215 L->getStartLoc(), LH.emitRemark());
1217 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1218 if (LH.getWidth() != 1)
1219 emitLoopVectorizeWarning(
1220 F->getContext(), *F, L->getStartLoc(),
1221 "failed explicitly specified loop vectorization");
1222 else if (LH.getInterleave() != 1)
1223 emitLoopInterleaveWarning(
1224 F->getContext(), *F, L->getStartLoc(),
1225 "failed explicitly specified loop interleaving");
1229 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1231 return V.push_back(&L);
1233 for (Loop *InnerL : L)
1234 addInnerLoop(*InnerL, V);
1237 /// The LoopVectorize Pass.
1238 struct LoopVectorize : public FunctionPass {
1239 /// Pass identification, replacement for typeid
1242 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1244 DisableUnrolling(NoUnrolling),
1245 AlwaysVectorize(AlwaysVectorize) {
1246 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1249 ScalarEvolution *SE;
1250 const DataLayout *DL;
1252 TargetTransformInfo *TTI;
1254 BlockFrequencyInfo *BFI;
1255 TargetLibraryInfo *TLI;
1257 AssumptionCache *AC;
1258 bool DisableUnrolling;
1259 bool AlwaysVectorize;
1261 BlockFrequency ColdEntryFreq;
1263 bool runOnFunction(Function &F) override {
1264 SE = &getAnalysis<ScalarEvolution>();
1265 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1266 DL = DLP ? &DLP->getDataLayout() : nullptr;
1267 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1268 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1269 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1270 BFI = &getAnalysis<BlockFrequencyInfo>();
1271 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1272 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1273 AA = &getAnalysis<AliasAnalysis>();
1274 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1276 // Compute some weights outside of the loop over the loops. Compute this
1277 // using a BranchProbability to re-use its scaling math.
1278 const BranchProbability ColdProb(1, 5); // 20%
1279 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1281 // If the target claims to have no vector registers don't attempt
1283 if (!TTI->getNumberOfRegisters(true))
1287 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1288 << ": Missing data layout\n");
1292 // Build up a worklist of inner-loops to vectorize. This is necessary as
1293 // the act of vectorizing or partially unrolling a loop creates new loops
1294 // and can invalidate iterators across the loops.
1295 SmallVector<Loop *, 8> Worklist;
1298 addInnerLoop(*L, Worklist);
1300 LoopsAnalyzed += Worklist.size();
1302 // Now walk the identified inner loops.
1303 bool Changed = false;
1304 while (!Worklist.empty())
1305 Changed |= processLoop(Worklist.pop_back_val());
1307 // Process each loop nest in the function.
1311 bool processLoop(Loop *L) {
1312 assert(L->empty() && "Only process inner loops.");
1315 const std::string DebugLocStr = getDebugLocString(L);
1318 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1319 << L->getHeader()->getParent()->getName() << "\" from "
1320 << DebugLocStr << "\n");
1322 LoopVectorizeHints Hints(L, DisableUnrolling);
1324 DEBUG(dbgs() << "LV: Loop hints:"
1326 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1328 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1330 : "?")) << " width=" << Hints.getWidth()
1331 << " unroll=" << Hints.getInterleave() << "\n");
1333 // Function containing loop
1334 Function *F = L->getHeader()->getParent();
1336 // Looking at the diagnostic output is the only way to determine if a loop
1337 // was vectorized (other than looking at the IR or machine code), so it
1338 // is important to generate an optimization remark for each loop. Most of
1339 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1340 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1341 // less verbose reporting vectorized loops and unvectorized loops that may
1342 // benefit from vectorization, respectively.
1344 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1345 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1346 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1347 L->getStartLoc(), Hints.emitRemark());
1351 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1352 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1353 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1354 L->getStartLoc(), Hints.emitRemark());
1358 if (Hints.getWidth() == 1 && Hints.getInterleave() == 1) {
1359 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1360 emitOptimizationRemarkAnalysis(
1361 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1362 "loop not vectorized: vector width and interleave count are "
1363 "explicitly set to 1");
1367 // Check the loop for a trip count threshold:
1368 // do not vectorize loops with a tiny trip count.
1369 const unsigned TC = SE->getSmallConstantTripCount(L);
1370 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1371 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1372 << "This loop is not worth vectorizing.");
1373 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1374 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1376 DEBUG(dbgs() << "\n");
1377 emitOptimizationRemarkAnalysis(
1378 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1379 "vectorization is not beneficial and is not explicitly forced");
1384 // Check if it is legal to vectorize the loop.
1385 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, AA, F, TTI);
1386 if (!LVL.canVectorize()) {
1387 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1388 emitMissedWarning(F, L, Hints);
1392 // Use the cost model.
1393 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI, AC, F,
1396 // Check the function attributes to find out if this function should be
1397 // optimized for size.
1398 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1399 F->hasFnAttribute(Attribute::OptimizeForSize);
1401 // Compute the weighted frequency of this loop being executed and see if it
1402 // is less than 20% of the function entry baseline frequency. Note that we
1403 // always have a canonical loop here because we think we *can* vectoriez.
1404 // FIXME: This is hidden behind a flag due to pervasive problems with
1405 // exactly what block frequency models.
1406 if (LoopVectorizeWithBlockFrequency) {
1407 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1408 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1409 LoopEntryFreq < ColdEntryFreq)
1413 // Check the function attributes to see if implicit floats are allowed.a
1414 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1415 // an integer loop and the vector instructions selected are purely integer
1416 // vector instructions?
1417 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1418 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1419 "attribute is used.\n");
1420 emitOptimizationRemarkAnalysis(
1421 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1422 "loop not vectorized due to NoImplicitFloat attribute");
1423 emitMissedWarning(F, L, Hints);
1427 // Select the optimal vectorization factor.
1428 const LoopVectorizationCostModel::VectorizationFactor VF =
1429 CM.selectVectorizationFactor(OptForSize);
1431 // Select the unroll factor.
1433 CM.selectUnrollFactor(OptForSize, VF.Width, VF.Cost);
1435 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1436 << DebugLocStr << '\n');
1437 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1439 if (VF.Width == 1) {
1440 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
1443 emitOptimizationRemarkAnalysis(
1444 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1445 "not beneficial to vectorize and user disabled interleaving");
1448 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1450 // Report the unrolling decision.
1451 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1452 Twine("unrolled with interleaving factor " +
1454 " (vectorization not beneficial)"));
1456 // We decided not to vectorize, but we may want to unroll.
1458 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1459 Unroller.vectorize(&LVL);
1461 // If we decided that it is *legal* to vectorize the loop then do it.
1462 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1466 // Report the vectorization decision.
1467 emitOptimizationRemark(
1468 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1469 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1470 ", unrolling interleave factor: " + Twine(UF) + ")");
1473 // Mark the loop as already vectorized to avoid vectorizing again.
1474 Hints.setAlreadyVectorized();
1476 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1480 void getAnalysisUsage(AnalysisUsage &AU) const override {
1481 AU.addRequired<AssumptionCacheTracker>();
1482 AU.addRequiredID(LoopSimplifyID);
1483 AU.addRequiredID(LCSSAID);
1484 AU.addRequired<BlockFrequencyInfo>();
1485 AU.addRequired<DominatorTreeWrapperPass>();
1486 AU.addRequired<LoopInfoWrapperPass>();
1487 AU.addRequired<ScalarEvolution>();
1488 AU.addRequired<TargetTransformInfoWrapperPass>();
1489 AU.addRequired<AliasAnalysis>();
1490 AU.addPreserved<LoopInfoWrapperPass>();
1491 AU.addPreserved<DominatorTreeWrapperPass>();
1492 AU.addPreserved<AliasAnalysis>();
1497 } // end anonymous namespace
1499 //===----------------------------------------------------------------------===//
1500 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1501 // LoopVectorizationCostModel.
1502 //===----------------------------------------------------------------------===//
1504 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1505 // We need to place the broadcast of invariant variables outside the loop.
1506 Instruction *Instr = dyn_cast<Instruction>(V);
1508 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1509 Instr->getParent()) != LoopVectorBody.end());
1510 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1512 // Place the code for broadcasting invariant variables in the new preheader.
1513 IRBuilder<>::InsertPointGuard Guard(Builder);
1515 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1517 // Broadcast the scalar into all locations in the vector.
1518 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1523 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
1525 assert(Val->getType()->isVectorTy() && "Must be a vector");
1526 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1527 "Elem must be an integer");
1528 assert(Step->getType() == Val->getType()->getScalarType() &&
1529 "Step has wrong type");
1530 // Create the types.
1531 Type *ITy = Val->getType()->getScalarType();
1532 VectorType *Ty = cast<VectorType>(Val->getType());
1533 int VLen = Ty->getNumElements();
1534 SmallVector<Constant*, 8> Indices;
1536 // Create a vector of consecutive numbers from zero to VF.
1537 for (int i = 0; i < VLen; ++i)
1538 Indices.push_back(ConstantInt::get(ITy, StartIdx + i));
1540 // Add the consecutive indices to the vector value.
1541 Constant *Cv = ConstantVector::get(Indices);
1542 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1543 Step = Builder.CreateVectorSplat(VLen, Step);
1544 assert(Step->getType() == Val->getType() && "Invalid step vec");
1545 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
1546 // which can be found from the original scalar operations.
1547 Step = Builder.CreateMul(Cv, Step);
1548 return Builder.CreateAdd(Val, Step, "induction");
1551 /// \brief Find the operand of the GEP that should be checked for consecutive
1552 /// stores. This ignores trailing indices that have no effect on the final
1554 static unsigned getGEPInductionOperand(const DataLayout *DL,
1555 const GetElementPtrInst *Gep) {
1556 unsigned LastOperand = Gep->getNumOperands() - 1;
1557 unsigned GEPAllocSize = DL->getTypeAllocSize(
1558 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1560 // Walk backwards and try to peel off zeros.
1561 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1562 // Find the type we're currently indexing into.
1563 gep_type_iterator GEPTI = gep_type_begin(Gep);
1564 std::advance(GEPTI, LastOperand - 1);
1566 // If it's a type with the same allocation size as the result of the GEP we
1567 // can peel off the zero index.
1568 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1576 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1577 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1578 // Make sure that the pointer does not point to structs.
1579 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1582 // If this value is a pointer induction variable we know it is consecutive.
1583 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1584 if (Phi && Inductions.count(Phi)) {
1585 InductionInfo II = Inductions[Phi];
1586 return II.getConsecutiveDirection();
1589 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1593 unsigned NumOperands = Gep->getNumOperands();
1594 Value *GpPtr = Gep->getPointerOperand();
1595 // If this GEP value is a consecutive pointer induction variable and all of
1596 // the indices are constant then we know it is consecutive. We can
1597 Phi = dyn_cast<PHINode>(GpPtr);
1598 if (Phi && Inductions.count(Phi)) {
1600 // Make sure that the pointer does not point to structs.
1601 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1602 if (GepPtrType->getElementType()->isAggregateType())
1605 // Make sure that all of the index operands are loop invariant.
1606 for (unsigned i = 1; i < NumOperands; ++i)
1607 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1610 InductionInfo II = Inductions[Phi];
1611 return II.getConsecutiveDirection();
1614 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1616 // Check that all of the gep indices are uniform except for our induction
1618 for (unsigned i = 0; i != NumOperands; ++i)
1619 if (i != InductionOperand &&
1620 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1623 // We can emit wide load/stores only if the last non-zero index is the
1624 // induction variable.
1625 const SCEV *Last = nullptr;
1626 if (!Strides.count(Gep))
1627 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1629 // Because of the multiplication by a stride we can have a s/zext cast.
1630 // We are going to replace this stride by 1 so the cast is safe to ignore.
1632 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1633 // %0 = trunc i64 %indvars.iv to i32
1634 // %mul = mul i32 %0, %Stride1
1635 // %idxprom = zext i32 %mul to i64 << Safe cast.
1636 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1638 Last = replaceSymbolicStrideSCEV(SE, Strides,
1639 Gep->getOperand(InductionOperand), Gep);
1640 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1642 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1646 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1647 const SCEV *Step = AR->getStepRecurrence(*SE);
1649 // The memory is consecutive because the last index is consecutive
1650 // and all other indices are loop invariant.
1653 if (Step->isAllOnesValue())
1660 bool LoopVectorizationLegality::isUniform(Value *V) {
1661 return LAA.isUniform(V);
1664 InnerLoopVectorizer::VectorParts&
1665 InnerLoopVectorizer::getVectorValue(Value *V) {
1666 assert(V != Induction && "The new induction variable should not be used.");
1667 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1669 // If we have a stride that is replaced by one, do it here.
1670 if (Legal->hasStride(V))
1671 V = ConstantInt::get(V->getType(), 1);
1673 // If we have this scalar in the map, return it.
1674 if (WidenMap.has(V))
1675 return WidenMap.get(V);
1677 // If this scalar is unknown, assume that it is a constant or that it is
1678 // loop invariant. Broadcast V and save the value for future uses.
1679 Value *B = getBroadcastInstrs(V);
1680 return WidenMap.splat(V, B);
1683 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1684 assert(Vec->getType()->isVectorTy() && "Invalid type");
1685 SmallVector<Constant*, 8> ShuffleMask;
1686 for (unsigned i = 0; i < VF; ++i)
1687 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1689 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1690 ConstantVector::get(ShuffleMask),
1694 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1695 // Attempt to issue a wide load.
1696 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1697 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1699 assert((LI || SI) && "Invalid Load/Store instruction");
1701 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1702 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1703 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1704 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1705 // An alignment of 0 means target abi alignment. We need to use the scalar's
1706 // target abi alignment in such a case.
