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 code that checks at runtime if the accessed arrays overlap.
276 /// Returns a pair of instructions where the first element is the first
277 /// instruction generated in possibly a sequence of instructions and the
278 /// second value is the final comparator value or NULL if no check is needed.
279 std::pair<Instruction *, Instruction *> addRuntimeCheck(Instruction *Loc);
281 /// \brief Add checks for strides that where assumed to be 1.
283 /// Returns the last check instruction and the first check instruction in the
284 /// pair as (first, last).
285 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
287 /// Create an empty loop, based on the loop ranges of the old loop.
288 void createEmptyLoop();
289 /// Copy and widen the instructions from the old loop.
290 virtual void vectorizeLoop();
292 /// \brief The Loop exit block may have single value PHI nodes where the
293 /// incoming value is 'Undef'. While vectorizing we only handled real values
294 /// that were defined inside the loop. Here we fix the 'undef case'.
298 /// A helper function that computes the predicate of the block BB, assuming
299 /// that the header block of the loop is set to True. It returns the *entry*
300 /// mask for the block BB.
301 VectorParts createBlockInMask(BasicBlock *BB);
302 /// A helper function that computes the predicate of the edge between SRC
304 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
306 /// A helper function to vectorize a single BB within the innermost loop.
307 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
309 /// Vectorize a single PHINode in a block. This method handles the induction
310 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
311 /// arbitrary length vectors.
312 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
313 unsigned UF, unsigned VF, PhiVector *PV);
315 /// Insert the new loop to the loop hierarchy and pass manager
316 /// and update the analysis passes.
317 void updateAnalysis();
319 /// This instruction is un-vectorizable. Implement it as a sequence
320 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
321 /// scalarized instruction behind an if block predicated on the control
322 /// dependence of the instruction.
323 virtual void scalarizeInstruction(Instruction *Instr,
324 bool IfPredicateStore=false);
326 /// Vectorize Load and Store instructions,
327 virtual void vectorizeMemoryInstruction(Instruction *Instr);
329 /// Create a broadcast instruction. This method generates a broadcast
330 /// instruction (shuffle) for loop invariant values and for the induction
331 /// value. If this is the induction variable then we extend it to N, N+1, ...
332 /// this is needed because each iteration in the loop corresponds to a SIMD
334 virtual Value *getBroadcastInstrs(Value *V);
336 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
337 /// to each vector element of Val. The sequence starts at StartIndex.
338 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step);
340 /// When we go over instructions in the basic block we rely on previous
341 /// values within the current basic block or on loop invariant values.
342 /// When we widen (vectorize) values we place them in the map. If the values
343 /// are not within the map, they have to be loop invariant, so we simply
344 /// broadcast them into a vector.
345 VectorParts &getVectorValue(Value *V);
347 /// Generate a shuffle sequence that will reverse the vector Vec.
348 virtual Value *reverseVector(Value *Vec);
350 /// This is a helper class that holds the vectorizer state. It maps scalar
351 /// instructions to vector instructions. When the code is 'unrolled' then
352 /// then a single scalar value is mapped to multiple vector parts. The parts
353 /// are stored in the VectorPart type.
355 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
357 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
359 /// \return True if 'Key' is saved in the Value Map.
360 bool has(Value *Key) const { return MapStorage.count(Key); }
362 /// Initializes a new entry in the map. Sets all of the vector parts to the
363 /// save value in 'Val'.
364 /// \return A reference to a vector with splat values.
365 VectorParts &splat(Value *Key, Value *Val) {
366 VectorParts &Entry = MapStorage[Key];
367 Entry.assign(UF, Val);
371 ///\return A reference to the value that is stored at 'Key'.
372 VectorParts &get(Value *Key) {
373 VectorParts &Entry = MapStorage[Key];
376 assert(Entry.size() == UF);
381 /// The unroll factor. Each entry in the map stores this number of vector
385 /// Map storage. We use std::map and not DenseMap because insertions to a
386 /// dense map invalidates its iterators.
387 std::map<Value *, VectorParts> MapStorage;
390 /// The original loop.
392 /// Scev analysis to use.
401 const DataLayout *DL;
402 /// Target Library Info.
403 const TargetLibraryInfo *TLI;
405 /// The vectorization SIMD factor to use. Each vector will have this many
410 /// The vectorization unroll factor to use. Each scalar is vectorized to this
411 /// many different vector instructions.
414 /// The builder that we use
417 // --- Vectorization state ---
419 /// The vector-loop preheader.
420 BasicBlock *LoopVectorPreHeader;
421 /// The scalar-loop preheader.
422 BasicBlock *LoopScalarPreHeader;
423 /// Middle Block between the vector and the scalar.
424 BasicBlock *LoopMiddleBlock;
425 ///The ExitBlock of the scalar loop.
426 BasicBlock *LoopExitBlock;
427 ///The vector loop body.
428 SmallVector<BasicBlock *, 4> LoopVectorBody;
429 ///The scalar loop body.
430 BasicBlock *LoopScalarBody;
431 /// A list of all bypass blocks. The first block is the entry of the loop.
432 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
434 /// The new Induction variable which was added to the new block.
436 /// The induction variable of the old basic block.
437 PHINode *OldInduction;
438 /// Holds the extended (to the widest induction type) start index.
440 /// Maps scalars to widened vectors.
442 EdgeMaskCache MaskCache;
444 LoopVectorizationLegality *Legal;
447 class InnerLoopUnroller : public InnerLoopVectorizer {
449 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
450 DominatorTree *DT, const DataLayout *DL,
451 const TargetLibraryInfo *TLI, unsigned UnrollFactor) :
452 InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { }
455 void scalarizeInstruction(Instruction *Instr,
456 bool IfPredicateStore = false) override;
457 void vectorizeMemoryInstruction(Instruction *Instr) override;
458 Value *getBroadcastInstrs(Value *V) override;
459 Value *getStepVector(Value *Val, int StartIdx, Value *Step) override;
460 Value *reverseVector(Value *Vec) override;
463 /// \brief Look for a meaningful debug location on the instruction or it's
465 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
470 if (I->getDebugLoc() != Empty)
473 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
474 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
475 if (OpInst->getDebugLoc() != Empty)
482 /// \brief Set the debug location in the builder using the debug location in the
484 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
485 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
486 B.SetCurrentDebugLocation(Inst->getDebugLoc());
488 B.SetCurrentDebugLocation(DebugLoc());
492 /// \return string containing a file name and a line # for the given loop.
493 static std::string getDebugLocString(const Loop *L) {
496 raw_string_ostream OS(Result);
497 const DebugLoc LoopDbgLoc = L->getStartLoc();
498 if (!LoopDbgLoc.isUnknown())
499 LoopDbgLoc.print(L->getHeader()->getContext(), OS);
501 // Just print the module name.
502 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
509 /// \brief Propagate known metadata from one instruction to another.
510 static void propagateMetadata(Instruction *To, const Instruction *From) {
511 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
512 From->getAllMetadataOtherThanDebugLoc(Metadata);
514 for (auto M : Metadata) {
515 unsigned Kind = M.first;
517 // These are safe to transfer (this is safe for TBAA, even when we
518 // if-convert, because should that metadata have had a control dependency
519 // on the condition, and thus actually aliased with some other
520 // non-speculated memory access when the condition was false, this would be
521 // caught by the runtime overlap checks).
522 if (Kind != LLVMContext::MD_tbaa &&
523 Kind != LLVMContext::MD_alias_scope &&
524 Kind != LLVMContext::MD_noalias &&
525 Kind != LLVMContext::MD_fpmath)
528 To->setMetadata(Kind, M.second);
532 /// \brief Propagate known metadata from one instruction to a vector of others.
533 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
535 if (Instruction *I = dyn_cast<Instruction>(V))
536 propagateMetadata(I, From);
539 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
540 /// to what vectorization factor.
541 /// This class does not look at the profitability of vectorization, only the
542 /// legality. This class has two main kinds of checks:
543 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
544 /// will change the order of memory accesses in a way that will change the
545 /// correctness of the program.
546 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
547 /// checks for a number of different conditions, such as the availability of a
548 /// single induction variable, that all types are supported and vectorize-able,
549 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
550 /// This class is also used by InnerLoopVectorizer for identifying
551 /// induction variable and the different reduction variables.
552 class LoopVectorizationLegality {
554 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, const DataLayout *DL,
555 DominatorTree *DT, TargetLibraryInfo *TLI,
556 AliasAnalysis *AA, Function *F,
557 const TargetTransformInfo *TTI)
558 : NumLoads(0), NumStores(0), NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
559 TLI(TLI), TheFunction(F), TTI(TTI), Induction(nullptr),
560 WidestIndTy(nullptr),
561 LAA(F, L, SE, DL, TLI, AA, DT,
562 LoopAccessAnalysis::VectorizerParams(
563 MaxVectorWidth, VectorizationFactor, VectorizationInterleave,
564 RuntimeMemoryCheckThreshold)),
565 HasFunNoNaNAttr(false) {}
567 /// This enum represents the kinds of reductions that we support.
569 RK_NoReduction, ///< Not a reduction.
570 RK_IntegerAdd, ///< Sum of integers.
571 RK_IntegerMult, ///< Product of integers.
572 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
573 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
574 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
575 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
576 RK_FloatAdd, ///< Sum of floats.
577 RK_FloatMult, ///< Product of floats.
578 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
581 /// This enum represents the kinds of inductions that we support.
583 IK_NoInduction, ///< Not an induction variable.
584 IK_IntInduction, ///< Integer induction variable. Step = C.
585 IK_PtrInduction ///< Pointer induction var. Step = C / sizeof(elem).
588 // This enum represents the kind of minmax reduction.
589 enum MinMaxReductionKind {
599 /// This struct holds information about reduction variables.
600 struct ReductionDescriptor {
601 ReductionDescriptor() : StartValue(nullptr), LoopExitInstr(nullptr),
602 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
604 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
605 MinMaxReductionKind MK)
606 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
608 // The starting value of the reduction.
609 // It does not have to be zero!
610 TrackingVH<Value> StartValue;
611 // The instruction who's value is used outside the loop.
612 Instruction *LoopExitInstr;
613 // The kind of the reduction.
615 // If this a min/max reduction the kind of reduction.
616 MinMaxReductionKind MinMaxKind;
619 /// This POD struct holds information about a potential reduction operation.
620 struct ReductionInstDesc {
621 ReductionInstDesc(bool IsRedux, Instruction *I) :
622 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
624 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
625 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
627 // Is this instruction a reduction candidate.
629 // The last instruction in a min/max pattern (select of the select(icmp())
630 // pattern), or the current reduction instruction otherwise.
631 Instruction *PatternLastInst;
632 // If this is a min/max pattern the comparison predicate.
633 MinMaxReductionKind MinMaxKind;
636 /// A struct for saving information about induction variables.
637 struct InductionInfo {
638 InductionInfo(Value *Start, InductionKind K, ConstantInt *Step)
639 : StartValue(Start), IK(K), StepValue(Step) {
640 assert(IK != IK_NoInduction && "Not an induction");
641 assert(StartValue && "StartValue is null");
642 assert(StepValue && !StepValue->isZero() && "StepValue is zero");
643 assert((IK != IK_PtrInduction || StartValue->getType()->isPointerTy()) &&
644 "StartValue is not a pointer for pointer induction");
645 assert((IK != IK_IntInduction || StartValue->getType()->isIntegerTy()) &&
646 "StartValue is not an integer for integer induction");
647 assert(StepValue->getType()->isIntegerTy() &&
648 "StepValue is not an integer");
651 : StartValue(nullptr), IK(IK_NoInduction), StepValue(nullptr) {}
653 /// Get the consecutive direction. Returns:
654 /// 0 - unknown or non-consecutive.
655 /// 1 - consecutive and increasing.
656 /// -1 - consecutive and decreasing.
657 int getConsecutiveDirection() const {
658 if (StepValue && (StepValue->isOne() || StepValue->isMinusOne()))
659 return StepValue->getSExtValue();
663 /// Compute the transformed value of Index at offset StartValue using step
665 /// For integer induction, returns StartValue + Index * StepValue.
666 /// For pointer induction, returns StartValue[Index * StepValue].
667 /// FIXME: The newly created binary instructions should contain nsw/nuw
668 /// flags, which can be found from the original scalar operations.
669 Value *transform(IRBuilder<> &B, Value *Index) const {
671 case IK_IntInduction:
672 assert(Index->getType() == StartValue->getType() &&
673 "Index type does not match StartValue type");
674 if (StepValue->isMinusOne())
675 return B.CreateSub(StartValue, Index);
676 if (!StepValue->isOne())
677 Index = B.CreateMul(Index, StepValue);
678 return B.CreateAdd(StartValue, Index);
680 case IK_PtrInduction:
681 if (StepValue->isMinusOne())
682 Index = B.CreateNeg(Index);
683 else if (!StepValue->isOne())
684 Index = B.CreateMul(Index, StepValue);
685 return B.CreateGEP(StartValue, Index);
690 llvm_unreachable("invalid enum");
694 TrackingVH<Value> StartValue;
698 ConstantInt *StepValue;
701 /// ReductionList contains the reduction descriptors for all
702 /// of the reductions that were found in the loop.
703 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
705 /// InductionList saves induction variables and maps them to the
706 /// induction descriptor.
707 typedef MapVector<PHINode*, InductionInfo> InductionList;
709 /// Returns true if it is legal to vectorize this loop.
710 /// This does not mean that it is profitable to vectorize this
711 /// loop, only that it is legal to do so.
714 /// Returns the Induction variable.
715 PHINode *getInduction() { return Induction; }
717 /// Returns the reduction variables found in the loop.
718 ReductionList *getReductionVars() { return &Reductions; }
720 /// Returns the induction variables found in the loop.
721 InductionList *getInductionVars() { return &Inductions; }
723 /// Returns the widest induction type.
724 Type *getWidestInductionType() { return WidestIndTy; }
726 /// Returns True if V is an induction variable in this loop.
727 bool isInductionVariable(const Value *V);
729 /// Return true if the block BB needs to be predicated in order for the loop
730 /// to be vectorized.
731 bool blockNeedsPredication(BasicBlock *BB);
733 /// Check if this pointer is consecutive when vectorizing. This happens
734 /// when the last index of the GEP is the induction variable, or that the
735 /// pointer itself is an induction variable.
736 /// This check allows us to vectorize A[idx] into a wide load/store.
738 /// 0 - Stride is unknown or non-consecutive.
739 /// 1 - Address is consecutive.
740 /// -1 - Address is consecutive, and decreasing.
741 int isConsecutivePtr(Value *Ptr);
743 /// Returns true if the value V is uniform within the loop.
744 bool isUniform(Value *V);
746 /// Returns true if this instruction will remain scalar after vectorization.
747 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
749 /// Returns the information that we collected about runtime memory check.
750 LoopAccessAnalysis::RuntimePointerCheck *getRuntimePointerCheck() {
751 return LAA.getRuntimePointerCheck();
754 /// This function returns the identity element (or neutral element) for
756 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
758 unsigned getMaxSafeDepDistBytes() { return LAA.getMaxSafeDepDistBytes(); }
760 bool hasStride(Value *V) { return StrideSet.count(V); }
761 bool mustCheckStrides() { return !StrideSet.empty(); }
762 SmallPtrSet<Value *, 8>::iterator strides_begin() {
763 return StrideSet.begin();
765 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
767 /// Returns true if the target machine supports masked store operation
768 /// for the given \p DataType and kind of access to \p Ptr.
769 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
770 return TTI->isLegalMaskedStore(DataType, isConsecutivePtr(Ptr));
772 /// Returns true if the target machine supports masked load operation
773 /// for the given \p DataType and kind of access to \p Ptr.
774 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
775 return TTI->isLegalMaskedLoad(DataType, isConsecutivePtr(Ptr));
777 /// Returns true if vector representation of the instruction \p I
779 bool isMaskRequired(const Instruction* I) {
780 return (MaskedOp.count(I) != 0);
782 unsigned getNumStores() const {
785 unsigned getNumLoads() const {
788 unsigned getNumPredStores() const {
789 return NumPredStores;
792 /// Check if a single basic block loop is vectorizable.
793 /// At this point we know that this is a loop with a constant trip count
794 /// and we only need to check individual instructions.
795 bool canVectorizeInstrs();
797 /// When we vectorize loops we may change the order in which
798 /// we read and write from memory. This method checks if it is
799 /// legal to vectorize the code, considering only memory constrains.
800 /// Returns true if the loop is vectorizable
801 bool canVectorizeMemory();
803 /// Return true if we can vectorize this loop using the IF-conversion
805 bool canVectorizeWithIfConvert();
807 /// Collect the variables that need to stay uniform after vectorization.
808 void collectLoopUniforms();
810 /// Return true if all of the instructions in the block can be speculatively
811 /// executed. \p SafePtrs is a list of addresses that are known to be legal
812 /// and we know that we can read from them without segfault.
813 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
815 /// Returns True, if 'Phi' is the kind of reduction variable for type
816 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
817 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
818 /// Returns a struct describing if the instruction 'I' can be a reduction
819 /// variable of type 'Kind'. If the reduction is a min/max pattern of
820 /// select(icmp()) this function advances the instruction pointer 'I' from the
821 /// compare instruction to the select instruction and stores this pointer in
822 /// 'PatternLastInst' member of the returned struct.
823 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
824 ReductionInstDesc &Desc);
825 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
826 /// pattern corresponding to a min(X, Y) or max(X, Y).
