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 : 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 {
783 return LAA.getNumStores();
785 unsigned getNumLoads() const {
786 return LAA.getNumLoads();
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);
845 unsigned NumPredStores;
847 /// The loop that we evaluate.
851 /// DataLayout analysis.
852 const DataLayout *DL;
853 /// Target Library Info.
854 TargetLibraryInfo *TLI;
856 Function *TheFunction;
857 /// Target Transform Info
858 const TargetTransformInfo *TTI;
860 // --- vectorization state --- //
862 /// Holds the integer induction variable. This is the counter of the
865 /// Holds the reduction variables.
866 ReductionList Reductions;
867 /// Holds all of the induction variables that we found in the loop.
868 /// Notice that inductions don't need to start at zero and that induction
869 /// variables can be pointers.
870 InductionList Inductions;
871 /// Holds the widest induction type encountered.
874 /// Allowed outside users. This holds the reduction
875 /// vars which can be accessed from outside the loop.
876 SmallPtrSet<Value*, 4> AllowedExit;
877 /// This set holds the variables which are known to be uniform after
879 SmallPtrSet<Instruction*, 4> Uniforms;
880 LoopAccessAnalysis LAA;
881 /// Can we assume the absence of NaNs.
882 bool HasFunNoNaNAttr;
884 ValueToValueMap Strides;
885 SmallPtrSet<Value *, 8> StrideSet;
887 /// While vectorizing these instructions we have to generate a
888 /// call to the appropriate masked intrinsic
889 SmallPtrSet<const Instruction*, 8> MaskedOp;
892 /// LoopVectorizationCostModel - estimates the expected speedups due to
894 /// In many cases vectorization is not profitable. This can happen because of
895 /// a number of reasons. In this class we mainly attempt to predict the
896 /// expected speedup/slowdowns due to the supported instruction set. We use the
897 /// TargetTransformInfo to query the different backends for the cost of
898 /// different operations.
899 class LoopVectorizationCostModel {
901 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
902 LoopVectorizationLegality *Legal,
903 const TargetTransformInfo &TTI,
904 const DataLayout *DL, const TargetLibraryInfo *TLI,
905 AssumptionCache *AC, const Function *F,
906 const LoopVectorizeHints *Hints)
907 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI),
908 TheFunction(F), Hints(Hints) {
909 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
912 /// Information about vectorization costs
913 struct VectorizationFactor {
914 unsigned Width; // Vector width with best cost
915 unsigned Cost; // Cost of the loop with that width
917 /// \return The most profitable vectorization factor and the cost of that VF.
918 /// This method checks every power of two up to VF. If UserVF is not ZERO
919 /// then this vectorization factor will be selected if vectorization is
921 VectorizationFactor selectVectorizationFactor(bool OptForSize);
923 /// \return The size (in bits) of the widest type in the code that
924 /// needs to be vectorized. We ignore values that remain scalar such as
925 /// 64 bit loop indices.
926 unsigned getWidestType();
928 /// \return The most profitable unroll factor.
929 /// If UserUF is non-zero then this method finds the best unroll-factor
930 /// based on register pressure and other parameters.
931 /// VF and LoopCost are the selected vectorization factor and the cost of the
933 unsigned selectUnrollFactor(bool OptForSize, unsigned VF, unsigned LoopCost);
935 /// \brief A struct that represents some properties of the register usage
937 struct RegisterUsage {
938 /// Holds the number of loop invariant values that are used in the loop.
939 unsigned LoopInvariantRegs;
940 /// Holds the maximum number of concurrent live intervals in the loop.
941 unsigned MaxLocalUsers;
942 /// Holds the number of instructions in the loop.
943 unsigned NumInstructions;
946 /// \return information about the register usage of the loop.
947 RegisterUsage calculateRegisterUsage();
950 /// Returns the expected execution cost. The unit of the cost does
951 /// not matter because we use the 'cost' units to compare different
952 /// vector widths. The cost that is returned is *not* normalized by
953 /// the factor width.
954 unsigned expectedCost(unsigned VF);
956 /// Returns the execution time cost of an instruction for a given vector
957 /// width. Vector width of one means scalar.
958 unsigned getInstructionCost(Instruction *I, unsigned VF);
960 /// A helper function for converting Scalar types to vector types.
961 /// If the incoming type is void, we return void. If the VF is 1, we return
963 static Type* ToVectorTy(Type *Scalar, unsigned VF);
965 /// Returns whether the instruction is a load or store and will be a emitted
966 /// as a vector operation.
967 bool isConsecutiveLoadOrStore(Instruction *I);
969 /// Report an analysis message to assist the user in diagnosing loops that are
971 void emitAnalysis(VectorizationReport &Message) {
972 VectorizationReport::emitAnalysis(Message, TheFunction, TheLoop);
975 /// Values used only by @llvm.assume calls.
976 SmallPtrSet<const Value *, 32> EphValues;
978 /// The loop that we evaluate.
982 /// Loop Info analysis.
984 /// Vectorization legality.
985 LoopVectorizationLegality *Legal;
986 /// Vector target information.
987 const TargetTransformInfo &TTI;
988 /// Target data layout information.
989 const DataLayout *DL;
990 /// Target Library Info.
991 const TargetLibraryInfo *TLI;
992 const Function *TheFunction;
993 // Loop Vectorize Hint.
994 const LoopVectorizeHints *Hints;
997 /// Utility class for getting and setting loop vectorizer hints in the form
998 /// of loop metadata.
999 /// This class keeps a number of loop annotations locally (as member variables)
1000 /// and can, upon request, write them back as metadata on the loop. It will
1001 /// initially scan the loop for existing metadata, and will update the local
1002 /// values based on information in the loop.
1003 /// We cannot write all values to metadata, as the mere presence of some info,
1004 /// for example 'force', means a decision has been made. So, we need to be
1005 /// careful NOT to add them if the user hasn't specifically asked so.
1006 class LoopVectorizeHints {
1013 /// Hint - associates name and validation with the hint value.
1016 unsigned Value; // This may have to change for non-numeric values.
1019 Hint(const char * Name, unsigned Value, HintKind Kind)
1020 : Name(Name), Value(Value), Kind(Kind) { }
1022 bool validate(unsigned Val) {
1025 return isPowerOf2_32(Val) && Val <= MaxVectorWidth;
1027 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1035 /// Vectorization width.
1037 /// Vectorization interleave factor.
1039 /// Vectorization forced
1042 /// Return the loop metadata prefix.
1043 static StringRef Prefix() { return "llvm.loop."; }
1047 FK_Undefined = -1, ///< Not selected.
1048 FK_Disabled = 0, ///< Forcing disabled.
1049 FK_Enabled = 1, ///< Forcing enabled.
1052 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
1053 : Width("vectorize.width", VectorizationFactor, HK_WIDTH),
1054 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1055 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1057 // Populate values with existing loop metadata.
1058 getHintsFromMetadata();
1060 // force-vector-interleave overrides DisableInterleaving.
1061 if (VectorizationInterleave.getNumOccurrences() > 0)
1062 Interleave.Value = VectorizationInterleave;
1064 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
1065 << "LV: Interleaving disabled by the pass manager\n");
1068 /// Mark the loop L as already vectorized by setting the width to 1.
1069 void setAlreadyVectorized() {
1070 Width.Value = Interleave.Value = 1;
1071 Hint Hints[] = {Width, Interleave};
1072 writeHintsToMetadata(Hints);
1075 /// Dumps all the hint information.
1076 std::string emitRemark() const {
1077 VectorizationReport R;
1078 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1079 R << "vectorization is explicitly disabled";
1081 R << "use -Rpass-analysis=loop-vectorize for more info";
1082 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1083 R << " (Force=true";
1084 if (Width.Value != 0)
1085 R << ", Vector Width=" << Width.Value;
1086 if (Interleave.Value != 0)
1087 R << ", Interleave Count=" << Interleave.Value;
1095 unsigned getWidth() const { return Width.Value; }
1096 unsigned getInterleave() const { return Interleave.Value; }
1097 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1100 /// Find hints specified in the loop metadata and update local values.
1101 void getHintsFromMetadata() {
1102 MDNode *LoopID = TheLoop->getLoopID();
1106 // First operand should refer to the loop id itself.
1107 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1108 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1110 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1111 const MDString *S = nullptr;
1112 SmallVector<Metadata *, 4> Args;
1114 // The expected hint is either a MDString or a MDNode with the first
1115 // operand a MDString.
1116 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1117 if (!MD || MD->getNumOperands() == 0)
1119 S = dyn_cast<MDString>(MD->getOperand(0));
1120 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1121 Args.push_back(MD->getOperand(i));
1123 S = dyn_cast<MDString>(LoopID->getOperand(i));
1124 assert(Args.size() == 0 && "too many arguments for MDString");
1130 // Check if the hint starts with the loop metadata prefix.
1131 StringRef Name = S->getString();
1132 if (Args.size() == 1)
1133 setHint(Name, Args[0]);
1137 /// Checks string hint with one operand and set value if valid.
1138 void setHint(StringRef Name, Metadata *Arg) {
1139 if (!Name.startswith(Prefix()))
1141 Name = Name.substr(Prefix().size(), StringRef::npos);
1143 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1145 unsigned Val = C->getZExtValue();
1147 Hint *Hints[] = {&Width, &Interleave, &Force};
1148 for (auto H : Hints) {
1149 if (Name == H->Name) {
1150 if (H->validate(Val))
1153 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1159 /// Create a new hint from name / value pair.
1160 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1161 LLVMContext &Context = TheLoop->getHeader()->getContext();
1162 Metadata *MDs[] = {MDString::get(Context, Name),
1163 ConstantAsMetadata::get(
1164 ConstantInt::get(Type::getInt32Ty(Context), V))};
1165 return MDNode::get(Context, MDs);
1168 /// Matches metadata with hint name.
1169 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1170 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1174 for (auto H : HintTypes)
1175 if (Name->getString().endswith(H.Name))
1180 /// Sets current hints into loop metadata, keeping other values intact.
1181 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1182 if (HintTypes.size() == 0)
1185 // Reserve the first element to LoopID (see below).
1186 SmallVector<Metadata *, 4> MDs(1);
1187 // If the loop already has metadata, then ignore the existing operands.
1188 MDNode *LoopID = TheLoop->getLoopID();
1190 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1191 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1192 // If node in update list, ignore old value.
1193 if (!matchesHintMetadataName(Node, HintTypes))
1194 MDs.push_back(Node);
1198 // Now, add the missing hints.
1199 for (auto H : HintTypes)
1200 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1202 // Replace current metadata node with new one.
1203 LLVMContext &Context = TheLoop->getHeader()->getContext();
1204 MDNode *NewLoopID = MDNode::get(Context, MDs);
1205 // Set operand 0 to refer to the loop id itself.
1206 NewLoopID->replaceOperandWith(0, NewLoopID);
1208 TheLoop->setLoopID(NewLoopID);
1211 /// The loop these hints belong to.
1212 const Loop *TheLoop;
1215 static void emitMissedWarning(Function *F, Loop *L,
1216 const LoopVectorizeHints &LH) {
1217 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1218 L->getStartLoc(), LH.emitRemark());
1220 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1221 if (LH.getWidth() != 1)
1222 emitLoopVectorizeWarning(
1223 F->getContext(), *F, L->getStartLoc(),
1224 "failed explicitly specified loop vectorization");
1225 else if (LH.getInterleave() != 1)
1226 emitLoopInterleaveWarning(
1227 F->getContext(), *F, L->getStartLoc(),
1228 "failed explicitly specified loop interleaving");
1232 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1234 return V.push_back(&L);
1236 for (Loop *InnerL : L)
1237 addInnerLoop(*InnerL, V);
1240 /// The LoopVectorize Pass.
1241 struct LoopVectorize : public FunctionPass {
1242 /// Pass identification, replacement for typeid
1245 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1247 DisableUnrolling(NoUnrolling),
1248 AlwaysVectorize(AlwaysVectorize) {
1249 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1252 ScalarEvolution *SE;
1253 const DataLayout *DL;
1255 TargetTransformInfo *TTI;
1257 BlockFrequencyInfo *BFI;
1258 TargetLibraryInfo *TLI;
1260 AssumptionCache *AC;
1261 bool DisableUnrolling;
1262 bool AlwaysVectorize;
1264 BlockFrequency ColdEntryFreq;
1266 bool runOnFunction(Function &F) override {
1267 SE = &getAnalysis<ScalarEvolution>();
1268 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1269 DL = DLP ? &DLP->getDataLayout() : nullptr;
1270 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1271 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1272 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1273 BFI = &getAnalysis<BlockFrequencyInfo>();
1274 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1275 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1276 AA = &getAnalysis<AliasAnalysis>();
1277 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1279 // Compute some weights outside of the loop over the loops. Compute this
1280 // using a BranchProbability to re-use its scaling math.