1708 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1709 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1710 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1711 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1713 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
1714 !Legal->isMaskRequired(SI))
1715 return scalarizeInstruction(Instr, true);
1717 if (ScalarAllocatedSize != VectorElementSize)
1718 return scalarizeInstruction(Instr);
1720 // If the pointer is loop invariant or if it is non-consecutive,
1721 // scalarize the load.
1722 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1723 bool Reverse = ConsecutiveStride < 0;
1724 bool UniformLoad = LI && Legal->isUniform(Ptr);
1725 if (!ConsecutiveStride || UniformLoad)
1726 return scalarizeInstruction(Instr);
1728 Constant *Zero = Builder.getInt32(0);
1729 VectorParts &Entry = WidenMap.get(Instr);
1731 // Handle consecutive loads/stores.
1732 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1733 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1734 setDebugLocFromInst(Builder, Gep);
1735 Value *PtrOperand = Gep->getPointerOperand();
1736 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1737 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1739 // Create the new GEP with the new induction variable.
1740 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1741 Gep2->setOperand(0, FirstBasePtr);
1742 Gep2->setName("gep.indvar.base");
1743 Ptr = Builder.Insert(Gep2);
1745 setDebugLocFromInst(Builder, Gep);
1746 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1747 OrigLoop) && "Base ptr must be invariant");
1749 // The last index does not have to be the induction. It can be
1750 // consecutive and be a function of the index. For example A[I+1];
1751 unsigned NumOperands = Gep->getNumOperands();
1752 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1753 // Create the new GEP with the new induction variable.
1754 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1756 for (unsigned i = 0; i < NumOperands; ++i) {
1757 Value *GepOperand = Gep->getOperand(i);
1758 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1760 // Update last index or loop invariant instruction anchored in loop.
1761 if (i == InductionOperand ||
1762 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1763 assert((i == InductionOperand ||
1764 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1765 "Must be last index or loop invariant");
1767 VectorParts &GEPParts = getVectorValue(GepOperand);
1768 Value *Index = GEPParts[0];
1769 Index = Builder.CreateExtractElement(Index, Zero);
1770 Gep2->setOperand(i, Index);
1771 Gep2->setName("gep.indvar.idx");
1774 Ptr = Builder.Insert(Gep2);
1776 // Use the induction element ptr.
1777 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1778 setDebugLocFromInst(Builder, Ptr);
1779 VectorParts &PtrVal = getVectorValue(Ptr);
1780 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1783 VectorParts Mask = createBlockInMask(Instr->getParent());
1786 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1787 "We do not allow storing to uniform addresses");
1788 setDebugLocFromInst(Builder, SI);
1789 // We don't want to update the value in the map as it might be used in
1790 // another expression. So don't use a reference type for "StoredVal".
1791 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1793 for (unsigned Part = 0; Part < UF; ++Part) {
1794 // Calculate the pointer for the specific unroll-part.
1795 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1798 // If we store to reverse consecutive memory locations then we need
1799 // to reverse the order of elements in the stored value.
1800 StoredVal[Part] = reverseVector(StoredVal[Part]);
1801 // If the address is consecutive but reversed, then the
1802 // wide store needs to start at the last vector element.
1803 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1804 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1805 Mask[Part] = reverseVector(Mask[Part]);
1808 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1809 DataTy->getPointerTo(AddressSpace));
1812 if (Legal->isMaskRequired(SI))
1813 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
1816 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
1817 propagateMetadata(NewSI, SI);
1823 assert(LI && "Must have a load instruction");
1824 setDebugLocFromInst(Builder, LI);
1825 for (unsigned Part = 0; Part < UF; ++Part) {
1826 // Calculate the pointer for the specific unroll-part.
1827 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1830 // If the address is consecutive but reversed, then the
1831 // wide load needs to start at the last vector element.
1832 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1833 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1834 Mask[Part] = reverseVector(Mask[Part]);
1838 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1839 DataTy->getPointerTo(AddressSpace));
1840 if (Legal->isMaskRequired(LI))
1841 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
1842 UndefValue::get(DataTy),
1843 "wide.masked.load");
1845 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
1846 propagateMetadata(NewLI, LI);
1847 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
1851 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1852 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1853 // Holds vector parameters or scalars, in case of uniform vals.
1854 SmallVector<VectorParts, 4> Params;
1856 setDebugLocFromInst(Builder, Instr);
1858 // Find all of the vectorized parameters.
1859 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1860 Value *SrcOp = Instr->getOperand(op);
1862 // If we are accessing the old induction variable, use the new one.
1863 if (SrcOp == OldInduction) {
1864 Params.push_back(getVectorValue(SrcOp));
1868 // Try using previously calculated values.
1869 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1871 // If the src is an instruction that appeared earlier in the basic block
1872 // then it should already be vectorized.
1873 if (SrcInst && OrigLoop->contains(SrcInst)) {
1874 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1875 // The parameter is a vector value from earlier.
1876 Params.push_back(WidenMap.get(SrcInst));
1878 // The parameter is a scalar from outside the loop. Maybe even a constant.
1879 VectorParts Scalars;
1880 Scalars.append(UF, SrcOp);
1881 Params.push_back(Scalars);
1885 assert(Params.size() == Instr->getNumOperands() &&
1886 "Invalid number of operands");
1888 // Does this instruction return a value ?
1889 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1891 Value *UndefVec = IsVoidRetTy ? nullptr :
1892 UndefValue::get(VectorType::get(Instr->getType(), VF));
1893 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1894 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1896 Instruction *InsertPt = Builder.GetInsertPoint();
1897 BasicBlock *IfBlock = Builder.GetInsertBlock();
1898 BasicBlock *CondBlock = nullptr;
1901 Loop *VectorLp = nullptr;
1902 if (IfPredicateStore) {
1903 assert(Instr->getParent()->getSinglePredecessor() &&
1904 "Only support single predecessor blocks");
1905 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1906 Instr->getParent());
1907 VectorLp = LI->getLoopFor(IfBlock);
1908 assert(VectorLp && "Must have a loop for this block");
1911 // For each vector unroll 'part':
1912 for (unsigned Part = 0; Part < UF; ++Part) {
1913 // For each scalar that we create:
1914 for (unsigned Width = 0; Width < VF; ++Width) {
1917 Value *Cmp = nullptr;
1918 if (IfPredicateStore) {
1919 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1920 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1921 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1922 LoopVectorBody.push_back(CondBlock);
1923 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
1924 // Update Builder with newly created basic block.
1925 Builder.SetInsertPoint(InsertPt);
1928 Instruction *Cloned = Instr->clone();
1930 Cloned->setName(Instr->getName() + ".cloned");
1931 // Replace the operands of the cloned instructions with extracted scalars.
1932 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1933 Value *Op = Params[op][Part];
1934 // Param is a vector. Need to extract the right lane.
1935 if (Op->getType()->isVectorTy())
1936 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1937 Cloned->setOperand(op, Op);
1940 // Place the cloned scalar in the new loop.
1941 Builder.Insert(Cloned);
1943 // If the original scalar returns a value we need to place it in a vector
1944 // so that future users will be able to use it.
1946 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1947 Builder.getInt32(Width));
1949 if (IfPredicateStore) {
1950 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1951 LoopVectorBody.push_back(NewIfBlock);
1952 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
1953 Builder.SetInsertPoint(InsertPt);
1954 Instruction *OldBr = IfBlock->getTerminator();
1955 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1956 OldBr->eraseFromParent();
1957 IfBlock = NewIfBlock;
1963 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1967 if (Instruction *I = dyn_cast<Instruction>(V))
1968 return I->getParent() == Loc->getParent() ? I : nullptr;
1972 std::pair<Instruction *, Instruction *>
1973 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
1974 Instruction *tnullptr = nullptr;
1975 if (!Legal->mustCheckStrides())
1976 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1978 IRBuilder<> ChkBuilder(Loc);
1981 Value *Check = nullptr;
1982 Instruction *FirstInst = nullptr;
1983 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
1984 SE = Legal->strides_end();
1986 Value *Ptr = stripIntegerCast(*SI);
1987 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
1989 // Store the first instruction we create.
1990 FirstInst = getFirstInst(FirstInst, C, Loc);
1992 Check = ChkBuilder.CreateOr(Check, C);
1997 // We have to do this trickery because the IRBuilder might fold the check to a
1998 // constant expression in which case there is no Instruction anchored in a
2000 LLVMContext &Ctx = Loc->getContext();
2001 Instruction *TheCheck =
2002 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2003 ChkBuilder.Insert(TheCheck, "stride.not.one");
2004 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2006 return std::make_pair(FirstInst, TheCheck);
2009 void InnerLoopVectorizer::createEmptyLoop() {
2011 In this function we generate a new loop. The new loop will contain
2012 the vectorized instructions while the old loop will continue to run the
2015 [ ] <-- Back-edge taken count overflow check.
2018 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2021 || [ ] <-- vector pre header.
2025 || [ ]_| <-- vector loop.
2028 | >[ ] <--- middle-block.
2031 -|- >[ ] <--- new preheader.
2035 | [ ]_| <-- old scalar loop to handle remainder.
2038 >[ ] <-- exit block.
2042 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2043 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2044 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2045 assert(BypassBlock && "Invalid loop structure");
2046 assert(ExitBlock && "Must have an exit block");
2048 // Some loops have a single integer induction variable, while other loops
2049 // don't. One example is c++ iterators that often have multiple pointer
2050 // induction variables. In the code below we also support a case where we
2051 // don't have a single induction variable.
2052 OldInduction = Legal->getInduction();
2053 Type *IdxTy = Legal->getWidestInductionType();
2055 // Find the loop boundaries.
2056 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2057 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2059 // The exit count might have the type of i64 while the phi is i32. This can
2060 // happen if we have an induction variable that is sign extended before the
2061 // compare. The only way that we get a backedge taken count is that the
2062 // induction variable was signed and as such will not overflow. In such a case
2063 // truncation is legal.
2064 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2065 IdxTy->getPrimitiveSizeInBits())
2066 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2068 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2069 // Get the total trip count from the count by adding 1.
2070 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2071 SE->getConstant(BackedgeTakeCount->getType(), 1));
2073 // Expand the trip count and place the new instructions in the preheader.
2074 // Notice that the pre-header does not change, only the loop body.
2075 SCEVExpander Exp(*SE, "induction");
2077 // We need to test whether the backedge-taken count is uint##_max. Adding one
2078 // to it will cause overflow and an incorrect loop trip count in the vector
2079 // body. In case of overflow we want to directly jump to the scalar remainder
2081 Value *BackedgeCount =
2082 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2083 BypassBlock->getTerminator());
2084 if (BackedgeCount->getType()->isPointerTy())
2085 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2086 "backedge.ptrcnt.to.int",
2087 BypassBlock->getTerminator());
2088 Instruction *CheckBCOverflow =
2089 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2090 Constant::getAllOnesValue(BackedgeCount->getType()),
2091 "backedge.overflow", BypassBlock->getTerminator());
2093 // The loop index does not have to start at Zero. Find the original start
2094 // value from the induction PHI node. If we don't have an induction variable
2095 // then we know that it starts at zero.
2096 Builder.SetInsertPoint(BypassBlock->getTerminator());
2097 Value *StartIdx = ExtendedIdx = OldInduction ?
2098 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2100 ConstantInt::get(IdxTy, 0);
2102 // We need an instruction to anchor the overflow check on. StartIdx needs to
2103 // be defined before the overflow check branch. Because the scalar preheader
2104 // is going to merge the start index and so the overflow branch block needs to
2105 // contain a definition of the start index.
2106 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2107 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2108 BypassBlock->getTerminator());
2110 // Count holds the overall loop count (N).
2111 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2112 BypassBlock->getTerminator());
2114 LoopBypassBlocks.push_back(BypassBlock);
2116 // Split the single block loop into the two loop structure described above.
2117 BasicBlock *VectorPH =
2118 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2119 BasicBlock *VecBody =
2120 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2121 BasicBlock *MiddleBlock =
2122 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2123 BasicBlock *ScalarPH =
2124 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2126 // Create and register the new vector loop.
2127 Loop* Lp = new Loop();
2128 Loop *ParentLoop = OrigLoop->getParentLoop();
2130 // Insert the new loop into the loop nest and register the new basic blocks
2131 // before calling any utilities such as SCEV that require valid LoopInfo.
2133 ParentLoop->addChildLoop(Lp);
2134 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
2135 ParentLoop->addBasicBlockToLoop(VectorPH, *LI);
2136 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
2138 LI->addTopLevelLoop(Lp);
2140 Lp->addBasicBlockToLoop(VecBody, *LI);
2142 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2144 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2146 // Generate the induction variable.
2147 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2148 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2149 // The loop step is equal to the vectorization factor (num of SIMD elements)
2150 // times the unroll factor (num of SIMD instructions).
2151 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2153 // This is the IR builder that we use to add all of the logic for bypassing
2154 // the new vector loop.
2155 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2156 setDebugLocFromInst(BypassBuilder,
2157 getDebugLocFromInstOrOperands(OldInduction));
2159 // We may need to extend the index in case there is a type mismatch.
2160 // We know that the count starts at zero and does not overflow.
2161 if (Count->getType() != IdxTy) {
2162 // The exit count can be of pointer type. Convert it to the correct
2164 if (ExitCount->getType()->isPointerTy())
2165 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2167 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2170 // Add the start index to the loop count to get the new end index.
2171 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2173 // Now we need to generate the expression for N - (N % VF), which is
2174 // the part that the vectorized body will execute.
2175 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2176 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2177 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2178 "end.idx.rnd.down");
2180 // Now, compare the new count to zero. If it is zero skip the vector loop and
2181 // jump to the scalar loop.