827 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
828 ReductionInstDesc &Prev);
829 /// Returns the induction kind of Phi and record the step. This function may
830 /// return NoInduction if the PHI is not an induction variable.
831 InductionKind isInductionVariable(PHINode *Phi, ConstantInt *&StepValue);
833 /// \brief Collect memory access with loop invariant strides.
835 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
837 void collectStridedAccess(Value *LoadOrStoreInst);
839 /// Report an analysis message to assist the user in diagnosing loops that are
841 void emitAnalysis(VectorizationReport &Message) {
842 VectorizationReport::emitAnalysis(Message, TheFunction, TheLoop);
847 unsigned NumPredStores;
849 /// The loop that we evaluate.
853 /// DataLayout analysis.
854 const DataLayout *DL;
855 /// Target Library Info.
856 TargetLibraryInfo *TLI;
858 Function *TheFunction;
859 /// Target Transform Info
860 const TargetTransformInfo *TTI;
862 // --- vectorization state --- //
864 /// Holds the integer induction variable. This is the counter of the
867 /// Holds the reduction variables.
868 ReductionList Reductions;
869 /// Holds all of the induction variables that we found in the loop.
870 /// Notice that inductions don't need to start at zero and that induction
871 /// variables can be pointers.
872 InductionList Inductions;
873 /// Holds the widest induction type encountered.
876 /// Allowed outside users. This holds the reduction
877 /// vars which can be accessed from outside the loop.
878 SmallPtrSet<Value*, 4> AllowedExit;
879 /// This set holds the variables which are known to be uniform after
881 SmallPtrSet<Instruction*, 4> Uniforms;
882 LoopAccessAnalysis LAA;
883 /// Can we assume the absence of NaNs.
884 bool HasFunNoNaNAttr;
886 ValueToValueMap Strides;
887 SmallPtrSet<Value *, 8> StrideSet;
889 /// While vectorizing these instructions we have to generate a
890 /// call to the appropriate masked intrinsic
891 SmallPtrSet<const Instruction*, 8> MaskedOp;
894 /// LoopVectorizationCostModel - estimates the expected speedups due to
896 /// In many cases vectorization is not profitable. This can happen because of
897 /// a number of reasons. In this class we mainly attempt to predict the
898 /// expected speedup/slowdowns due to the supported instruction set. We use the
899 /// TargetTransformInfo to query the different backends for the cost of
900 /// different operations.
901 class LoopVectorizationCostModel {
903 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
904 LoopVectorizationLegality *Legal,
905 const TargetTransformInfo &TTI,
906 const DataLayout *DL, const TargetLibraryInfo *TLI,
907 AssumptionCache *AC, const Function *F,
908 const LoopVectorizeHints *Hints)
909 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI),
910 TheFunction(F), Hints(Hints) {
911 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
914 /// Information about vectorization costs
915 struct VectorizationFactor {
916 unsigned Width; // Vector width with best cost
917 unsigned Cost; // Cost of the loop with that width
919 /// \return The most profitable vectorization factor and the cost of that VF.
920 /// This method checks every power of two up to VF. If UserVF is not ZERO
921 /// then this vectorization factor will be selected if vectorization is
923 VectorizationFactor selectVectorizationFactor(bool OptForSize);
925 /// \return The size (in bits) of the widest type in the code that
926 /// needs to be vectorized. We ignore values that remain scalar such as
927 /// 64 bit loop indices.
928 unsigned getWidestType();
930 /// \return The most profitable unroll factor.
931 /// If UserUF is non-zero then this method finds the best unroll-factor
932 /// based on register pressure and other parameters.
933 /// VF and LoopCost are the selected vectorization factor and the cost of the
935 unsigned selectUnrollFactor(bool OptForSize, unsigned VF, unsigned LoopCost);
937 /// \brief A struct that represents some properties of the register usage
939 struct RegisterUsage {
940 /// Holds the number of loop invariant values that are used in the loop.
941 unsigned LoopInvariantRegs;
942 /// Holds the maximum number of concurrent live intervals in the loop.
943 unsigned MaxLocalUsers;
944 /// Holds the number of instructions in the loop.
945 unsigned NumInstructions;
948 /// \return information about the register usage of the loop.
949 RegisterUsage calculateRegisterUsage();
952 /// Returns the expected execution cost. The unit of the cost does
953 /// not matter because we use the 'cost' units to compare different
954 /// vector widths. The cost that is returned is *not* normalized by
955 /// the factor width.
956 unsigned expectedCost(unsigned VF);
958 /// Returns the execution time cost of an instruction for a given vector
959 /// width. Vector width of one means scalar.
960 unsigned getInstructionCost(Instruction *I, unsigned VF);
962 /// A helper function for converting Scalar types to vector types.
963 /// If the incoming type is void, we return void. If the VF is 1, we return
965 static Type* ToVectorTy(Type *Scalar, unsigned VF);
967 /// Returns whether the instruction is a load or store and will be a emitted
968 /// as a vector operation.
969 bool isConsecutiveLoadOrStore(Instruction *I);
971 /// Report an analysis message to assist the user in diagnosing loops that are
973 void emitAnalysis(VectorizationReport &Message) {
974 VectorizationReport::emitAnalysis(Message, TheFunction, TheLoop);
977 /// Values used only by @llvm.assume calls.
978 SmallPtrSet<const Value *, 32> EphValues;
980 /// The loop that we evaluate.
984 /// Loop Info analysis.
986 /// Vectorization legality.
987 LoopVectorizationLegality *Legal;
988 /// Vector target information.
989 const TargetTransformInfo &TTI;
990 /// Target data layout information.
991 const DataLayout *DL;
992 /// Target Library Info.
993 const TargetLibraryInfo *TLI;
994 const Function *TheFunction;
995 // Loop Vectorize Hint.
996 const LoopVectorizeHints *Hints;
999 /// Utility class for getting and setting loop vectorizer hints in the form
1000 /// of loop metadata.
1001 /// This class keeps a number of loop annotations locally (as member variables)
1002 /// and can, upon request, write them back as metadata on the loop. It will
1003 /// initially scan the loop for existing metadata, and will update the local
1004 /// values based on information in the loop.
1005 /// We cannot write all values to metadata, as the mere presence of some info,
1006 /// for example 'force', means a decision has been made. So, we need to be
1007 /// careful NOT to add them if the user hasn't specifically asked so.
1008 class LoopVectorizeHints {
1015 /// Hint - associates name and validation with the hint value.
1018 unsigned Value; // This may have to change for non-numeric values.
1021 Hint(const char * Name, unsigned Value, HintKind Kind)
1022 : Name(Name), Value(Value), Kind(Kind) { }
1024 bool validate(unsigned Val) {
1027 return isPowerOf2_32(Val) && Val <= MaxVectorWidth;
1029 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1037 /// Vectorization width.
1039 /// Vectorization interleave factor.
1041 /// Vectorization forced
1044 /// Return the loop metadata prefix.
1045 static StringRef Prefix() { return "llvm.loop."; }
1049 FK_Undefined = -1, ///< Not selected.
1050 FK_Disabled = 0, ///< Forcing disabled.
1051 FK_Enabled = 1, ///< Forcing enabled.
1054 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
1055 : Width("vectorize.width", VectorizationFactor, HK_WIDTH),
1056 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1057 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1059 // Populate values with existing loop metadata.
1060 getHintsFromMetadata();
1062 // force-vector-interleave overrides DisableInterleaving.
1063 if (VectorizationInterleave.getNumOccurrences() > 0)
1064 Interleave.Value = VectorizationInterleave;
1066 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
1067 << "LV: Interleaving disabled by the pass manager\n");
1070 /// Mark the loop L as already vectorized by setting the width to 1.
1071 void setAlreadyVectorized() {
1072 Width.Value = Interleave.Value = 1;
1073 Hint Hints[] = {Width, Interleave};
1074 writeHintsToMetadata(Hints);
1077 /// Dumps all the hint information.
1078 std::string emitRemark() const {
1079 VectorizationReport R;
1080 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1081 R << "vectorization is explicitly disabled";
1083 R << "use -Rpass-analysis=loop-vectorize for more info";
1084 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1085 R << " (Force=true";
1086 if (Width.Value != 0)
1087 R << ", Vector Width=" << Width.Value;
1088 if (Interleave.Value != 0)
1089 R << ", Interleave Count=" << Interleave.Value;
1097 unsigned getWidth() const { return Width.Value; }
1098 unsigned getInterleave() const { return Interleave.Value; }
1099 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1102 /// Find hints specified in the loop metadata and update local values.
1103 void getHintsFromMetadata() {
1104 MDNode *LoopID = TheLoop->getLoopID();
1108 // First operand should refer to the loop id itself.
1109 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1110 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1112 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1113 const MDString *S = nullptr;
1114 SmallVector<Metadata *, 4> Args;
1116 // The expected hint is either a MDString or a MDNode with the first
1117 // operand a MDString.
1118 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1119 if (!MD || MD->getNumOperands() == 0)
1121 S = dyn_cast<MDString>(MD->getOperand(0));
1122 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1123 Args.push_back(MD->getOperand(i));
1125 S = dyn_cast<MDString>(LoopID->getOperand(i));
1126 assert(Args.size() == 0 && "too many arguments for MDString");
1132 // Check if the hint starts with the loop metadata prefix.
1133 StringRef Name = S->getString();
1134 if (Args.size() == 1)
1135 setHint(Name, Args[0]);
1139 /// Checks string hint with one operand and set value if valid.
1140 void setHint(StringRef Name, Metadata *Arg) {
1141 if (!Name.startswith(Prefix()))
1143 Name = Name.substr(Prefix().size(), StringRef::npos);
1145 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1147 unsigned Val = C->getZExtValue();
1149 Hint *Hints[] = {&Width, &Interleave, &Force};
1150 for (auto H : Hints) {
1151 if (Name == H->Name) {
1152 if (H->validate(Val))
1155 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1161 /// Create a new hint from name / value pair.
1162 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1163 LLVMContext &Context = TheLoop->getHeader()->getContext();
1164 Metadata *MDs[] = {MDString::get(Context, Name),
1165 ConstantAsMetadata::get(
1166 ConstantInt::get(Type::getInt32Ty(Context), V))};
1167 return MDNode::get(Context, MDs);
1170 /// Matches metadata with hint name.
1171 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1172 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1176 for (auto H : HintTypes)
1177 if (Name->getString().endswith(H.Name))
1182 /// Sets current hints into loop metadata, keeping other values intact.
1183 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1184 if (HintTypes.size() == 0)
1187 // Reserve the first element to LoopID (see below).
1188 SmallVector<Metadata *, 4> MDs(1);
1189 // If the loop already has metadata, then ignore the existing operands.
1190 MDNode *LoopID = TheLoop->getLoopID();
1192 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1193 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1194 // If node in update list, ignore old value.
1195 if (!matchesHintMetadataName(Node, HintTypes))
1196 MDs.push_back(Node);
1200 // Now, add the missing hints.
1201 for (auto H : HintTypes)
1202 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1204 // Replace current metadata node with new one.
1205 LLVMContext &Context = TheLoop->getHeader()->getContext();
1206 MDNode *NewLoopID = MDNode::get(Context, MDs);
1207 // Set operand 0 to refer to the loop id itself.
1208 NewLoopID->replaceOperandWith(0, NewLoopID);
1210 TheLoop->setLoopID(NewLoopID);
1213 /// The loop these hints belong to.
1214 const Loop *TheLoop;
1217 static void emitMissedWarning(Function *F, Loop *L,
1218 const LoopVectorizeHints &LH) {
1219 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1220 L->getStartLoc(), LH.emitRemark());
1222 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1223 if (LH.getWidth() != 1)
1224 emitLoopVectorizeWarning(
1225 F->getContext(), *F, L->getStartLoc(),
1226 "failed explicitly specified loop vectorization");
1227 else if (LH.getInterleave() != 1)
1228 emitLoopInterleaveWarning(
1229 F->getContext(), *F, L->getStartLoc(),
1230 "failed explicitly specified loop interleaving");
1234 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1236 return V.push_back(&L);
1238 for (Loop *InnerL : L)
1239 addInnerLoop(*InnerL, V);
1242 /// The LoopVectorize Pass.
1243 struct LoopVectorize : public FunctionPass {
1244 /// Pass identification, replacement for typeid
1247 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1249 DisableUnrolling(NoUnrolling),
1250 AlwaysVectorize(AlwaysVectorize) {
1251 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1254 ScalarEvolution *SE;
1255 const DataLayout *DL;
1257 TargetTransformInfo *TTI;
1259 BlockFrequencyInfo *BFI;
1260 TargetLibraryInfo *TLI;
1262 AssumptionCache *AC;
1263 bool DisableUnrolling;
1264 bool AlwaysVectorize;
1266 BlockFrequency ColdEntryFreq;
1268 bool runOnFunction(Function &F) override {
1269 SE = &getAnalysis<ScalarEvolution>();
1270 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1271 DL = DLP ? &DLP->getDataLayout() : nullptr;
1272 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1273 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1274 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1275 BFI = &getAnalysis<BlockFrequencyInfo>();
1276 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1277 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1278 AA = &getAnalysis<AliasAnalysis>();
1279 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1281 // Compute some weights outside of the loop over the loops. Compute this
1282 // using a BranchProbability to re-use its scaling math.
1283 const BranchProbability ColdProb(1, 5); // 20%
1284 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1286 // If the target claims to have no vector registers don't attempt
1288 if (!TTI->getNumberOfRegisters(true))
1292 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1293 << ": Missing data layout\n");
1297 // Build up a worklist of inner-loops to vectorize. This is necessary as
1298 // the act of vectorizing or partially unrolling a loop creates new loops
1299 // and can invalidate iterators across the loops.
1300 SmallVector<Loop *, 8> Worklist;
1303 addInnerLoop(*L, Worklist);
1305 LoopsAnalyzed += Worklist.size();
1307 // Now walk the identified inner loops.
1308 bool Changed = false;
1309 while (!Worklist.empty())
1310 Changed |= processLoop(Worklist.pop_back_val());
1312 // Process each loop nest in the function.
1316 bool processLoop(Loop *L) {
1317 assert(L->empty() && "Only process inner loops.");
1320 const std::string DebugLocStr = getDebugLocString(L);
1323 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1324 << L->getHeader()->getParent()->getName() << "\" from "
1325 << DebugLocStr << "\n");
1327 LoopVectorizeHints Hints(L, DisableUnrolling);
1329 DEBUG(dbgs() << "LV: Loop hints:"
1331 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1333 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1335 : "?")) << " width=" << Hints.getWidth()
1336 << " unroll=" << Hints.getInterleave() << "\n");
1338 // Function containing loop
1339 Function *F = L->getHeader()->getParent();
1341 // Looking at the diagnostic output is the only way to determine if a loop
1342 // was vectorized (other than looking at the IR or machine code), so it
1343 // is important to generate an optimization remark for each loop. Most of
1344 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1345 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1346 // less verbose reporting vectorized loops and unvectorized loops that may
1347 // benefit from vectorization, respectively.
1349 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1350 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1351 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1352 L->getStartLoc(), Hints.emitRemark());
1356 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1357 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1358 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1359 L->getStartLoc(), Hints.emitRemark());
1363 if (Hints.getWidth() == 1 && Hints.getInterleave() == 1) {
1364 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1365 emitOptimizationRemarkAnalysis(
1366 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1367 "loop not vectorized: vector width and interleave count are "
1368 "explicitly set to 1");
1372 // Check the loop for a trip count threshold:
1373 // do not vectorize loops with a tiny trip count.
1374 const unsigned TC = SE->getSmallConstantTripCount(L);
1375 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1376 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1377 << "This loop is not worth vectorizing.");
1378 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1379 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1381 DEBUG(dbgs() << "\n");
1382 emitOptimizationRemarkAnalysis(
1383 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1384 "vectorization is not beneficial and is not explicitly forced");
1389 // Check if it is legal to vectorize the loop.
1390 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, AA, F, TTI);
1391 if (!LVL.canVectorize()) {
1392 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1393 emitMissedWarning(F, L, Hints);
1397 // Use the cost model.
1398 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI, AC, F,
1401 // Check the function attributes to find out if this function should be
1402 // optimized for size.
1403 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1404 F->hasFnAttribute(Attribute::OptimizeForSize);
1406 // Compute the weighted frequency of this loop being executed and see if it
1407 // is less than 20% of the function entry baseline frequency. Note that we
1408 // always have a canonical loop here because we think we *can* vectoriez.
1409 // FIXME: This is hidden behind a flag due to pervasive problems with
1410 // exactly what block frequency models.
1411 if (LoopVectorizeWithBlockFrequency) {
1412 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1413 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1414 LoopEntryFreq < ColdEntryFreq)
1418 // Check the function attributes to see if implicit floats are allowed.a
1419 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1420 // an integer loop and the vector instructions selected are purely integer
1421 // vector instructions?
1422 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1423 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1424 "attribute is used.\n");
1425 emitOptimizationRemarkAnalysis(
1426 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1427 "loop not vectorized due to NoImplicitFloat attribute");
1428 emitMissedWarning(F, L, Hints);
1432 // Select the optimal vectorization factor.
1433 const LoopVectorizationCostModel::VectorizationFactor VF =
1434 CM.selectVectorizationFactor(OptForSize);
1436 // Select the unroll factor.
1438 CM.selectUnrollFactor(OptForSize, VF.Width, VF.Cost);
1440 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1441 << DebugLocStr << '\n');
1442 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1444 if (VF.Width == 1) {
1445 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
1448 emitOptimizationRemarkAnalysis(
1449 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1450 "not beneficial to vectorize and user disabled interleaving");
1453 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1455 // Report the unrolling decision.