1281 const BranchProbability ColdProb(1, 5); // 20%
1282 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1284 // If the target claims to have no vector registers don't attempt
1286 if (!TTI->getNumberOfRegisters(true))
1290 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1291 << ": Missing data layout\n");
1295 // Build up a worklist of inner-loops to vectorize. This is necessary as
1296 // the act of vectorizing or partially unrolling a loop creates new loops
1297 // and can invalidate iterators across the loops.
1298 SmallVector<Loop *, 8> Worklist;
1301 addInnerLoop(*L, Worklist);
1303 LoopsAnalyzed += Worklist.size();
1305 // Now walk the identified inner loops.
1306 bool Changed = false;
1307 while (!Worklist.empty())
1308 Changed |= processLoop(Worklist.pop_back_val());
1310 // Process each loop nest in the function.
1314 bool processLoop(Loop *L) {
1315 assert(L->empty() && "Only process inner loops.");
1318 const std::string DebugLocStr = getDebugLocString(L);
1321 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1322 << L->getHeader()->getParent()->getName() << "\" from "
1323 << DebugLocStr << "\n");
1325 LoopVectorizeHints Hints(L, DisableUnrolling);
1327 DEBUG(dbgs() << "LV: Loop hints:"
1329 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1331 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1333 : "?")) << " width=" << Hints.getWidth()
1334 << " unroll=" << Hints.getInterleave() << "\n");
1336 // Function containing loop
1337 Function *F = L->getHeader()->getParent();
1339 // Looking at the diagnostic output is the only way to determine if a loop
1340 // was vectorized (other than looking at the IR or machine code), so it
1341 // is important to generate an optimization remark for each loop. Most of
1342 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1343 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1344 // less verbose reporting vectorized loops and unvectorized loops that may
1345 // benefit from vectorization, respectively.
1347 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1348 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1349 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1350 L->getStartLoc(), Hints.emitRemark());
1354 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1355 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1356 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1357 L->getStartLoc(), Hints.emitRemark());
1361 if (Hints.getWidth() == 1 && Hints.getInterleave() == 1) {
1362 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1363 emitOptimizationRemarkAnalysis(
1364 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1365 "loop not vectorized: vector width and interleave count are "
1366 "explicitly set to 1");
1370 // Check the loop for a trip count threshold:
1371 // do not vectorize loops with a tiny trip count.
1372 const unsigned TC = SE->getSmallConstantTripCount(L);
1373 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1374 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1375 << "This loop is not worth vectorizing.");
1376 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1377 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1379 DEBUG(dbgs() << "\n");
1380 emitOptimizationRemarkAnalysis(
1381 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1382 "vectorization is not beneficial and is not explicitly forced");
1387 // Check if it is legal to vectorize the loop.
1388 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, AA, F, TTI);
1389 if (!LVL.canVectorize()) {
1390 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1391 emitMissedWarning(F, L, Hints);
1395 // Use the cost model.
1396 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI, AC, F,
1399 // Check the function attributes to find out if this function should be
1400 // optimized for size.
1401 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1402 F->hasFnAttribute(Attribute::OptimizeForSize);
1404 // Compute the weighted frequency of this loop being executed and see if it
1405 // is less than 20% of the function entry baseline frequency. Note that we
1406 // always have a canonical loop here because we think we *can* vectoriez.
1407 // FIXME: This is hidden behind a flag due to pervasive problems with
1408 // exactly what block frequency models.
1409 if (LoopVectorizeWithBlockFrequency) {
1410 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1411 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1412 LoopEntryFreq < ColdEntryFreq)
1416 // Check the function attributes to see if implicit floats are allowed.a
1417 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1418 // an integer loop and the vector instructions selected are purely integer
1419 // vector instructions?
1420 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1421 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1422 "attribute is used.\n");
1423 emitOptimizationRemarkAnalysis(
1424 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1425 "loop not vectorized due to NoImplicitFloat attribute");
1426 emitMissedWarning(F, L, Hints);
1430 // Select the optimal vectorization factor.
1431 const LoopVectorizationCostModel::VectorizationFactor VF =
1432 CM.selectVectorizationFactor(OptForSize);
1434 // Select the unroll factor.
1436 CM.selectUnrollFactor(OptForSize, VF.Width, VF.Cost);
1438 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1439 << DebugLocStr << '\n');
1440 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1442 if (VF.Width == 1) {
1443 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
1446 emitOptimizationRemarkAnalysis(
1447 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1448 "not beneficial to vectorize and user disabled interleaving");
1451 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1453 // Report the unrolling decision.
1454 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1455 Twine("unrolled with interleaving factor " +
1457 " (vectorization not beneficial)"));
1459 // We decided not to vectorize, but we may want to unroll.
1461 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1462 Unroller.vectorize(&LVL);
1464 // If we decided that it is *legal* to vectorize the loop then do it.
1465 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1469 // Report the vectorization decision.
1470 emitOptimizationRemark(
1471 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1472 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1473 ", unrolling interleave factor: " + Twine(UF) + ")");
1476 // Mark the loop as already vectorized to avoid vectorizing again.
1477 Hints.setAlreadyVectorized();
1479 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1483 void getAnalysisUsage(AnalysisUsage &AU) const override {
1484 AU.addRequired<AssumptionCacheTracker>();
1485 AU.addRequiredID(LoopSimplifyID);
1486 AU.addRequiredID(LCSSAID);
1487 AU.addRequired<BlockFrequencyInfo>();
1488 AU.addRequired<DominatorTreeWrapperPass>();
1489 AU.addRequired<LoopInfoWrapperPass>();
1490 AU.addRequired<ScalarEvolution>();
1491 AU.addRequired<TargetTransformInfoWrapperPass>();
1492 AU.addRequired<AliasAnalysis>();
1493 AU.addPreserved<LoopInfoWrapperPass>();
1494 AU.addPreserved<DominatorTreeWrapperPass>();
1495 AU.addPreserved<AliasAnalysis>();
1500 } // end anonymous namespace
1502 //===----------------------------------------------------------------------===//
1503 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1504 // LoopVectorizationCostModel.
1505 //===----------------------------------------------------------------------===//
1507 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1508 // We need to place the broadcast of invariant variables outside the loop.
1509 Instruction *Instr = dyn_cast<Instruction>(V);
1511 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1512 Instr->getParent()) != LoopVectorBody.end());
1513 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1515 // Place the code for broadcasting invariant variables in the new preheader.
1516 IRBuilder<>::InsertPointGuard Guard(Builder);
1518 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1520 // Broadcast the scalar into all locations in the vector.
1521 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1526 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
1528 assert(Val->getType()->isVectorTy() && "Must be a vector");
1529 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1530 "Elem must be an integer");
1531 assert(Step->getType() == Val->getType()->getScalarType() &&
1532 "Step has wrong type");
1533 // Create the types.
1534 Type *ITy = Val->getType()->getScalarType();
1535 VectorType *Ty = cast<VectorType>(Val->getType());
1536 int VLen = Ty->getNumElements();
1537 SmallVector<Constant*, 8> Indices;
1539 // Create a vector of consecutive numbers from zero to VF.
1540 for (int i = 0; i < VLen; ++i)
1541 Indices.push_back(ConstantInt::get(ITy, StartIdx + i));
1543 // Add the consecutive indices to the vector value.
1544 Constant *Cv = ConstantVector::get(Indices);
1545 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1546 Step = Builder.CreateVectorSplat(VLen, Step);
1547 assert(Step->getType() == Val->getType() && "Invalid step vec");
1548 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
1549 // which can be found from the original scalar operations.
1550 Step = Builder.CreateMul(Cv, Step);
1551 return Builder.CreateAdd(Val, Step, "induction");
1554 /// \brief Find the operand of the GEP that should be checked for consecutive
1555 /// stores. This ignores trailing indices that have no effect on the final
1557 static unsigned getGEPInductionOperand(const DataLayout *DL,
1558 const GetElementPtrInst *Gep) {
1559 unsigned LastOperand = Gep->getNumOperands() - 1;
1560 unsigned GEPAllocSize = DL->getTypeAllocSize(
1561 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1563 // Walk backwards and try to peel off zeros.
1564 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1565 // Find the type we're currently indexing into.
1566 gep_type_iterator GEPTI = gep_type_begin(Gep);
1567 std::advance(GEPTI, LastOperand - 1);
1569 // If it's a type with the same allocation size as the result of the GEP we
1570 // can peel off the zero index.
1571 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1579 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1580 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1581 // Make sure that the pointer does not point to structs.
1582 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1585 // If this value is a pointer induction variable we know it is consecutive.
1586 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1587 if (Phi && Inductions.count(Phi)) {
1588 InductionInfo II = Inductions[Phi];
1589 return II.getConsecutiveDirection();
1592 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1596 unsigned NumOperands = Gep->getNumOperands();
1597 Value *GpPtr = Gep->getPointerOperand();
1598 // If this GEP value is a consecutive pointer induction variable and all of
1599 // the indices are constant then we know it is consecutive. We can
1600 Phi = dyn_cast<PHINode>(GpPtr);
1601 if (Phi && Inductions.count(Phi)) {
1603 // Make sure that the pointer does not point to structs.
1604 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1605 if (GepPtrType->getElementType()->isAggregateType())
1608 // Make sure that all of the index operands are loop invariant.
1609 for (unsigned i = 1; i < NumOperands; ++i)
1610 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1613 InductionInfo II = Inductions[Phi];
1614 return II.getConsecutiveDirection();
1617 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1619 // Check that all of the gep indices are uniform except for our induction
1621 for (unsigned i = 0; i != NumOperands; ++i)
1622 if (i != InductionOperand &&
1623 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1626 // We can emit wide load/stores only if the last non-zero index is the
1627 // induction variable.
1628 const SCEV *Last = nullptr;
1629 if (!Strides.count(Gep))
1630 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1632 // Because of the multiplication by a stride we can have a s/zext cast.
1633 // We are going to replace this stride by 1 so the cast is safe to ignore.
1635 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1636 // %0 = trunc i64 %indvars.iv to i32
1637 // %mul = mul i32 %0, %Stride1
1638 // %idxprom = zext i32 %mul to i64 << Safe cast.
1639 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1641 Last = replaceSymbolicStrideSCEV(SE, Strides,
1642 Gep->getOperand(InductionOperand), Gep);
1643 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1645 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1649 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1650 const SCEV *Step = AR->getStepRecurrence(*SE);
1652 // The memory is consecutive because the last index is consecutive
1653 // and all other indices are loop invariant.
1656 if (Step->isAllOnesValue())
1663 bool LoopVectorizationLegality::isUniform(Value *V) {
1664 return LAA.isUniform(V);
1667 InnerLoopVectorizer::VectorParts&
1668 InnerLoopVectorizer::getVectorValue(Value *V) {
1669 assert(V != Induction && "The new induction variable should not be used.");
1670 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1672 // If we have a stride that is replaced by one, do it here.
1673 if (Legal->hasStride(V))
1674 V = ConstantInt::get(V->getType(), 1);
1676 // If we have this scalar in the map, return it.
1677 if (WidenMap.has(V))
1678 return WidenMap.get(V);
1680 // If this scalar is unknown, assume that it is a constant or that it is
1681 // loop invariant. Broadcast V and save the value for future uses.
1682 Value *B = getBroadcastInstrs(V);
1683 return WidenMap.splat(V, B);
1686 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1687 assert(Vec->getType()->isVectorTy() && "Invalid type");
1688 SmallVector<Constant*, 8> ShuffleMask;
1689 for (unsigned i = 0; i < VF; ++i)
1690 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1692 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1693 ConstantVector::get(ShuffleMask),
1697 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1698 // Attempt to issue a wide load.
1699 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1700 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1702 assert((LI || SI) && "Invalid Load/Store instruction");
1704 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1705 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1706 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1707 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1708 // An alignment of 0 means target abi alignment. We need to use the scalar's
1709 // target abi alignment in such a case.
1711 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1712 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1713 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1714 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1716 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
1717 !Legal->isMaskRequired(SI))
1718 return scalarizeInstruction(Instr, true);
1720 if (ScalarAllocatedSize != VectorElementSize)
1721 return scalarizeInstruction(Instr);
1723 // If the pointer is loop invariant or if it is non-consecutive,
1724 // scalarize the load.
1725 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1726 bool Reverse = ConsecutiveStride < 0;
1727 bool UniformLoad = LI && Legal->isUniform(Ptr);
1728 if (!ConsecutiveStride || UniformLoad)
1729 return scalarizeInstruction(Instr);
1731 Constant *Zero = Builder.getInt32(0);
1732 VectorParts &Entry = WidenMap.get(Instr);
1734 // Handle consecutive loads/stores.