2183 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2185 BasicBlock *LastBypassBlock = BypassBlock;
2187 // Generate code to check that the loops trip count that we computed by adding
2188 // one to the backedge-taken count will not overflow.
2190 auto PastOverflowCheck =
2191 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2192 BasicBlock *CheckBlock =
2193 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2195 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2196 LoopBypassBlocks.push_back(CheckBlock);
2197 Instruction *OldTerm = LastBypassBlock->getTerminator();
2198 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2199 OldTerm->eraseFromParent();
2200 LastBypassBlock = CheckBlock;
2203 // Generate the code to check that the strides we assumed to be one are really
2204 // one. We want the new basic block to start at the first instruction in a
2205 // sequence of instructions that form a check.
2206 Instruction *StrideCheck;
2207 Instruction *FirstCheckInst;
2208 std::tie(FirstCheckInst, StrideCheck) =
2209 addStrideCheck(LastBypassBlock->getTerminator());
2211 // Create a new block containing the stride check.
2212 BasicBlock *CheckBlock =
2213 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2215 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2216 LoopBypassBlocks.push_back(CheckBlock);
2218 // Replace the branch into the memory check block with a conditional branch
2219 // for the "few elements case".
2220 Instruction *OldTerm = LastBypassBlock->getTerminator();
2221 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2222 OldTerm->eraseFromParent();
2225 LastBypassBlock = CheckBlock;
2228 // Generate the code that checks in runtime if arrays overlap. We put the
2229 // checks into a separate block to make the more common case of few elements
2231 Instruction *MemRuntimeCheck;
2232 std::tie(FirstCheckInst, MemRuntimeCheck) =
2233 Legal->getLAA()->addRuntimeCheck(LastBypassBlock->getTerminator());
2234 if (MemRuntimeCheck) {
2235 // Create a new block containing the memory check.
2236 BasicBlock *CheckBlock =
2237 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.memcheck");
2239 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2240 LoopBypassBlocks.push_back(CheckBlock);
2242 // Replace the branch into the memory check block with a conditional branch
2243 // for the "few elements case".
2244 Instruction *OldTerm = LastBypassBlock->getTerminator();
2245 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2246 OldTerm->eraseFromParent();
2248 Cmp = MemRuntimeCheck;
2249 LastBypassBlock = CheckBlock;
2252 LastBypassBlock->getTerminator()->eraseFromParent();
2253 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2256 // We are going to resume the execution of the scalar loop.
2257 // Go over all of the induction variables that we found and fix the
2258 // PHIs that are left in the scalar version of the loop.
2259 // The starting values of PHI nodes depend on the counter of the last
2260 // iteration in the vectorized loop.
2261 // If we come from a bypass edge then we need to start from the original
2264 // This variable saves the new starting index for the scalar loop.
2265 PHINode *ResumeIndex = nullptr;
2266 LoopVectorizationLegality::InductionList::iterator I, E;
2267 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2268 // Set builder to point to last bypass block.
2269 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2270 for (I = List->begin(), E = List->end(); I != E; ++I) {
2271 PHINode *OrigPhi = I->first;
2272 LoopVectorizationLegality::InductionInfo II = I->second;
2274 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2275 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2276 MiddleBlock->getTerminator());
2277 // We might have extended the type of the induction variable but we need a
2278 // truncated version for the scalar loop.
2279 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2280 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2281 MiddleBlock->getTerminator()) : nullptr;
2283 // Create phi nodes to merge from the backedge-taken check block.
2284 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2285 ScalarPH->getTerminator());
2286 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2288 PHINode *BCTruncResumeVal = nullptr;
2289 if (OrigPhi == OldInduction) {
2291 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2292 ScalarPH->getTerminator());
2293 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2296 Value *EndValue = nullptr;
2298 case LoopVectorizationLegality::IK_NoInduction:
2299 llvm_unreachable("Unknown induction");
2300 case LoopVectorizationLegality::IK_IntInduction: {
2301 // Handle the integer induction counter.
2302 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2304 // We have the canonical induction variable.
2305 if (OrigPhi == OldInduction) {
2306 // Create a truncated version of the resume value for the scalar loop,
2307 // we might have promoted the type to a larger width.
2309 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2310 // The new PHI merges the original incoming value, in case of a bypass,
2311 // or the value at the end of the vectorized loop.
2312 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2313 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2314 TruncResumeVal->addIncoming(EndValue, VecBody);
2316 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2318 // We know what the end value is.
2319 EndValue = IdxEndRoundDown;
2320 // We also know which PHI node holds it.
2321 ResumeIndex = ResumeVal;
2325 // Not the canonical induction variable - add the vector loop count to the
2327 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2328 II.StartValue->getType(),
2330 EndValue = II.transform(BypassBuilder, CRD);
2331 EndValue->setName("ind.end");
2334 case LoopVectorizationLegality::IK_PtrInduction: {
2335 EndValue = II.transform(BypassBuilder, CountRoundDown);
2336 EndValue->setName("ptr.ind.end");
2341 // The new PHI merges the original incoming value, in case of a bypass,
2342 // or the value at the end of the vectorized loop.
2343 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2344 if (OrigPhi == OldInduction)
2345 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2347 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2349 ResumeVal->addIncoming(EndValue, VecBody);
2351 // Fix the scalar body counter (PHI node).
2352 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2354 // The old induction's phi node in the scalar body needs the truncated
2356 if (OrigPhi == OldInduction) {
2357 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2358 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2360 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2361 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2365 // If we are generating a new induction variable then we also need to
2366 // generate the code that calculates the exit value. This value is not
2367 // simply the end of the counter because we may skip the vectorized body
2368 // in case of a runtime check.
2370 assert(!ResumeIndex && "Unexpected resume value found");
2371 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2372 MiddleBlock->getTerminator());
2373 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2374 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2375 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2378 // Make sure that we found the index where scalar loop needs to continue.
2379 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2380 "Invalid resume Index");
2382 // Add a check in the middle block to see if we have completed
2383 // all of the iterations in the first vector loop.
2384 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2385 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2386 ResumeIndex, "cmp.n",
2387 MiddleBlock->getTerminator());
2389 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2390 // Remove the old terminator.
2391 MiddleBlock->getTerminator()->eraseFromParent();
2393 // Create i+1 and fill the PHINode.
2394 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2395 Induction->addIncoming(StartIdx, VectorPH);
2396 Induction->addIncoming(NextIdx, VecBody);
2397 // Create the compare.
2398 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2399 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2401 // Now we have two terminators. Remove the old one from the block.
2402 VecBody->getTerminator()->eraseFromParent();
2404 // Get ready to start creating new instructions into the vectorized body.
2405 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2408 LoopVectorPreHeader = VectorPH;
2409 LoopScalarPreHeader = ScalarPH;
2410 LoopMiddleBlock = MiddleBlock;
2411 LoopExitBlock = ExitBlock;
2412 LoopVectorBody.push_back(VecBody);
2413 LoopScalarBody = OldBasicBlock;
2415 LoopVectorizeHints Hints(Lp, true);
2416 Hints.setAlreadyVectorized();
2419 /// This function returns the identity element (or neutral element) for
2420 /// the operation K.
2422 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2427 // Adding, Xoring, Oring zero to a number does not change it.
2428 return ConstantInt::get(Tp, 0);
2429 case RK_IntegerMult:
2430 // Multiplying a number by 1 does not change it.
2431 return ConstantInt::get(Tp, 1);
2433 // AND-ing a number with an all-1 value does not change it.
2434 return ConstantInt::get(Tp, -1, true);
2436 // Multiplying a number by 1 does not change it.
2437 return ConstantFP::get(Tp, 1.0L);
2439 // Adding zero to a number does not change it.
2440 return ConstantFP::get(Tp, 0.0L);
2442 llvm_unreachable("Unknown reduction kind");
2446 /// This function translates the reduction kind to an LLVM binary operator.
2448 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2450 case LoopVectorizationLegality::RK_IntegerAdd:
2451 return Instruction::Add;
2452 case LoopVectorizationLegality::RK_IntegerMult:
2453 return Instruction::Mul;
2454 case LoopVectorizationLegality::RK_IntegerOr:
2455 return Instruction::Or;
2456 case LoopVectorizationLegality::RK_IntegerAnd:
2457 return Instruction::And;
2458 case LoopVectorizationLegality::RK_IntegerXor:
2459 return Instruction::Xor;
2460 case LoopVectorizationLegality::RK_FloatMult:
2461 return Instruction::FMul;
2462 case LoopVectorizationLegality::RK_FloatAdd:
2463 return Instruction::FAdd;
2464 case LoopVectorizationLegality::RK_IntegerMinMax:
2465 return Instruction::ICmp;
2466 case LoopVectorizationLegality::RK_FloatMinMax:
2467 return Instruction::FCmp;
2469 llvm_unreachable("Unknown reduction operation");
2473 Value *createMinMaxOp(IRBuilder<> &Builder,
2474 LoopVectorizationLegality::MinMaxReductionKind RK,
2477 CmpInst::Predicate P = CmpInst::ICMP_NE;
2480 llvm_unreachable("Unknown min/max reduction kind");
2481 case LoopVectorizationLegality::MRK_UIntMin:
2482 P = CmpInst::ICMP_ULT;
2484 case LoopVectorizationLegality::MRK_UIntMax:
2485 P = CmpInst::ICMP_UGT;
2487 case LoopVectorizationLegality::MRK_SIntMin:
2488 P = CmpInst::ICMP_SLT;
2490 case LoopVectorizationLegality::MRK_SIntMax:
2491 P = CmpInst::ICMP_SGT;
2493 case LoopVectorizationLegality::MRK_FloatMin:
2494 P = CmpInst::FCMP_OLT;
2496 case LoopVectorizationLegality::MRK_FloatMax:
2497 P = CmpInst::FCMP_OGT;
2502 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2503 RK == LoopVectorizationLegality::MRK_FloatMax)
2504 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2506 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2508 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2513 struct CSEDenseMapInfo {
2514 static bool canHandle(Instruction *I) {
2515 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2516 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2518 static inline Instruction *getEmptyKey() {
2519 return DenseMapInfo<Instruction *>::getEmptyKey();
2521 static inline Instruction *getTombstoneKey() {
2522 return DenseMapInfo<Instruction *>::getTombstoneKey();
2524 static unsigned getHashValue(Instruction *I) {
2525 assert(canHandle(I) && "Unknown instruction!");
2526 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2527 I->value_op_end()));
2529 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2530 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2531 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2533 return LHS->isIdenticalTo(RHS);
2538 /// \brief Check whether this block is a predicated block.
2539 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2540 /// = ...; " blocks. We start with one vectorized basic block. For every
2541 /// conditional block we split this vectorized block. Therefore, every second
2542 /// block will be a predicated one.
2543 static bool isPredicatedBlock(unsigned BlockNum) {
2544 return BlockNum % 2;
2547 ///\brief Perform cse of induction variable instructions.
2548 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2549 // Perform simple cse.
2550 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2551 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2552 BasicBlock *BB = BBs[i];
2553 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2554 Instruction *In = I++;
2556 if (!CSEDenseMapInfo::canHandle(In))
2559 // Check if we can replace this instruction with any of the
2560 // visited instructions.
2561 if (Instruction *V = CSEMap.lookup(In)) {
2562 In->replaceAllUsesWith(V);
2563 In->eraseFromParent();
2566 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2567 // ...;" blocks for predicated stores. Every second block is a predicated
2569 if (isPredicatedBlock(i))
2577 /// \brief Adds a 'fast' flag to floating point operations.
2578 static Value *addFastMathFlag(Value *V) {
2579 if (isa<FPMathOperator>(V)){
2580 FastMathFlags Flags;
2581 Flags.setUnsafeAlgebra();
2582 cast<Instruction>(V)->setFastMathFlags(Flags);
2587 void InnerLoopVectorizer::vectorizeLoop() {
2588 //===------------------------------------------------===//
2590 // Notice: any optimization or new instruction that go
2591 // into the code below should be also be implemented in
2594 //===------------------------------------------------===//
2595 Constant *Zero = Builder.getInt32(0);
2597 // In order to support reduction variables we need to be able to vectorize
2598 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2599 // stages. First, we create a new vector PHI node with no incoming edges.
2600 // We use this value when we vectorize all of the instructions that use the
2601 // PHI. Next, after all of the instructions in the block are complete we
2602 // add the new incoming edges to the PHI. At this point all of the
2603 // instructions in the basic block are vectorized, so we can use them to
2604 // construct the PHI.
2605 PhiVector RdxPHIsToFix;
2607 // Scan the loop in a topological order to ensure that defs are vectorized
2609 LoopBlocksDFS DFS(OrigLoop);
2612 // Vectorize all of the blocks in the original loop.
2613 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2614 be = DFS.endRPO(); bb != be; ++bb)
2615 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2617 // At this point every instruction in the original loop is widened to
2618 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2619 // that we vectorized. The PHI nodes are currently empty because we did
2620 // not want to introduce cycles. Notice that the remaining PHI nodes
2621 // that we need to fix are reduction variables.
2623 // Create the 'reduced' values for each of the induction vars.
2624 // The reduced values are the vector values that we scalarize and combine
2625 // after the loop is finished.
2626 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2628 PHINode *RdxPhi = *it;
2629 assert(RdxPhi && "Unable to recover vectorized PHI");
2631 // Find the reduction variable descriptor.
2632 assert(Legal->getReductionVars()->count(RdxPhi) &&
2633 "Unable to find the reduction variable");
2634 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2635 (*Legal->getReductionVars())[RdxPhi];
2637 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2639 // We need to generate a reduction vector from the incoming scalar.