1456 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1457 Twine("unrolled with interleaving factor " +
1459 " (vectorization not beneficial)"));
1461 // We decided not to vectorize, but we may want to unroll.
1463 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1464 Unroller.vectorize(&LVL);
1466 // If we decided that it is *legal* to vectorize the loop then do it.
1467 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1471 // Report the vectorization decision.
1472 emitOptimizationRemark(
1473 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1474 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1475 ", unrolling interleave factor: " + Twine(UF) + ")");
1478 // Mark the loop as already vectorized to avoid vectorizing again.
1479 Hints.setAlreadyVectorized();
1481 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1485 void getAnalysisUsage(AnalysisUsage &AU) const override {
1486 AU.addRequired<AssumptionCacheTracker>();
1487 AU.addRequiredID(LoopSimplifyID);
1488 AU.addRequiredID(LCSSAID);
1489 AU.addRequired<BlockFrequencyInfo>();
1490 AU.addRequired<DominatorTreeWrapperPass>();
1491 AU.addRequired<LoopInfoWrapperPass>();
1492 AU.addRequired<ScalarEvolution>();
1493 AU.addRequired<TargetTransformInfoWrapperPass>();
1494 AU.addRequired<AliasAnalysis>();
1495 AU.addPreserved<LoopInfoWrapperPass>();
1496 AU.addPreserved<DominatorTreeWrapperPass>();
1497 AU.addPreserved<AliasAnalysis>();
1502 } // end anonymous namespace
1504 //===----------------------------------------------------------------------===//
1505 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1506 // LoopVectorizationCostModel.
1507 //===----------------------------------------------------------------------===//
1509 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1510 // We need to place the broadcast of invariant variables outside the loop.
1511 Instruction *Instr = dyn_cast<Instruction>(V);
1513 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1514 Instr->getParent()) != LoopVectorBody.end());
1515 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1517 // Place the code for broadcasting invariant variables in the new preheader.
1518 IRBuilder<>::InsertPointGuard Guard(Builder);
1520 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1522 // Broadcast the scalar into all locations in the vector.
1523 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1528 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
1530 assert(Val->getType()->isVectorTy() && "Must be a vector");
1531 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1532 "Elem must be an integer");
1533 assert(Step->getType() == Val->getType()->getScalarType() &&
1534 "Step has wrong type");
1535 // Create the types.
1536 Type *ITy = Val->getType()->getScalarType();
1537 VectorType *Ty = cast<VectorType>(Val->getType());
1538 int VLen = Ty->getNumElements();
1539 SmallVector<Constant*, 8> Indices;
1541 // Create a vector of consecutive numbers from zero to VF.
1542 for (int i = 0; i < VLen; ++i)
1543 Indices.push_back(ConstantInt::get(ITy, StartIdx + i));
1545 // Add the consecutive indices to the vector value.
1546 Constant *Cv = ConstantVector::get(Indices);
1547 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1548 Step = Builder.CreateVectorSplat(VLen, Step);
1549 assert(Step->getType() == Val->getType() && "Invalid step vec");
1550 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
1551 // which can be found from the original scalar operations.
1552 Step = Builder.CreateMul(Cv, Step);
1553 return Builder.CreateAdd(Val, Step, "induction");
1556 /// \brief Find the operand of the GEP that should be checked for consecutive
1557 /// stores. This ignores trailing indices that have no effect on the final
1559 static unsigned getGEPInductionOperand(const DataLayout *DL,
1560 const GetElementPtrInst *Gep) {
1561 unsigned LastOperand = Gep->getNumOperands() - 1;
1562 unsigned GEPAllocSize = DL->getTypeAllocSize(
1563 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1565 // Walk backwards and try to peel off zeros.
1566 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1567 // Find the type we're currently indexing into.
1568 gep_type_iterator GEPTI = gep_type_begin(Gep);
1569 std::advance(GEPTI, LastOperand - 1);
1571 // If it's a type with the same allocation size as the result of the GEP we
1572 // can peel off the zero index.
1573 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1581 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1582 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1583 // Make sure that the pointer does not point to structs.
1584 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1587 // If this value is a pointer induction variable we know it is consecutive.
1588 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1589 if (Phi && Inductions.count(Phi)) {
1590 InductionInfo II = Inductions[Phi];
1591 return II.getConsecutiveDirection();
1594 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1598 unsigned NumOperands = Gep->getNumOperands();
1599 Value *GpPtr = Gep->getPointerOperand();
1600 // If this GEP value is a consecutive pointer induction variable and all of
1601 // the indices are constant then we know it is consecutive. We can
1602 Phi = dyn_cast<PHINode>(GpPtr);
1603 if (Phi && Inductions.count(Phi)) {
1605 // Make sure that the pointer does not point to structs.
1606 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1607 if (GepPtrType->getElementType()->isAggregateType())
1610 // Make sure that all of the index operands are loop invariant.
1611 for (unsigned i = 1; i < NumOperands; ++i)
1612 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1615 InductionInfo II = Inductions[Phi];
1616 return II.getConsecutiveDirection();
1619 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1621 // Check that all of the gep indices are uniform except for our induction
1623 for (unsigned i = 0; i != NumOperands; ++i)
1624 if (i != InductionOperand &&
1625 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1628 // We can emit wide load/stores only if the last non-zero index is the
1629 // induction variable.
1630 const SCEV *Last = nullptr;
1631 if (!Strides.count(Gep))
1632 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1634 // Because of the multiplication by a stride we can have a s/zext cast.
1635 // We are going to replace this stride by 1 so the cast is safe to ignore.
1637 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1638 // %0 = trunc i64 %indvars.iv to i32
1639 // %mul = mul i32 %0, %Stride1
1640 // %idxprom = zext i32 %mul to i64 << Safe cast.
1641 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1643 Last = replaceSymbolicStrideSCEV(SE, Strides,
1644 Gep->getOperand(InductionOperand), Gep);
1645 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1647 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1651 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1652 const SCEV *Step = AR->getStepRecurrence(*SE);
1654 // The memory is consecutive because the last index is consecutive
1655 // and all other indices are loop invariant.
1658 if (Step->isAllOnesValue())
1665 bool LoopVectorizationLegality::isUniform(Value *V) {
1666 return LAA.isUniform(V);
1669 InnerLoopVectorizer::VectorParts&
1670 InnerLoopVectorizer::getVectorValue(Value *V) {
1671 assert(V != Induction && "The new induction variable should not be used.");
1672 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1674 // If we have a stride that is replaced by one, do it here.
1675 if (Legal->hasStride(V))
1676 V = ConstantInt::get(V->getType(), 1);
1678 // If we have this scalar in the map, return it.
1679 if (WidenMap.has(V))
1680 return WidenMap.get(V);
1682 // If this scalar is unknown, assume that it is a constant or that it is
1683 // loop invariant. Broadcast V and save the value for future uses.
1684 Value *B = getBroadcastInstrs(V);
1685 return WidenMap.splat(V, B);
1688 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1689 assert(Vec->getType()->isVectorTy() && "Invalid type");
1690 SmallVector<Constant*, 8> ShuffleMask;
1691 for (unsigned i = 0; i < VF; ++i)
1692 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1694 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1695 ConstantVector::get(ShuffleMask),
1699 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1700 // Attempt to issue a wide load.
1701 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1702 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1704 assert((LI || SI) && "Invalid Load/Store instruction");
1706 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1707 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1708 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1709 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1710 // An alignment of 0 means target abi alignment. We need to use the scalar's
1711 // target abi alignment in such a case.
1713 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1714 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1715 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1716 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1718 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
1719 !Legal->isMaskRequired(SI))
1720 return scalarizeInstruction(Instr, true);
1722 if (ScalarAllocatedSize != VectorElementSize)
1723 return scalarizeInstruction(Instr);
1725 // If the pointer is loop invariant or if it is non-consecutive,
1726 // scalarize the load.
1727 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1728 bool Reverse = ConsecutiveStride < 0;
1729 bool UniformLoad = LI && Legal->isUniform(Ptr);
1730 if (!ConsecutiveStride || UniformLoad)
1731 return scalarizeInstruction(Instr);
1733 Constant *Zero = Builder.getInt32(0);
1734 VectorParts &Entry = WidenMap.get(Instr);
1736 // Handle consecutive loads/stores.
1737 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1738 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1739 setDebugLocFromInst(Builder, Gep);
1740 Value *PtrOperand = Gep->getPointerOperand();
1741 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1742 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1744 // Create the new GEP with the new induction variable.
1745 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1746 Gep2->setOperand(0, FirstBasePtr);
1747 Gep2->setName("gep.indvar.base");
1748 Ptr = Builder.Insert(Gep2);
1750 setDebugLocFromInst(Builder, Gep);
1751 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1752 OrigLoop) && "Base ptr must be invariant");
1754 // The last index does not have to be the induction. It can be
1755 // consecutive and be a function of the index. For example A[I+1];
1756 unsigned NumOperands = Gep->getNumOperands();
1757 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1758 // Create the new GEP with the new induction variable.
1759 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1761 for (unsigned i = 0; i < NumOperands; ++i) {
1762 Value *GepOperand = Gep->getOperand(i);
1763 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1765 // Update last index or loop invariant instruction anchored in loop.
1766 if (i == InductionOperand ||
1767 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1768 assert((i == InductionOperand ||
1769 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1770 "Must be last index or loop invariant");
1772 VectorParts &GEPParts = getVectorValue(GepOperand);
1773 Value *Index = GEPParts[0];
1774 Index = Builder.CreateExtractElement(Index, Zero);
1775 Gep2->setOperand(i, Index);
1776 Gep2->setName("gep.indvar.idx");
1779 Ptr = Builder.Insert(Gep2);
1781 // Use the induction element ptr.
1782 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1783 setDebugLocFromInst(Builder, Ptr);
1784 VectorParts &PtrVal = getVectorValue(Ptr);
1785 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1788 VectorParts Mask = createBlockInMask(Instr->getParent());
1791 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1792 "We do not allow storing to uniform addresses");
1793 setDebugLocFromInst(Builder, SI);
1794 // We don't want to update the value in the map as it might be used in
1795 // another expression. So don't use a reference type for "StoredVal".
1796 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1798 for (unsigned Part = 0; Part < UF; ++Part) {
1799 // Calculate the pointer for the specific unroll-part.
1800 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1803 // If we store to reverse consecutive memory locations then we need
1804 // to reverse the order of elements in the stored value.
1805 StoredVal[Part] = reverseVector(StoredVal[Part]);
1806 // If the address is consecutive but reversed, then the
1807 // wide store needs to start at the last vector element.
1808 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1809 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1810 Mask[Part] = reverseVector(Mask[Part]);
1813 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1814 DataTy->getPointerTo(AddressSpace));
1817 if (Legal->isMaskRequired(SI))
1818 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
1821 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
1822 propagateMetadata(NewSI, SI);
1828 assert(LI && "Must have a load instruction");
1829 setDebugLocFromInst(Builder, LI);
1830 for (unsigned Part = 0; Part < UF; ++Part) {
1831 // Calculate the pointer for the specific unroll-part.
1832 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1835 // If the address is consecutive but reversed, then the
1836 // wide load needs to start at the last vector element.
1837 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1838 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1839 Mask[Part] = reverseVector(Mask[Part]);
1843 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1844 DataTy->getPointerTo(AddressSpace));
1845 if (Legal->isMaskRequired(LI))
1846 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
1847 UndefValue::get(DataTy),
1848 "wide.masked.load");
1850 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
1851 propagateMetadata(NewLI, LI);
1852 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
1856 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1857 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1858 // Holds vector parameters or scalars, in case of uniform vals.
1859 SmallVector<VectorParts, 4> Params;
1861 setDebugLocFromInst(Builder, Instr);
1863 // Find all of the vectorized parameters.
1864 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1865 Value *SrcOp = Instr->getOperand(op);
1867 // If we are accessing the old induction variable, use the new one.
1868 if (SrcOp == OldInduction) {
1869 Params.push_back(getVectorValue(SrcOp));
1873 // Try using previously calculated values.
1874 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1876 // If the src is an instruction that appeared earlier in the basic block
1877 // then it should already be vectorized.
1878 if (SrcInst && OrigLoop->contains(SrcInst)) {
1879 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1880 // The parameter is a vector value from earlier.
1881 Params.push_back(WidenMap.get(SrcInst));
1883 // The parameter is a scalar from outside the loop. Maybe even a constant.
1884 VectorParts Scalars;
1885 Scalars.append(UF, SrcOp);
1886 Params.push_back(Scalars);
1890 assert(Params.size() == Instr->getNumOperands() &&
1891 "Invalid number of operands");
1893 // Does this instruction return a value ?
1894 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1896 Value *UndefVec = IsVoidRetTy ? nullptr :
1897 UndefValue::get(VectorType::get(Instr->getType(), VF));
1898 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1899 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1901 Instruction *InsertPt = Builder.GetInsertPoint();
1902 BasicBlock *IfBlock = Builder.GetInsertBlock();
1903 BasicBlock *CondBlock = nullptr;
1906 Loop *VectorLp = nullptr;
1907 if (IfPredicateStore) {
1908 assert(Instr->getParent()->getSinglePredecessor() &&
1909 "Only support single predecessor blocks");
1910 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1911 Instr->getParent());
1912 VectorLp = LI->getLoopFor(IfBlock);
1913 assert(VectorLp && "Must have a loop for this block");
1916 // For each vector unroll 'part':
1917 for (unsigned Part = 0; Part < UF; ++Part) {
1918 // For each scalar that we create:
1919 for (unsigned Width = 0; Width < VF; ++Width) {
1922 Value *Cmp = nullptr;
1923 if (IfPredicateStore) {
1924 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1925 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1926 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1927 LoopVectorBody.push_back(CondBlock);
1928 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
1929 // Update Builder with newly created basic block.
1930 Builder.SetInsertPoint(InsertPt);
1933 Instruction *Cloned = Instr->clone();
1935 Cloned->setName(Instr->getName() + ".cloned");
1936 // Replace the operands of the cloned instructions with extracted scalars.
1937 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1938 Value *Op = Params[op][Part];
1939 // Param is a vector. Need to extract the right lane.
1940 if (Op->getType()->isVectorTy())
1941 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1942 Cloned->setOperand(op, Op);
1945 // Place the cloned scalar in the new loop.
1946 Builder.Insert(Cloned);
1948 // If the original scalar returns a value we need to place it in a vector
1949 // so that future users will be able to use it.
1951 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1952 Builder.getInt32(Width));
1954 if (IfPredicateStore) {
1955 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1956 LoopVectorBody.push_back(NewIfBlock);
1957 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
1958 Builder.SetInsertPoint(InsertPt);
1959 Instruction *OldBr = IfBlock->getTerminator();
1960 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1961 OldBr->eraseFromParent();
1962 IfBlock = NewIfBlock;
1968 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1972 if (Instruction *I = dyn_cast<Instruction>(V))
1973 return I->getParent() == Loc->getParent() ? I : nullptr;
1977 std::pair<Instruction *, Instruction *>
1978 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
1979 Instruction *tnullptr = nullptr;
1980 if (!Legal->mustCheckStrides())
1981 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1983 IRBuilder<> ChkBuilder(Loc);
1986 Value *Check = nullptr;
1987 Instruction *FirstInst = nullptr;
1988 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
1989 SE = Legal->strides_end();
1991 Value *Ptr = stripIntegerCast(*SI);
1992 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
1994 // Store the first instruction we create.
1995 FirstInst = getFirstInst(FirstInst, C, Loc);
1997 Check = ChkBuilder.CreateOr(Check, C);
2002 // We have to do this trickery because the IRBuilder might fold the check to a
2003 // constant expression in which case there is no Instruction anchored in a
2005 LLVMContext &Ctx = Loc->getContext();
2006 Instruction *TheCheck =
2007 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2008 ChkBuilder.Insert(TheCheck, "stride.not.one");
2009 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2011 return std::make_pair(FirstInst, TheCheck);
2014 std::pair<Instruction *, Instruction *>
2015 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
2016 LoopAccessAnalysis::RuntimePointerCheck *PtrRtCheck =
2017 Legal->getRuntimePointerCheck();
2019 Instruction *tnullptr = nullptr;
2020 if (!PtrRtCheck->Need)
2021 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2023 unsigned NumPointers = PtrRtCheck->Pointers.size();
2024 SmallVector<TrackingVH<Value> , 2> Starts;
2025 SmallVector<TrackingVH<Value> , 2> Ends;
2027 LLVMContext &Ctx = Loc->getContext();
2028 SCEVExpander Exp(*SE, "induction");
2029 Instruction *FirstInst = nullptr;
2031 for (unsigned i = 0; i < NumPointers; ++i) {
2032 Value *Ptr = PtrRtCheck->Pointers[i];
2033 const SCEV *Sc = SE->getSCEV(Ptr);
2035 if (SE->isLoopInvariant(Sc, OrigLoop)) {
2036 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
2038 Starts.push_back(Ptr);
2039 Ends.push_back(Ptr);
2041 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
2042 unsigned AS = Ptr->getType()->getPointerAddressSpace();
2044 // Use this type for pointer arithmetic.
2045 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
2047 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
2048 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
2049 Starts.push_back(Start);
2050 Ends.push_back(End);
2054 IRBuilder<> ChkBuilder(Loc);
2055 // Our instructions might fold to a constant.