1735 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1736 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1737 setDebugLocFromInst(Builder, Gep);
1738 Value *PtrOperand = Gep->getPointerOperand();
1739 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1740 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1742 // Create the new GEP with the new induction variable.
1743 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1744 Gep2->setOperand(0, FirstBasePtr);
1745 Gep2->setName("gep.indvar.base");
1746 Ptr = Builder.Insert(Gep2);
1748 setDebugLocFromInst(Builder, Gep);
1749 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1750 OrigLoop) && "Base ptr must be invariant");
1752 // The last index does not have to be the induction. It can be
1753 // consecutive and be a function of the index. For example A[I+1];
1754 unsigned NumOperands = Gep->getNumOperands();
1755 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1756 // Create the new GEP with the new induction variable.
1757 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1759 for (unsigned i = 0; i < NumOperands; ++i) {
1760 Value *GepOperand = Gep->getOperand(i);
1761 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1763 // Update last index or loop invariant instruction anchored in loop.
1764 if (i == InductionOperand ||
1765 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1766 assert((i == InductionOperand ||
1767 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1768 "Must be last index or loop invariant");
1770 VectorParts &GEPParts = getVectorValue(GepOperand);
1771 Value *Index = GEPParts[0];
1772 Index = Builder.CreateExtractElement(Index, Zero);
1773 Gep2->setOperand(i, Index);
1774 Gep2->setName("gep.indvar.idx");
1777 Ptr = Builder.Insert(Gep2);
1779 // Use the induction element ptr.
1780 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1781 setDebugLocFromInst(Builder, Ptr);
1782 VectorParts &PtrVal = getVectorValue(Ptr);
1783 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1786 VectorParts Mask = createBlockInMask(Instr->getParent());
1789 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1790 "We do not allow storing to uniform addresses");
1791 setDebugLocFromInst(Builder, SI);
1792 // We don't want to update the value in the map as it might be used in
1793 // another expression. So don't use a reference type for "StoredVal".
1794 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1796 for (unsigned Part = 0; Part < UF; ++Part) {
1797 // Calculate the pointer for the specific unroll-part.
1798 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1801 // If we store to reverse consecutive memory locations then we need
1802 // to reverse the order of elements in the stored value.
1803 StoredVal[Part] = reverseVector(StoredVal[Part]);
1804 // If the address is consecutive but reversed, then the
1805 // wide store needs to start at the last vector element.
1806 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1807 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1808 Mask[Part] = reverseVector(Mask[Part]);
1811 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1812 DataTy->getPointerTo(AddressSpace));
1815 if (Legal->isMaskRequired(SI))
1816 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
1819 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
1820 propagateMetadata(NewSI, SI);
1826 assert(LI && "Must have a load instruction");
1827 setDebugLocFromInst(Builder, LI);
1828 for (unsigned Part = 0; Part < UF; ++Part) {
1829 // Calculate the pointer for the specific unroll-part.
1830 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1833 // If the address is consecutive but reversed, then the
1834 // wide load needs to start at the last vector element.
1835 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1836 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1837 Mask[Part] = reverseVector(Mask[Part]);
1841 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1842 DataTy->getPointerTo(AddressSpace));
1843 if (Legal->isMaskRequired(LI))
1844 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
1845 UndefValue::get(DataTy),
1846 "wide.masked.load");
1848 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
1849 propagateMetadata(NewLI, LI);
1850 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
1854 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1855 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1856 // Holds vector parameters or scalars, in case of uniform vals.
1857 SmallVector<VectorParts, 4> Params;
1859 setDebugLocFromInst(Builder, Instr);
1861 // Find all of the vectorized parameters.
1862 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1863 Value *SrcOp = Instr->getOperand(op);
1865 // If we are accessing the old induction variable, use the new one.
1866 if (SrcOp == OldInduction) {
1867 Params.push_back(getVectorValue(SrcOp));
1871 // Try using previously calculated values.
1872 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1874 // If the src is an instruction that appeared earlier in the basic block
1875 // then it should already be vectorized.
1876 if (SrcInst && OrigLoop->contains(SrcInst)) {
1877 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1878 // The parameter is a vector value from earlier.
1879 Params.push_back(WidenMap.get(SrcInst));
1881 // The parameter is a scalar from outside the loop. Maybe even a constant.
1882 VectorParts Scalars;
1883 Scalars.append(UF, SrcOp);
1884 Params.push_back(Scalars);
1888 assert(Params.size() == Instr->getNumOperands() &&
1889 "Invalid number of operands");
1891 // Does this instruction return a value ?
1892 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1894 Value *UndefVec = IsVoidRetTy ? nullptr :
1895 UndefValue::get(VectorType::get(Instr->getType(), VF));
1896 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1897 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1899 Instruction *InsertPt = Builder.GetInsertPoint();
1900 BasicBlock *IfBlock = Builder.GetInsertBlock();
1901 BasicBlock *CondBlock = nullptr;
1904 Loop *VectorLp = nullptr;
1905 if (IfPredicateStore) {
1906 assert(Instr->getParent()->getSinglePredecessor() &&
1907 "Only support single predecessor blocks");
1908 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1909 Instr->getParent());
1910 VectorLp = LI->getLoopFor(IfBlock);
1911 assert(VectorLp && "Must have a loop for this block");
1914 // For each vector unroll 'part':
1915 for (unsigned Part = 0; Part < UF; ++Part) {
1916 // For each scalar that we create:
1917 for (unsigned Width = 0; Width < VF; ++Width) {
1920 Value *Cmp = nullptr;
1921 if (IfPredicateStore) {
1922 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1923 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1924 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1925 LoopVectorBody.push_back(CondBlock);
1926 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
1927 // Update Builder with newly created basic block.
1928 Builder.SetInsertPoint(InsertPt);
1931 Instruction *Cloned = Instr->clone();
1933 Cloned->setName(Instr->getName() + ".cloned");
1934 // Replace the operands of the cloned instructions with extracted scalars.
1935 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1936 Value *Op = Params[op][Part];
1937 // Param is a vector. Need to extract the right lane.
1938 if (Op->getType()->isVectorTy())
1939 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1940 Cloned->setOperand(op, Op);
1943 // Place the cloned scalar in the new loop.
1944 Builder.Insert(Cloned);
1946 // If the original scalar returns a value we need to place it in a vector
1947 // so that future users will be able to use it.
1949 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1950 Builder.getInt32(Width));
1952 if (IfPredicateStore) {
1953 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1954 LoopVectorBody.push_back(NewIfBlock);
1955 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
1956 Builder.SetInsertPoint(InsertPt);
1957 Instruction *OldBr = IfBlock->getTerminator();
1958 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1959 OldBr->eraseFromParent();
1960 IfBlock = NewIfBlock;
1966 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1970 if (Instruction *I = dyn_cast<Instruction>(V))
1971 return I->getParent() == Loc->getParent() ? I : nullptr;
1975 std::pair<Instruction *, Instruction *>
1976 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
1977 Instruction *tnullptr = nullptr;
1978 if (!Legal->mustCheckStrides())
1979 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1981 IRBuilder<> ChkBuilder(Loc);
1984 Value *Check = nullptr;
1985 Instruction *FirstInst = nullptr;
1986 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
1987 SE = Legal->strides_end();
1989 Value *Ptr = stripIntegerCast(*SI);
1990 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
1992 // Store the first instruction we create.
1993 FirstInst = getFirstInst(FirstInst, C, Loc);
1995 Check = ChkBuilder.CreateOr(Check, C);
2000 // We have to do this trickery because the IRBuilder might fold the check to a
2001 // constant expression in which case there is no Instruction anchored in a
2003 LLVMContext &Ctx = Loc->getContext();
2004 Instruction *TheCheck =
2005 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2006 ChkBuilder.Insert(TheCheck, "stride.not.one");
2007 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2009 return std::make_pair(FirstInst, TheCheck);
2012 std::pair<Instruction *, Instruction *>
2013 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
2014 LoopAccessAnalysis::RuntimePointerCheck *PtrRtCheck =
2015 Legal->getRuntimePointerCheck();
2017 Instruction *tnullptr = nullptr;
2018 if (!PtrRtCheck->Need)
2019 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2021 unsigned NumPointers = PtrRtCheck->Pointers.size();
2022 SmallVector<TrackingVH<Value> , 2> Starts;
2023 SmallVector<TrackingVH<Value> , 2> Ends;
2025 LLVMContext &Ctx = Loc->getContext();
2026 SCEVExpander Exp(*SE, "induction");
2027 Instruction *FirstInst = nullptr;
2029 for (unsigned i = 0; i < NumPointers; ++i) {
2030 Value *Ptr = PtrRtCheck->Pointers[i];
2031 const SCEV *Sc = SE->getSCEV(Ptr);
2033 if (SE->isLoopInvariant(Sc, OrigLoop)) {
2034 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
2036 Starts.push_back(Ptr);
2037 Ends.push_back(Ptr);
2039 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
2040 unsigned AS = Ptr->getType()->getPointerAddressSpace();
2042 // Use this type for pointer arithmetic.
2043 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
2045 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
2046 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
2047 Starts.push_back(Start);
2048 Ends.push_back(End);
2052 IRBuilder<> ChkBuilder(Loc);
2053 // Our instructions might fold to a constant.
2054 Value *MemoryRuntimeCheck = nullptr;
2055 for (unsigned i = 0; i < NumPointers; ++i) {
2056 for (unsigned j = i+1; j < NumPointers; ++j) {
2057 // No need to check if two readonly pointers intersect.
2058 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
2061 // Only need to check pointers between two different dependency sets.
2062 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
2064 // Only need to check pointers in the same alias set.
2065 if (PtrRtCheck->AliasSetId[i] != PtrRtCheck->AliasSetId[j])
2068 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
2069 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
2071 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
2072 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
2073 "Trying to bounds check pointers with different address spaces");
2075 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
2076 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
2078 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
2079 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
2080 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
2081 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
2083 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
2084 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
2085 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
2086 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
2087 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
2088 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2089 if (MemoryRuntimeCheck) {
2090 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
2092 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2094 MemoryRuntimeCheck = IsConflict;
2098 // We have to do this trickery because the IRBuilder might fold the check to a
2099 // constant expression in which case there is no Instruction anchored in a
2101 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
2102 ConstantInt::getTrue(Ctx));
2103 ChkBuilder.Insert(Check, "memcheck.conflict");
2104 FirstInst = getFirstInst(FirstInst, Check, Loc);
2105 return std::make_pair(FirstInst, Check);
2108 void InnerLoopVectorizer::createEmptyLoop() {
2110 In this function we generate a new loop. The new loop will contain
2111 the vectorized instructions while the old loop will continue to run the
2114 [ ] <-- Back-edge taken count overflow check.
2117 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2120 || [ ] <-- vector pre header.
2124 || [ ]_| <-- vector loop.
2127 | >[ ] <--- middle-block.
2130 -|- >[ ] <--- new preheader.
2134 | [ ]_| <-- old scalar loop to handle remainder.
2137 >[ ] <-- exit block.
2141 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2142 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2143 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2144 assert(BypassBlock && "Invalid loop structure");
2145 assert(ExitBlock && "Must have an exit block");
2147 // Some loops have a single integer induction variable, while other loops
2148 // don't. One example is c++ iterators that often have multiple pointer
2149 // induction variables. In the code below we also support a case where we
2150 // don't have a single induction variable.
2151 OldInduction = Legal->getInduction();
2152 Type *IdxTy = Legal->getWidestInductionType();
2154 // Find the loop boundaries.
2155 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2156 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2158 // The exit count might have the type of i64 while the phi is i32. This can
2159 // happen if we have an induction variable that is sign extended before the
2160 // compare. The only way that we get a backedge taken count is that the
2161 // induction variable was signed and as such will not overflow. In such a case
2162 // truncation is legal.
2163 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2164 IdxTy->getPrimitiveSizeInBits())
2165 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2167 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2168 // Get the total trip count from the count by adding 1.
2169 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2170 SE->getConstant(BackedgeTakeCount->getType(), 1));
2172 // Expand the trip count and place the new instructions in the preheader.
2173 // Notice that the pre-header does not change, only the loop body.
2174 SCEVExpander Exp(*SE, "induction");
2176 // We need to test whether the backedge-taken count is uint##_max. Adding one
2177 // to it will cause overflow and an incorrect loop trip count in the vector
2178 // body. In case of overflow we want to directly jump to the scalar remainder
2180 Value *BackedgeCount =
2181 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2182 BypassBlock->getTerminator());
2183 if (BackedgeCount->getType()->isPointerTy())
2184 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2185 "backedge.ptrcnt.to.int",
2186 BypassBlock->getTerminator());
2187 Instruction *CheckBCOverflow =
2188 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2189 Constant::getAllOnesValue(BackedgeCount->getType()),
2190 "backedge.overflow", BypassBlock->getTerminator());
2192 // The loop index does not have to start at Zero. Find the original start
2193 // value from the induction PHI node. If we don't have an induction variable
2194 // then we know that it starts at zero.