2640 // To do so, we need to generate the 'identity' vector and override
2641 // one of the elements with the incoming scalar reduction. We need
2642 // to do it in the vector-loop preheader.
2643 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
2645 // This is the vector-clone of the value that leaves the loop.
2646 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2647 Type *VecTy = VectorExit[0]->getType();
2649 // Find the reduction identity variable. Zero for addition, or, xor,
2650 // one for multiplication, -1 for And.
2653 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2654 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2655 // MinMax reduction have the start value as their identify.
2657 VectorStart = Identity = RdxDesc.StartValue;
2659 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2664 // Handle other reduction kinds:
2666 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2667 VecTy->getScalarType());
2670 // This vector is the Identity vector where the first element is the
2671 // incoming scalar reduction.
2672 VectorStart = RdxDesc.StartValue;
2674 Identity = ConstantVector::getSplat(VF, Iden);
2676 // This vector is the Identity vector where the first element is the
2677 // incoming scalar reduction.
2678 VectorStart = Builder.CreateInsertElement(Identity,
2679 RdxDesc.StartValue, Zero);
2683 // Fix the vector-loop phi.
2685 // Reductions do not have to start at zero. They can start with
2686 // any loop invariant values.
2687 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2688 BasicBlock *Latch = OrigLoop->getLoopLatch();
2689 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2690 VectorParts &Val = getVectorValue(LoopVal);
2691 for (unsigned part = 0; part < UF; ++part) {
2692 // Make sure to add the reduction stat value only to the
2693 // first unroll part.
2694 Value *StartVal = (part == 0) ? VectorStart : Identity;
2695 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
2696 LoopVectorPreHeader);
2697 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2698 LoopVectorBody.back());
2701 // Before each round, move the insertion point right between
2702 // the PHIs and the values we are going to write.
2703 // This allows us to write both PHINodes and the extractelement
2705 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2707 VectorParts RdxParts;
2708 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2709 for (unsigned part = 0; part < UF; ++part) {
2710 // This PHINode contains the vectorized reduction variable, or
2711 // the initial value vector, if we bypass the vector loop.
2712 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2713 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2714 Value *StartVal = (part == 0) ? VectorStart : Identity;
2715 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2716 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2717 NewPhi->addIncoming(RdxExitVal[part],
2718 LoopVectorBody.back());
2719 RdxParts.push_back(NewPhi);
2722 // Reduce all of the unrolled parts into a single vector.
2723 Value *ReducedPartRdx = RdxParts[0];
2724 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2725 setDebugLocFromInst(Builder, ReducedPartRdx);
2726 for (unsigned part = 1; part < UF; ++part) {
2727 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2728 // Floating point operations had to be 'fast' to enable the reduction.
2729 ReducedPartRdx = addFastMathFlag(
2730 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2731 ReducedPartRdx, "bin.rdx"));
2733 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2734 ReducedPartRdx, RdxParts[part]);
2738 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2739 // and vector ops, reducing the set of values being computed by half each
2741 assert(isPowerOf2_32(VF) &&
2742 "Reduction emission only supported for pow2 vectors!");
2743 Value *TmpVec = ReducedPartRdx;
2744 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2745 for (unsigned i = VF; i != 1; i >>= 1) {
2746 // Move the upper half of the vector to the lower half.
2747 for (unsigned j = 0; j != i/2; ++j)
2748 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2750 // Fill the rest of the mask with undef.
2751 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2752 UndefValue::get(Builder.getInt32Ty()));
2755 Builder.CreateShuffleVector(TmpVec,
2756 UndefValue::get(TmpVec->getType()),
2757 ConstantVector::get(ShuffleMask),
2760 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2761 // Floating point operations had to be 'fast' to enable the reduction.
2762 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2763 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2765 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2768 // The result is in the first element of the vector.
2769 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2770 Builder.getInt32(0));
2773 // Create a phi node that merges control-flow from the backedge-taken check
2774 // block and the middle block.
2775 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
2776 LoopScalarPreHeader->getTerminator());
2777 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
2778 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2780 // Now, we need to fix the users of the reduction variable
2781 // inside and outside of the scalar remainder loop.
2782 // We know that the loop is in LCSSA form. We need to update the
2783 // PHI nodes in the exit blocks.
2784 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2785 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2786 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2787 if (!LCSSAPhi) break;
2789 // All PHINodes need to have a single entry edge, or two if
2790 // we already fixed them.
2791 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2793 // We found our reduction value exit-PHI. Update it with the
2794 // incoming bypass edge.
2795 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2796 // Add an edge coming from the bypass.
2797 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2800 }// end of the LCSSA phi scan.
2802 // Fix the scalar loop reduction variable with the incoming reduction sum
2803 // from the vector body and from the backedge value.
2804 int IncomingEdgeBlockIdx =
2805 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2806 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2807 // Pick the other block.
2808 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2809 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
2810 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2811 }// end of for each redux variable.
2815 // Remove redundant induction instructions.
2816 cse(LoopVectorBody);
2819 void InnerLoopVectorizer::fixLCSSAPHIs() {
2820 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2821 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2822 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2823 if (!LCSSAPhi) break;
2824 if (LCSSAPhi->getNumIncomingValues() == 1)
2825 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2830 InnerLoopVectorizer::VectorParts
2831 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2832 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2835 // Look for cached value.
2836 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2837 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2838 if (ECEntryIt != MaskCache.end())
2839 return ECEntryIt->second;
2841 VectorParts SrcMask = createBlockInMask(Src);
2843 // The terminator has to be a branch inst!
2844 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2845 assert(BI && "Unexpected terminator found");
2847 if (BI->isConditional()) {
2848 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2850 if (BI->getSuccessor(0) != Dst)
2851 for (unsigned part = 0; part < UF; ++part)
2852 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2854 for (unsigned part = 0; part < UF; ++part)
2855 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2857 MaskCache[Edge] = EdgeMask;
2861 MaskCache[Edge] = SrcMask;
2865 InnerLoopVectorizer::VectorParts
2866 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2867 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2869 // Loop incoming mask is all-one.
2870 if (OrigLoop->getHeader() == BB) {
2871 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2872 return getVectorValue(C);
2875 // This is the block mask. We OR all incoming edges, and with zero.
2876 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2877 VectorParts BlockMask = getVectorValue(Zero);
2880 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2881 VectorParts EM = createEdgeMask(*it, BB);
2882 for (unsigned part = 0; part < UF; ++part)
2883 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2889 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2890 InnerLoopVectorizer::VectorParts &Entry,
2891 unsigned UF, unsigned VF, PhiVector *PV) {
2892 PHINode* P = cast<PHINode>(PN);
2893 // Handle reduction variables:
2894 if (Legal->getReductionVars()->count(P)) {
2895 for (unsigned part = 0; part < UF; ++part) {
2896 // This is phase one of vectorizing PHIs.
2897 Type *VecTy = (VF == 1) ? PN->getType() :
2898 VectorType::get(PN->getType(), VF);
2899 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2900 LoopVectorBody.back()-> getFirstInsertionPt());
2906 setDebugLocFromInst(Builder, P);
2907 // Check for PHI nodes that are lowered to vector selects.
2908 if (P->getParent() != OrigLoop->getHeader()) {
2909 // We know that all PHIs in non-header blocks are converted into
2910 // selects, so we don't have to worry about the insertion order and we
2911 // can just use the builder.
2912 // At this point we generate the predication tree. There may be
2913 // duplications since this is a simple recursive scan, but future
2914 // optimizations will clean it up.
2916 unsigned NumIncoming = P->getNumIncomingValues();
2918 // Generate a sequence of selects of the form:
2919 // SELECT(Mask3, In3,
2920 // SELECT(Mask2, In2,
2922 for (unsigned In = 0; In < NumIncoming; In++) {
2923 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2925 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2927 for (unsigned part = 0; part < UF; ++part) {
2928 // We might have single edge PHIs (blocks) - use an identity
2929 // 'select' for the first PHI operand.
2931 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2934 // Select between the current value and the previous incoming edge
2935 // based on the incoming mask.
2936 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2937 Entry[part], "predphi");
2943 // This PHINode must be an induction variable.
2944 // Make sure that we know about it.
2945 assert(Legal->getInductionVars()->count(P) &&
2946 "Not an induction variable");
2948 LoopVectorizationLegality::InductionInfo II =
2949 Legal->getInductionVars()->lookup(P);
2951 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
2952 // which can be found from the original scalar operations.
2954 case LoopVectorizationLegality::IK_NoInduction:
2955 llvm_unreachable("Unknown induction");
2956 case LoopVectorizationLegality::IK_IntInduction: {
2957 assert(P->getType() == II.StartValue->getType() && "Types must match");
2958 Type *PhiTy = P->getType();
2960 if (P == OldInduction) {
2961 // Handle the canonical induction variable. We might have had to
2963 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2965 // Handle other induction variables that are now based on the
2967 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2969 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2970 Broadcasted = II.transform(Builder, NormalizedIdx);
2971 Broadcasted->setName("offset.idx");
2973 Broadcasted = getBroadcastInstrs(Broadcasted);
2974 // After broadcasting the induction variable we need to make the vector
2975 // consecutive by adding 0, 1, 2, etc.
2976 for (unsigned part = 0; part < UF; ++part)
2977 Entry[part] = getStepVector(Broadcasted, VF * part, II.StepValue);
2980 case LoopVectorizationLegality::IK_PtrInduction:
2981 // Handle the pointer induction variable case.
2982 assert(P->getType()->isPointerTy() && "Unexpected type.");
2983 // This is the normalized GEP that starts counting at zero.
2984 Value *NormalizedIdx =
2985 Builder.CreateSub(Induction, ExtendedIdx, "normalized.idx");
2986 // This is the vector of results. Notice that we don't generate
2987 // vector geps because scalar geps result in better code.
2988 for (unsigned part = 0; part < UF; ++part) {
2990 int EltIndex = part;
2991 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
2992 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
2993 Value *SclrGep = II.transform(Builder, GlobalIdx);
2994 SclrGep->setName("next.gep");
2995 Entry[part] = SclrGep;
2999 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3000 for (unsigned int i = 0; i < VF; ++i) {
3001 int EltIndex = i + part * VF;
3002 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3003 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3004 Value *SclrGep = II.transform(Builder, GlobalIdx);
3005 SclrGep->setName("next.gep");
3006 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3007 Builder.getInt32(i),
3010 Entry[part] = VecVal;
3016 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3017 // For each instruction in the old loop.
3018 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3019 VectorParts &Entry = WidenMap.get(it);
3020 switch (it->getOpcode()) {
3021 case Instruction::Br:
3022 // Nothing to do for PHIs and BR, since we already took care of the
3023 // loop control flow instructions.
3025 case Instruction::PHI: {
3026 // Vectorize PHINodes.
3027 widenPHIInstruction(it, Entry, UF, VF, PV);
3031 case Instruction::Add:
3032 case Instruction::FAdd:
3033 case Instruction::Sub:
3034 case Instruction::FSub:
3035 case Instruction::Mul:
3036 case Instruction::FMul:
3037 case Instruction::UDiv:
3038 case Instruction::SDiv:
3039 case Instruction::FDiv:
3040 case Instruction::URem:
3041 case Instruction::SRem:
3042 case Instruction::FRem:
3043 case Instruction::Shl:
3044 case Instruction::LShr:
3045 case Instruction::AShr:
3046 case Instruction::And:
3047 case Instruction::Or:
3048 case Instruction::Xor: {
3049 // Just widen binops.
3050 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3051 setDebugLocFromInst(Builder, BinOp);
3052 VectorParts &A = getVectorValue(it->getOperand(0));
3053 VectorParts &B = getVectorValue(it->getOperand(1));
3055 // Use this vector value for all users of the original instruction.
3056 for (unsigned Part = 0; Part < UF; ++Part) {
3057 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3059 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3060 VecOp->copyIRFlags(BinOp);
3065 propagateMetadata(Entry, it);
3068 case Instruction::Select: {
3070 // If the selector is loop invariant we can create a select
3071 // instruction with a scalar condition. Otherwise, use vector-select.
3072 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3074 setDebugLocFromInst(Builder, it);
3076 // The condition can be loop invariant but still defined inside the
3077 // loop. This means that we can't just use the original 'cond' value.
3078 // We have to take the 'vectorized' value and pick the first lane.
3079 // Instcombine will make this a no-op.
3080 VectorParts &Cond = getVectorValue(it->getOperand(0));
3081 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3082 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3084 Value *ScalarCond = (VF == 1) ? Cond[0] :
3085 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3087 for (unsigned Part = 0; Part < UF; ++Part) {
3088 Entry[Part] = Builder.CreateSelect(
3089 InvariantCond ? ScalarCond : Cond[Part],
3094 propagateMetadata(Entry, it);
3098 case Instruction::ICmp:
3099 case Instruction::FCmp: {
3100 // Widen compares. Generate vector compares.
3101 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3102 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3103 setDebugLocFromInst(Builder, it);
3104 VectorParts &A = getVectorValue(it->getOperand(0));
3105 VectorParts &B = getVectorValue(it->getOperand(1));
3106 for (unsigned Part = 0; Part < UF; ++Part) {
3109 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3111 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3115 propagateMetadata(Entry, it);
3119 case Instruction::Store:
3120 case Instruction::Load:
3121 vectorizeMemoryInstruction(it);
3123 case Instruction::ZExt:
3124 case Instruction::SExt:
3125 case Instruction::FPToUI:
3126 case Instruction::FPToSI:
3127 case Instruction::FPExt:
3128 case Instruction::PtrToInt:
3129 case Instruction::IntToPtr:
3130 case Instruction::SIToFP:
3131 case Instruction::UIToFP:
3132 case Instruction::Trunc:
3133 case Instruction::FPTrunc:
3134 case Instruction::BitCast: {
3135 CastInst *CI = dyn_cast<CastInst>(it);
3136 setDebugLocFromInst(Builder, it);
3137 /// Optimize the special case where the source is the induction
3138 /// variable. Notice that we can only optimize the 'trunc' case
3139 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3140 /// c. other casts depend on pointer size.