2056 Value *MemoryRuntimeCheck = nullptr;
2057 for (unsigned i = 0; i < NumPointers; ++i) {
2058 for (unsigned j = i+1; j < NumPointers; ++j) {
2059 // No need to check if two readonly pointers intersect.
2060 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
2063 // Only need to check pointers between two different dependency sets.
2064 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
2066 // Only need to check pointers in the same alias set.
2067 if (PtrRtCheck->AliasSetId[i] != PtrRtCheck->AliasSetId[j])
2070 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
2071 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
2073 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
2074 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
2075 "Trying to bounds check pointers with different address spaces");
2077 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
2078 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
2080 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
2081 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
2082 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
2083 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
2085 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
2086 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
2087 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
2088 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
2089 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
2090 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2091 if (MemoryRuntimeCheck) {
2092 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
2094 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2096 MemoryRuntimeCheck = IsConflict;
2100 // We have to do this trickery because the IRBuilder might fold the check to a
2101 // constant expression in which case there is no Instruction anchored in a
2103 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
2104 ConstantInt::getTrue(Ctx));
2105 ChkBuilder.Insert(Check, "memcheck.conflict");
2106 FirstInst = getFirstInst(FirstInst, Check, Loc);
2107 return std::make_pair(FirstInst, Check);
2110 void InnerLoopVectorizer::createEmptyLoop() {
2112 In this function we generate a new loop. The new loop will contain
2113 the vectorized instructions while the old loop will continue to run the
2116 [ ] <-- Back-edge taken count overflow check.
2119 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2122 || [ ] <-- vector pre header.
2126 || [ ]_| <-- vector loop.
2129 | >[ ] <--- middle-block.
2132 -|- >[ ] <--- new preheader.
2136 | [ ]_| <-- old scalar loop to handle remainder.
2139 >[ ] <-- exit block.
2143 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2144 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2145 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2146 assert(BypassBlock && "Invalid loop structure");
2147 assert(ExitBlock && "Must have an exit block");
2149 // Some loops have a single integer induction variable, while other loops
2150 // don't. One example is c++ iterators that often have multiple pointer
2151 // induction variables. In the code below we also support a case where we
2152 // don't have a single induction variable.
2153 OldInduction = Legal->getInduction();
2154 Type *IdxTy = Legal->getWidestInductionType();
2156 // Find the loop boundaries.
2157 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2158 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2160 // The exit count might have the type of i64 while the phi is i32. This can
2161 // happen if we have an induction variable that is sign extended before the
2162 // compare. The only way that we get a backedge taken count is that the
2163 // induction variable was signed and as such will not overflow. In such a case
2164 // truncation is legal.
2165 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2166 IdxTy->getPrimitiveSizeInBits())
2167 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2169 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2170 // Get the total trip count from the count by adding 1.
2171 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2172 SE->getConstant(BackedgeTakeCount->getType(), 1));
2174 // Expand the trip count and place the new instructions in the preheader.
2175 // Notice that the pre-header does not change, only the loop body.
2176 SCEVExpander Exp(*SE, "induction");
2178 // We need to test whether the backedge-taken count is uint##_max. Adding one
2179 // to it will cause overflow and an incorrect loop trip count in the vector
2180 // body. In case of overflow we want to directly jump to the scalar remainder
2182 Value *BackedgeCount =
2183 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2184 BypassBlock->getTerminator());
2185 if (BackedgeCount->getType()->isPointerTy())
2186 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2187 "backedge.ptrcnt.to.int",
2188 BypassBlock->getTerminator());
2189 Instruction *CheckBCOverflow =
2190 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2191 Constant::getAllOnesValue(BackedgeCount->getType()),
2192 "backedge.overflow", BypassBlock->getTerminator());
2194 // The loop index does not have to start at Zero. Find the original start
2195 // value from the induction PHI node. If we don't have an induction variable
2196 // then we know that it starts at zero.
2197 Builder.SetInsertPoint(BypassBlock->getTerminator());
2198 Value *StartIdx = ExtendedIdx = OldInduction ?
2199 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2201 ConstantInt::get(IdxTy, 0);
2203 // We need an instruction to anchor the overflow check on. StartIdx needs to
2204 // be defined before the overflow check branch. Because the scalar preheader
2205 // is going to merge the start index and so the overflow branch block needs to
2206 // contain a definition of the start index.
2207 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2208 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2209 BypassBlock->getTerminator());
2211 // Count holds the overall loop count (N).
2212 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2213 BypassBlock->getTerminator());
2215 LoopBypassBlocks.push_back(BypassBlock);
2217 // Split the single block loop into the two loop structure described above.
2218 BasicBlock *VectorPH =
2219 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2220 BasicBlock *VecBody =
2221 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2222 BasicBlock *MiddleBlock =
2223 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2224 BasicBlock *ScalarPH =
2225 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2227 // Create and register the new vector loop.
2228 Loop* Lp = new Loop();
2229 Loop *ParentLoop = OrigLoop->getParentLoop();
2231 // Insert the new loop into the loop nest and register the new basic blocks
2232 // before calling any utilities such as SCEV that require valid LoopInfo.
2234 ParentLoop->addChildLoop(Lp);
2235 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
2236 ParentLoop->addBasicBlockToLoop(VectorPH, *LI);
2237 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
2239 LI->addTopLevelLoop(Lp);
2241 Lp->addBasicBlockToLoop(VecBody, *LI);
2243 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2245 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2247 // Generate the induction variable.
2248 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2249 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2250 // The loop step is equal to the vectorization factor (num of SIMD elements)
2251 // times the unroll factor (num of SIMD instructions).
2252 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2254 // This is the IR builder that we use to add all of the logic for bypassing
2255 // the new vector loop.
2256 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2257 setDebugLocFromInst(BypassBuilder,
2258 getDebugLocFromInstOrOperands(OldInduction));
2260 // We may need to extend the index in case there is a type mismatch.
2261 // We know that the count starts at zero and does not overflow.
2262 if (Count->getType() != IdxTy) {
2263 // The exit count can be of pointer type. Convert it to the correct
2265 if (ExitCount->getType()->isPointerTy())
2266 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2268 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2271 // Add the start index to the loop count to get the new end index.
2272 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2274 // Now we need to generate the expression for N - (N % VF), which is
2275 // the part that the vectorized body will execute.
2276 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2277 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2278 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2279 "end.idx.rnd.down");
2281 // Now, compare the new count to zero. If it is zero skip the vector loop and
2282 // jump to the scalar loop.
2284 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2286 BasicBlock *LastBypassBlock = BypassBlock;
2288 // Generate code to check that the loops trip count that we computed by adding
2289 // one to the backedge-taken count will not overflow.
2291 auto PastOverflowCheck =
2292 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2293 BasicBlock *CheckBlock =
2294 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2296 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2297 LoopBypassBlocks.push_back(CheckBlock);
2298 Instruction *OldTerm = LastBypassBlock->getTerminator();
2299 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2300 OldTerm->eraseFromParent();
2301 LastBypassBlock = CheckBlock;
2304 // Generate the code to check that the strides we assumed to be one are really
2305 // one. We want the new basic block to start at the first instruction in a
2306 // sequence of instructions that form a check.
2307 Instruction *StrideCheck;
2308 Instruction *FirstCheckInst;
2309 std::tie(FirstCheckInst, StrideCheck) =
2310 addStrideCheck(LastBypassBlock->getTerminator());
2312 // Create a new block containing the stride check.
2313 BasicBlock *CheckBlock =
2314 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2316 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2317 LoopBypassBlocks.push_back(CheckBlock);
2319 // Replace the branch into the memory check block with a conditional branch
2320 // for the "few elements case".
2321 Instruction *OldTerm = LastBypassBlock->getTerminator();
2322 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2323 OldTerm->eraseFromParent();
2326 LastBypassBlock = CheckBlock;
2329 // Generate the code that checks in runtime if arrays overlap. We put the
2330 // checks into a separate block to make the more common case of few elements
2332 Instruction *MemRuntimeCheck;
2333 std::tie(FirstCheckInst, MemRuntimeCheck) =
2334 addRuntimeCheck(LastBypassBlock->getTerminator());
2335 if (MemRuntimeCheck) {
2336 // Create a new block containing the memory check.
2337 BasicBlock *CheckBlock =
2338 LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
2340 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2341 LoopBypassBlocks.push_back(CheckBlock);
2343 // Replace the branch into the memory check block with a conditional branch
2344 // for the "few elements case".
2345 Instruction *OldTerm = LastBypassBlock->getTerminator();
2346 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2347 OldTerm->eraseFromParent();
2349 Cmp = MemRuntimeCheck;
2350 LastBypassBlock = CheckBlock;
2353 LastBypassBlock->getTerminator()->eraseFromParent();
2354 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2357 // We are going to resume the execution of the scalar loop.
2358 // Go over all of the induction variables that we found and fix the
2359 // PHIs that are left in the scalar version of the loop.
2360 // The starting values of PHI nodes depend on the counter of the last
2361 // iteration in the vectorized loop.
2362 // If we come from a bypass edge then we need to start from the original
2365 // This variable saves the new starting index for the scalar loop.
2366 PHINode *ResumeIndex = nullptr;
2367 LoopVectorizationLegality::InductionList::iterator I, E;
2368 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2369 // Set builder to point to last bypass block.
2370 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2371 for (I = List->begin(), E = List->end(); I != E; ++I) {
2372 PHINode *OrigPhi = I->first;
2373 LoopVectorizationLegality::InductionInfo II = I->second;
2375 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2376 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2377 MiddleBlock->getTerminator());
2378 // We might have extended the type of the induction variable but we need a
2379 // truncated version for the scalar loop.
2380 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2381 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2382 MiddleBlock->getTerminator()) : nullptr;
2384 // Create phi nodes to merge from the backedge-taken check block.
2385 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2386 ScalarPH->getTerminator());
2387 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2389 PHINode *BCTruncResumeVal = nullptr;
2390 if (OrigPhi == OldInduction) {
2392 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2393 ScalarPH->getTerminator());
2394 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2397 Value *EndValue = nullptr;
2399 case LoopVectorizationLegality::IK_NoInduction:
2400 llvm_unreachable("Unknown induction");
2401 case LoopVectorizationLegality::IK_IntInduction: {
2402 // Handle the integer induction counter.
2403 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2405 // We have the canonical induction variable.
2406 if (OrigPhi == OldInduction) {
2407 // Create a truncated version of the resume value for the scalar loop,
2408 // we might have promoted the type to a larger width.
2410 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2411 // The new PHI merges the original incoming value, in case of a bypass,
2412 // or the value at the end of the vectorized loop.
2413 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2414 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2415 TruncResumeVal->addIncoming(EndValue, VecBody);
2417 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2419 // We know what the end value is.
2420 EndValue = IdxEndRoundDown;
2421 // We also know which PHI node holds it.
2422 ResumeIndex = ResumeVal;
2426 // Not the canonical induction variable - add the vector loop count to the
2428 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2429 II.StartValue->getType(),
2431 EndValue = II.transform(BypassBuilder, CRD);
2432 EndValue->setName("ind.end");
2435 case LoopVectorizationLegality::IK_PtrInduction: {
2436 EndValue = II.transform(BypassBuilder, CountRoundDown);
2437 EndValue->setName("ptr.ind.end");
2442 // The new PHI merges the original incoming value, in case of a bypass,
2443 // or the value at the end of the vectorized loop.
2444 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2445 if (OrigPhi == OldInduction)
2446 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2448 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2450 ResumeVal->addIncoming(EndValue, VecBody);
2452 // Fix the scalar body counter (PHI node).
2453 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2455 // The old induction's phi node in the scalar body needs the truncated
2457 if (OrigPhi == OldInduction) {
2458 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2459 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2461 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2462 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2466 // If we are generating a new induction variable then we also need to
2467 // generate the code that calculates the exit value. This value is not
2468 // simply the end of the counter because we may skip the vectorized body
2469 // in case of a runtime check.
2471 assert(!ResumeIndex && "Unexpected resume value found");
2472 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2473 MiddleBlock->getTerminator());
2474 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2475 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2476 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2479 // Make sure that we found the index where scalar loop needs to continue.
2480 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2481 "Invalid resume Index");
2483 // Add a check in the middle block to see if we have completed
2484 // all of the iterations in the first vector loop.
2485 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2486 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2487 ResumeIndex, "cmp.n",
2488 MiddleBlock->getTerminator());
2490 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2491 // Remove the old terminator.
2492 MiddleBlock->getTerminator()->eraseFromParent();
2494 // Create i+1 and fill the PHINode.
2495 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2496 Induction->addIncoming(StartIdx, VectorPH);
2497 Induction->addIncoming(NextIdx, VecBody);
2498 // Create the compare.
2499 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2500 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2502 // Now we have two terminators. Remove the old one from the block.
2503 VecBody->getTerminator()->eraseFromParent();
2505 // Get ready to start creating new instructions into the vectorized body.
2506 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2509 LoopVectorPreHeader = VectorPH;
2510 LoopScalarPreHeader = ScalarPH;
2511 LoopMiddleBlock = MiddleBlock;
2512 LoopExitBlock = ExitBlock;
2513 LoopVectorBody.push_back(VecBody);
2514 LoopScalarBody = OldBasicBlock;
2516 LoopVectorizeHints Hints(Lp, true);
2517 Hints.setAlreadyVectorized();
2520 /// This function returns the identity element (or neutral element) for
2521 /// the operation K.
2523 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2528 // Adding, Xoring, Oring zero to a number does not change it.
2529 return ConstantInt::get(Tp, 0);
2530 case RK_IntegerMult:
2531 // Multiplying a number by 1 does not change it.
2532 return ConstantInt::get(Tp, 1);
2534 // AND-ing a number with an all-1 value does not change it.
2535 return ConstantInt::get(Tp, -1, true);
2537 // Multiplying a number by 1 does not change it.
2538 return ConstantFP::get(Tp, 1.0L);
2540 // Adding zero to a number does not change it.
2541 return ConstantFP::get(Tp, 0.0L);
2543 llvm_unreachable("Unknown reduction kind");
2547 /// This function translates the reduction kind to an LLVM binary operator.
2549 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2551 case LoopVectorizationLegality::RK_IntegerAdd:
2552 return Instruction::Add;
2553 case LoopVectorizationLegality::RK_IntegerMult:
2554 return Instruction::Mul;
2555 case LoopVectorizationLegality::RK_IntegerOr:
2556 return Instruction::Or;
2557 case LoopVectorizationLegality::RK_IntegerAnd:
2558 return Instruction::And;
2559 case LoopVectorizationLegality::RK_IntegerXor:
2560 return Instruction::Xor;
2561 case LoopVectorizationLegality::RK_FloatMult:
2562 return Instruction::FMul;
2563 case LoopVectorizationLegality::RK_FloatAdd:
2564 return Instruction::FAdd;
2565 case LoopVectorizationLegality::RK_IntegerMinMax:
2566 return Instruction::ICmp;
2567 case LoopVectorizationLegality::RK_FloatMinMax:
2568 return Instruction::FCmp;
2570 llvm_unreachable("Unknown reduction operation");
2574 Value *createMinMaxOp(IRBuilder<> &Builder,
2575 LoopVectorizationLegality::MinMaxReductionKind RK,
2578 CmpInst::Predicate P = CmpInst::ICMP_NE;
2581 llvm_unreachable("Unknown min/max reduction kind");
2582 case LoopVectorizationLegality::MRK_UIntMin:
2583 P = CmpInst::ICMP_ULT;
2585 case LoopVectorizationLegality::MRK_UIntMax:
2586 P = CmpInst::ICMP_UGT;
2588 case LoopVectorizationLegality::MRK_SIntMin:
2589 P = CmpInst::ICMP_SLT;
2591 case LoopVectorizationLegality::MRK_SIntMax:
2592 P = CmpInst::ICMP_SGT;
2594 case LoopVectorizationLegality::MRK_FloatMin:
2595 P = CmpInst::FCMP_OLT;
2597 case LoopVectorizationLegality::MRK_FloatMax:
2598 P = CmpInst::FCMP_OGT;
2603 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2604 RK == LoopVectorizationLegality::MRK_FloatMax)
2605 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2607 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2609 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2614 struct CSEDenseMapInfo {
2615 static bool canHandle(Instruction *I) {
2616 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2617 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2619 static inline Instruction *getEmptyKey() {
2620 return DenseMapInfo<Instruction *>::getEmptyKey();
2622 static inline Instruction *getTombstoneKey() {
2623 return DenseMapInfo<Instruction *>::getTombstoneKey();
2625 static unsigned getHashValue(Instruction *I) {
2626 assert(canHandle(I) && "Unknown instruction!");
2627 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2628 I->value_op_end()));
2630 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2631 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2632 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2634 return LHS->isIdenticalTo(RHS);
2639 /// \brief Check whether this block is a predicated block.
2640 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2641 /// = ...; " blocks. We start with one vectorized basic block. For every
2642 /// conditional block we split this vectorized block. Therefore, every second
2643 /// block will be a predicated one.
2644 static bool isPredicatedBlock(unsigned BlockNum) {
2645 return BlockNum % 2;
2648 ///\brief Perform cse of induction variable instructions.
2649 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2650 // Perform simple cse.
2651 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2652 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2653 BasicBlock *BB = BBs[i];
2654 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2655 Instruction *In = I++;
2657 if (!CSEDenseMapInfo::canHandle(In))
2660 // Check if we can replace this instruction with any of the
2661 // visited instructions.