2195 Builder.SetInsertPoint(BypassBlock->getTerminator());
2196 Value *StartIdx = ExtendedIdx = OldInduction ?
2197 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2199 ConstantInt::get(IdxTy, 0);
2201 // We need an instruction to anchor the overflow check on. StartIdx needs to
2202 // be defined before the overflow check branch. Because the scalar preheader
2203 // is going to merge the start index and so the overflow branch block needs to
2204 // contain a definition of the start index.
2205 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2206 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2207 BypassBlock->getTerminator());
2209 // Count holds the overall loop count (N).
2210 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2211 BypassBlock->getTerminator());
2213 LoopBypassBlocks.push_back(BypassBlock);
2215 // Split the single block loop into the two loop structure described above.
2216 BasicBlock *VectorPH =
2217 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2218 BasicBlock *VecBody =
2219 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2220 BasicBlock *MiddleBlock =
2221 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2222 BasicBlock *ScalarPH =
2223 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2225 // Create and register the new vector loop.
2226 Loop* Lp = new Loop();
2227 Loop *ParentLoop = OrigLoop->getParentLoop();
2229 // Insert the new loop into the loop nest and register the new basic blocks
2230 // before calling any utilities such as SCEV that require valid LoopInfo.
2232 ParentLoop->addChildLoop(Lp);
2233 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
2234 ParentLoop->addBasicBlockToLoop(VectorPH, *LI);
2235 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
2237 LI->addTopLevelLoop(Lp);
2239 Lp->addBasicBlockToLoop(VecBody, *LI);
2241 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2243 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2245 // Generate the induction variable.
2246 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2247 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2248 // The loop step is equal to the vectorization factor (num of SIMD elements)
2249 // times the unroll factor (num of SIMD instructions).
2250 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2252 // This is the IR builder that we use to add all of the logic for bypassing
2253 // the new vector loop.
2254 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2255 setDebugLocFromInst(BypassBuilder,
2256 getDebugLocFromInstOrOperands(OldInduction));
2258 // We may need to extend the index in case there is a type mismatch.
2259 // We know that the count starts at zero and does not overflow.
2260 if (Count->getType() != IdxTy) {
2261 // The exit count can be of pointer type. Convert it to the correct
2263 if (ExitCount->getType()->isPointerTy())
2264 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2266 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2269 // Add the start index to the loop count to get the new end index.
2270 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2272 // Now we need to generate the expression for N - (N % VF), which is
2273 // the part that the vectorized body will execute.
2274 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2275 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2276 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2277 "end.idx.rnd.down");
2279 // Now, compare the new count to zero. If it is zero skip the vector loop and
2280 // jump to the scalar loop.
2282 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2284 BasicBlock *LastBypassBlock = BypassBlock;
2286 // Generate code to check that the loops trip count that we computed by adding
2287 // one to the backedge-taken count will not overflow.
2289 auto PastOverflowCheck =
2290 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2291 BasicBlock *CheckBlock =
2292 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2294 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2295 LoopBypassBlocks.push_back(CheckBlock);
2296 Instruction *OldTerm = LastBypassBlock->getTerminator();
2297 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2298 OldTerm->eraseFromParent();
2299 LastBypassBlock = CheckBlock;
2302 // Generate the code to check that the strides we assumed to be one are really
2303 // one. We want the new basic block to start at the first instruction in a
2304 // sequence of instructions that form a check.
2305 Instruction *StrideCheck;
2306 Instruction *FirstCheckInst;
2307 std::tie(FirstCheckInst, StrideCheck) =
2308 addStrideCheck(LastBypassBlock->getTerminator());
2310 // Create a new block containing the stride check.
2311 BasicBlock *CheckBlock =
2312 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2314 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2315 LoopBypassBlocks.push_back(CheckBlock);
2317 // Replace the branch into the memory check block with a conditional branch
2318 // for the "few elements case".
2319 Instruction *OldTerm = LastBypassBlock->getTerminator();
2320 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2321 OldTerm->eraseFromParent();
2324 LastBypassBlock = CheckBlock;
2327 // Generate the code that checks in runtime if arrays overlap. We put the
2328 // checks into a separate block to make the more common case of few elements
2330 Instruction *MemRuntimeCheck;
2331 std::tie(FirstCheckInst, MemRuntimeCheck) =
2332 addRuntimeCheck(LastBypassBlock->getTerminator());
2333 if (MemRuntimeCheck) {
2334 // Create a new block containing the memory check.
2335 BasicBlock *CheckBlock =
2336 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.memcheck");
2338 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2339 LoopBypassBlocks.push_back(CheckBlock);
2341 // Replace the branch into the memory check block with a conditional branch
2342 // for the "few elements case".
2343 Instruction *OldTerm = LastBypassBlock->getTerminator();
2344 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2345 OldTerm->eraseFromParent();
2347 Cmp = MemRuntimeCheck;
2348 LastBypassBlock = CheckBlock;
2351 LastBypassBlock->getTerminator()->eraseFromParent();
2352 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2355 // We are going to resume the execution of the scalar loop.
2356 // Go over all of the induction variables that we found and fix the
2357 // PHIs that are left in the scalar version of the loop.
2358 // The starting values of PHI nodes depend on the counter of the last
2359 // iteration in the vectorized loop.
2360 // If we come from a bypass edge then we need to start from the original
2363 // This variable saves the new starting index for the scalar loop.
2364 PHINode *ResumeIndex = nullptr;
2365 LoopVectorizationLegality::InductionList::iterator I, E;
2366 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2367 // Set builder to point to last bypass block.
2368 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2369 for (I = List->begin(), E = List->end(); I != E; ++I) {
2370 PHINode *OrigPhi = I->first;
2371 LoopVectorizationLegality::InductionInfo II = I->second;
2373 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2374 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2375 MiddleBlock->getTerminator());
2376 // We might have extended the type of the induction variable but we need a
2377 // truncated version for the scalar loop.
2378 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2379 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2380 MiddleBlock->getTerminator()) : nullptr;
2382 // Create phi nodes to merge from the backedge-taken check block.
2383 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2384 ScalarPH->getTerminator());
2385 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2387 PHINode *BCTruncResumeVal = nullptr;
2388 if (OrigPhi == OldInduction) {
2390 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2391 ScalarPH->getTerminator());
2392 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2395 Value *EndValue = nullptr;
2397 case LoopVectorizationLegality::IK_NoInduction:
2398 llvm_unreachable("Unknown induction");
2399 case LoopVectorizationLegality::IK_IntInduction: {
2400 // Handle the integer induction counter.
2401 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2403 // We have the canonical induction variable.
2404 if (OrigPhi == OldInduction) {
2405 // Create a truncated version of the resume value for the scalar loop,
2406 // we might have promoted the type to a larger width.
2408 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2409 // The new PHI merges the original incoming value, in case of a bypass,
2410 // or the value at the end of the vectorized loop.
2411 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2412 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2413 TruncResumeVal->addIncoming(EndValue, VecBody);
2415 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2417 // We know what the end value is.
2418 EndValue = IdxEndRoundDown;
2419 // We also know which PHI node holds it.
2420 ResumeIndex = ResumeVal;
2424 // Not the canonical induction variable - add the vector loop count to the
2426 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2427 II.StartValue->getType(),
2429 EndValue = II.transform(BypassBuilder, CRD);
2430 EndValue->setName("ind.end");
2433 case LoopVectorizationLegality::IK_PtrInduction: {
2434 EndValue = II.transform(BypassBuilder, CountRoundDown);
2435 EndValue->setName("ptr.ind.end");
2440 // The new PHI merges the original incoming value, in case of a bypass,
2441 // or the value at the end of the vectorized loop.
2442 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2443 if (OrigPhi == OldInduction)
2444 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2446 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2448 ResumeVal->addIncoming(EndValue, VecBody);
2450 // Fix the scalar body counter (PHI node).
2451 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2453 // The old induction's phi node in the scalar body needs the truncated
2455 if (OrigPhi == OldInduction) {
2456 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2457 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2459 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2460 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2464 // If we are generating a new induction variable then we also need to
2465 // generate the code that calculates the exit value. This value is not
2466 // simply the end of the counter because we may skip the vectorized body
2467 // in case of a runtime check.
2469 assert(!ResumeIndex && "Unexpected resume value found");
2470 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2471 MiddleBlock->getTerminator());
2472 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2473 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2474 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2477 // Make sure that we found the index where scalar loop needs to continue.
2478 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2479 "Invalid resume Index");
2481 // Add a check in the middle block to see if we have completed
2482 // all of the iterations in the first vector loop.
2483 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2484 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2485 ResumeIndex, "cmp.n",
2486 MiddleBlock->getTerminator());
2488 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2489 // Remove the old terminator.
2490 MiddleBlock->getTerminator()->eraseFromParent();
2492 // Create i+1 and fill the PHINode.
2493 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2494 Induction->addIncoming(StartIdx, VectorPH);
2495 Induction->addIncoming(NextIdx, VecBody);
2496 // Create the compare.
2497 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2498 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2500 // Now we have two terminators. Remove the old one from the block.
2501 VecBody->getTerminator()->eraseFromParent();
2503 // Get ready to start creating new instructions into the vectorized body.
2504 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2507 LoopVectorPreHeader = VectorPH;
2508 LoopScalarPreHeader = ScalarPH;
2509 LoopMiddleBlock = MiddleBlock;
2510 LoopExitBlock = ExitBlock;
2511 LoopVectorBody.push_back(VecBody);
2512 LoopScalarBody = OldBasicBlock;
2514 LoopVectorizeHints Hints(Lp, true);
2515 Hints.setAlreadyVectorized();
2518 /// This function returns the identity element (or neutral element) for
2519 /// the operation K.
2521 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2526 // Adding, Xoring, Oring zero to a number does not change it.
2527 return ConstantInt::get(Tp, 0);
2528 case RK_IntegerMult:
2529 // Multiplying a number by 1 does not change it.
2530 return ConstantInt::get(Tp, 1);
2532 // AND-ing a number with an all-1 value does not change it.
2533 return ConstantInt::get(Tp, -1, true);
2535 // Multiplying a number by 1 does not change it.
2536 return ConstantFP::get(Tp, 1.0L);
2538 // Adding zero to a number does not change it.
2539 return ConstantFP::get(Tp, 0.0L);
2541 llvm_unreachable("Unknown reduction kind");
2545 /// This function translates the reduction kind to an LLVM binary operator.
2547 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2549 case LoopVectorizationLegality::RK_IntegerAdd:
2550 return Instruction::Add;
2551 case LoopVectorizationLegality::RK_IntegerMult:
2552 return Instruction::Mul;
2553 case LoopVectorizationLegality::RK_IntegerOr:
2554 return Instruction::Or;
2555 case LoopVectorizationLegality::RK_IntegerAnd:
2556 return Instruction::And;
2557 case LoopVectorizationLegality::RK_IntegerXor:
2558 return Instruction::Xor;
2559 case LoopVectorizationLegality::RK_FloatMult:
2560 return Instruction::FMul;
2561 case LoopVectorizationLegality::RK_FloatAdd:
2562 return Instruction::FAdd;
2563 case LoopVectorizationLegality::RK_IntegerMinMax:
2564 return Instruction::ICmp;
2565 case LoopVectorizationLegality::RK_FloatMinMax:
2566 return Instruction::FCmp;
2568 llvm_unreachable("Unknown reduction operation");
2572 Value *createMinMaxOp(IRBuilder<> &Builder,
2573 LoopVectorizationLegality::MinMaxReductionKind RK,
2576 CmpInst::Predicate P = CmpInst::ICMP_NE;
2579 llvm_unreachable("Unknown min/max reduction kind");
2580 case LoopVectorizationLegality::MRK_UIntMin:
2581 P = CmpInst::ICMP_ULT;
2583 case LoopVectorizationLegality::MRK_UIntMax:
2584 P = CmpInst::ICMP_UGT;
2586 case LoopVectorizationLegality::MRK_SIntMin:
2587 P = CmpInst::ICMP_SLT;
2589 case LoopVectorizationLegality::MRK_SIntMax:
2590 P = CmpInst::ICMP_SGT;
2592 case LoopVectorizationLegality::MRK_FloatMin:
2593 P = CmpInst::FCMP_OLT;
2595 case LoopVectorizationLegality::MRK_FloatMax:
2596 P = CmpInst::FCMP_OGT;
2601 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2602 RK == LoopVectorizationLegality::MRK_FloatMax)
2603 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2605 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2607 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2612 struct CSEDenseMapInfo {
2613 static bool canHandle(Instruction *I) {
2614 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2615 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2617 static inline Instruction *getEmptyKey() {
2618 return DenseMapInfo<Instruction *>::getEmptyKey();
2620 static inline Instruction *getTombstoneKey() {
2621 return DenseMapInfo<Instruction *>::getTombstoneKey();
2623 static unsigned getHashValue(Instruction *I) {
2624 assert(canHandle(I) && "Unknown instruction!");
2625 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2626 I->value_op_end()));
2628 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2629 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2630 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2632 return LHS->isIdenticalTo(RHS);
2637 /// \brief Check whether this block is a predicated block.