3141 if (CI->getOperand(0) == OldInduction &&
3142 it->getOpcode() == Instruction::Trunc) {
3143 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3145 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3146 LoopVectorizationLegality::InductionInfo II =
3147 Legal->getInductionVars()->lookup(OldInduction);
3149 ConstantInt::getSigned(CI->getType(), II.StepValue->getSExtValue());
3150 for (unsigned Part = 0; Part < UF; ++Part)
3151 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
3152 propagateMetadata(Entry, it);
3155 /// Vectorize casts.
3156 Type *DestTy = (VF == 1) ? CI->getType() :
3157 VectorType::get(CI->getType(), VF);
3159 VectorParts &A = getVectorValue(it->getOperand(0));
3160 for (unsigned Part = 0; Part < UF; ++Part)
3161 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3162 propagateMetadata(Entry, it);
3166 case Instruction::Call: {
3167 // Ignore dbg intrinsics.
3168 if (isa<DbgInfoIntrinsic>(it))
3170 setDebugLocFromInst(Builder, it);
3172 Module *M = BB->getParent()->getParent();
3173 CallInst *CI = cast<CallInst>(it);
3174 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3175 assert(ID && "Not an intrinsic call!");
3177 case Intrinsic::assume:
3178 case Intrinsic::lifetime_end:
3179 case Intrinsic::lifetime_start:
3180 scalarizeInstruction(it);
3183 bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
3184 for (unsigned Part = 0; Part < UF; ++Part) {
3185 SmallVector<Value *, 4> Args;
3186 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3187 if (HasScalarOpd && i == 1) {
3188 Args.push_back(CI->getArgOperand(i));
3191 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3192 Args.push_back(Arg[Part]);
3194 Type *Tys[] = {CI->getType()};
3196 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3198 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3199 Entry[Part] = Builder.CreateCall(F, Args);
3202 propagateMetadata(Entry, it);
3209 // All other instructions are unsupported. Scalarize them.
3210 scalarizeInstruction(it);
3213 }// end of for_each instr.
3216 void InnerLoopVectorizer::updateAnalysis() {
3217 // Forget the original basic block.
3218 SE->forgetLoop(OrigLoop);
3220 // Update the dominator tree information.
3221 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3222 "Entry does not dominate exit.");
3224 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3225 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3226 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3228 // Due to if predication of stores we might create a sequence of "if(pred)
3229 // a[i] = ...; " blocks.
3230 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3232 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3233 else if (isPredicatedBlock(i)) {
3234 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3236 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3240 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3241 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3242 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3243 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3245 DEBUG(DT->verifyDomTree());
3248 /// \brief Check whether it is safe to if-convert this phi node.
3250 /// Phi nodes with constant expressions that can trap are not safe to if
3252 static bool canIfConvertPHINodes(BasicBlock *BB) {
3253 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3254 PHINode *Phi = dyn_cast<PHINode>(I);
3257 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3258 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3265 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3266 if (!EnableIfConversion) {
3267 emitAnalysis(VectorizationReport() << "if-conversion is disabled");
3271 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3273 // A list of pointers that we can safely read and write to.
3274 SmallPtrSet<Value *, 8> SafePointes;
3276 // Collect safe addresses.
3277 for (Loop::block_iterator BI = TheLoop->block_begin(),
3278 BE = TheLoop->block_end(); BI != BE; ++BI) {
3279 BasicBlock *BB = *BI;
3281 if (blockNeedsPredication(BB))
3284 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3285 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3286 SafePointes.insert(LI->getPointerOperand());
3287 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3288 SafePointes.insert(SI->getPointerOperand());
3292 // Collect the blocks that need predication.
3293 BasicBlock *Header = TheLoop->getHeader();
3294 for (Loop::block_iterator BI = TheLoop->block_begin(),
3295 BE = TheLoop->block_end(); BI != BE; ++BI) {
3296 BasicBlock *BB = *BI;
3298 // We don't support switch statements inside loops.
3299 if (!isa<BranchInst>(BB->getTerminator())) {
3300 emitAnalysis(VectorizationReport(BB->getTerminator())
3301 << "loop contains a switch statement");
3305 // We must be able to predicate all blocks that need to be predicated.
3306 if (blockNeedsPredication(BB)) {
3307 if (!blockCanBePredicated(BB, SafePointes)) {
3308 emitAnalysis(VectorizationReport(BB->getTerminator())
3309 << "control flow cannot be substituted for a select");
3312 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3313 emitAnalysis(VectorizationReport(BB->getTerminator())
3314 << "control flow cannot be substituted for a select");
3319 // We can if-convert this loop.
3323 bool LoopVectorizationLegality::canVectorize() {
3324 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3325 // be canonicalized.
3326 if (!TheLoop->getLoopPreheader()) {
3328 VectorizationReport() <<
3329 "loop control flow is not understood by vectorizer");
3333 // We can only vectorize innermost loops.
3334 if (!TheLoop->getSubLoopsVector().empty()) {
3335 emitAnalysis(VectorizationReport() << "loop is not the innermost loop");
3339 // We must have a single backedge.
3340 if (TheLoop->getNumBackEdges() != 1) {
3342 VectorizationReport() <<
3343 "loop control flow is not understood by vectorizer");
3347 // We must have a single exiting block.
3348 if (!TheLoop->getExitingBlock()) {
3350 VectorizationReport() <<
3351 "loop control flow is not understood by vectorizer");
3355 // We only handle bottom-tested loops, i.e. loop in which the condition is
3356 // checked at the end of each iteration. With that we can assume that all
3357 // instructions in the loop are executed the same number of times.
3358 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
3360 VectorizationReport() <<
3361 "loop control flow is not understood by vectorizer");
3365 // We need to have a loop header.
3366 DEBUG(dbgs() << "LV: Found a loop: " <<
3367 TheLoop->getHeader()->getName() << '\n');
3369 // Check if we can if-convert non-single-bb loops.
3370 unsigned NumBlocks = TheLoop->getNumBlocks();
3371 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3372 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3376 // ScalarEvolution needs to be able to find the exit count.
3377 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3378 if (ExitCount == SE->getCouldNotCompute()) {
3379 emitAnalysis(VectorizationReport() <<
3380 "could not determine number of loop iterations");
3381 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3385 // Check if we can vectorize the instructions and CFG in this loop.
3386 if (!canVectorizeInstrs()) {
3387 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3391 // Go over each instruction and look at memory deps.
3392 if (!canVectorizeMemory()) {
3393 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3397 // Collect all of the variables that remain uniform after vectorization.
3398 collectLoopUniforms();
3400 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3401 (LAA.getRuntimePointerCheck()->Need ? " (with a runtime bound check)" :
3405 // Okay! We can vectorize. At this point we don't have any other mem analysis
3406 // which may limit our maximum vectorization factor, so just return true with
3411 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3412 if (Ty->isPointerTy())
3413 return DL.getIntPtrType(Ty);
3415 // It is possible that char's or short's overflow when we ask for the loop's
3416 // trip count, work around this by changing the type size.
3417 if (Ty->getScalarSizeInBits() < 32)
3418 return Type::getInt32Ty(Ty->getContext());
3423 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3424 Ty0 = convertPointerToIntegerType(DL, Ty0);
3425 Ty1 = convertPointerToIntegerType(DL, Ty1);
3426 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3431 /// \brief Check that the instruction has outside loop users and is not an
3432 /// identified reduction variable.
3433 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3434 SmallPtrSetImpl<Value *> &Reductions) {
3435 // Reduction instructions are allowed to have exit users. All other
3436 // instructions must not have external users.
3437 if (!Reductions.count(Inst))
3438 //Check that all of the users of the loop are inside the BB.
3439 for (User *U : Inst->users()) {
3440 Instruction *UI = cast<Instruction>(U);
3441 // This user may be a reduction exit value.
3442 if (!TheLoop->contains(UI)) {
3443 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3450 bool LoopVectorizationLegality::canVectorizeInstrs() {
3451 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3452 BasicBlock *Header = TheLoop->getHeader();
3454 // Look for the attribute signaling the absence of NaNs.
3455 Function &F = *Header->getParent();
3456 if (F.hasFnAttribute("no-nans-fp-math"))
3457 HasFunNoNaNAttr = F.getAttributes().getAttribute(
3458 AttributeSet::FunctionIndex,
3459 "no-nans-fp-math").getValueAsString() == "true";
3461 // For each block in the loop.
3462 for (Loop::block_iterator bb = TheLoop->block_begin(),
3463 be = TheLoop->block_end(); bb != be; ++bb) {
3465 // Scan the instructions in the block and look for hazards.
3466 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3469 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3470 Type *PhiTy = Phi->getType();
3471 // Check that this PHI type is allowed.
3472 if (!PhiTy->isIntegerTy() &&
3473 !PhiTy->isFloatingPointTy() &&
3474 !PhiTy->isPointerTy()) {
3475 emitAnalysis(VectorizationReport(it)
3476 << "loop control flow is not understood by vectorizer");
3477 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3481 // If this PHINode is not in the header block, then we know that we
3482 // can convert it to select during if-conversion. No need to check if
3483 // the PHIs in this block are induction or reduction variables.
3484 if (*bb != Header) {
3485 // Check that this instruction has no outside users or is an
3486 // identified reduction value with an outside user.
3487 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3489 emitAnalysis(VectorizationReport(it) <<
3490 "value could not be identified as "
3491 "an induction or reduction variable");
3495 // We only allow if-converted PHIs with exactly two incoming values.
3496 if (Phi->getNumIncomingValues() != 2) {
3497 emitAnalysis(VectorizationReport(it)
3498 << "control flow not understood by vectorizer");
3499 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3503 // This is the value coming from the preheader.
3504 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3505 ConstantInt *StepValue = nullptr;
3506 // Check if this is an induction variable.
3507 InductionKind IK = isInductionVariable(Phi, StepValue);
3509 if (IK_NoInduction != IK) {
3510 // Get the widest type.
3512 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3514 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3516 // Int inductions are special because we only allow one IV.
3517 if (IK == IK_IntInduction && StepValue->isOne()) {
3518 // Use the phi node with the widest type as induction. Use the last
3519 // one if there are multiple (no good reason for doing this other
3520 // than it is expedient).
3521 if (!Induction || PhiTy == WidestIndTy)
3525 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3526 Inductions[Phi] = InductionInfo(StartValue, IK, StepValue);
3528 // Until we explicitly handle the case of an induction variable with
3529 // an outside loop user we have to give up vectorizing this loop.
3530 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3531 emitAnalysis(VectorizationReport(it) <<
3532 "use of induction value outside of the "
3533 "loop is not handled by vectorizer");
3540 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3541 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3544 if (AddReductionVar(Phi, RK_IntegerMult)) {
3545 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3548 if (AddReductionVar(Phi, RK_IntegerOr)) {
3549 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3552 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3553 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3556 if (AddReductionVar(Phi, RK_IntegerXor)) {
3557 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3560 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3561 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3564 if (AddReductionVar(Phi, RK_FloatMult)) {
3565 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3568 if (AddReductionVar(Phi, RK_FloatAdd)) {
3569 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3572 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3573 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3578 emitAnalysis(VectorizationReport(it) <<
3579 "value that could not be identified as "
3580 "reduction is used outside the loop");
3581 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3583 }// end of PHI handling
3585 // We still don't handle functions. However, we can ignore dbg intrinsic
3586 // calls and we do handle certain intrinsic and libm functions.
3587 CallInst *CI = dyn_cast<CallInst>(it);
3588 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3589 emitAnalysis(VectorizationReport(it) <<
3590 "call instruction cannot be vectorized");
3591 DEBUG(dbgs() << "LV: Found a call site.\n");
3595 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
3596 // second argument is the same (i.e. loop invariant)
3598 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
3599 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
3600 emitAnalysis(VectorizationReport(it)
3601 << "intrinsic instruction cannot be vectorized");
3602 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
3607 // Check that the instruction return type is vectorizable.
3608 // Also, we can't vectorize extractelement instructions.
3609 if ((!VectorType::isValidElementType(it->getType()) &&
3610 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3611 emitAnalysis(VectorizationReport(it)
3612 << "instruction return type cannot be vectorized");
3613 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3617 // Check that the stored type is vectorizable.
3618 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3619 Type *T = ST->getValueOperand()->getType();
3620 if (!VectorType::isValidElementType(T)) {
3621 emitAnalysis(VectorizationReport(ST) <<
3622 "store instruction cannot be vectorized");
3625 if (EnableMemAccessVersioning)
3626 collectStridedAccess(ST);
3629 if (EnableMemAccessVersioning)
3630 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3631 collectStridedAccess(LI);
3633 // Reduction instructions are allowed to have exit users.
3634 // All other instructions must not have external users.
3635 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3636 emitAnalysis(VectorizationReport(it) <<
3637 "value cannot be used outside the loop");
3646 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3647 if (Inductions.empty()) {
3648 emitAnalysis(VectorizationReport()
3649 << "loop induction variable could not be identified");
3657 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3658 /// return the induction operand of the gep pointer.