2662 if (Instruction *V = CSEMap.lookup(In)) {
2663 In->replaceAllUsesWith(V);
2664 In->eraseFromParent();
2667 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2668 // ...;" blocks for predicated stores. Every second block is a predicated
2670 if (isPredicatedBlock(i))
2678 /// \brief Adds a 'fast' flag to floating point operations.
2679 static Value *addFastMathFlag(Value *V) {
2680 if (isa<FPMathOperator>(V)){
2681 FastMathFlags Flags;
2682 Flags.setUnsafeAlgebra();
2683 cast<Instruction>(V)->setFastMathFlags(Flags);
2688 void InnerLoopVectorizer::vectorizeLoop() {
2689 //===------------------------------------------------===//
2691 // Notice: any optimization or new instruction that go
2692 // into the code below should be also be implemented in
2695 //===------------------------------------------------===//
2696 Constant *Zero = Builder.getInt32(0);
2698 // In order to support reduction variables we need to be able to vectorize
2699 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2700 // stages. First, we create a new vector PHI node with no incoming edges.
2701 // We use this value when we vectorize all of the instructions that use the
2702 // PHI. Next, after all of the instructions in the block are complete we
2703 // add the new incoming edges to the PHI. At this point all of the
2704 // instructions in the basic block are vectorized, so we can use them to
2705 // construct the PHI.
2706 PhiVector RdxPHIsToFix;
2708 // Scan the loop in a topological order to ensure that defs are vectorized
2710 LoopBlocksDFS DFS(OrigLoop);
2713 // Vectorize all of the blocks in the original loop.
2714 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2715 be = DFS.endRPO(); bb != be; ++bb)
2716 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2718 // At this point every instruction in the original loop is widened to
2719 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2720 // that we vectorized. The PHI nodes are currently empty because we did
2721 // not want to introduce cycles. Notice that the remaining PHI nodes
2722 // that we need to fix are reduction variables.
2724 // Create the 'reduced' values for each of the induction vars.
2725 // The reduced values are the vector values that we scalarize and combine
2726 // after the loop is finished.
2727 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2729 PHINode *RdxPhi = *it;
2730 assert(RdxPhi && "Unable to recover vectorized PHI");
2732 // Find the reduction variable descriptor.
2733 assert(Legal->getReductionVars()->count(RdxPhi) &&
2734 "Unable to find the reduction variable");
2735 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2736 (*Legal->getReductionVars())[RdxPhi];
2738 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2740 // We need to generate a reduction vector from the incoming scalar.
2741 // To do so, we need to generate the 'identity' vector and override
2742 // one of the elements with the incoming scalar reduction. We need
2743 // to do it in the vector-loop preheader.
2744 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
2746 // This is the vector-clone of the value that leaves the loop.
2747 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2748 Type *VecTy = VectorExit[0]->getType();
2750 // Find the reduction identity variable. Zero for addition, or, xor,
2751 // one for multiplication, -1 for And.
2754 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2755 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2756 // MinMax reduction have the start value as their identify.
2758 VectorStart = Identity = RdxDesc.StartValue;
2760 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2765 // Handle other reduction kinds:
2767 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2768 VecTy->getScalarType());
2771 // This vector is the Identity vector where the first element is the
2772 // incoming scalar reduction.
2773 VectorStart = RdxDesc.StartValue;
2775 Identity = ConstantVector::getSplat(VF, Iden);
2777 // This vector is the Identity vector where the first element is the
2778 // incoming scalar reduction.
2779 VectorStart = Builder.CreateInsertElement(Identity,
2780 RdxDesc.StartValue, Zero);
2784 // Fix the vector-loop phi.
2786 // Reductions do not have to start at zero. They can start with
2787 // any loop invariant values.
2788 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2789 BasicBlock *Latch = OrigLoop->getLoopLatch();
2790 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2791 VectorParts &Val = getVectorValue(LoopVal);
2792 for (unsigned part = 0; part < UF; ++part) {
2793 // Make sure to add the reduction stat value only to the
2794 // first unroll part.
2795 Value *StartVal = (part == 0) ? VectorStart : Identity;
2796 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
2797 LoopVectorPreHeader);
2798 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2799 LoopVectorBody.back());
2802 // Before each round, move the insertion point right between
2803 // the PHIs and the values we are going to write.
2804 // This allows us to write both PHINodes and the extractelement
2806 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2808 VectorParts RdxParts;
2809 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2810 for (unsigned part = 0; part < UF; ++part) {
2811 // This PHINode contains the vectorized reduction variable, or
2812 // the initial value vector, if we bypass the vector loop.
2813 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2814 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2815 Value *StartVal = (part == 0) ? VectorStart : Identity;
2816 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2817 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2818 NewPhi->addIncoming(RdxExitVal[part],
2819 LoopVectorBody.back());
2820 RdxParts.push_back(NewPhi);
2823 // Reduce all of the unrolled parts into a single vector.
2824 Value *ReducedPartRdx = RdxParts[0];
2825 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2826 setDebugLocFromInst(Builder, ReducedPartRdx);
2827 for (unsigned part = 1; part < UF; ++part) {
2828 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2829 // Floating point operations had to be 'fast' to enable the reduction.
2830 ReducedPartRdx = addFastMathFlag(
2831 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2832 ReducedPartRdx, "bin.rdx"));
2834 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2835 ReducedPartRdx, RdxParts[part]);
2839 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2840 // and vector ops, reducing the set of values being computed by half each
2842 assert(isPowerOf2_32(VF) &&
2843 "Reduction emission only supported for pow2 vectors!");
2844 Value *TmpVec = ReducedPartRdx;
2845 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2846 for (unsigned i = VF; i != 1; i >>= 1) {
2847 // Move the upper half of the vector to the lower half.
2848 for (unsigned j = 0; j != i/2; ++j)
2849 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2851 // Fill the rest of the mask with undef.
2852 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2853 UndefValue::get(Builder.getInt32Ty()));
2856 Builder.CreateShuffleVector(TmpVec,
2857 UndefValue::get(TmpVec->getType()),
2858 ConstantVector::get(ShuffleMask),
2861 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2862 // Floating point operations had to be 'fast' to enable the reduction.
2863 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2864 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2866 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2869 // The result is in the first element of the vector.
2870 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2871 Builder.getInt32(0));
2874 // Create a phi node that merges control-flow from the backedge-taken check
2875 // block and the middle block.
2876 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
2877 LoopScalarPreHeader->getTerminator());
2878 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
2879 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2881 // Now, we need to fix the users of the reduction variable
2882 // inside and outside of the scalar remainder loop.
2883 // We know that the loop is in LCSSA form. We need to update the
2884 // PHI nodes in the exit blocks.
2885 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2886 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2887 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2888 if (!LCSSAPhi) break;
2890 // All PHINodes need to have a single entry edge, or two if
2891 // we already fixed them.
2892 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2894 // We found our reduction value exit-PHI. Update it with the
2895 // incoming bypass edge.
2896 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2897 // Add an edge coming from the bypass.
2898 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2901 }// end of the LCSSA phi scan.
2903 // Fix the scalar loop reduction variable with the incoming reduction sum
2904 // from the vector body and from the backedge value.
2905 int IncomingEdgeBlockIdx =
2906 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2907 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2908 // Pick the other block.
2909 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2910 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
2911 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2912 }// end of for each redux variable.
2916 // Remove redundant induction instructions.
2917 cse(LoopVectorBody);
2920 void InnerLoopVectorizer::fixLCSSAPHIs() {
2921 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2922 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2923 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2924 if (!LCSSAPhi) break;
2925 if (LCSSAPhi->getNumIncomingValues() == 1)
2926 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2931 InnerLoopVectorizer::VectorParts
2932 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2933 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2936 // Look for cached value.
2937 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2938 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2939 if (ECEntryIt != MaskCache.end())
2940 return ECEntryIt->second;
2942 VectorParts SrcMask = createBlockInMask(Src);
2944 // The terminator has to be a branch inst!
2945 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2946 assert(BI && "Unexpected terminator found");
2948 if (BI->isConditional()) {
2949 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2951 if (BI->getSuccessor(0) != Dst)
2952 for (unsigned part = 0; part < UF; ++part)
2953 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2955 for (unsigned part = 0; part < UF; ++part)
2956 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2958 MaskCache[Edge] = EdgeMask;
2962 MaskCache[Edge] = SrcMask;
2966 InnerLoopVectorizer::VectorParts
2967 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2968 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2970 // Loop incoming mask is all-one.
2971 if (OrigLoop->getHeader() == BB) {
2972 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2973 return getVectorValue(C);
2976 // This is the block mask. We OR all incoming edges, and with zero.
2977 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2978 VectorParts BlockMask = getVectorValue(Zero);
2981 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2982 VectorParts EM = createEdgeMask(*it, BB);
2983 for (unsigned part = 0; part < UF; ++part)
2984 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2990 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2991 InnerLoopVectorizer::VectorParts &Entry,
2992 unsigned UF, unsigned VF, PhiVector *PV) {
2993 PHINode* P = cast<PHINode>(PN);
2994 // Handle reduction variables:
2995 if (Legal->getReductionVars()->count(P)) {
2996 for (unsigned part = 0; part < UF; ++part) {
2997 // This is phase one of vectorizing PHIs.
2998 Type *VecTy = (VF == 1) ? PN->getType() :
2999 VectorType::get(PN->getType(), VF);
3000 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
3001 LoopVectorBody.back()-> getFirstInsertionPt());
3007 setDebugLocFromInst(Builder, P);
3008 // Check for PHI nodes that are lowered to vector selects.
3009 if (P->getParent() != OrigLoop->getHeader()) {
3010 // We know that all PHIs in non-header blocks are converted into
3011 // selects, so we don't have to worry about the insertion order and we
3012 // can just use the builder.
3013 // At this point we generate the predication tree. There may be
3014 // duplications since this is a simple recursive scan, but future
3015 // optimizations will clean it up.
3017 unsigned NumIncoming = P->getNumIncomingValues();
3019 // Generate a sequence of selects of the form:
3020 // SELECT(Mask3, In3,
3021 // SELECT(Mask2, In2,
3023 for (unsigned In = 0; In < NumIncoming; In++) {
3024 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
3026 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3028 for (unsigned part = 0; part < UF; ++part) {
3029 // We might have single edge PHIs (blocks) - use an identity
3030 // 'select' for the first PHI operand.
3032 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3035 // Select between the current value and the previous incoming edge
3036 // based on the incoming mask.
3037 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3038 Entry[part], "predphi");
3044 // This PHINode must be an induction variable.
3045 // Make sure that we know about it.
3046 assert(Legal->getInductionVars()->count(P) &&
3047 "Not an induction variable");
3049 LoopVectorizationLegality::InductionInfo II =
3050 Legal->getInductionVars()->lookup(P);
3052 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
3053 // which can be found from the original scalar operations.
3055 case LoopVectorizationLegality::IK_NoInduction:
3056 llvm_unreachable("Unknown induction");
3057 case LoopVectorizationLegality::IK_IntInduction: {
3058 assert(P->getType() == II.StartValue->getType() && "Types must match");
3059 Type *PhiTy = P->getType();
3061 if (P == OldInduction) {
3062 // Handle the canonical induction variable. We might have had to
3064 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
3066 // Handle other induction variables that are now based on the
3068 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
3070 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
3071 Broadcasted = II.transform(Builder, NormalizedIdx);
3072 Broadcasted->setName("offset.idx");
3074 Broadcasted = getBroadcastInstrs(Broadcasted);
3075 // After broadcasting the induction variable we need to make the vector
3076 // consecutive by adding 0, 1, 2, etc.
3077 for (unsigned part = 0; part < UF; ++part)
3078 Entry[part] = getStepVector(Broadcasted, VF * part, II.StepValue);
3081 case LoopVectorizationLegality::IK_PtrInduction:
3082 // Handle the pointer induction variable case.
3083 assert(P->getType()->isPointerTy() && "Unexpected type.");
3084 // This is the normalized GEP that starts counting at zero.
3085 Value *NormalizedIdx =
3086 Builder.CreateSub(Induction, ExtendedIdx, "normalized.idx");
3087 // This is the vector of results. Notice that we don't generate
3088 // vector geps because scalar geps result in better code.
3089 for (unsigned part = 0; part < UF; ++part) {
3091 int EltIndex = part;
3092 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3093 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3094 Value *SclrGep = II.transform(Builder, GlobalIdx);
3095 SclrGep->setName("next.gep");
3096 Entry[part] = SclrGep;
3100 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3101 for (unsigned int i = 0; i < VF; ++i) {
3102 int EltIndex = i + part * VF;
3103 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3104 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3105 Value *SclrGep = II.transform(Builder, GlobalIdx);
3106 SclrGep->setName("next.gep");
3107 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3108 Builder.getInt32(i),
3111 Entry[part] = VecVal;
3117 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3118 // For each instruction in the old loop.
3119 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3120 VectorParts &Entry = WidenMap.get(it);
3121 switch (it->getOpcode()) {
3122 case Instruction::Br:
3123 // Nothing to do for PHIs and BR, since we already took care of the
3124 // loop control flow instructions.
3126 case Instruction::PHI: {
3127 // Vectorize PHINodes.
3128 widenPHIInstruction(it, Entry, UF, VF, PV);
3132 case Instruction::Add:
3133 case Instruction::FAdd:
3134 case Instruction::Sub:
3135 case Instruction::FSub:
3136 case Instruction::Mul:
3137 case Instruction::FMul:
3138 case Instruction::UDiv:
3139 case Instruction::SDiv:
3140 case Instruction::FDiv:
3141 case Instruction::URem:
3142 case Instruction::SRem:
3143 case Instruction::FRem:
3144 case Instruction::Shl:
3145 case Instruction::LShr:
3146 case Instruction::AShr:
3147 case Instruction::And:
3148 case Instruction::Or:
3149 case Instruction::Xor: {
3150 // Just widen binops.
3151 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3152 setDebugLocFromInst(Builder, BinOp);
3153 VectorParts &A = getVectorValue(it->getOperand(0));
3154 VectorParts &B = getVectorValue(it->getOperand(1));
3156 // Use this vector value for all users of the original instruction.
3157 for (unsigned Part = 0; Part < UF; ++Part) {
3158 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3160 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3161 VecOp->copyIRFlags(BinOp);
3166 propagateMetadata(Entry, it);
3169 case Instruction::Select: {
3171 // If the selector is loop invariant we can create a select
3172 // instruction with a scalar condition. Otherwise, use vector-select.
3173 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3175 setDebugLocFromInst(Builder, it);
3177 // The condition can be loop invariant but still defined inside the
3178 // loop. This means that we can't just use the original 'cond' value.
3179 // We have to take the 'vectorized' value and pick the first lane.
3180 // Instcombine will make this a no-op.
3181 VectorParts &Cond = getVectorValue(it->getOperand(0));
3182 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3183 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3185 Value *ScalarCond = (VF == 1) ? Cond[0] :
3186 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3188 for (unsigned Part = 0; Part < UF; ++Part) {
3189 Entry[Part] = Builder.CreateSelect(
3190 InvariantCond ? ScalarCond : Cond[Part],
3195 propagateMetadata(Entry, it);
3199 case Instruction::ICmp:
3200 case Instruction::FCmp: {
3201 // Widen compares. Generate vector compares.
3202 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3203 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3204 setDebugLocFromInst(Builder, it);
3205 VectorParts &A = getVectorValue(it->getOperand(0));
3206 VectorParts &B = getVectorValue(it->getOperand(1));
3207 for (unsigned Part = 0; Part < UF; ++Part) {
3210 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3212 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3216 propagateMetadata(Entry, it);
3220 case Instruction::Store:
3221 case Instruction::Load:
3222 vectorizeMemoryInstruction(it);
3224 case Instruction::ZExt:
3225 case Instruction::SExt:
3226 case Instruction::FPToUI:
3227 case Instruction::FPToSI:
3228 case Instruction::FPExt:
3229 case Instruction::PtrToInt:
3230 case Instruction::IntToPtr:
3231 case Instruction::SIToFP:
3232 case Instruction::UIToFP:
3233 case Instruction::Trunc:
3234 case Instruction::FPTrunc:
3235 case Instruction::BitCast: {
3236 CastInst *CI = dyn_cast<CastInst>(it);
3237 setDebugLocFromInst(Builder, it);
3238 /// Optimize the special case where the source is the induction
3239 /// variable. Notice that we can only optimize the 'trunc' case
3240 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3241 /// c. other casts depend on pointer size.
3242 if (CI->getOperand(0) == OldInduction &&
3243 it->getOpcode() == Instruction::Trunc) {
3244 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3246 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3247 LoopVectorizationLegality::InductionInfo II =
3248 Legal->getInductionVars()->lookup(OldInduction);
3250 ConstantInt::getSigned(CI->getType(), II.StepValue->getSExtValue());
3251 for (unsigned Part = 0; Part < UF; ++Part)
3252 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
3253 propagateMetadata(Entry, it);
3256 /// Vectorize casts.
3257 Type *DestTy = (VF == 1) ? CI->getType() :
3258 VectorType::get(CI->getType(), VF);
3260 VectorParts &A = getVectorValue(it->getOperand(0));
3261 for (unsigned Part = 0; Part < UF; ++Part)
3262 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3263 propagateMetadata(Entry, it);
3267 case Instruction::Call: {
3268 // Ignore dbg intrinsics.