2638 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2639 /// = ...; " blocks. We start with one vectorized basic block. For every
2640 /// conditional block we split this vectorized block. Therefore, every second
2641 /// block will be a predicated one.
2642 static bool isPredicatedBlock(unsigned BlockNum) {
2643 return BlockNum % 2;
2646 ///\brief Perform cse of induction variable instructions.
2647 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2648 // Perform simple cse.
2649 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2650 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2651 BasicBlock *BB = BBs[i];
2652 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2653 Instruction *In = I++;
2655 if (!CSEDenseMapInfo::canHandle(In))
2658 // Check if we can replace this instruction with any of the
2659 // visited instructions.
2660 if (Instruction *V = CSEMap.lookup(In)) {
2661 In->replaceAllUsesWith(V);
2662 In->eraseFromParent();
2665 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2666 // ...;" blocks for predicated stores. Every second block is a predicated
2668 if (isPredicatedBlock(i))
2676 /// \brief Adds a 'fast' flag to floating point operations.
2677 static Value *addFastMathFlag(Value *V) {
2678 if (isa<FPMathOperator>(V)){
2679 FastMathFlags Flags;
2680 Flags.setUnsafeAlgebra();
2681 cast<Instruction>(V)->setFastMathFlags(Flags);
2686 void InnerLoopVectorizer::vectorizeLoop() {
2687 //===------------------------------------------------===//
2689 // Notice: any optimization or new instruction that go
2690 // into the code below should be also be implemented in
2693 //===------------------------------------------------===//
2694 Constant *Zero = Builder.getInt32(0);
2696 // In order to support reduction variables we need to be able to vectorize
2697 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2698 // stages. First, we create a new vector PHI node with no incoming edges.
2699 // We use this value when we vectorize all of the instructions that use the
2700 // PHI. Next, after all of the instructions in the block are complete we
2701 // add the new incoming edges to the PHI. At this point all of the
2702 // instructions in the basic block are vectorized, so we can use them to
2703 // construct the PHI.
2704 PhiVector RdxPHIsToFix;
2706 // Scan the loop in a topological order to ensure that defs are vectorized
2708 LoopBlocksDFS DFS(OrigLoop);
2711 // Vectorize all of the blocks in the original loop.
2712 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2713 be = DFS.endRPO(); bb != be; ++bb)
2714 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2716 // At this point every instruction in the original loop is widened to
2717 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2718 // that we vectorized. The PHI nodes are currently empty because we did
2719 // not want to introduce cycles. Notice that the remaining PHI nodes
2720 // that we need to fix are reduction variables.
2722 // Create the 'reduced' values for each of the induction vars.
2723 // The reduced values are the vector values that we scalarize and combine
2724 // after the loop is finished.
2725 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2727 PHINode *RdxPhi = *it;
2728 assert(RdxPhi && "Unable to recover vectorized PHI");
2730 // Find the reduction variable descriptor.
2731 assert(Legal->getReductionVars()->count(RdxPhi) &&
2732 "Unable to find the reduction variable");
2733 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2734 (*Legal->getReductionVars())[RdxPhi];
2736 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2738 // We need to generate a reduction vector from the incoming scalar.
2739 // To do so, we need to generate the 'identity' vector and override
2740 // one of the elements with the incoming scalar reduction. We need
2741 // to do it in the vector-loop preheader.
2742 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
2744 // This is the vector-clone of the value that leaves the loop.
2745 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2746 Type *VecTy = VectorExit[0]->getType();
2748 // Find the reduction identity variable. Zero for addition, or, xor,
2749 // one for multiplication, -1 for And.
2752 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2753 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2754 // MinMax reduction have the start value as their identify.
2756 VectorStart = Identity = RdxDesc.StartValue;
2758 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2763 // Handle other reduction kinds:
2765 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2766 VecTy->getScalarType());
2769 // This vector is the Identity vector where the first element is the
2770 // incoming scalar reduction.
2771 VectorStart = RdxDesc.StartValue;
2773 Identity = ConstantVector::getSplat(VF, Iden);
2775 // This vector is the Identity vector where the first element is the
2776 // incoming scalar reduction.
2777 VectorStart = Builder.CreateInsertElement(Identity,
2778 RdxDesc.StartValue, Zero);
2782 // Fix the vector-loop phi.
2784 // Reductions do not have to start at zero. They can start with
2785 // any loop invariant values.
2786 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2787 BasicBlock *Latch = OrigLoop->getLoopLatch();
2788 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2789 VectorParts &Val = getVectorValue(LoopVal);
2790 for (unsigned part = 0; part < UF; ++part) {
2791 // Make sure to add the reduction stat value only to the
2792 // first unroll part.
2793 Value *StartVal = (part == 0) ? VectorStart : Identity;
2794 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
2795 LoopVectorPreHeader);
2796 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2797 LoopVectorBody.back());
2800 // Before each round, move the insertion point right between
2801 // the PHIs and the values we are going to write.
2802 // This allows us to write both PHINodes and the extractelement
2804 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2806 VectorParts RdxParts;
2807 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2808 for (unsigned part = 0; part < UF; ++part) {
2809 // This PHINode contains the vectorized reduction variable, or
2810 // the initial value vector, if we bypass the vector loop.
2811 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2812 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2813 Value *StartVal = (part == 0) ? VectorStart : Identity;
2814 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2815 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2816 NewPhi->addIncoming(RdxExitVal[part],
2817 LoopVectorBody.back());
2818 RdxParts.push_back(NewPhi);
2821 // Reduce all of the unrolled parts into a single vector.
2822 Value *ReducedPartRdx = RdxParts[0];
2823 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2824 setDebugLocFromInst(Builder, ReducedPartRdx);
2825 for (unsigned part = 1; part < UF; ++part) {
2826 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2827 // Floating point operations had to be 'fast' to enable the reduction.
2828 ReducedPartRdx = addFastMathFlag(
2829 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2830 ReducedPartRdx, "bin.rdx"));
2832 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2833 ReducedPartRdx, RdxParts[part]);
2837 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2838 // and vector ops, reducing the set of values being computed by half each
2840 assert(isPowerOf2_32(VF) &&
2841 "Reduction emission only supported for pow2 vectors!");
2842 Value *TmpVec = ReducedPartRdx;
2843 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2844 for (unsigned i = VF; i != 1; i >>= 1) {
2845 // Move the upper half of the vector to the lower half.
2846 for (unsigned j = 0; j != i/2; ++j)
2847 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2849 // Fill the rest of the mask with undef.
2850 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2851 UndefValue::get(Builder.getInt32Ty()));
2854 Builder.CreateShuffleVector(TmpVec,
2855 UndefValue::get(TmpVec->getType()),
2856 ConstantVector::get(ShuffleMask),
2859 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2860 // Floating point operations had to be 'fast' to enable the reduction.
2861 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2862 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2864 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2867 // The result is in the first element of the vector.
2868 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2869 Builder.getInt32(0));
2872 // Create a phi node that merges control-flow from the backedge-taken check
2873 // block and the middle block.
2874 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
2875 LoopScalarPreHeader->getTerminator());
2876 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
2877 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2879 // Now, we need to fix the users of the reduction variable
2880 // inside and outside of the scalar remainder loop.
2881 // We know that the loop is in LCSSA form. We need to update the
2882 // PHI nodes in the exit blocks.
2883 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2884 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2885 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2886 if (!LCSSAPhi) break;
2888 // All PHINodes need to have a single entry edge, or two if
2889 // we already fixed them.
2890 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2892 // We found our reduction value exit-PHI. Update it with the
2893 // incoming bypass edge.
2894 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2895 // Add an edge coming from the bypass.
2896 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2899 }// end of the LCSSA phi scan.
2901 // Fix the scalar loop reduction variable with the incoming reduction sum
2902 // from the vector body and from the backedge value.
2903 int IncomingEdgeBlockIdx =
2904 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2905 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2906 // Pick the other block.
2907 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2908 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
2909 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2910 }// end of for each redux variable.
2914 // Remove redundant induction instructions.
2915 cse(LoopVectorBody);
2918 void InnerLoopVectorizer::fixLCSSAPHIs() {
2919 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2920 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2921 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2922 if (!LCSSAPhi) break;
2923 if (LCSSAPhi->getNumIncomingValues() == 1)
2924 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2929 InnerLoopVectorizer::VectorParts
2930 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2931 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2934 // Look for cached value.
2935 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2936 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2937 if (ECEntryIt != MaskCache.end())
2938 return ECEntryIt->second;
2940 VectorParts SrcMask = createBlockInMask(Src);
2942 // The terminator has to be a branch inst!
2943 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2944 assert(BI && "Unexpected terminator found");
2946 if (BI->isConditional()) {
2947 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2949 if (BI->getSuccessor(0) != Dst)
2950 for (unsigned part = 0; part < UF; ++part)
2951 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2953 for (unsigned part = 0; part < UF; ++part)
2954 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2956 MaskCache[Edge] = EdgeMask;
2960 MaskCache[Edge] = SrcMask;
2964 InnerLoopVectorizer::VectorParts
2965 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2966 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2968 // Loop incoming mask is all-one.
2969 if (OrigLoop->getHeader() == BB) {
2970 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2971 return getVectorValue(C);
2974 // This is the block mask. We OR all incoming edges, and with zero.
2975 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2976 VectorParts BlockMask = getVectorValue(Zero);
2979 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2980 VectorParts EM = createEdgeMask(*it, BB);
2981 for (unsigned part = 0; part < UF; ++part)
2982 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2988 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2989 InnerLoopVectorizer::VectorParts &Entry,
2990 unsigned UF, unsigned VF, PhiVector *PV) {
2991 PHINode* P = cast<PHINode>(PN);
2992 // Handle reduction variables:
2993 if (Legal->getReductionVars()->count(P)) {
2994 for (unsigned part = 0; part < UF; ++part) {
2995 // This is phase one of vectorizing PHIs.
2996 Type *VecTy = (VF == 1) ? PN->getType() :
2997 VectorType::get(PN->getType(), VF);
2998 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2999 LoopVectorBody.back()-> getFirstInsertionPt());
3005 setDebugLocFromInst(Builder, P);
3006 // Check for PHI nodes that are lowered to vector selects.
3007 if (P->getParent() != OrigLoop->getHeader()) {
3008 // We know that all PHIs in non-header blocks are converted into
3009 // selects, so we don't have to worry about the insertion order and we
3010 // can just use the builder.
3011 // At this point we generate the predication tree. There may be
3012 // duplications since this is a simple recursive scan, but future
3013 // optimizations will clean it up.
3015 unsigned NumIncoming = P->getNumIncomingValues();
3017 // Generate a sequence of selects of the form:
3018 // SELECT(Mask3, In3,
3019 // SELECT(Mask2, In2,
3021 for (unsigned In = 0; In < NumIncoming; In++) {
3022 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
3024 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3026 for (unsigned part = 0; part < UF; ++part) {
3027 // We might have single edge PHIs (blocks) - use an identity
3028 // 'select' for the first PHI operand.
3030 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3033 // Select between the current value and the previous incoming edge
3034 // based on the incoming mask.
3035 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3036 Entry[part], "predphi");
3042 // This PHINode must be an induction variable.
3043 // Make sure that we know about it.
3044 assert(Legal->getInductionVars()->count(P) &&
3045 "Not an induction variable");
3047 LoopVectorizationLegality::InductionInfo II =
3048 Legal->getInductionVars()->lookup(P);
3050 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
3051 // which can be found from the original scalar operations.
3053 case LoopVectorizationLegality::IK_NoInduction:
3054 llvm_unreachable("Unknown induction");
3055 case LoopVectorizationLegality::IK_IntInduction: {
3056 assert(P->getType() == II.StartValue->getType() && "Types must match");
3057 Type *PhiTy = P->getType();
3059 if (P == OldInduction) {
3060 // Handle the canonical induction variable. We might have had to
3062 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
3064 // Handle other induction variables that are now based on the
3066 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
3068 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
3069 Broadcasted = II.transform(Builder, NormalizedIdx);
3070 Broadcasted->setName("offset.idx");
3072 Broadcasted = getBroadcastInstrs(Broadcasted);
3073 // After broadcasting the induction variable we need to make the vector
3074 // consecutive by adding 0, 1, 2, etc.
3075 for (unsigned part = 0; part < UF; ++part)
3076 Entry[part] = getStepVector(Broadcasted, VF * part, II.StepValue);
3079 case LoopVectorizationLegality::IK_PtrInduction:
3080 // Handle the pointer induction variable case.