3659 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3660 const DataLayout *DL, Loop *Lp) {
3661 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3665 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3667 // Check that all of the gep indices are uniform except for our induction
3669 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3670 if (i != InductionOperand &&
3671 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3673 return GEP->getOperand(InductionOperand);
3676 ///\brief Look for a cast use of the passed value.
3677 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3678 Value *UniqueCast = nullptr;
3679 for (User *U : Ptr->users()) {
3680 CastInst *CI = dyn_cast<CastInst>(U);
3681 if (CI && CI->getType() == Ty) {
3691 ///\brief Get the stride of a pointer access in a loop.
3692 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3693 /// pointer to the Value, or null otherwise.
3694 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3695 const DataLayout *DL, Loop *Lp) {
3696 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3697 if (!PtrTy || PtrTy->isAggregateType())
3700 // Try to remove a gep instruction to make the pointer (actually index at this
3701 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3702 // pointer, otherwise, we are analyzing the index.
3703 Value *OrigPtr = Ptr;
3705 // The size of the pointer access.
3706 int64_t PtrAccessSize = 1;
3708 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3709 const SCEV *V = SE->getSCEV(Ptr);
3713 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3714 V = C->getOperand();
3716 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3720 V = S->getStepRecurrence(*SE);
3724 // Strip off the size of access multiplication if we are still analyzing the
3726 if (OrigPtr == Ptr) {
3727 DL->getTypeAllocSize(PtrTy->getElementType());
3728 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3729 if (M->getOperand(0)->getSCEVType() != scConstant)
3732 const APInt &APStepVal =
3733 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3735 // Huge step value - give up.
3736 if (APStepVal.getBitWidth() > 64)
3739 int64_t StepVal = APStepVal.getSExtValue();
3740 if (PtrAccessSize != StepVal)
3742 V = M->getOperand(1);
3747 Type *StripedOffRecurrenceCast = nullptr;
3748 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3749 StripedOffRecurrenceCast = C->getType();
3750 V = C->getOperand();
3753 // Look for the loop invariant symbolic value.
3754 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3758 Value *Stride = U->getValue();
3759 if (!Lp->isLoopInvariant(Stride))
3762 // If we have stripped off the recurrence cast we have to make sure that we
3763 // return the value that is used in this loop so that we can replace it later.
3764 if (StripedOffRecurrenceCast)
3765 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3770 void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) {
3771 Value *Ptr = nullptr;
3772 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3773 Ptr = LI->getPointerOperand();
3774 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3775 Ptr = SI->getPointerOperand();
3779 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3783 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3784 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3785 Strides[Ptr] = Stride;
3786 StrideSet.insert(Stride);
3789 void LoopVectorizationLegality::collectLoopUniforms() {
3790 // We now know that the loop is vectorizable!
3791 // Collect variables that will remain uniform after vectorization.
3792 std::vector<Value*> Worklist;
3793 BasicBlock *Latch = TheLoop->getLoopLatch();
3795 // Start with the conditional branch and walk up the block.
3796 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3798 // Also add all consecutive pointer values; these values will be uniform
3799 // after vectorization (and subsequent cleanup) and, until revectorization is
3800 // supported, all dependencies must also be uniform.
3801 for (Loop::block_iterator B = TheLoop->block_begin(),
3802 BE = TheLoop->block_end(); B != BE; ++B)
3803 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
3805 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
3806 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3808 while (!Worklist.empty()) {
3809 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3810 Worklist.pop_back();
3812 // Look at instructions inside this loop.
3813 // Stop when reaching PHI nodes.
3814 // TODO: we need to follow values all over the loop, not only in this block.
3815 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3818 // This is a known uniform.
3821 // Insert all operands.
3822 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3826 bool LoopVectorizationLegality::canVectorizeMemory() {
3827 return LAA.canVectorizeMemory(Strides);
3830 static bool hasMultipleUsesOf(Instruction *I,
3831 SmallPtrSetImpl<Instruction *> &Insts) {
3832 unsigned NumUses = 0;
3833 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
3834 if (Insts.count(dyn_cast<Instruction>(*Use)))
3843 static bool areAllUsesIn(Instruction *I, SmallPtrSetImpl<Instruction *> &Set) {
3844 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
3845 if (!Set.count(dyn_cast<Instruction>(*Use)))
3850 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
3851 ReductionKind Kind) {
3852 if (Phi->getNumIncomingValues() != 2)
3855 // Reduction variables are only found in the loop header block.
3856 if (Phi->getParent() != TheLoop->getHeader())
3859 // Obtain the reduction start value from the value that comes from the loop
3861 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
3863 // ExitInstruction is the single value which is used outside the loop.
3864 // We only allow for a single reduction value to be used outside the loop.
3865 // This includes users of the reduction, variables (which form a cycle
3866 // which ends in the phi node).
3867 Instruction *ExitInstruction = nullptr;
3868 // Indicates that we found a reduction operation in our scan.
3869 bool FoundReduxOp = false;
3871 // We start with the PHI node and scan for all of the users of this
3872 // instruction. All users must be instructions that can be used as reduction
3873 // variables (such as ADD). We must have a single out-of-block user. The cycle
3874 // must include the original PHI.
3875 bool FoundStartPHI = false;
3877 // To recognize min/max patterns formed by a icmp select sequence, we store
3878 // the number of instruction we saw from the recognized min/max pattern,
3879 // to make sure we only see exactly the two instructions.
3880 unsigned NumCmpSelectPatternInst = 0;
3881 ReductionInstDesc ReduxDesc(false, nullptr);
3883 SmallPtrSet<Instruction *, 8> VisitedInsts;
3884 SmallVector<Instruction *, 8> Worklist;
3885 Worklist.push_back(Phi);
3886 VisitedInsts.insert(Phi);
3888 // A value in the reduction can be used:
3889 // - By the reduction:
3890 // - Reduction operation:
3891 // - One use of reduction value (safe).
3892 // - Multiple use of reduction value (not safe).
3894 // - All uses of the PHI must be the reduction (safe).
3895 // - Otherwise, not safe.
3896 // - By one instruction outside of the loop (safe).
3897 // - By further instructions outside of the loop (not safe).
3898 // - By an instruction that is not part of the reduction (not safe).
3900 // * An instruction type other than PHI or the reduction operation.
3901 // * A PHI in the header other than the initial PHI.
3902 while (!Worklist.empty()) {
3903 Instruction *Cur = Worklist.back();
3904 Worklist.pop_back();
3907 // If the instruction has no users then this is a broken chain and can't be
3908 // a reduction variable.
3909 if (Cur->use_empty())
3912 bool IsAPhi = isa<PHINode>(Cur);
3914 // A header PHI use other than the original PHI.
3915 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
3918 // Reductions of instructions such as Div, and Sub is only possible if the
3919 // LHS is the reduction variable.
3920 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
3921 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
3922 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
3925 // Any reduction instruction must be of one of the allowed kinds.
3926 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
3927 if (!ReduxDesc.IsReduction)
3930 // A reduction operation must only have one use of the reduction value.
3931 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
3932 hasMultipleUsesOf(Cur, VisitedInsts))
3935 // All inputs to a PHI node must be a reduction value.
3936 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
3939 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
3940 isa<SelectInst>(Cur)))
3941 ++NumCmpSelectPatternInst;
3942 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
3943 isa<SelectInst>(Cur)))
3944 ++NumCmpSelectPatternInst;
3946 // Check whether we found a reduction operator.
3947 FoundReduxOp |= !IsAPhi;
3949 // Process users of current instruction. Push non-PHI nodes after PHI nodes
3950 // onto the stack. This way we are going to have seen all inputs to PHI
3951 // nodes once we get to them.
3952 SmallVector<Instruction *, 8> NonPHIs;
3953 SmallVector<Instruction *, 8> PHIs;
3954 for (User *U : Cur->users()) {
3955 Instruction *UI = cast<Instruction>(U);
3957 // Check if we found the exit user.
3958 BasicBlock *Parent = UI->getParent();
3959 if (!TheLoop->contains(Parent)) {
3960 // Exit if you find multiple outside users or if the header phi node is
3961 // being used. In this case the user uses the value of the previous
3962 // iteration, in which case we would loose "VF-1" iterations of the
3963 // reduction operation if we vectorize.
3964 if (ExitInstruction != nullptr || Cur == Phi)
3967 // The instruction used by an outside user must be the last instruction
3968 // before we feed back to the reduction phi. Otherwise, we loose VF-1
3969 // operations on the value.
3970 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
3973 ExitInstruction = Cur;
3977 // Process instructions only once (termination). Each reduction cycle
3978 // value must only be used once, except by phi nodes and min/max
3979 // reductions which are represented as a cmp followed by a select.
3980 ReductionInstDesc IgnoredVal(false, nullptr);
3981 if (VisitedInsts.insert(UI).second) {
3982 if (isa<PHINode>(UI))
3985 NonPHIs.push_back(UI);
3986 } else if (!isa<PHINode>(UI) &&
3987 ((!isa<FCmpInst>(UI) &&
3988 !isa<ICmpInst>(UI) &&
3989 !isa<SelectInst>(UI)) ||
3990 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
3993 // Remember that we completed the cycle.
3995 FoundStartPHI = true;
3997 Worklist.append(PHIs.begin(), PHIs.end());
3998 Worklist.append(NonPHIs.begin(), NonPHIs.end());
4001 // This means we have seen one but not the other instruction of the
4002 // pattern or more than just a select and cmp.
4003 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
4004 NumCmpSelectPatternInst != 2)
4007 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
4010 // We found a reduction var if we have reached the original phi node and we
4011 // only have a single instruction with out-of-loop users.
4013 // This instruction is allowed to have out-of-loop users.
4014 AllowedExit.insert(ExitInstruction);
4016 // Save the description of this reduction variable.
4017 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
4018 ReduxDesc.MinMaxKind);
4019 Reductions[Phi] = RD;
4020 // We've ended the cycle. This is a reduction variable if we have an
4021 // outside user and it has a binary op.
4026 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
4027 /// pattern corresponding to a min(X, Y) or max(X, Y).
4028 LoopVectorizationLegality::ReductionInstDesc
4029 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
4030 ReductionInstDesc &Prev) {
4032 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
4033 "Expect a select instruction");
4034 Instruction *Cmp = nullptr;
4035 SelectInst *Select = nullptr;
4037 // We must handle the select(cmp()) as a single instruction. Advance to the
4039 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
4040 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
4041 return ReductionInstDesc(false, I);
4042 return ReductionInstDesc(Select, Prev.MinMaxKind);
4045 // Only handle single use cases for now.
4046 if (!(Select = dyn_cast<SelectInst>(I)))
4047 return ReductionInstDesc(false, I);
4048 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
4049 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
4050 return ReductionInstDesc(false, I);
4051 if (!Cmp->hasOneUse())
4052 return ReductionInstDesc(false, I);
4057 // Look for a min/max pattern.
4058 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4059 return ReductionInstDesc(Select, MRK_UIntMin);
4060 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4061 return ReductionInstDesc(Select, MRK_UIntMax);
4062 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4063 return ReductionInstDesc(Select, MRK_SIntMax);
4064 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4065 return ReductionInstDesc(Select, MRK_SIntMin);
4066 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4067 return ReductionInstDesc(Select, MRK_FloatMin);
4068 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4069 return ReductionInstDesc(Select, MRK_FloatMax);
4070 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4071 return ReductionInstDesc(Select, MRK_FloatMin);
4072 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4073 return ReductionInstDesc(Select, MRK_FloatMax);
4075 return ReductionInstDesc(false, I);
4078 LoopVectorizationLegality::ReductionInstDesc
4079 LoopVectorizationLegality::isReductionInstr(Instruction *I,
4081 ReductionInstDesc &Prev) {
4082 bool FP = I->getType()->isFloatingPointTy();
4083 bool FastMath = FP && I->hasUnsafeAlgebra();
4084 switch (I->getOpcode()) {
4086 return ReductionInstDesc(false, I);
4087 case Instruction::PHI:
4088 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
4089 Kind != RK_FloatMinMax))
4090 return ReductionInstDesc(false, I);
4091 return ReductionInstDesc(I, Prev.MinMaxKind);
4092 case Instruction::Sub:
4093 case Instruction::Add:
4094 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4095 case Instruction::Mul:
4096 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4097 case Instruction::And:
4098 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4099 case Instruction::Or:
4100 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4101 case Instruction::Xor:
4102 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4103 case Instruction::FMul:
4104 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4105 case Instruction::FSub:
4106 case Instruction::FAdd:
4107 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4108 case Instruction::FCmp:
4109 case Instruction::ICmp:
4110 case Instruction::Select:
4111 if (Kind != RK_IntegerMinMax &&
4112 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4113 return ReductionInstDesc(false, I);
4114 return isMinMaxSelectCmpPattern(I, Prev);
4118 LoopVectorizationLegality::InductionKind
4119 LoopVectorizationLegality::isInductionVariable(PHINode *Phi,
4120 ConstantInt *&StepValue) {
4121 Type *PhiTy = Phi->getType();
4122 // We only handle integer and pointer inductions variables.
4123 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4124 return IK_NoInduction;
4126 // Check that the PHI is consecutive.
4127 const SCEV *PhiScev = SE->getSCEV(Phi);
4128 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4130 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4131 return IK_NoInduction;
4134 const SCEV *Step = AR->getStepRecurrence(*SE);
4135 // Calculate the pointer stride and check if it is consecutive.