3269 if (isa<DbgInfoIntrinsic>(it))
3271 setDebugLocFromInst(Builder, it);
3273 Module *M = BB->getParent()->getParent();
3274 CallInst *CI = cast<CallInst>(it);
3275 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3276 assert(ID && "Not an intrinsic call!");
3278 case Intrinsic::assume:
3279 case Intrinsic::lifetime_end:
3280 case Intrinsic::lifetime_start:
3281 scalarizeInstruction(it);
3284 bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
3285 for (unsigned Part = 0; Part < UF; ++Part) {
3286 SmallVector<Value *, 4> Args;
3287 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3288 if (HasScalarOpd && i == 1) {
3289 Args.push_back(CI->getArgOperand(i));
3292 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3293 Args.push_back(Arg[Part]);
3295 Type *Tys[] = {CI->getType()};
3297 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3299 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3300 Entry[Part] = Builder.CreateCall(F, Args);
3303 propagateMetadata(Entry, it);
3310 // All other instructions are unsupported. Scalarize them.
3311 scalarizeInstruction(it);
3314 }// end of for_each instr.
3317 void InnerLoopVectorizer::updateAnalysis() {
3318 // Forget the original basic block.
3319 SE->forgetLoop(OrigLoop);
3321 // Update the dominator tree information.
3322 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3323 "Entry does not dominate exit.");
3325 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3326 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3327 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3329 // Due to if predication of stores we might create a sequence of "if(pred)
3330 // a[i] = ...; " blocks.
3331 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3333 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3334 else if (isPredicatedBlock(i)) {
3335 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3337 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3341 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3342 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3343 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3344 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3346 DEBUG(DT->verifyDomTree());
3349 /// \brief Check whether it is safe to if-convert this phi node.
3351 /// Phi nodes with constant expressions that can trap are not safe to if
3353 static bool canIfConvertPHINodes(BasicBlock *BB) {
3354 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3355 PHINode *Phi = dyn_cast<PHINode>(I);
3358 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3359 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3366 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3367 if (!EnableIfConversion) {
3368 emitAnalysis(VectorizationReport() << "if-conversion is disabled");
3372 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3374 // A list of pointers that we can safely read and write to.
3375 SmallPtrSet<Value *, 8> SafePointes;
3377 // Collect safe addresses.
3378 for (Loop::block_iterator BI = TheLoop->block_begin(),
3379 BE = TheLoop->block_end(); BI != BE; ++BI) {
3380 BasicBlock *BB = *BI;
3382 if (blockNeedsPredication(BB))
3385 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3386 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3387 SafePointes.insert(LI->getPointerOperand());
3388 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3389 SafePointes.insert(SI->getPointerOperand());
3393 // Collect the blocks that need predication.
3394 BasicBlock *Header = TheLoop->getHeader();
3395 for (Loop::block_iterator BI = TheLoop->block_begin(),
3396 BE = TheLoop->block_end(); BI != BE; ++BI) {
3397 BasicBlock *BB = *BI;
3399 // We don't support switch statements inside loops.
3400 if (!isa<BranchInst>(BB->getTerminator())) {
3401 emitAnalysis(VectorizationReport(BB->getTerminator())
3402 << "loop contains a switch statement");
3406 // We must be able to predicate all blocks that need to be predicated.
3407 if (blockNeedsPredication(BB)) {
3408 if (!blockCanBePredicated(BB, SafePointes)) {
3409 emitAnalysis(VectorizationReport(BB->getTerminator())
3410 << "control flow cannot be substituted for a select");
3413 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3414 emitAnalysis(VectorizationReport(BB->getTerminator())
3415 << "control flow cannot be substituted for a select");
3420 // We can if-convert this loop.
3424 bool LoopVectorizationLegality::canVectorize() {
3425 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3426 // be canonicalized.
3427 if (!TheLoop->getLoopPreheader()) {
3429 VectorizationReport() <<
3430 "loop control flow is not understood by vectorizer");
3434 // We can only vectorize innermost loops.
3435 if (!TheLoop->getSubLoopsVector().empty()) {
3436 emitAnalysis(VectorizationReport() << "loop is not the innermost loop");
3440 // We must have a single backedge.
3441 if (TheLoop->getNumBackEdges() != 1) {
3443 VectorizationReport() <<
3444 "loop control flow is not understood by vectorizer");
3448 // We must have a single exiting block.
3449 if (!TheLoop->getExitingBlock()) {
3451 VectorizationReport() <<
3452 "loop control flow is not understood by vectorizer");
3456 // We only handle bottom-tested loops, i.e. loop in which the condition is
3457 // checked at the end of each iteration. With that we can assume that all
3458 // instructions in the loop are executed the same number of times.
3459 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
3461 VectorizationReport() <<
3462 "loop control flow is not understood by vectorizer");
3466 // We need to have a loop header.
3467 DEBUG(dbgs() << "LV: Found a loop: " <<
3468 TheLoop->getHeader()->getName() << '\n');
3470 // Check if we can if-convert non-single-bb loops.
3471 unsigned NumBlocks = TheLoop->getNumBlocks();
3472 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3473 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3477 // ScalarEvolution needs to be able to find the exit count.
3478 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3479 if (ExitCount == SE->getCouldNotCompute()) {
3480 emitAnalysis(VectorizationReport() <<
3481 "could not determine number of loop iterations");
3482 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3486 // Check if we can vectorize the instructions and CFG in this loop.
3487 if (!canVectorizeInstrs()) {
3488 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3492 // Go over each instruction and look at memory deps.
3493 if (!canVectorizeMemory()) {
3494 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3498 // Collect all of the variables that remain uniform after vectorization.
3499 collectLoopUniforms();
3501 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3502 (LAA.getRuntimePointerCheck()->Need ? " (with a runtime bound check)" :
3506 // Okay! We can vectorize. At this point we don't have any other mem analysis
3507 // which may limit our maximum vectorization factor, so just return true with
3512 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3513 if (Ty->isPointerTy())
3514 return DL.getIntPtrType(Ty);
3516 // It is possible that char's or short's overflow when we ask for the loop's
3517 // trip count, work around this by changing the type size.
3518 if (Ty->getScalarSizeInBits() < 32)
3519 return Type::getInt32Ty(Ty->getContext());
3524 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3525 Ty0 = convertPointerToIntegerType(DL, Ty0);
3526 Ty1 = convertPointerToIntegerType(DL, Ty1);
3527 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3532 /// \brief Check that the instruction has outside loop users and is not an
3533 /// identified reduction variable.
3534 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3535 SmallPtrSetImpl<Value *> &Reductions) {
3536 // Reduction instructions are allowed to have exit users. All other
3537 // instructions must not have external users.
3538 if (!Reductions.count(Inst))
3539 //Check that all of the users of the loop are inside the BB.
3540 for (User *U : Inst->users()) {
3541 Instruction *UI = cast<Instruction>(U);
3542 // This user may be a reduction exit value.
3543 if (!TheLoop->contains(UI)) {
3544 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3551 bool LoopVectorizationLegality::canVectorizeInstrs() {
3552 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3553 BasicBlock *Header = TheLoop->getHeader();
3555 // Look for the attribute signaling the absence of NaNs.
3556 Function &F = *Header->getParent();
3557 if (F.hasFnAttribute("no-nans-fp-math"))
3558 HasFunNoNaNAttr = F.getAttributes().getAttribute(
3559 AttributeSet::FunctionIndex,
3560 "no-nans-fp-math").getValueAsString() == "true";
3562 // For each block in the loop.
3563 for (Loop::block_iterator bb = TheLoop->block_begin(),
3564 be = TheLoop->block_end(); bb != be; ++bb) {
3566 // Scan the instructions in the block and look for hazards.
3567 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3570 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3571 Type *PhiTy = Phi->getType();
3572 // Check that this PHI type is allowed.
3573 if (!PhiTy->isIntegerTy() &&
3574 !PhiTy->isFloatingPointTy() &&
3575 !PhiTy->isPointerTy()) {
3576 emitAnalysis(VectorizationReport(it)
3577 << "loop control flow is not understood by vectorizer");
3578 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3582 // If this PHINode is not in the header block, then we know that we
3583 // can convert it to select during if-conversion. No need to check if
3584 // the PHIs in this block are induction or reduction variables.
3585 if (*bb != Header) {
3586 // Check that this instruction has no outside users or is an
3587 // identified reduction value with an outside user.
3588 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3590 emitAnalysis(VectorizationReport(it) <<
3591 "value could not be identified as "
3592 "an induction or reduction variable");
3596 // We only allow if-converted PHIs with exactly two incoming values.
3597 if (Phi->getNumIncomingValues() != 2) {
3598 emitAnalysis(VectorizationReport(it)
3599 << "control flow not understood by vectorizer");
3600 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3604 // This is the value coming from the preheader.
3605 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3606 ConstantInt *StepValue = nullptr;
3607 // Check if this is an induction variable.
3608 InductionKind IK = isInductionVariable(Phi, StepValue);
3610 if (IK_NoInduction != IK) {
3611 // Get the widest type.
3613 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3615 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3617 // Int inductions are special because we only allow one IV.
3618 if (IK == IK_IntInduction && StepValue->isOne()) {
3619 // Use the phi node with the widest type as induction. Use the last
3620 // one if there are multiple (no good reason for doing this other
3621 // than it is expedient).
3622 if (!Induction || PhiTy == WidestIndTy)
3626 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3627 Inductions[Phi] = InductionInfo(StartValue, IK, StepValue);
3629 // Until we explicitly handle the case of an induction variable with
3630 // an outside loop user we have to give up vectorizing this loop.
3631 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3632 emitAnalysis(VectorizationReport(it) <<
3633 "use of induction value outside of the "
3634 "loop is not handled by vectorizer");
3641 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3642 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3645 if (AddReductionVar(Phi, RK_IntegerMult)) {
3646 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3649 if (AddReductionVar(Phi, RK_IntegerOr)) {
3650 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3653 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3654 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3657 if (AddReductionVar(Phi, RK_IntegerXor)) {
3658 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3661 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3662 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3665 if (AddReductionVar(Phi, RK_FloatMult)) {
3666 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3669 if (AddReductionVar(Phi, RK_FloatAdd)) {
3670 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3673 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3674 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3679 emitAnalysis(VectorizationReport(it) <<
3680 "value that could not be identified as "
3681 "reduction is used outside the loop");
3682 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3684 }// end of PHI handling
3686 // We still don't handle functions. However, we can ignore dbg intrinsic
3687 // calls and we do handle certain intrinsic and libm functions.
3688 CallInst *CI = dyn_cast<CallInst>(it);
3689 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3690 emitAnalysis(VectorizationReport(it) <<
3691 "call instruction cannot be vectorized");
3692 DEBUG(dbgs() << "LV: Found a call site.\n");
3696 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
3697 // second argument is the same (i.e. loop invariant)
3699 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
3700 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
3701 emitAnalysis(VectorizationReport(it)
3702 << "intrinsic instruction cannot be vectorized");
3703 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
3708 // Check that the instruction return type is vectorizable.
3709 // Also, we can't vectorize extractelement instructions.
3710 if ((!VectorType::isValidElementType(it->getType()) &&
3711 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3712 emitAnalysis(VectorizationReport(it)
3713 << "instruction return type cannot be vectorized");
3714 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3718 // Check that the stored type is vectorizable.
3719 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3720 Type *T = ST->getValueOperand()->getType();
3721 if (!VectorType::isValidElementType(T)) {
3722 emitAnalysis(VectorizationReport(ST) <<
3723 "store instruction cannot be vectorized");
3726 if (EnableMemAccessVersioning)
3727 collectStridedAccess(ST);
3730 if (EnableMemAccessVersioning)
3731 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3732 collectStridedAccess(LI);
3734 // Reduction instructions are allowed to have exit users.
3735 // All other instructions must not have external users.
3736 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3737 emitAnalysis(VectorizationReport(it) <<
3738 "value cannot be used outside the loop");
3747 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3748 if (Inductions.empty()) {
3749 emitAnalysis(VectorizationReport()
3750 << "loop induction variable could not be identified");
3758 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3759 /// return the induction operand of the gep pointer.
3760 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3761 const DataLayout *DL, Loop *Lp) {
3762 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3766 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3768 // Check that all of the gep indices are uniform except for our induction
3770 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3771 if (i != InductionOperand &&
3772 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3774 return GEP->getOperand(InductionOperand);
3777 ///\brief Look for a cast use of the passed value.
3778 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3779 Value *UniqueCast = nullptr;
3780 for (User *U : Ptr->users()) {
3781 CastInst *CI = dyn_cast<CastInst>(U);
3782 if (CI && CI->getType() == Ty) {
3792 ///\brief Get the stride of a pointer access in a loop.
3793 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3794 /// pointer to the Value, or null otherwise.
3795 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3796 const DataLayout *DL, Loop *Lp) {
3797 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3798 if (!PtrTy || PtrTy->isAggregateType())
3801 // Try to remove a gep instruction to make the pointer (actually index at this
3802 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3803 // pointer, otherwise, we are analyzing the index.
3804 Value *OrigPtr = Ptr;
3806 // The size of the pointer access.
3807 int64_t PtrAccessSize = 1;
3809 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3810 const SCEV *V = SE->getSCEV(Ptr);
3814 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3815 V = C->getOperand();
3817 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3821 V = S->getStepRecurrence(*SE);
3825 // Strip off the size of access multiplication if we are still analyzing the
3827 if (OrigPtr == Ptr) {
3828 DL->getTypeAllocSize(PtrTy->getElementType());
3829 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3830 if (M->getOperand(0)->getSCEVType() != scConstant)
3833 const APInt &APStepVal =
3834 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3836 // Huge step value - give up.
3837 if (APStepVal.getBitWidth() > 64)
3840 int64_t StepVal = APStepVal.getSExtValue();
3841 if (PtrAccessSize != StepVal)
3843 V = M->getOperand(1);
3848 Type *StripedOffRecurrenceCast = nullptr;
3849 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3850 StripedOffRecurrenceCast = C->getType();
3851 V = C->getOperand();
3854 // Look for the loop invariant symbolic value.
3855 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3859 Value *Stride = U->getValue();
3860 if (!Lp->isLoopInvariant(Stride))
3863 // If we have stripped off the recurrence cast we have to make sure that we
3864 // return the value that is used in this loop so that we can replace it later.
3865 if (StripedOffRecurrenceCast)
3866 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3871 void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) {
3872 Value *Ptr = nullptr;
3873 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3874 Ptr = LI->getPointerOperand();
3875 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3876 Ptr = SI->getPointerOperand();
3880 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3884 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3885 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3886 Strides[Ptr] = Stride;
3887 StrideSet.insert(Stride);
3890 void LoopVectorizationLegality::collectLoopUniforms() {
3891 // We now know that the loop is vectorizable!
3892 // Collect variables that will remain uniform after vectorization.
3893 std::vector<Value*> Worklist;
3894 BasicBlock *Latch = TheLoop->getLoopLatch();
3896 // Start with the conditional branch and walk up the block.
3897 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3899 // Also add all consecutive pointer values; these values will be uniform
3900 // after vectorization (and subsequent cleanup) and, until revectorization is
3901 // supported, all dependencies must also be uniform.
3902 for (Loop::block_iterator B = TheLoop->block_begin(),
3903 BE = TheLoop->block_end(); B != BE; ++B)
3904 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
3906 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
3907 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3909 while (!Worklist.empty()) {
3910 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3911 Worklist.pop_back();
3913 // Look at instructions inside this loop.
3914 // Stop when reaching PHI nodes.
3915 // TODO: we need to follow values all over the loop, not only in this block.
3916 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3919 // This is a known uniform.
3922 // Insert all operands.
3923 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3927 bool LoopVectorizationLegality::canVectorizeMemory() {
3928 return LAA.canVectorizeMemory(Strides);
3931 static bool hasMultipleUsesOf(Instruction *I,
3932 SmallPtrSetImpl<Instruction *> &Insts) {
3933 unsigned NumUses = 0;
3934 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
3935 if (Insts.count(dyn_cast<Instruction>(*Use)))
3944 static bool areAllUsesIn(Instruction *I, SmallPtrSetImpl<Instruction *> &Set) {
3945 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
3946 if (!Set.count(dyn_cast<Instruction>(*Use)))
3951 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
3952 ReductionKind Kind) {
3953 if (Phi->getNumIncomingValues() != 2)
3956 // Reduction variables are only found in the loop header block.
3957 if (Phi->getParent() != TheLoop->getHeader())
3960 // Obtain the reduction start value from the value that comes from the loop
3962 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
3964 // ExitInstruction is the single value which is used outside the loop.
3965 // We only allow for a single reduction value to be used outside the loop.
3966 // This includes users of the reduction, variables (which form a cycle
3967 // which ends in the phi node).
3968 Instruction *ExitInstruction = nullptr;
3969 // Indicates that we found a reduction operation in our scan.
3970 bool FoundReduxOp = false;
3972 // We start with the PHI node and scan for all of the users of this
3973 // instruction. All users must be instructions that can be used as reduction
3974 // variables (such as ADD). We must have a single out-of-block user. The cycle
3975 // must include the original PHI.
3976 bool FoundStartPHI = false;
3978 // To recognize min/max patterns formed by a icmp select sequence, we store
3979 // the number of instruction we saw from the recognized min/max pattern,
3980 // to make sure we only see exactly the two instructions.
3981 unsigned NumCmpSelectPatternInst = 0;
3982 ReductionInstDesc ReduxDesc(false, nullptr);
3984 SmallPtrSet<Instruction *, 8> VisitedInsts;
3985 SmallVector<Instruction *, 8> Worklist;
3986 Worklist.push_back(Phi);
3987 VisitedInsts.insert(Phi);
3989 // A value in the reduction can be used:
3990 // - By the reduction:
3991 // - Reduction operation:
3992 // - One use of reduction value (safe).