3081 assert(P->getType()->isPointerTy() && "Unexpected type.");
3082 // This is the normalized GEP that starts counting at zero.
3083 Value *NormalizedIdx =
3084 Builder.CreateSub(Induction, ExtendedIdx, "normalized.idx");
3085 // This is the vector of results. Notice that we don't generate
3086 // vector geps because scalar geps result in better code.
3087 for (unsigned part = 0; part < UF; ++part) {
3089 int EltIndex = part;
3090 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3091 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3092 Value *SclrGep = II.transform(Builder, GlobalIdx);
3093 SclrGep->setName("next.gep");
3094 Entry[part] = SclrGep;
3098 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3099 for (unsigned int i = 0; i < VF; ++i) {
3100 int EltIndex = i + part * VF;
3101 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3102 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3103 Value *SclrGep = II.transform(Builder, GlobalIdx);
3104 SclrGep->setName("next.gep");
3105 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3106 Builder.getInt32(i),
3109 Entry[part] = VecVal;
3115 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3116 // For each instruction in the old loop.
3117 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3118 VectorParts &Entry = WidenMap.get(it);
3119 switch (it->getOpcode()) {
3120 case Instruction::Br:
3121 // Nothing to do for PHIs and BR, since we already took care of the
3122 // loop control flow instructions.
3124 case Instruction::PHI: {
3125 // Vectorize PHINodes.
3126 widenPHIInstruction(it, Entry, UF, VF, PV);
3130 case Instruction::Add:
3131 case Instruction::FAdd:
3132 case Instruction::Sub:
3133 case Instruction::FSub:
3134 case Instruction::Mul:
3135 case Instruction::FMul:
3136 case Instruction::UDiv:
3137 case Instruction::SDiv:
3138 case Instruction::FDiv:
3139 case Instruction::URem:
3140 case Instruction::SRem:
3141 case Instruction::FRem:
3142 case Instruction::Shl:
3143 case Instruction::LShr:
3144 case Instruction::AShr:
3145 case Instruction::And:
3146 case Instruction::Or:
3147 case Instruction::Xor: {
3148 // Just widen binops.
3149 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3150 setDebugLocFromInst(Builder, BinOp);
3151 VectorParts &A = getVectorValue(it->getOperand(0));
3152 VectorParts &B = getVectorValue(it->getOperand(1));
3154 // Use this vector value for all users of the original instruction.
3155 for (unsigned Part = 0; Part < UF; ++Part) {
3156 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3158 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3159 VecOp->copyIRFlags(BinOp);
3164 propagateMetadata(Entry, it);
3167 case Instruction::Select: {
3169 // If the selector is loop invariant we can create a select
3170 // instruction with a scalar condition. Otherwise, use vector-select.
3171 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3173 setDebugLocFromInst(Builder, it);
3175 // The condition can be loop invariant but still defined inside the
3176 // loop. This means that we can't just use the original 'cond' value.
3177 // We have to take the 'vectorized' value and pick the first lane.
3178 // Instcombine will make this a no-op.
3179 VectorParts &Cond = getVectorValue(it->getOperand(0));
3180 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3181 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3183 Value *ScalarCond = (VF == 1) ? Cond[0] :
3184 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3186 for (unsigned Part = 0; Part < UF; ++Part) {
3187 Entry[Part] = Builder.CreateSelect(
3188 InvariantCond ? ScalarCond : Cond[Part],
3193 propagateMetadata(Entry, it);
3197 case Instruction::ICmp:
3198 case Instruction::FCmp: {
3199 // Widen compares. Generate vector compares.
3200 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3201 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3202 setDebugLocFromInst(Builder, it);
3203 VectorParts &A = getVectorValue(it->getOperand(0));
3204 VectorParts &B = getVectorValue(it->getOperand(1));
3205 for (unsigned Part = 0; Part < UF; ++Part) {
3208 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3210 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3214 propagateMetadata(Entry, it);
3218 case Instruction::Store:
3219 case Instruction::Load:
3220 vectorizeMemoryInstruction(it);
3222 case Instruction::ZExt:
3223 case Instruction::SExt:
3224 case Instruction::FPToUI:
3225 case Instruction::FPToSI:
3226 case Instruction::FPExt:
3227 case Instruction::PtrToInt:
3228 case Instruction::IntToPtr:
3229 case Instruction::SIToFP:
3230 case Instruction::UIToFP:
3231 case Instruction::Trunc:
3232 case Instruction::FPTrunc:
3233 case Instruction::BitCast: {
3234 CastInst *CI = dyn_cast<CastInst>(it);
3235 setDebugLocFromInst(Builder, it);
3236 /// Optimize the special case where the source is the induction
3237 /// variable. Notice that we can only optimize the 'trunc' case
3238 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3239 /// c. other casts depend on pointer size.
3240 if (CI->getOperand(0) == OldInduction &&
3241 it->getOpcode() == Instruction::Trunc) {
3242 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3244 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3245 LoopVectorizationLegality::InductionInfo II =
3246 Legal->getInductionVars()->lookup(OldInduction);
3248 ConstantInt::getSigned(CI->getType(), II.StepValue->getSExtValue());
3249 for (unsigned Part = 0; Part < UF; ++Part)
3250 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
3251 propagateMetadata(Entry, it);
3254 /// Vectorize casts.
3255 Type *DestTy = (VF == 1) ? CI->getType() :
3256 VectorType::get(CI->getType(), VF);
3258 VectorParts &A = getVectorValue(it->getOperand(0));
3259 for (unsigned Part = 0; Part < UF; ++Part)
3260 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3261 propagateMetadata(Entry, it);
3265 case Instruction::Call: {
3266 // Ignore dbg intrinsics.
3267 if (isa<DbgInfoIntrinsic>(it))
3269 setDebugLocFromInst(Builder, it);
3271 Module *M = BB->getParent()->getParent();
3272 CallInst *CI = cast<CallInst>(it);
3273 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3274 assert(ID && "Not an intrinsic call!");
3276 case Intrinsic::assume:
3277 case Intrinsic::lifetime_end:
3278 case Intrinsic::lifetime_start:
3279 scalarizeInstruction(it);
3282 bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
3283 for (unsigned Part = 0; Part < UF; ++Part) {
3284 SmallVector<Value *, 4> Args;
3285 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3286 if (HasScalarOpd && i == 1) {
3287 Args.push_back(CI->getArgOperand(i));
3290 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3291 Args.push_back(Arg[Part]);
3293 Type *Tys[] = {CI->getType()};
3295 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3297 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3298 Entry[Part] = Builder.CreateCall(F, Args);
3301 propagateMetadata(Entry, it);
3308 // All other instructions are unsupported. Scalarize them.
3309 scalarizeInstruction(it);
3312 }// end of for_each instr.
3315 void InnerLoopVectorizer::updateAnalysis() {
3316 // Forget the original basic block.
3317 SE->forgetLoop(OrigLoop);
3319 // Update the dominator tree information.
3320 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3321 "Entry does not dominate exit.");
3323 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3324 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3325 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3327 // Due to if predication of stores we might create a sequence of "if(pred)
3328 // a[i] = ...; " blocks.
3329 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3331 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3332 else if (isPredicatedBlock(i)) {
3333 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3335 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3339 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3340 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3341 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3342 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3344 DEBUG(DT->verifyDomTree());
3347 /// \brief Check whether it is safe to if-convert this phi node.
3349 /// Phi nodes with constant expressions that can trap are not safe to if
3351 static bool canIfConvertPHINodes(BasicBlock *BB) {
3352 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3353 PHINode *Phi = dyn_cast<PHINode>(I);
3356 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3357 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3364 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3365 if (!EnableIfConversion) {
3366 emitAnalysis(VectorizationReport() << "if-conversion is disabled");
3370 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3372 // A list of pointers that we can safely read and write to.
3373 SmallPtrSet<Value *, 8> SafePointes;
3375 // Collect safe addresses.
3376 for (Loop::block_iterator BI = TheLoop->block_begin(),
3377 BE = TheLoop->block_end(); BI != BE; ++BI) {
3378 BasicBlock *BB = *BI;
3380 if (blockNeedsPredication(BB))
3383 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3384 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3385 SafePointes.insert(LI->getPointerOperand());
3386 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3387 SafePointes.insert(SI->getPointerOperand());
3391 // Collect the blocks that need predication.
3392 BasicBlock *Header = TheLoop->getHeader();
3393 for (Loop::block_iterator BI = TheLoop->block_begin(),
3394 BE = TheLoop->block_end(); BI != BE; ++BI) {
3395 BasicBlock *BB = *BI;
3397 // We don't support switch statements inside loops.
3398 if (!isa<BranchInst>(BB->getTerminator())) {
3399 emitAnalysis(VectorizationReport(BB->getTerminator())
3400 << "loop contains a switch statement");
3404 // We must be able to predicate all blocks that need to be predicated.
3405 if (blockNeedsPredication(BB)) {
3406 if (!blockCanBePredicated(BB, SafePointes)) {
3407 emitAnalysis(VectorizationReport(BB->getTerminator())
3408 << "control flow cannot be substituted for a select");
3411 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3412 emitAnalysis(VectorizationReport(BB->getTerminator())
3413 << "control flow cannot be substituted for a select");
3418 // We can if-convert this loop.
3422 bool LoopVectorizationLegality::canVectorize() {
3423 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3424 // be canonicalized.
3425 if (!TheLoop->getLoopPreheader()) {
3427 VectorizationReport() <<
3428 "loop control flow is not understood by vectorizer");
3432 // We can only vectorize innermost loops.
3433 if (!TheLoop->getSubLoopsVector().empty()) {
3434 emitAnalysis(VectorizationReport() << "loop is not the innermost loop");
3438 // We must have a single backedge.
3439 if (TheLoop->getNumBackEdges() != 1) {
3441 VectorizationReport() <<
3442 "loop control flow is not understood by vectorizer");
3446 // We must have a single exiting block.
3447 if (!TheLoop->getExitingBlock()) {
3449 VectorizationReport() <<
3450 "loop control flow is not understood by vectorizer");
3454 // We only handle bottom-tested loops, i.e. loop in which the condition is
3455 // checked at the end of each iteration. With that we can assume that all
3456 // instructions in the loop are executed the same number of times.
3457 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
3459 VectorizationReport() <<
3460 "loop control flow is not understood by vectorizer");
3464 // We need to have a loop header.
3465 DEBUG(dbgs() << "LV: Found a loop: " <<
3466 TheLoop->getHeader()->getName() << '\n');
3468 // Check if we can if-convert non-single-bb loops.
3469 unsigned NumBlocks = TheLoop->getNumBlocks();
3470 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3471 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3475 // ScalarEvolution needs to be able to find the exit count.
3476 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3477 if (ExitCount == SE->getCouldNotCompute()) {
3478 emitAnalysis(VectorizationReport() <<
3479 "could not determine number of loop iterations");
3480 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3484 // Check if we can vectorize the instructions and CFG in this loop.
3485 if (!canVectorizeInstrs()) {
3486 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3490 // Go over each instruction and look at memory deps.
3491 if (!canVectorizeMemory()) {
3492 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3496 // Collect all of the variables that remain uniform after vectorization.
3497 collectLoopUniforms();
3499 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3500 (LAA.getRuntimePointerCheck()->Need ? " (with a runtime bound check)" :
3504 // Okay! We can vectorize. At this point we don't have any other mem analysis
3505 // which may limit our maximum vectorization factor, so just return true with
3510 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3511 if (Ty->isPointerTy())
3512 return DL.getIntPtrType(Ty);
3514 // It is possible that char's or short's overflow when we ask for the loop's
3515 // trip count, work around this by changing the type size.
3516 if (Ty->getScalarSizeInBits() < 32)
3517 return Type::getInt32Ty(Ty->getContext());
3522 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3523 Ty0 = convertPointerToIntegerType(DL, Ty0);
3524 Ty1 = convertPointerToIntegerType(DL, Ty1);
3525 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3530 /// \brief Check that the instruction has outside loop users and is not an
3531 /// identified reduction variable.
3532 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3533 SmallPtrSetImpl<Value *> &Reductions) {
3534 // Reduction instructions are allowed to have exit users. All other
3535 // instructions must not have external users.
3536 if (!Reductions.count(Inst))
3537 //Check that all of the users of the loop are inside the BB.
3538 for (User *U : Inst->users()) {
3539 Instruction *UI = cast<Instruction>(U);
3540 // This user may be a reduction exit value.
3541 if (!TheLoop->contains(UI)) {
3542 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3549 bool LoopVectorizationLegality::canVectorizeInstrs() {
3550 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3551 BasicBlock *Header = TheLoop->getHeader();
3553 // Look for the attribute signaling the absence of NaNs.