4136 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4138 return IK_NoInduction;
4140 ConstantInt *CV = C->getValue();
4141 if (PhiTy->isIntegerTy()) {
4143 return IK_IntInduction;
4146 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4147 Type *PointerElementType = PhiTy->getPointerElementType();
4148 // The pointer stride cannot be determined if the pointer element type is not
4150 if (!PointerElementType->isSized())
4151 return IK_NoInduction;
4153 int64_t Size = static_cast<int64_t>(DL->getTypeAllocSize(PointerElementType));
4154 int64_t CVSize = CV->getSExtValue();
4156 return IK_NoInduction;
4157 StepValue = ConstantInt::getSigned(CV->getType(), CVSize / Size);
4158 return IK_PtrInduction;
4161 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4162 Value *In0 = const_cast<Value*>(V);
4163 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4167 return Inductions.count(PN);
4170 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4171 return LAA.blockNeedsPredication(BB);
4174 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4175 SmallPtrSetImpl<Value *> &SafePtrs) {
4177 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4178 // Check that we don't have a constant expression that can trap as operand.
4179 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4181 if (Constant *C = dyn_cast<Constant>(*OI))
4185 // We might be able to hoist the load.
4186 if (it->mayReadFromMemory()) {
4187 LoadInst *LI = dyn_cast<LoadInst>(it);
4190 if (!SafePtrs.count(LI->getPointerOperand())) {
4191 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
4192 MaskedOp.insert(LI);
4199 // We don't predicate stores at the moment.
4200 if (it->mayWriteToMemory()) {
4201 StoreInst *SI = dyn_cast<StoreInst>(it);
4202 // We only support predication of stores in basic blocks with one
4207 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
4208 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
4210 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
4211 !isSinglePredecessor) {
4212 // Build a masked store if it is legal for the target, otherwise scalarize
4214 bool isLegalMaskedOp =
4215 isLegalMaskedStore(SI->getValueOperand()->getType(),
4216 SI->getPointerOperand());
4217 if (isLegalMaskedOp) {
4219 MaskedOp.insert(SI);
4228 // The instructions below can trap.
4229 switch (it->getOpcode()) {
4231 case Instruction::UDiv:
4232 case Instruction::SDiv:
4233 case Instruction::URem:
4234 case Instruction::SRem:
4242 LoopVectorizationCostModel::VectorizationFactor
4243 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
4244 // Width 1 means no vectorize
4245 VectorizationFactor Factor = { 1U, 0U };
4246 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4247 emitAnalysis(VectorizationReport() <<
4248 "runtime pointer checks needed. Enable vectorization of this "
4249 "loop with '#pragma clang loop vectorize(enable)' when "
4250 "compiling with -Os");
4251 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4255 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
4256 emitAnalysis(VectorizationReport() <<
4257 "store that is conditionally executed prevents vectorization");
4258 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
4262 // Find the trip count.
4263 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4264 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4266 unsigned WidestType = getWidestType();
4267 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4268 unsigned MaxSafeDepDist = -1U;
4269 if (Legal->getMaxSafeDepDistBytes() != -1U)
4270 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4271 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4272 WidestRegister : MaxSafeDepDist);
4273 unsigned MaxVectorSize = WidestRegister / WidestType;
4274 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4275 DEBUG(dbgs() << "LV: The Widest register is: "
4276 << WidestRegister << " bits.\n");
4278 if (MaxVectorSize == 0) {
4279 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4283 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
4284 " into one vector!");
4286 unsigned VF = MaxVectorSize;
4288 // If we optimize the program for size, avoid creating the tail loop.
4290 // If we are unable to calculate the trip count then don't try to vectorize.
4293 (VectorizationReport() <<
4294 "unable to calculate the loop count due to complex control flow");
4295 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4299 // Find the maximum SIMD width that can fit within the trip count.
4300 VF = TC % MaxVectorSize;
4305 // If the trip count that we found modulo the vectorization factor is not
4306 // zero then we require a tail.
4308 emitAnalysis(VectorizationReport() <<
4309 "cannot optimize for size and vectorize at the "
4310 "same time. Enable vectorization of this loop "
4311 "with '#pragma clang loop vectorize(enable)' "
4312 "when compiling with -Os");
4313 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4318 int UserVF = Hints->getWidth();
4320 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4321 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4323 Factor.Width = UserVF;
4327 float Cost = expectedCost(1);
4329 const float ScalarCost = Cost;
4332 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
4334 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
4335 // Ignore scalar width, because the user explicitly wants vectorization.
4336 if (ForceVectorization && VF > 1) {
4338 Cost = expectedCost(Width) / (float)Width;
4341 for (unsigned i=2; i <= VF; i*=2) {
4342 // Notice that the vector loop needs to be executed less times, so
4343 // we need to divide the cost of the vector loops by the width of
4344 // the vector elements.
4345 float VectorCost = expectedCost(i) / (float)i;
4346 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
4347 (int)VectorCost << ".\n");
4348 if (VectorCost < Cost) {
4354 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
4355 << "LV: Vectorization seems to be not beneficial, "
4356 << "but was forced by a user.\n");
4357 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
4358 Factor.Width = Width;
4359 Factor.Cost = Width * Cost;
4363 unsigned LoopVectorizationCostModel::getWidestType() {
4364 unsigned MaxWidth = 8;
4367 for (Loop::block_iterator bb = TheLoop->block_begin(),
4368 be = TheLoop->block_end(); bb != be; ++bb) {
4369 BasicBlock *BB = *bb;
4371 // For each instruction in the loop.
4372 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4373 Type *T = it->getType();
4375 // Ignore ephemeral values.
4376 if (EphValues.count(it))
4379 // Only examine Loads, Stores and PHINodes.
4380 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4383 // Examine PHI nodes that are reduction variables.
4384 if (PHINode *PN = dyn_cast<PHINode>(it))
4385 if (!Legal->getReductionVars()->count(PN))
4388 // Examine the stored values.
4389 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4390 T = ST->getValueOperand()->getType();
4392 // Ignore loaded pointer types and stored pointer types that are not
4393 // consecutive. However, we do want to take consecutive stores/loads of
4394 // pointer vectors into account.
4395 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4398 MaxWidth = std::max(MaxWidth,
4399 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
4407 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
4409 unsigned LoopCost) {
4411 // -- The unroll heuristics --
4412 // We unroll the loop in order to expose ILP and reduce the loop overhead.
4413 // There are many micro-architectural considerations that we can't predict
4414 // at this level. For example, frontend pressure (on decode or fetch) due to
4415 // code size, or the number and capabilities of the execution ports.
4417 // We use the following heuristics to select the unroll factor:
4418 // 1. If the code has reductions, then we unroll in order to break the cross
4419 // iteration dependency.
4420 // 2. If the loop is really small, then we unroll in order to reduce the loop
4422 // 3. We don't unroll if we think that we will spill registers to memory due
4423 // to the increased register pressure.
4425 // Use the user preference, unless 'auto' is selected.
4426 int UserUF = Hints->getInterleave();
4430 // When we optimize for size, we don't unroll.
4434 // We used the distance for the unroll factor.
4435 if (Legal->getMaxSafeDepDistBytes() != -1U)
4438 // Do not unroll loops with a relatively small trip count.
4439 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4440 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
4443 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
4444 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
4448 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
4449 TargetNumRegisters = ForceTargetNumScalarRegs;
4451 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
4452 TargetNumRegisters = ForceTargetNumVectorRegs;
4455 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4456 // We divide by these constants so assume that we have at least one
4457 // instruction that uses at least one register.
4458 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4459 R.NumInstructions = std::max(R.NumInstructions, 1U);
4461 // We calculate the unroll factor using the following formula.
4462 // Subtract the number of loop invariants from the number of available
4463 // registers. These registers are used by all of the unrolled instances.
4464 // Next, divide the remaining registers by the number of registers that is
4465 // required by the loop, in order to estimate how many parallel instances
4466 // fit without causing spills. All of this is rounded down if necessary to be
4467 // a power of two. We want power of two unroll factors to simplify any
4468 // addressing operations or alignment considerations.
4469 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
4472 // Don't count the induction variable as unrolled.
4473 if (EnableIndVarRegisterHeur)
4474 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
4475 std::max(1U, (R.MaxLocalUsers - 1)));
4477 // Clamp the unroll factor ranges to reasonable factors.
4478 unsigned MaxInterleaveSize = TTI.getMaxInterleaveFactor();
4480 // Check if the user has overridden the unroll max.
4482 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
4483 MaxInterleaveSize = ForceTargetMaxScalarInterleaveFactor;
4485 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
4486 MaxInterleaveSize = ForceTargetMaxVectorInterleaveFactor;
4489 // If we did not calculate the cost for VF (because the user selected the VF)
4490 // then we calculate the cost of VF here.
4492 LoopCost = expectedCost(VF);
4494 // Clamp the calculated UF to be between the 1 and the max unroll factor
4495 // that the target allows.
4496 if (UF > MaxInterleaveSize)
4497 UF = MaxInterleaveSize;
4501 // Unroll if we vectorized this loop and there is a reduction that could
4502 // benefit from unrolling.
4503 if (VF > 1 && Legal->getReductionVars()->size()) {
4504 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
4508 // Note that if we've already vectorized the loop we will have done the
4509 // runtime check and so unrolling won't require further checks.
4510 bool UnrollingRequiresRuntimePointerCheck =
4511 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
4513 // We want to unroll small loops in order to reduce the loop overhead and
4514 // potentially expose ILP opportunities.
4515 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
4516 if (!UnrollingRequiresRuntimePointerCheck &&
4517 LoopCost < SmallLoopCost) {
4518 // We assume that the cost overhead is 1 and we use the cost model
4519 // to estimate the cost of the loop and unroll until the cost of the
4520 // loop overhead is about 5% of the cost of the loop.
4521 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
4523 // Unroll until store/load ports (estimated by max unroll factor) are
4525 unsigned NumStores = Legal->getNumStores();
4526 unsigned NumLoads = Legal->getNumLoads();
4527 unsigned StoresUF = UF / (NumStores ? NumStores : 1);
4528 unsigned LoadsUF = UF / (NumLoads ? NumLoads : 1);
4530 // If we have a scalar reduction (vector reductions are already dealt with
4531 // by this point), we can increase the critical path length if the loop
4532 // we're unrolling is inside another loop. Limit, by default to 2, so the
4533 // critical path only gets increased by one reduction operation.
4534 if (Legal->getReductionVars()->size() &&
4535 TheLoop->getLoopDepth() > 1) {
4536 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionUF);
4537 SmallUF = std::min(SmallUF, F);
4538 StoresUF = std::min(StoresUF, F);
4539 LoadsUF = std::min(LoadsUF, F);
4542 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
4543 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
4544 return std::max(StoresUF, LoadsUF);
4547 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
4551 DEBUG(dbgs() << "LV: Not Unrolling.\n");
4555 LoopVectorizationCostModel::RegisterUsage
4556 LoopVectorizationCostModel::calculateRegisterUsage() {
4557 // This function calculates the register usage by measuring the highest number
4558 // of values that are alive at a single location. Obviously, this is a very
4559 // rough estimation. We scan the loop in a topological order in order and
4560 // assign a number to each instruction. We use RPO to ensure that defs are
4561 // met before their users. We assume that each instruction that has in-loop
4562 // users starts an interval. We record every time that an in-loop value is
4563 // used, so we have a list of the first and last occurrences of each
4564 // instruction. Next, we transpose this data structure into a multi map that
4565 // holds the list of intervals that *end* at a specific location. This multi
4566 // map allows us to perform a linear search. We scan the instructions linearly
4567 // and record each time that a new interval starts, by placing it in a set.
4568 // If we find this value in the multi-map then we remove it from the set.
4569 // The max register usage is the maximum size of the set.
4570 // We also search for instructions that are defined outside the loop, but are
4571 // used inside the loop. We need this number separately from the max-interval
4572 // usage number because when we unroll, loop-invariant values do not take
4574 LoopBlocksDFS DFS(TheLoop);
4578 R.NumInstructions = 0;
4580 // Each 'key' in the map opens a new interval. The values
4581 // of the map are the index of the 'last seen' usage of the
4582 // instruction that is the key.
4583 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4584 // Maps instruction to its index.
4585 DenseMap<unsigned, Instruction*> IdxToInstr;
4586 // Marks the end of each interval.
4587 IntervalMap EndPoint;
4588 // Saves the list of instruction indices that are used in the loop.
4589 SmallSet<Instruction*, 8> Ends;
4590 // Saves the list of values that are used in the loop but are
4591 // defined outside the loop, such as arguments and constants.
4592 SmallPtrSet<Value*, 8> LoopInvariants;
4595 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4596 be = DFS.endRPO(); bb != be; ++bb) {
4597 R.NumInstructions += (*bb)->size();
4598 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4600 Instruction *I = it;
4601 IdxToInstr[Index++] = I;
4603 // Save the end location of each USE.
4604 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4605 Value *U = I->getOperand(i);
4606 Instruction *Instr = dyn_cast<Instruction>(U);
4608 // Ignore non-instruction values such as arguments, constants, etc.
4609 if (!Instr) continue;
4611 // If this instruction is outside the loop then record it and continue.
4612 if (!TheLoop->contains(Instr)) {
4613 LoopInvariants.insert(Instr);
4617 // Overwrite previous end points.
4618 EndPoint[Instr] = Index;
4624 // Saves the list of intervals that end with the index in 'key'.
4625 typedef SmallVector<Instruction*, 2> InstrList;
4626 DenseMap<unsigned, InstrList> TransposeEnds;
4628 // Transpose the EndPoints to a list of values that end at each index.
4629 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4631 TransposeEnds[it->second].push_back(it->first);
4633 SmallSet<Instruction*, 8> OpenIntervals;
4634 unsigned MaxUsage = 0;
4637 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4638 for (unsigned int i = 0; i < Index; ++i) {
4639 Instruction *I = IdxToInstr[i];
4640 // Ignore instructions that are never used within the loop.