3993 // - Multiple use of reduction value (not safe).
3995 // - All uses of the PHI must be the reduction (safe).
3996 // - Otherwise, not safe.
3997 // - By one instruction outside of the loop (safe).
3998 // - By further instructions outside of the loop (not safe).
3999 // - By an instruction that is not part of the reduction (not safe).
4001 // * An instruction type other than PHI or the reduction operation.
4002 // * A PHI in the header other than the initial PHI.
4003 while (!Worklist.empty()) {
4004 Instruction *Cur = Worklist.back();
4005 Worklist.pop_back();
4008 // If the instruction has no users then this is a broken chain and can't be
4009 // a reduction variable.
4010 if (Cur->use_empty())
4013 bool IsAPhi = isa<PHINode>(Cur);
4015 // A header PHI use other than the original PHI.
4016 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
4019 // Reductions of instructions such as Div, and Sub is only possible if the
4020 // LHS is the reduction variable.
4021 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
4022 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
4023 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
4026 // Any reduction instruction must be of one of the allowed kinds.
4027 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
4028 if (!ReduxDesc.IsReduction)
4031 // A reduction operation must only have one use of the reduction value.
4032 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
4033 hasMultipleUsesOf(Cur, VisitedInsts))
4036 // All inputs to a PHI node must be a reduction value.
4037 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
4040 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
4041 isa<SelectInst>(Cur)))
4042 ++NumCmpSelectPatternInst;
4043 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
4044 isa<SelectInst>(Cur)))
4045 ++NumCmpSelectPatternInst;
4047 // Check whether we found a reduction operator.
4048 FoundReduxOp |= !IsAPhi;
4050 // Process users of current instruction. Push non-PHI nodes after PHI nodes
4051 // onto the stack. This way we are going to have seen all inputs to PHI
4052 // nodes once we get to them.
4053 SmallVector<Instruction *, 8> NonPHIs;
4054 SmallVector<Instruction *, 8> PHIs;
4055 for (User *U : Cur->users()) {
4056 Instruction *UI = cast<Instruction>(U);
4058 // Check if we found the exit user.
4059 BasicBlock *Parent = UI->getParent();
4060 if (!TheLoop->contains(Parent)) {
4061 // Exit if you find multiple outside users or if the header phi node is
4062 // being used. In this case the user uses the value of the previous
4063 // iteration, in which case we would loose "VF-1" iterations of the
4064 // reduction operation if we vectorize.
4065 if (ExitInstruction != nullptr || Cur == Phi)
4068 // The instruction used by an outside user must be the last instruction
4069 // before we feed back to the reduction phi. Otherwise, we loose VF-1
4070 // operations on the value.
4071 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
4074 ExitInstruction = Cur;
4078 // Process instructions only once (termination). Each reduction cycle
4079 // value must only be used once, except by phi nodes and min/max
4080 // reductions which are represented as a cmp followed by a select.
4081 ReductionInstDesc IgnoredVal(false, nullptr);
4082 if (VisitedInsts.insert(UI).second) {
4083 if (isa<PHINode>(UI))
4086 NonPHIs.push_back(UI);
4087 } else if (!isa<PHINode>(UI) &&
4088 ((!isa<FCmpInst>(UI) &&
4089 !isa<ICmpInst>(UI) &&
4090 !isa<SelectInst>(UI)) ||
4091 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
4094 // Remember that we completed the cycle.
4096 FoundStartPHI = true;
4098 Worklist.append(PHIs.begin(), PHIs.end());
4099 Worklist.append(NonPHIs.begin(), NonPHIs.end());
4102 // This means we have seen one but not the other instruction of the
4103 // pattern or more than just a select and cmp.
4104 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
4105 NumCmpSelectPatternInst != 2)
4108 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
4111 // We found a reduction var if we have reached the original phi node and we
4112 // only have a single instruction with out-of-loop users.
4114 // This instruction is allowed to have out-of-loop users.
4115 AllowedExit.insert(ExitInstruction);
4117 // Save the description of this reduction variable.
4118 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
4119 ReduxDesc.MinMaxKind);
4120 Reductions[Phi] = RD;
4121 // We've ended the cycle. This is a reduction variable if we have an
4122 // outside user and it has a binary op.
4127 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
4128 /// pattern corresponding to a min(X, Y) or max(X, Y).
4129 LoopVectorizationLegality::ReductionInstDesc
4130 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
4131 ReductionInstDesc &Prev) {
4133 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
4134 "Expect a select instruction");
4135 Instruction *Cmp = nullptr;
4136 SelectInst *Select = nullptr;
4138 // We must handle the select(cmp()) as a single instruction. Advance to the
4140 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
4141 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
4142 return ReductionInstDesc(false, I);
4143 return ReductionInstDesc(Select, Prev.MinMaxKind);
4146 // Only handle single use cases for now.
4147 if (!(Select = dyn_cast<SelectInst>(I)))
4148 return ReductionInstDesc(false, I);
4149 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
4150 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
4151 return ReductionInstDesc(false, I);
4152 if (!Cmp->hasOneUse())
4153 return ReductionInstDesc(false, I);
4158 // Look for a min/max pattern.
4159 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4160 return ReductionInstDesc(Select, MRK_UIntMin);
4161 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4162 return ReductionInstDesc(Select, MRK_UIntMax);
4163 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4164 return ReductionInstDesc(Select, MRK_SIntMax);
4165 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4166 return ReductionInstDesc(Select, MRK_SIntMin);
4167 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4168 return ReductionInstDesc(Select, MRK_FloatMin);
4169 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4170 return ReductionInstDesc(Select, MRK_FloatMax);
4171 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4172 return ReductionInstDesc(Select, MRK_FloatMin);
4173 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4174 return ReductionInstDesc(Select, MRK_FloatMax);
4176 return ReductionInstDesc(false, I);
4179 LoopVectorizationLegality::ReductionInstDesc
4180 LoopVectorizationLegality::isReductionInstr(Instruction *I,
4182 ReductionInstDesc &Prev) {
4183 bool FP = I->getType()->isFloatingPointTy();
4184 bool FastMath = FP && I->hasUnsafeAlgebra();
4185 switch (I->getOpcode()) {
4187 return ReductionInstDesc(false, I);
4188 case Instruction::PHI:
4189 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
4190 Kind != RK_FloatMinMax))
4191 return ReductionInstDesc(false, I);
4192 return ReductionInstDesc(I, Prev.MinMaxKind);
4193 case Instruction::Sub:
4194 case Instruction::Add:
4195 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4196 case Instruction::Mul:
4197 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4198 case Instruction::And:
4199 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4200 case Instruction::Or:
4201 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4202 case Instruction::Xor:
4203 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4204 case Instruction::FMul:
4205 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4206 case Instruction::FSub:
4207 case Instruction::FAdd:
4208 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4209 case Instruction::FCmp:
4210 case Instruction::ICmp:
4211 case Instruction::Select:
4212 if (Kind != RK_IntegerMinMax &&
4213 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4214 return ReductionInstDesc(false, I);
4215 return isMinMaxSelectCmpPattern(I, Prev);
4219 LoopVectorizationLegality::InductionKind
4220 LoopVectorizationLegality::isInductionVariable(PHINode *Phi,
4221 ConstantInt *&StepValue) {
4222 Type *PhiTy = Phi->getType();
4223 // We only handle integer and pointer inductions variables.
4224 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4225 return IK_NoInduction;
4227 // Check that the PHI is consecutive.
4228 const SCEV *PhiScev = SE->getSCEV(Phi);
4229 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4231 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4232 return IK_NoInduction;
4235 const SCEV *Step = AR->getStepRecurrence(*SE);
4236 // Calculate the pointer stride and check if it is consecutive.
4237 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4239 return IK_NoInduction;
4241 ConstantInt *CV = C->getValue();
4242 if (PhiTy->isIntegerTy()) {
4244 return IK_IntInduction;
4247 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4248 Type *PointerElementType = PhiTy->getPointerElementType();
4249 // The pointer stride cannot be determined if the pointer element type is not
4251 if (!PointerElementType->isSized())
4252 return IK_NoInduction;
4254 int64_t Size = static_cast<int64_t>(DL->getTypeAllocSize(PointerElementType));
4255 int64_t CVSize = CV->getSExtValue();
4257 return IK_NoInduction;
4258 StepValue = ConstantInt::getSigned(CV->getType(), CVSize / Size);
4259 return IK_PtrInduction;
4262 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4263 Value *In0 = const_cast<Value*>(V);
4264 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4268 return Inductions.count(PN);
4271 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4272 return LAA.blockNeedsPredication(BB);
4275 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4276 SmallPtrSetImpl<Value *> &SafePtrs) {
4278 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4279 // Check that we don't have a constant expression that can trap as operand.
4280 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4282 if (Constant *C = dyn_cast<Constant>(*OI))
4286 // We might be able to hoist the load.
4287 if (it->mayReadFromMemory()) {
4288 LoadInst *LI = dyn_cast<LoadInst>(it);
4291 if (!SafePtrs.count(LI->getPointerOperand())) {
4292 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
4293 MaskedOp.insert(LI);
4300 // We don't predicate stores at the moment.
4301 if (it->mayWriteToMemory()) {
4302 StoreInst *SI = dyn_cast<StoreInst>(it);
4303 // We only support predication of stores in basic blocks with one
4308 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
4309 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
4311 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
4312 !isSinglePredecessor) {
4313 // Build a masked store if it is legal for the target, otherwise scalarize
4315 bool isLegalMaskedOp =
4316 isLegalMaskedStore(SI->getValueOperand()->getType(),
4317 SI->getPointerOperand());
4318 if (isLegalMaskedOp) {
4320 MaskedOp.insert(SI);
4329 // The instructions below can trap.
4330 switch (it->getOpcode()) {
4332 case Instruction::UDiv:
4333 case Instruction::SDiv:
4334 case Instruction::URem:
4335 case Instruction::SRem:
4343 LoopVectorizationCostModel::VectorizationFactor
4344 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
4345 // Width 1 means no vectorize
4346 VectorizationFactor Factor = { 1U, 0U };
4347 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4348 emitAnalysis(VectorizationReport() <<
4349 "runtime pointer checks needed. Enable vectorization of this "
4350 "loop with '#pragma clang loop vectorize(enable)' when "
4351 "compiling with -Os");
4352 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4356 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
4357 emitAnalysis(VectorizationReport() <<
4358 "store that is conditionally executed prevents vectorization");
4359 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
4363 // Find the trip count.
4364 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4365 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4367 unsigned WidestType = getWidestType();
4368 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4369 unsigned MaxSafeDepDist = -1U;
4370 if (Legal->getMaxSafeDepDistBytes() != -1U)
4371 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4372 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4373 WidestRegister : MaxSafeDepDist);
4374 unsigned MaxVectorSize = WidestRegister / WidestType;
4375 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4376 DEBUG(dbgs() << "LV: The Widest register is: "
4377 << WidestRegister << " bits.\n");
4379 if (MaxVectorSize == 0) {
4380 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4384 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
4385 " into one vector!");
4387 unsigned VF = MaxVectorSize;
4389 // If we optimize the program for size, avoid creating the tail loop.
4391 // If we are unable to calculate the trip count then don't try to vectorize.
4394 (VectorizationReport() <<
4395 "unable to calculate the loop count due to complex control flow");
4396 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4400 // Find the maximum SIMD width that can fit within the trip count.
4401 VF = TC % MaxVectorSize;
4406 // If the trip count that we found modulo the vectorization factor is not
4407 // zero then we require a tail.
4409 emitAnalysis(VectorizationReport() <<
4410 "cannot optimize for size and vectorize at the "
4411 "same time. Enable vectorization of this loop "
4412 "with '#pragma clang loop vectorize(enable)' "
4413 "when compiling with -Os");
4414 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4419 int UserVF = Hints->getWidth();
4421 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4422 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4424 Factor.Width = UserVF;
4428 float Cost = expectedCost(1);
4430 const float ScalarCost = Cost;
4433 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
4435 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
4436 // Ignore scalar width, because the user explicitly wants vectorization.
4437 if (ForceVectorization && VF > 1) {
4439 Cost = expectedCost(Width) / (float)Width;
4442 for (unsigned i=2; i <= VF; i*=2) {
4443 // Notice that the vector loop needs to be executed less times, so
4444 // we need to divide the cost of the vector loops by the width of
4445 // the vector elements.
4446 float VectorCost = expectedCost(i) / (float)i;
4447 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
4448 (int)VectorCost << ".\n");
4449 if (VectorCost < Cost) {
4455 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
4456 << "LV: Vectorization seems to be not beneficial, "
4457 << "but was forced by a user.\n");
4458 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
4459 Factor.Width = Width;
4460 Factor.Cost = Width * Cost;
4464 unsigned LoopVectorizationCostModel::getWidestType() {
4465 unsigned MaxWidth = 8;
4468 for (Loop::block_iterator bb = TheLoop->block_begin(),
4469 be = TheLoop->block_end(); bb != be; ++bb) {
4470 BasicBlock *BB = *bb;
4472 // For each instruction in the loop.
4473 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4474 Type *T = it->getType();
4476 // Ignore ephemeral values.
4477 if (EphValues.count(it))
4480 // Only examine Loads, Stores and PHINodes.
4481 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4484 // Examine PHI nodes that are reduction variables.
4485 if (PHINode *PN = dyn_cast<PHINode>(it))
4486 if (!Legal->getReductionVars()->count(PN))
4489 // Examine the stored values.
4490 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4491 T = ST->getValueOperand()->getType();
4493 // Ignore loaded pointer types and stored pointer types that are not
4494 // consecutive. However, we do want to take consecutive stores/loads of
4495 // pointer vectors into account.
4496 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4499 MaxWidth = std::max(MaxWidth,
4500 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
4508 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
4510 unsigned LoopCost) {
4512 // -- The unroll heuristics --
4513 // We unroll the loop in order to expose ILP and reduce the loop overhead.
4514 // There are many micro-architectural considerations that we can't predict
4515 // at this level. For example, frontend pressure (on decode or fetch) due to
4516 // code size, or the number and capabilities of the execution ports.
4518 // We use the following heuristics to select the unroll factor:
4519 // 1. If the code has reductions, then we unroll in order to break the cross
4520 // iteration dependency.
4521 // 2. If the loop is really small, then we unroll in order to reduce the loop
4523 // 3. We don't unroll if we think that we will spill registers to memory due
4524 // to the increased register pressure.
4526 // Use the user preference, unless 'auto' is selected.
4527 int UserUF = Hints->getInterleave();
4531 // When we optimize for size, we don't unroll.
4535 // We used the distance for the unroll factor.
4536 if (Legal->getMaxSafeDepDistBytes() != -1U)
4539 // Do not unroll loops with a relatively small trip count.
4540 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4541 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
4544 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
4545 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
4549 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
4550 TargetNumRegisters = ForceTargetNumScalarRegs;
4552 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
4553 TargetNumRegisters = ForceTargetNumVectorRegs;
4556 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4557 // We divide by these constants so assume that we have at least one
4558 // instruction that uses at least one register.
4559 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4560 R.NumInstructions = std::max(R.NumInstructions, 1U);
4562 // We calculate the unroll factor using the following formula.
4563 // Subtract the number of loop invariants from the number of available
4564 // registers. These registers are used by all of the unrolled instances.
4565 // Next, divide the remaining registers by the number of registers that is
4566 // required by the loop, in order to estimate how many parallel instances
4567 // fit without causing spills. All of this is rounded down if necessary to be
4568 // a power of two. We want power of two unroll factors to simplify any
4569 // addressing operations or alignment considerations.
4570 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
4573 // Don't count the induction variable as unrolled.
4574 if (EnableIndVarRegisterHeur)
4575 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
4576 std::max(1U, (R.MaxLocalUsers - 1)));
4578 // Clamp the unroll factor ranges to reasonable factors.
4579 unsigned MaxInterleaveSize = TTI.getMaxInterleaveFactor();
4581 // Check if the user has overridden the unroll max.
4583 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
4584 MaxInterleaveSize = ForceTargetMaxScalarInterleaveFactor;
4586 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
4587 MaxInterleaveSize = ForceTargetMaxVectorInterleaveFactor;
4590 // If we did not calculate the cost for VF (because the user selected the VF)
4591 // then we calculate the cost of VF here.
4593 LoopCost = expectedCost(VF);
4595 // Clamp the calculated UF to be between the 1 and the max unroll factor
4596 // that the target allows.
4597 if (UF > MaxInterleaveSize)
4598 UF = MaxInterleaveSize;
4602 // Unroll if we vectorized this loop and there is a reduction that could
4603 // benefit from unrolling.
4604 if (VF > 1 && Legal->getReductionVars()->size()) {
4605 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
4609 // Note that if we've already vectorized the loop we will have done the
4610 // runtime check and so unrolling won't require further checks.
4611 bool UnrollingRequiresRuntimePointerCheck =
4612 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
4614 // We want to unroll small loops in order to reduce the loop overhead and
4615 // potentially expose ILP opportunities.
4616 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
4617 if (!UnrollingRequiresRuntimePointerCheck &&
4618 LoopCost < SmallLoopCost) {
4619 // We assume that the cost overhead is 1 and we use the cost model
4620 // to estimate the cost of the loop and unroll until the cost of the
4621 // loop overhead is about 5% of the cost of the loop.