3554 Function &F = *Header->getParent();
3555 if (F.hasFnAttribute("no-nans-fp-math"))
3556 HasFunNoNaNAttr = F.getAttributes().getAttribute(
3557 AttributeSet::FunctionIndex,
3558 "no-nans-fp-math").getValueAsString() == "true";
3560 // For each block in the loop.
3561 for (Loop::block_iterator bb = TheLoop->block_begin(),
3562 be = TheLoop->block_end(); bb != be; ++bb) {
3564 // Scan the instructions in the block and look for hazards.
3565 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3568 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3569 Type *PhiTy = Phi->getType();
3570 // Check that this PHI type is allowed.
3571 if (!PhiTy->isIntegerTy() &&
3572 !PhiTy->isFloatingPointTy() &&
3573 !PhiTy->isPointerTy()) {
3574 emitAnalysis(VectorizationReport(it)
3575 << "loop control flow is not understood by vectorizer");
3576 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3580 // If this PHINode is not in the header block, then we know that we
3581 // can convert it to select during if-conversion. No need to check if
3582 // the PHIs in this block are induction or reduction variables.
3583 if (*bb != Header) {
3584 // Check that this instruction has no outside users or is an
3585 // identified reduction value with an outside user.
3586 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3588 emitAnalysis(VectorizationReport(it) <<
3589 "value could not be identified as "
3590 "an induction or reduction variable");
3594 // We only allow if-converted PHIs with exactly two incoming values.
3595 if (Phi->getNumIncomingValues() != 2) {
3596 emitAnalysis(VectorizationReport(it)
3597 << "control flow not understood by vectorizer");
3598 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3602 // This is the value coming from the preheader.
3603 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3604 ConstantInt *StepValue = nullptr;
3605 // Check if this is an induction variable.
3606 InductionKind IK = isInductionVariable(Phi, StepValue);
3608 if (IK_NoInduction != IK) {
3609 // Get the widest type.
3611 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3613 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3615 // Int inductions are special because we only allow one IV.
3616 if (IK == IK_IntInduction && StepValue->isOne()) {
3617 // Use the phi node with the widest type as induction. Use the last
3618 // one if there are multiple (no good reason for doing this other
3619 // than it is expedient).
3620 if (!Induction || PhiTy == WidestIndTy)
3624 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3625 Inductions[Phi] = InductionInfo(StartValue, IK, StepValue);
3627 // Until we explicitly handle the case of an induction variable with
3628 // an outside loop user we have to give up vectorizing this loop.
3629 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3630 emitAnalysis(VectorizationReport(it) <<
3631 "use of induction value outside of the "
3632 "loop is not handled by vectorizer");
3639 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3640 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3643 if (AddReductionVar(Phi, RK_IntegerMult)) {
3644 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3647 if (AddReductionVar(Phi, RK_IntegerOr)) {
3648 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3651 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3652 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3655 if (AddReductionVar(Phi, RK_IntegerXor)) {
3656 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3659 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3660 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3663 if (AddReductionVar(Phi, RK_FloatMult)) {
3664 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3667 if (AddReductionVar(Phi, RK_FloatAdd)) {
3668 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3671 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3672 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3677 emitAnalysis(VectorizationReport(it) <<
3678 "value that could not be identified as "
3679 "reduction is used outside the loop");
3680 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3682 }// end of PHI handling
3684 // We still don't handle functions. However, we can ignore dbg intrinsic
3685 // calls and we do handle certain intrinsic and libm functions.
3686 CallInst *CI = dyn_cast<CallInst>(it);
3687 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3688 emitAnalysis(VectorizationReport(it) <<
3689 "call instruction cannot be vectorized");
3690 DEBUG(dbgs() << "LV: Found a call site.\n");
3694 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
3695 // second argument is the same (i.e. loop invariant)
3697 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
3698 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
3699 emitAnalysis(VectorizationReport(it)
3700 << "intrinsic instruction cannot be vectorized");
3701 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
3706 // Check that the instruction return type is vectorizable.
3707 // Also, we can't vectorize extractelement instructions.
3708 if ((!VectorType::isValidElementType(it->getType()) &&
3709 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3710 emitAnalysis(VectorizationReport(it)
3711 << "instruction return type cannot be vectorized");
3712 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3716 // Check that the stored type is vectorizable.
3717 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3718 Type *T = ST->getValueOperand()->getType();
3719 if (!VectorType::isValidElementType(T)) {
3720 emitAnalysis(VectorizationReport(ST) <<
3721 "store instruction cannot be vectorized");
3724 if (EnableMemAccessVersioning)
3725 collectStridedAccess(ST);
3728 if (EnableMemAccessVersioning)
3729 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3730 collectStridedAccess(LI);
3732 // Reduction instructions are allowed to have exit users.
3733 // All other instructions must not have external users.
3734 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3735 emitAnalysis(VectorizationReport(it) <<
3736 "value cannot be used outside the loop");
3745 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3746 if (Inductions.empty()) {
3747 emitAnalysis(VectorizationReport()
3748 << "loop induction variable could not be identified");
3756 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3757 /// return the induction operand of the gep pointer.
3758 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3759 const DataLayout *DL, Loop *Lp) {
3760 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3764 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3766 // Check that all of the gep indices are uniform except for our induction
3768 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3769 if (i != InductionOperand &&
3770 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3772 return GEP->getOperand(InductionOperand);
3775 ///\brief Look for a cast use of the passed value.
3776 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3777 Value *UniqueCast = nullptr;
3778 for (User *U : Ptr->users()) {
3779 CastInst *CI = dyn_cast<CastInst>(U);
3780 if (CI && CI->getType() == Ty) {
3790 ///\brief Get the stride of a pointer access in a loop.
3791 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3792 /// pointer to the Value, or null otherwise.
3793 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3794 const DataLayout *DL, Loop *Lp) {
3795 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3796 if (!PtrTy || PtrTy->isAggregateType())
3799 // Try to remove a gep instruction to make the pointer (actually index at this
3800 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3801 // pointer, otherwise, we are analyzing the index.
3802 Value *OrigPtr = Ptr;
3804 // The size of the pointer access.
3805 int64_t PtrAccessSize = 1;
3807 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3808 const SCEV *V = SE->getSCEV(Ptr);
3812 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3813 V = C->getOperand();
3815 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3819 V = S->getStepRecurrence(*SE);
3823 // Strip off the size of access multiplication if we are still analyzing the
3825 if (OrigPtr == Ptr) {
3826 DL->getTypeAllocSize(PtrTy->getElementType());
3827 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3828 if (M->getOperand(0)->getSCEVType() != scConstant)
3831 const APInt &APStepVal =
3832 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3834 // Huge step value - give up.
3835 if (APStepVal.getBitWidth() > 64)
3838 int64_t StepVal = APStepVal.getSExtValue();
3839 if (PtrAccessSize != StepVal)
3841 V = M->getOperand(1);
3846 Type *StripedOffRecurrenceCast = nullptr;
3847 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3848 StripedOffRecurrenceCast = C->getType();
3849 V = C->getOperand();
3852 // Look for the loop invariant symbolic value.
3853 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3857 Value *Stride = U->getValue();
3858 if (!Lp->isLoopInvariant(Stride))
3861 // If we have stripped off the recurrence cast we have to make sure that we
3862 // return the value that is used in this loop so that we can replace it later.
3863 if (StripedOffRecurrenceCast)
3864 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3869 void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) {
3870 Value *Ptr = nullptr;
3871 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3872 Ptr = LI->getPointerOperand();
3873 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3874 Ptr = SI->getPointerOperand();
3878 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3882 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3883 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3884 Strides[Ptr] = Stride;
3885 StrideSet.insert(Stride);
3888 void LoopVectorizationLegality::collectLoopUniforms() {
3889 // We now know that the loop is vectorizable!
3890 // Collect variables that will remain uniform after vectorization.
3891 std::vector<Value*> Worklist;
3892 BasicBlock *Latch = TheLoop->getLoopLatch();
3894 // Start with the conditional branch and walk up the block.
3895 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3897 // Also add all consecutive pointer values; these values will be uniform
3898 // after vectorization (and subsequent cleanup) and, until revectorization is
3899 // supported, all dependencies must also be uniform.
3900 for (Loop::block_iterator B = TheLoop->block_begin(),
3901 BE = TheLoop->block_end(); B != BE; ++B)
3902 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
3904 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
3905 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3907 while (!Worklist.empty()) {
3908 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3909 Worklist.pop_back();
3911 // Look at instructions inside this loop.
3912 // Stop when reaching PHI nodes.
3913 // TODO: we need to follow values all over the loop, not only in this block.
3914 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3917 // This is a known uniform.
3920 // Insert all operands.
3921 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3925 bool LoopVectorizationLegality::canVectorizeMemory() {
3926 return LAA.canVectorizeMemory(Strides);
3929 static bool hasMultipleUsesOf(Instruction *I,
3930 SmallPtrSetImpl<Instruction *> &Insts) {
3931 unsigned NumUses = 0;
3932 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
3933 if (Insts.count(dyn_cast<Instruction>(*Use)))
3942 static bool areAllUsesIn(Instruction *I, SmallPtrSetImpl<Instruction *> &Set) {
3943 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
3944 if (!Set.count(dyn_cast<Instruction>(*Use)))
3949 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
3950 ReductionKind Kind) {
3951 if (Phi->getNumIncomingValues() != 2)
3954 // Reduction variables are only found in the loop header block.
3955 if (Phi->getParent() != TheLoop->getHeader())
3958 // Obtain the reduction start value from the value that comes from the loop
3960 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
3962 // ExitInstruction is the single value which is used outside the loop.
3963 // We only allow for a single reduction value to be used outside the loop.
3964 // This includes users of the reduction, variables (which form a cycle
3965 // which ends in the phi node).
3966 Instruction *ExitInstruction = nullptr;
3967 // Indicates that we found a reduction operation in our scan.
3968 bool FoundReduxOp = false;
3970 // We start with the PHI node and scan for all of the users of this
3971 // instruction. All users must be instructions that can be used as reduction
3972 // variables (such as ADD). We must have a single out-of-block user. The cycle
3973 // must include the original PHI.
3974 bool FoundStartPHI = false;
3976 // To recognize min/max patterns formed by a icmp select sequence, we store
3977 // the number of instruction we saw from the recognized min/max pattern,
3978 // to make sure we only see exactly the two instructions.
3979 unsigned NumCmpSelectPatternInst = 0;
3980 ReductionInstDesc ReduxDesc(false, nullptr);
3982 SmallPtrSet<Instruction *, 8> VisitedInsts;
3983 SmallVector<Instruction *, 8> Worklist;
3984 Worklist.push_back(Phi);
3985 VisitedInsts.insert(Phi);
3987 // A value in the reduction can be used:
3988 // - By the reduction:
3989 // - Reduction operation:
3990 // - One use of reduction value (safe).
3991 // - Multiple use of reduction value (not safe).
3993 // - All uses of the PHI must be the reduction (safe).
3994 // - Otherwise, not safe.
3995 // - By one instruction outside of the loop (safe).
3996 // - By further instructions outside of the loop (not safe).
3997 // - By an instruction that is not part of the reduction (not safe).
3999 // * An instruction type other than PHI or the reduction operation.
4000 // * A PHI in the header other than the initial PHI.
4001 while (!Worklist.empty()) {
4002 Instruction *Cur = Worklist.back();
4003 Worklist.pop_back();
4006 // If the instruction has no users then this is a broken chain and can't be
4007 // a reduction variable.
4008 if (Cur->use_empty())
4011 bool IsAPhi = isa<PHINode>(Cur);
4013 // A header PHI use other than the original PHI.
4014 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
4017 // Reductions of instructions such as Div, and Sub is only possible if the
4018 // LHS is the reduction variable.
4019 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
4020 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
4021 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
4024 // Any reduction instruction must be of one of the allowed kinds.
4025 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
4026 if (!ReduxDesc.IsReduction)
4029 // A reduction operation must only have one use of the reduction value.
4030 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
4031 hasMultipleUsesOf(Cur, VisitedInsts))
4034 // All inputs to a PHI node must be a reduction value.
4035 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
4038 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
4039 isa<SelectInst>(Cur)))
4040 ++NumCmpSelectPatternInst;
4041 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
4042 isa<SelectInst>(Cur)))
4043 ++NumCmpSelectPatternInst;
4045 // Check whether we found a reduction operator.
4046 FoundReduxOp |= !IsAPhi;
4048 // Process users of current instruction. Push non-PHI nodes after PHI nodes
4049 // onto the stack. This way we are going to have seen all inputs to PHI
4050 // nodes once we get to them.
4051 SmallVector<Instruction *, 8> NonPHIs;
4052 SmallVector<Instruction *, 8> PHIs;
4053 for (User *U : Cur->users()) {
4054 Instruction *UI = cast<Instruction>(U);
4056 // Check if we found the exit user.