4641 if (!Ends.count(I)) continue;
4643 // Ignore ephemeral values.
4644 if (EphValues.count(I))
4647 // Remove all of the instructions that end at this location.
4648 InstrList &List = TransposeEnds[i];
4649 for (unsigned int j=0, e = List.size(); j < e; ++j)
4650 OpenIntervals.erase(List[j]);
4652 // Count the number of live interals.
4653 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4655 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4656 OpenIntervals.size() << '\n');
4658 // Add the current instruction to the list of open intervals.
4659 OpenIntervals.insert(I);
4662 unsigned Invariant = LoopInvariants.size();
4663 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
4664 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
4665 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
4667 R.LoopInvariantRegs = Invariant;
4668 R.MaxLocalUsers = MaxUsage;
4672 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4676 for (Loop::block_iterator bb = TheLoop->block_begin(),
4677 be = TheLoop->block_end(); bb != be; ++bb) {
4678 unsigned BlockCost = 0;
4679 BasicBlock *BB = *bb;
4681 // For each instruction in the old loop.
4682 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4683 // Skip dbg intrinsics.
4684 if (isa<DbgInfoIntrinsic>(it))
4687 // Ignore ephemeral values.
4688 if (EphValues.count(it))
4691 unsigned C = getInstructionCost(it, VF);
4693 // Check if we should override the cost.
4694 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
4695 C = ForceTargetInstructionCost;
4698 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
4699 VF << " For instruction: " << *it << '\n');
4702 // We assume that if-converted blocks have a 50% chance of being executed.
4703 // When the code is scalar then some of the blocks are avoided due to CF.
4704 // When the code is vectorized we execute all code paths.
4705 if (VF == 1 && Legal->blockNeedsPredication(*bb))
4714 /// \brief Check whether the address computation for a non-consecutive memory
4715 /// access looks like an unlikely candidate for being merged into the indexing
4718 /// We look for a GEP which has one index that is an induction variable and all
4719 /// other indices are loop invariant. If the stride of this access is also
4720 /// within a small bound we decide that this address computation can likely be
4721 /// merged into the addressing mode.
4722 /// In all other cases, we identify the address computation as complex.
4723 static bool isLikelyComplexAddressComputation(Value *Ptr,
4724 LoopVectorizationLegality *Legal,
4725 ScalarEvolution *SE,
4726 const Loop *TheLoop) {
4727 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
4731 // We are looking for a gep with all loop invariant indices except for one
4732 // which should be an induction variable.
4733 unsigned NumOperands = Gep->getNumOperands();
4734 for (unsigned i = 1; i < NumOperands; ++i) {
4735 Value *Opd = Gep->getOperand(i);
4736 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
4737 !Legal->isInductionVariable(Opd))
4741 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
4742 // can likely be merged into the address computation.
4743 unsigned MaxMergeDistance = 64;
4745 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
4749 // Check the step is constant.
4750 const SCEV *Step = AddRec->getStepRecurrence(*SE);
4751 // Calculate the pointer stride and check if it is consecutive.
4752 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4756 const APInt &APStepVal = C->getValue()->getValue();
4758 // Huge step value - give up.
4759 if (APStepVal.getBitWidth() > 64)
4762 int64_t StepVal = APStepVal.getSExtValue();
4764 return StepVal > MaxMergeDistance;
4767 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
4768 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
4774 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
4775 // If we know that this instruction will remain uniform, check the cost of
4776 // the scalar version.
4777 if (Legal->isUniformAfterVectorization(I))
4780 Type *RetTy = I->getType();
4781 Type *VectorTy = ToVectorTy(RetTy, VF);
4783 // TODO: We need to estimate the cost of intrinsic calls.
4784 switch (I->getOpcode()) {
4785 case Instruction::GetElementPtr:
4786 // We mark this instruction as zero-cost because the cost of GEPs in
4787 // vectorized code depends on whether the corresponding memory instruction
4788 // is scalarized or not. Therefore, we handle GEPs with the memory
4789 // instruction cost.
4791 case Instruction::Br: {
4792 return TTI.getCFInstrCost(I->getOpcode());
4794 case Instruction::PHI:
4795 //TODO: IF-converted IFs become selects.
4797 case Instruction::Add:
4798 case Instruction::FAdd:
4799 case Instruction::Sub:
4800 case Instruction::FSub:
4801 case Instruction::Mul:
4802 case Instruction::FMul:
4803 case Instruction::UDiv:
4804 case Instruction::SDiv:
4805 case Instruction::FDiv:
4806 case Instruction::URem:
4807 case Instruction::SRem:
4808 case Instruction::FRem:
4809 case Instruction::Shl:
4810 case Instruction::LShr:
4811 case Instruction::AShr:
4812 case Instruction::And:
4813 case Instruction::Or:
4814 case Instruction::Xor: {
4815 // Since we will replace the stride by 1 the multiplication should go away.
4816 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
4818 // Certain instructions can be cheaper to vectorize if they have a constant
4819 // second vector operand. One example of this are shifts on x86.
4820 TargetTransformInfo::OperandValueKind Op1VK =
4821 TargetTransformInfo::OK_AnyValue;
4822 TargetTransformInfo::OperandValueKind Op2VK =
4823 TargetTransformInfo::OK_AnyValue;
4824 TargetTransformInfo::OperandValueProperties Op1VP =
4825 TargetTransformInfo::OP_None;
4826 TargetTransformInfo::OperandValueProperties Op2VP =
4827 TargetTransformInfo::OP_None;
4828 Value *Op2 = I->getOperand(1);
4830 // Check for a splat of a constant or for a non uniform vector of constants.
4831 if (isa<ConstantInt>(Op2)) {
4832 ConstantInt *CInt = cast<ConstantInt>(Op2);
4833 if (CInt && CInt->getValue().isPowerOf2())
4834 Op2VP = TargetTransformInfo::OP_PowerOf2;
4835 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4836 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
4837 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
4838 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
4840 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
4841 if (CInt && CInt->getValue().isPowerOf2())
4842 Op2VP = TargetTransformInfo::OP_PowerOf2;
4843 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4847 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
4850 case Instruction::Select: {
4851 SelectInst *SI = cast<SelectInst>(I);
4852 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
4853 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
4854 Type *CondTy = SI->getCondition()->getType();
4856 CondTy = VectorType::get(CondTy, VF);
4858 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
4860 case Instruction::ICmp:
4861 case Instruction::FCmp: {
4862 Type *ValTy = I->getOperand(0)->getType();
4863 VectorTy = ToVectorTy(ValTy, VF);
4864 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
4866 case Instruction::Store:
4867 case Instruction::Load: {
4868 StoreInst *SI = dyn_cast<StoreInst>(I);
4869 LoadInst *LI = dyn_cast<LoadInst>(I);
4870 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
4872 VectorTy = ToVectorTy(ValTy, VF);
4874 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
4875 unsigned AS = SI ? SI->getPointerAddressSpace() :
4876 LI->getPointerAddressSpace();
4877 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
4878 // We add the cost of address computation here instead of with the gep
4879 // instruction because only here we know whether the operation is
4882 return TTI.getAddressComputationCost(VectorTy) +
4883 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4885 // Scalarized loads/stores.
4886 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
4887 bool Reverse = ConsecutiveStride < 0;
4888 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
4889 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
4890 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
4891 bool IsComplexComputation =
4892 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
4894 // The cost of extracting from the value vector and pointer vector.
4895 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
4896 for (unsigned i = 0; i < VF; ++i) {
4897 // The cost of extracting the pointer operand.
4898 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
4899 // In case of STORE, the cost of ExtractElement from the vector.
4900 // In case of LOAD, the cost of InsertElement into the returned
4902 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
4903 Instruction::InsertElement,
4907 // The cost of the scalar loads/stores.
4908 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
4909 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
4914 // Wide load/stores.
4915 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
4916 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4919 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
4923 case Instruction::ZExt:
4924 case Instruction::SExt:
4925 case Instruction::FPToUI:
4926 case Instruction::FPToSI:
4927 case Instruction::FPExt:
4928 case Instruction::PtrToInt:
4929 case Instruction::IntToPtr:
4930 case Instruction::SIToFP:
4931 case Instruction::UIToFP:
4932 case Instruction::Trunc:
4933 case Instruction::FPTrunc:
4934 case Instruction::BitCast: {
4935 // We optimize the truncation of induction variable.
4936 // The cost of these is the same as the scalar operation.
4937 if (I->getOpcode() == Instruction::Trunc &&
4938 Legal->isInductionVariable(I->getOperand(0)))
4939 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
4940 I->getOperand(0)->getType());
4942 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
4943 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
4945 case Instruction::Call: {
4946 CallInst *CI = cast<CallInst>(I);
4947 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
4948 assert(ID && "Not an intrinsic call!");
4949 Type *RetTy = ToVectorTy(CI->getType(), VF);
4950 SmallVector<Type*, 4> Tys;
4951 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
4952 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
4953 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
4956 // We are scalarizing the instruction. Return the cost of the scalar
4957 // instruction, plus the cost of insert and extract into vector
4958 // elements, times the vector width.
4961 if (!RetTy->isVoidTy() && VF != 1) {
4962 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
4964 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
4967 // The cost of inserting the results plus extracting each one of the
4969 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
4972 // The cost of executing VF copies of the scalar instruction. This opcode
4973 // is unknown. Assume that it is the same as 'mul'.
4974 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
4980 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
4981 if (Scalar->isVoidTy() || VF == 1)
4983 return VectorType::get(Scalar, VF);
4986 char LoopVectorize::ID = 0;
4987 static const char lv_name[] = "Loop Vectorization";
4988 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
4989 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
4990 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
4991 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
4992 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
4993 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
4994 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4995 INITIALIZE_PASS_DEPENDENCY(LCSSA)
4996 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
4997 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4998 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5001 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5002 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5006 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5007 // Check for a store.
5008 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5009 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5011 // Check for a load.
5012 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5013 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5019 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5020 bool IfPredicateStore) {
5021 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5022 // Holds vector parameters or scalars, in case of uniform vals.
5023 SmallVector<VectorParts, 4> Params;
5025 setDebugLocFromInst(Builder, Instr);
5027 // Find all of the vectorized parameters.
5028 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5029 Value *SrcOp = Instr->getOperand(op);
5031 // If we are accessing the old induction variable, use the new one.
5032 if (SrcOp == OldInduction) {
5033 Params.push_back(getVectorValue(SrcOp));
5037 // Try using previously calculated values.
5038 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5040 // If the src is an instruction that appeared earlier in the basic block
5041 // then it should already be vectorized.
5042 if (SrcInst && OrigLoop->contains(SrcInst)) {
5043 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5044 // The parameter is a vector value from earlier.
5045 Params.push_back(WidenMap.get(SrcInst));
5047 // The parameter is a scalar from outside the loop. Maybe even a constant.
5048 VectorParts Scalars;
5049 Scalars.append(UF, SrcOp);
5050 Params.push_back(Scalars);
5054 assert(Params.size() == Instr->getNumOperands() &&
5055 "Invalid number of operands");
5057 // Does this instruction return a value ?
5058 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5060 Value *UndefVec = IsVoidRetTy ? nullptr :
5061 UndefValue::get(Instr->getType());
5062 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5063 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5065 Instruction *InsertPt = Builder.GetInsertPoint();
5066 BasicBlock *IfBlock = Builder.GetInsertBlock();
5067 BasicBlock *CondBlock = nullptr;
5070 Loop *VectorLp = nullptr;
5071 if (IfPredicateStore) {
5072 assert(Instr->getParent()->getSinglePredecessor() &&
5073 "Only support single predecessor blocks");
5074 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5075 Instr->getParent());
5076 VectorLp = LI->getLoopFor(IfBlock);
5077 assert(VectorLp && "Must have a loop for this block");
5080 // For each vector unroll 'part':
5081 for (unsigned Part = 0; Part < UF; ++Part) {
5082 // For each scalar that we create:
5084 // Start an "if (pred) a[i] = ..." block.
5085 Value *Cmp = nullptr;
5086 if (IfPredicateStore) {
5087 if (Cond[Part]->getType()->isVectorTy())
5089 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5090 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5091 ConstantInt::get(Cond[Part]->getType(), 1));
5092 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5093 LoopVectorBody.push_back(CondBlock);
5094 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
5095 // Update Builder with newly created basic block.
5096 Builder.SetInsertPoint(InsertPt);
5099 Instruction *Cloned = Instr->clone();
5101 Cloned->setName(Instr->getName() + ".cloned");
5102 // Replace the operands of the cloned instructions with extracted scalars.
5103 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5104 Value *Op = Params[op][Part];
5105 Cloned->setOperand(op, Op);
5108 // Place the cloned scalar in the new loop.
5109 Builder.Insert(Cloned);
5111 // If the original scalar returns a value we need to place it in a vector
5112 // so that future users will be able to use it.
5114 VecResults[Part] = Cloned;
5117 if (IfPredicateStore) {
5118 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
5119 LoopVectorBody.push_back(NewIfBlock);
5120 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
5121 Builder.SetInsertPoint(InsertPt);
5122 Instruction *OldBr = IfBlock->getTerminator();
5123 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
5124 OldBr->eraseFromParent();
5125 IfBlock = NewIfBlock;
5130 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5131 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5132 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5134 return scalarizeInstruction(Instr, IfPredicateStore);
5137 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5141 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5145 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) {
5146 // When unrolling and the VF is 1, we only need to add a simple scalar.
5147 Type *ITy = Val->getType();
5148 assert(!ITy->isVectorTy() && "Val must be a scalar");
5149 Constant *C = ConstantInt::get(ITy, StartIdx);
5150 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");