4622 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
4624 // Unroll until store/load ports (estimated by max unroll factor) are
4626 unsigned NumStores = Legal->getNumStores();
4627 unsigned NumLoads = Legal->getNumLoads();
4628 unsigned StoresUF = UF / (NumStores ? NumStores : 1);
4629 unsigned LoadsUF = UF / (NumLoads ? NumLoads : 1);
4631 // If we have a scalar reduction (vector reductions are already dealt with
4632 // by this point), we can increase the critical path length if the loop
4633 // we're unrolling is inside another loop. Limit, by default to 2, so the
4634 // critical path only gets increased by one reduction operation.
4635 if (Legal->getReductionVars()->size() &&
4636 TheLoop->getLoopDepth() > 1) {
4637 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionUF);
4638 SmallUF = std::min(SmallUF, F);
4639 StoresUF = std::min(StoresUF, F);
4640 LoadsUF = std::min(LoadsUF, F);
4643 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
4644 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
4645 return std::max(StoresUF, LoadsUF);
4648 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
4652 DEBUG(dbgs() << "LV: Not Unrolling.\n");
4656 LoopVectorizationCostModel::RegisterUsage
4657 LoopVectorizationCostModel::calculateRegisterUsage() {
4658 // This function calculates the register usage by measuring the highest number
4659 // of values that are alive at a single location. Obviously, this is a very
4660 // rough estimation. We scan the loop in a topological order in order and
4661 // assign a number to each instruction. We use RPO to ensure that defs are
4662 // met before their users. We assume that each instruction that has in-loop
4663 // users starts an interval. We record every time that an in-loop value is
4664 // used, so we have a list of the first and last occurrences of each
4665 // instruction. Next, we transpose this data structure into a multi map that
4666 // holds the list of intervals that *end* at a specific location. This multi
4667 // map allows us to perform a linear search. We scan the instructions linearly
4668 // and record each time that a new interval starts, by placing it in a set.
4669 // If we find this value in the multi-map then we remove it from the set.
4670 // The max register usage is the maximum size of the set.
4671 // We also search for instructions that are defined outside the loop, but are
4672 // used inside the loop. We need this number separately from the max-interval
4673 // usage number because when we unroll, loop-invariant values do not take
4675 LoopBlocksDFS DFS(TheLoop);
4679 R.NumInstructions = 0;
4681 // Each 'key' in the map opens a new interval. The values
4682 // of the map are the index of the 'last seen' usage of the
4683 // instruction that is the key.
4684 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4685 // Maps instruction to its index.
4686 DenseMap<unsigned, Instruction*> IdxToInstr;
4687 // Marks the end of each interval.
4688 IntervalMap EndPoint;
4689 // Saves the list of instruction indices that are used in the loop.
4690 SmallSet<Instruction*, 8> Ends;
4691 // Saves the list of values that are used in the loop but are
4692 // defined outside the loop, such as arguments and constants.
4693 SmallPtrSet<Value*, 8> LoopInvariants;
4696 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4697 be = DFS.endRPO(); bb != be; ++bb) {
4698 R.NumInstructions += (*bb)->size();
4699 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4701 Instruction *I = it;
4702 IdxToInstr[Index++] = I;
4704 // Save the end location of each USE.
4705 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4706 Value *U = I->getOperand(i);
4707 Instruction *Instr = dyn_cast<Instruction>(U);
4709 // Ignore non-instruction values such as arguments, constants, etc.
4710 if (!Instr) continue;
4712 // If this instruction is outside the loop then record it and continue.
4713 if (!TheLoop->contains(Instr)) {
4714 LoopInvariants.insert(Instr);
4718 // Overwrite previous end points.
4719 EndPoint[Instr] = Index;
4725 // Saves the list of intervals that end with the index in 'key'.
4726 typedef SmallVector<Instruction*, 2> InstrList;
4727 DenseMap<unsigned, InstrList> TransposeEnds;
4729 // Transpose the EndPoints to a list of values that end at each index.
4730 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4732 TransposeEnds[it->second].push_back(it->first);
4734 SmallSet<Instruction*, 8> OpenIntervals;
4735 unsigned MaxUsage = 0;
4738 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4739 for (unsigned int i = 0; i < Index; ++i) {
4740 Instruction *I = IdxToInstr[i];
4741 // Ignore instructions that are never used within the loop.
4742 if (!Ends.count(I)) continue;
4744 // Ignore ephemeral values.
4745 if (EphValues.count(I))
4748 // Remove all of the instructions that end at this location.
4749 InstrList &List = TransposeEnds[i];
4750 for (unsigned int j=0, e = List.size(); j < e; ++j)
4751 OpenIntervals.erase(List[j]);
4753 // Count the number of live interals.
4754 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4756 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4757 OpenIntervals.size() << '\n');
4759 // Add the current instruction to the list of open intervals.
4760 OpenIntervals.insert(I);
4763 unsigned Invariant = LoopInvariants.size();
4764 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
4765 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
4766 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
4768 R.LoopInvariantRegs = Invariant;
4769 R.MaxLocalUsers = MaxUsage;
4773 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4777 for (Loop::block_iterator bb = TheLoop->block_begin(),
4778 be = TheLoop->block_end(); bb != be; ++bb) {
4779 unsigned BlockCost = 0;
4780 BasicBlock *BB = *bb;
4782 // For each instruction in the old loop.
4783 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4784 // Skip dbg intrinsics.
4785 if (isa<DbgInfoIntrinsic>(it))
4788 // Ignore ephemeral values.
4789 if (EphValues.count(it))
4792 unsigned C = getInstructionCost(it, VF);
4794 // Check if we should override the cost.
4795 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
4796 C = ForceTargetInstructionCost;
4799 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
4800 VF << " For instruction: " << *it << '\n');
4803 // We assume that if-converted blocks have a 50% chance of being executed.
4804 // When the code is scalar then some of the blocks are avoided due to CF.
4805 // When the code is vectorized we execute all code paths.
4806 if (VF == 1 && Legal->blockNeedsPredication(*bb))
4815 /// \brief Check whether the address computation for a non-consecutive memory
4816 /// access looks like an unlikely candidate for being merged into the indexing
4819 /// We look for a GEP which has one index that is an induction variable and all
4820 /// other indices are loop invariant. If the stride of this access is also
4821 /// within a small bound we decide that this address computation can likely be
4822 /// merged into the addressing mode.
4823 /// In all other cases, we identify the address computation as complex.
4824 static bool isLikelyComplexAddressComputation(Value *Ptr,
4825 LoopVectorizationLegality *Legal,
4826 ScalarEvolution *SE,
4827 const Loop *TheLoop) {
4828 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
4832 // We are looking for a gep with all loop invariant indices except for one
4833 // which should be an induction variable.
4834 unsigned NumOperands = Gep->getNumOperands();
4835 for (unsigned i = 1; i < NumOperands; ++i) {
4836 Value *Opd = Gep->getOperand(i);
4837 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
4838 !Legal->isInductionVariable(Opd))
4842 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
4843 // can likely be merged into the address computation.
4844 unsigned MaxMergeDistance = 64;
4846 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
4850 // Check the step is constant.
4851 const SCEV *Step = AddRec->getStepRecurrence(*SE);
4852 // Calculate the pointer stride and check if it is consecutive.
4853 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4857 const APInt &APStepVal = C->getValue()->getValue();
4859 // Huge step value - give up.
4860 if (APStepVal.getBitWidth() > 64)
4863 int64_t StepVal = APStepVal.getSExtValue();
4865 return StepVal > MaxMergeDistance;
4868 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
4869 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
4875 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
4876 // If we know that this instruction will remain uniform, check the cost of
4877 // the scalar version.
4878 if (Legal->isUniformAfterVectorization(I))
4881 Type *RetTy = I->getType();
4882 Type *VectorTy = ToVectorTy(RetTy, VF);
4884 // TODO: We need to estimate the cost of intrinsic calls.
4885 switch (I->getOpcode()) {
4886 case Instruction::GetElementPtr:
4887 // We mark this instruction as zero-cost because the cost of GEPs in
4888 // vectorized code depends on whether the corresponding memory instruction
4889 // is scalarized or not. Therefore, we handle GEPs with the memory
4890 // instruction cost.
4892 case Instruction::Br: {
4893 return TTI.getCFInstrCost(I->getOpcode());
4895 case Instruction::PHI:
4896 //TODO: IF-converted IFs become selects.
4898 case Instruction::Add:
4899 case Instruction::FAdd:
4900 case Instruction::Sub:
4901 case Instruction::FSub:
4902 case Instruction::Mul:
4903 case Instruction::FMul:
4904 case Instruction::UDiv:
4905 case Instruction::SDiv:
4906 case Instruction::FDiv:
4907 case Instruction::URem:
4908 case Instruction::SRem:
4909 case Instruction::FRem:
4910 case Instruction::Shl:
4911 case Instruction::LShr:
4912 case Instruction::AShr:
4913 case Instruction::And:
4914 case Instruction::Or:
4915 case Instruction::Xor: {
4916 // Since we will replace the stride by 1 the multiplication should go away.
4917 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
4919 // Certain instructions can be cheaper to vectorize if they have a constant
4920 // second vector operand. One example of this are shifts on x86.
4921 TargetTransformInfo::OperandValueKind Op1VK =
4922 TargetTransformInfo::OK_AnyValue;
4923 TargetTransformInfo::OperandValueKind Op2VK =
4924 TargetTransformInfo::OK_AnyValue;
4925 TargetTransformInfo::OperandValueProperties Op1VP =
4926 TargetTransformInfo::OP_None;
4927 TargetTransformInfo::OperandValueProperties Op2VP =
4928 TargetTransformInfo::OP_None;
4929 Value *Op2 = I->getOperand(1);
4931 // Check for a splat of a constant or for a non uniform vector of constants.
4932 if (isa<ConstantInt>(Op2)) {
4933 ConstantInt *CInt = cast<ConstantInt>(Op2);
4934 if (CInt && CInt->getValue().isPowerOf2())
4935 Op2VP = TargetTransformInfo::OP_PowerOf2;
4936 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4937 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
4938 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
4939 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
4941 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
4942 if (CInt && CInt->getValue().isPowerOf2())
4943 Op2VP = TargetTransformInfo::OP_PowerOf2;
4944 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4948 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
4951 case Instruction::Select: {
4952 SelectInst *SI = cast<SelectInst>(I);
4953 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
4954 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
4955 Type *CondTy = SI->getCondition()->getType();
4957 CondTy = VectorType::get(CondTy, VF);
4959 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
4961 case Instruction::ICmp:
4962 case Instruction::FCmp: {
4963 Type *ValTy = I->getOperand(0)->getType();
4964 VectorTy = ToVectorTy(ValTy, VF);
4965 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
4967 case Instruction::Store:
4968 case Instruction::Load: {
4969 StoreInst *SI = dyn_cast<StoreInst>(I);
4970 LoadInst *LI = dyn_cast<LoadInst>(I);
4971 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
4973 VectorTy = ToVectorTy(ValTy, VF);
4975 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
4976 unsigned AS = SI ? SI->getPointerAddressSpace() :
4977 LI->getPointerAddressSpace();
4978 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
4979 // We add the cost of address computation here instead of with the gep
4980 // instruction because only here we know whether the operation is
4983 return TTI.getAddressComputationCost(VectorTy) +
4984 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4986 // Scalarized loads/stores.
4987 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
4988 bool Reverse = ConsecutiveStride < 0;
4989 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
4990 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
4991 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
4992 bool IsComplexComputation =
4993 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
4995 // The cost of extracting from the value vector and pointer vector.
4996 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
4997 for (unsigned i = 0; i < VF; ++i) {
4998 // The cost of extracting the pointer operand.
4999 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5000 // In case of STORE, the cost of ExtractElement from the vector.
5001 // In case of LOAD, the cost of InsertElement into the returned
5003 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5004 Instruction::InsertElement,
5008 // The cost of the scalar loads/stores.
5009 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5010 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5015 // Wide load/stores.
5016 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5017 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5020 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5024 case Instruction::ZExt:
5025 case Instruction::SExt:
5026 case Instruction::FPToUI:
5027 case Instruction::FPToSI:
5028 case Instruction::FPExt:
5029 case Instruction::PtrToInt:
5030 case Instruction::IntToPtr:
5031 case Instruction::SIToFP:
5032 case Instruction::UIToFP:
5033 case Instruction::Trunc:
5034 case Instruction::FPTrunc:
5035 case Instruction::BitCast: {
5036 // We optimize the truncation of induction variable.
5037 // The cost of these is the same as the scalar operation.
5038 if (I->getOpcode() == Instruction::Trunc &&
5039 Legal->isInductionVariable(I->getOperand(0)))
5040 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5041 I->getOperand(0)->getType());
5043 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5044 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5046 case Instruction::Call: {
5047 CallInst *CI = cast<CallInst>(I);
5048 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
5049 assert(ID && "Not an intrinsic call!");
5050 Type *RetTy = ToVectorTy(CI->getType(), VF);
5051 SmallVector<Type*, 4> Tys;
5052 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
5053 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
5054 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
5057 // We are scalarizing the instruction. Return the cost of the scalar
5058 // instruction, plus the cost of insert and extract into vector
5059 // elements, times the vector width.
5062 if (!RetTy->isVoidTy() && VF != 1) {
5063 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5065 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5068 // The cost of inserting the results plus extracting each one of the
5070 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5073 // The cost of executing VF copies of the scalar instruction. This opcode
5074 // is unknown. Assume that it is the same as 'mul'.
5075 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5081 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
5082 if (Scalar->isVoidTy() || VF == 1)
5084 return VectorType::get(Scalar, VF);
5087 char LoopVectorize::ID = 0;
5088 static const char lv_name[] = "Loop Vectorization";
5089 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5090 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
5091 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
5092 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
5093 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
5094 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5095 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5096 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5097 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
5098 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5099 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5102 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5103 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5107 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5108 // Check for a store.
5109 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5110 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5112 // Check for a load.
5113 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5114 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5120 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5121 bool IfPredicateStore) {
5122 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5123 // Holds vector parameters or scalars, in case of uniform vals.
5124 SmallVector<VectorParts, 4> Params;
5126 setDebugLocFromInst(Builder, Instr);
5128 // Find all of the vectorized parameters.
5129 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5130 Value *SrcOp = Instr->getOperand(op);
5132 // If we are accessing the old induction variable, use the new one.
5133 if (SrcOp == OldInduction) {
5134 Params.push_back(getVectorValue(SrcOp));
5138 // Try using previously calculated values.
5139 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5141 // If the src is an instruction that appeared earlier in the basic block
5142 // then it should already be vectorized.
5143 if (SrcInst && OrigLoop->contains(SrcInst)) {
5144 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5145 // The parameter is a vector value from earlier.
5146 Params.push_back(WidenMap.get(SrcInst));
5148 // The parameter is a scalar from outside the loop. Maybe even a constant.
5149 VectorParts Scalars;
5150 Scalars.append(UF, SrcOp);
5151 Params.push_back(Scalars);
5155 assert(Params.size() == Instr->getNumOperands() &&
5156 "Invalid number of operands");
5158 // Does this instruction return a value ?
5159 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5161 Value *UndefVec = IsVoidRetTy ? nullptr :
5162 UndefValue::get(Instr->getType());
5163 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5164 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5166 Instruction *InsertPt = Builder.GetInsertPoint();
5167 BasicBlock *IfBlock = Builder.GetInsertBlock();
5168 BasicBlock *CondBlock = nullptr;
5171 Loop *VectorLp = nullptr;
5172 if (IfPredicateStore) {
5173 assert(Instr->getParent()->getSinglePredecessor() &&
5174 "Only support single predecessor blocks");
5175 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5176 Instr->getParent());
5177 VectorLp = LI->getLoopFor(IfBlock);
5178 assert(VectorLp && "Must have a loop for this block");
5181 // For each vector unroll 'part':
5182 for (unsigned Part = 0; Part < UF; ++Part) {
5183 // For each scalar that we create:
5185 // Start an "if (pred) a[i] = ..." block.
5186 Value *Cmp = nullptr;
5187 if (IfPredicateStore) {
5188 if (Cond[Part]->getType()->isVectorTy())
5190 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5191 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5192 ConstantInt::get(Cond[Part]->getType(), 1));
5193 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5194 LoopVectorBody.push_back(CondBlock);
5195 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
5196 // Update Builder with newly created basic block.
5197 Builder.SetInsertPoint(InsertPt);
5200 Instruction *Cloned = Instr->clone();
5202 Cloned->setName(Instr->getName() + ".cloned");
5203 // Replace the operands of the cloned instructions with extracted scalars.
5204 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5205 Value *Op = Params[op][Part];
5206 Cloned->setOperand(op, Op);
5209 // Place the cloned scalar in the new loop.
5210 Builder.Insert(Cloned);
5212 // If the original scalar returns a value we need to place it in a vector
5213 // so that future users will be able to use it.
5215 VecResults[Part] = Cloned;
5218 if (IfPredicateStore) {
5219 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
5220 LoopVectorBody.push_back(NewIfBlock);
5221 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
5222 Builder.SetInsertPoint(InsertPt);
5223 Instruction *OldBr = IfBlock->getTerminator();
5224 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
5225 OldBr->eraseFromParent();
5226 IfBlock = NewIfBlock;
5231 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5232 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5233 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5235 return scalarizeInstruction(Instr, IfPredicateStore);
5238 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5242 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5246 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) {
5247 // When unrolling and the VF is 1, we only need to add a simple scalar.
5248 Type *ITy = Val->getType();
5249 assert(!ITy->isVectorTy() && "Val must be a scalar");
5250 Constant *C = ConstantInt::get(ITy, StartIdx);
5251 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");