4057 BasicBlock *Parent = UI->getParent();
4058 if (!TheLoop->contains(Parent)) {
4059 // Exit if you find multiple outside users or if the header phi node is
4060 // being used. In this case the user uses the value of the previous
4061 // iteration, in which case we would loose "VF-1" iterations of the
4062 // reduction operation if we vectorize.
4063 if (ExitInstruction != nullptr || Cur == Phi)
4066 // The instruction used by an outside user must be the last instruction
4067 // before we feed back to the reduction phi. Otherwise, we loose VF-1
4068 // operations on the value.
4069 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
4072 ExitInstruction = Cur;
4076 // Process instructions only once (termination). Each reduction cycle
4077 // value must only be used once, except by phi nodes and min/max
4078 // reductions which are represented as a cmp followed by a select.
4079 ReductionInstDesc IgnoredVal(false, nullptr);
4080 if (VisitedInsts.insert(UI).second) {
4081 if (isa<PHINode>(UI))
4084 NonPHIs.push_back(UI);
4085 } else if (!isa<PHINode>(UI) &&
4086 ((!isa<FCmpInst>(UI) &&
4087 !isa<ICmpInst>(UI) &&
4088 !isa<SelectInst>(UI)) ||
4089 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
4092 // Remember that we completed the cycle.
4094 FoundStartPHI = true;
4096 Worklist.append(PHIs.begin(), PHIs.end());
4097 Worklist.append(NonPHIs.begin(), NonPHIs.end());
4100 // This means we have seen one but not the other instruction of the
4101 // pattern or more than just a select and cmp.
4102 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
4103 NumCmpSelectPatternInst != 2)
4106 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
4109 // We found a reduction var if we have reached the original phi node and we
4110 // only have a single instruction with out-of-loop users.
4112 // This instruction is allowed to have out-of-loop users.
4113 AllowedExit.insert(ExitInstruction);
4115 // Save the description of this reduction variable.
4116 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
4117 ReduxDesc.MinMaxKind);
4118 Reductions[Phi] = RD;
4119 // We've ended the cycle. This is a reduction variable if we have an
4120 // outside user and it has a binary op.
4125 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
4126 /// pattern corresponding to a min(X, Y) or max(X, Y).
4127 LoopVectorizationLegality::ReductionInstDesc
4128 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
4129 ReductionInstDesc &Prev) {
4131 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
4132 "Expect a select instruction");
4133 Instruction *Cmp = nullptr;
4134 SelectInst *Select = nullptr;
4136 // We must handle the select(cmp()) as a single instruction. Advance to the
4138 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
4139 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
4140 return ReductionInstDesc(false, I);
4141 return ReductionInstDesc(Select, Prev.MinMaxKind);
4144 // Only handle single use cases for now.
4145 if (!(Select = dyn_cast<SelectInst>(I)))
4146 return ReductionInstDesc(false, I);
4147 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
4148 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
4149 return ReductionInstDesc(false, I);
4150 if (!Cmp->hasOneUse())
4151 return ReductionInstDesc(false, I);
4156 // Look for a min/max pattern.
4157 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4158 return ReductionInstDesc(Select, MRK_UIntMin);
4159 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4160 return ReductionInstDesc(Select, MRK_UIntMax);
4161 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4162 return ReductionInstDesc(Select, MRK_SIntMax);
4163 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4164 return ReductionInstDesc(Select, MRK_SIntMin);
4165 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4166 return ReductionInstDesc(Select, MRK_FloatMin);
4167 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4168 return ReductionInstDesc(Select, MRK_FloatMax);
4169 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4170 return ReductionInstDesc(Select, MRK_FloatMin);
4171 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4172 return ReductionInstDesc(Select, MRK_FloatMax);
4174 return ReductionInstDesc(false, I);
4177 LoopVectorizationLegality::ReductionInstDesc
4178 LoopVectorizationLegality::isReductionInstr(Instruction *I,
4180 ReductionInstDesc &Prev) {
4181 bool FP = I->getType()->isFloatingPointTy();
4182 bool FastMath = FP && I->hasUnsafeAlgebra();
4183 switch (I->getOpcode()) {
4185 return ReductionInstDesc(false, I);
4186 case Instruction::PHI:
4187 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
4188 Kind != RK_FloatMinMax))
4189 return ReductionInstDesc(false, I);
4190 return ReductionInstDesc(I, Prev.MinMaxKind);
4191 case Instruction::Sub:
4192 case Instruction::Add:
4193 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4194 case Instruction::Mul:
4195 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4196 case Instruction::And:
4197 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4198 case Instruction::Or:
4199 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4200 case Instruction::Xor:
4201 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4202 case Instruction::FMul:
4203 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4204 case Instruction::FSub:
4205 case Instruction::FAdd:
4206 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4207 case Instruction::FCmp:
4208 case Instruction::ICmp:
4209 case Instruction::Select:
4210 if (Kind != RK_IntegerMinMax &&
4211 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4212 return ReductionInstDesc(false, I);
4213 return isMinMaxSelectCmpPattern(I, Prev);
4217 LoopVectorizationLegality::InductionKind
4218 LoopVectorizationLegality::isInductionVariable(PHINode *Phi,
4219 ConstantInt *&StepValue) {
4220 Type *PhiTy = Phi->getType();
4221 // We only handle integer and pointer inductions variables.
4222 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4223 return IK_NoInduction;
4225 // Check that the PHI is consecutive.
4226 const SCEV *PhiScev = SE->getSCEV(Phi);
4227 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4229 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4230 return IK_NoInduction;
4233 const SCEV *Step = AR->getStepRecurrence(*SE);
4234 // Calculate the pointer stride and check if it is consecutive.
4235 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4237 return IK_NoInduction;
4239 ConstantInt *CV = C->getValue();
4240 if (PhiTy->isIntegerTy()) {
4242 return IK_IntInduction;
4245 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4246 Type *PointerElementType = PhiTy->getPointerElementType();
4247 // The pointer stride cannot be determined if the pointer element type is not
4249 if (!PointerElementType->isSized())
4250 return IK_NoInduction;
4252 int64_t Size = static_cast<int64_t>(DL->getTypeAllocSize(PointerElementType));
4253 int64_t CVSize = CV->getSExtValue();
4255 return IK_NoInduction;
4256 StepValue = ConstantInt::getSigned(CV->getType(), CVSize / Size);
4257 return IK_PtrInduction;
4260 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4261 Value *In0 = const_cast<Value*>(V);
4262 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4266 return Inductions.count(PN);
4269 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4270 return LAA.blockNeedsPredication(BB);
4273 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4274 SmallPtrSetImpl<Value *> &SafePtrs) {
4276 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4277 // Check that we don't have a constant expression that can trap as operand.
4278 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4280 if (Constant *C = dyn_cast<Constant>(*OI))
4284 // We might be able to hoist the load.
4285 if (it->mayReadFromMemory()) {
4286 LoadInst *LI = dyn_cast<LoadInst>(it);
4289 if (!SafePtrs.count(LI->getPointerOperand())) {
4290 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
4291 MaskedOp.insert(LI);
4298 // We don't predicate stores at the moment.
4299 if (it->mayWriteToMemory()) {
4300 StoreInst *SI = dyn_cast<StoreInst>(it);
4301 // We only support predication of stores in basic blocks with one
4306 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
4307 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
4309 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
4310 !isSinglePredecessor) {
4311 // Build a masked store if it is legal for the target, otherwise scalarize
4313 bool isLegalMaskedOp =
4314 isLegalMaskedStore(SI->getValueOperand()->getType(),
4315 SI->getPointerOperand());
4316 if (isLegalMaskedOp) {
4318 MaskedOp.insert(SI);
4327 // The instructions below can trap.
4328 switch (it->getOpcode()) {
4330 case Instruction::UDiv:
4331 case Instruction::SDiv:
4332 case Instruction::URem:
4333 case Instruction::SRem:
4341 LoopVectorizationCostModel::VectorizationFactor
4342 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
4343 // Width 1 means no vectorize
4344 VectorizationFactor Factor = { 1U, 0U };
4345 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4346 emitAnalysis(VectorizationReport() <<
4347 "runtime pointer checks needed. Enable vectorization of this "
4348 "loop with '#pragma clang loop vectorize(enable)' when "
4349 "compiling with -Os");
4350 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4354 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
4355 emitAnalysis(VectorizationReport() <<
4356 "store that is conditionally executed prevents vectorization");
4357 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
4361 // Find the trip count.
4362 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4363 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4365 unsigned WidestType = getWidestType();
4366 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4367 unsigned MaxSafeDepDist = -1U;
4368 if (Legal->getMaxSafeDepDistBytes() != -1U)
4369 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4370 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4371 WidestRegister : MaxSafeDepDist);
4372 unsigned MaxVectorSize = WidestRegister / WidestType;
4373 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4374 DEBUG(dbgs() << "LV: The Widest register is: "
4375 << WidestRegister << " bits.\n");
4377 if (MaxVectorSize == 0) {
4378 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4382 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
4383 " into one vector!");
4385 unsigned VF = MaxVectorSize;
4387 // If we optimize the program for size, avoid creating the tail loop.
4389 // If we are unable to calculate the trip count then don't try to vectorize.
4392 (VectorizationReport() <<
4393 "unable to calculate the loop count due to complex control flow");
4394 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4398 // Find the maximum SIMD width that can fit within the trip count.
4399 VF = TC % MaxVectorSize;
4404 // If the trip count that we found modulo the vectorization factor is not
4405 // zero then we require a tail.
4407 emitAnalysis(VectorizationReport() <<
4408 "cannot optimize for size and vectorize at the "
4409 "same time. Enable vectorization of this loop "
4410 "with '#pragma clang loop vectorize(enable)' "
4411 "when compiling with -Os");
4412 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4417 int UserVF = Hints->getWidth();
4419 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4420 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4422 Factor.Width = UserVF;
4426 float Cost = expectedCost(1);
4428 const float ScalarCost = Cost;
4431 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
4433 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
4434 // Ignore scalar width, because the user explicitly wants vectorization.
4435 if (ForceVectorization && VF > 1) {
4437 Cost = expectedCost(Width) / (float)Width;
4440 for (unsigned i=2; i <= VF; i*=2) {
4441 // Notice that the vector loop needs to be executed less times, so
4442 // we need to divide the cost of the vector loops by the width of
4443 // the vector elements.
4444 float VectorCost = expectedCost(i) / (float)i;
4445 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
4446 (int)VectorCost << ".\n");
4447 if (VectorCost < Cost) {
4453 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
4454 << "LV: Vectorization seems to be not beneficial, "
4455 << "but was forced by a user.\n");
4456 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
4457 Factor.Width = Width;
4458 Factor.Cost = Width * Cost;
4462 unsigned LoopVectorizationCostModel::getWidestType() {
4463 unsigned MaxWidth = 8;
4466 for (Loop::block_iterator bb = TheLoop->block_begin(),
4467 be = TheLoop->block_end(); bb != be; ++bb) {
4468 BasicBlock *BB = *bb;
4470 // For each instruction in the loop.
4471 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4472 Type *T = it->getType();
4474 // Ignore ephemeral values.
4475 if (EphValues.count(it))
4478 // Only examine Loads, Stores and PHINodes.
4479 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4482 // Examine PHI nodes that are reduction variables.
4483 if (PHINode *PN = dyn_cast<PHINode>(it))
4484 if (!Legal->getReductionVars()->count(PN))
4487 // Examine the stored values.
4488 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4489 T = ST->getValueOperand()->getType();
4491 // Ignore loaded pointer types and stored pointer types that are not
4492 // consecutive. However, we do want to take consecutive stores/loads of
4493 // pointer vectors into account.
4494 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4497 MaxWidth = std::max(MaxWidth,
4498 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
4506 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
4508 unsigned LoopCost) {
4510 // -- The unroll heuristics --
4511 // We unroll the loop in order to expose ILP and reduce the loop overhead.
4512 // There are many micro-architectural considerations that we can't predict
4513 // at this level. For example, frontend pressure (on decode or fetch) due to
4514 // code size, or the number and capabilities of the execution ports.
4516 // We use the following heuristics to select the unroll factor:
4517 // 1. If the code has reductions, then we unroll in order to break the cross
4518 // iteration dependency.
4519 // 2. If the loop is really small, then we unroll in order to reduce the loop
4521 // 3. We don't unroll if we think that we will spill registers to memory due
4522 // to the increased register pressure.
4524 // Use the user preference, unless 'auto' is selected.
4525 int UserUF = Hints->getInterleave();
4529 // When we optimize for size, we don't unroll.
4533 // We used the distance for the unroll factor.
4534 if (Legal->getMaxSafeDepDistBytes() != -1U)
4537 // Do not unroll loops with a relatively small trip count.
4538 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4539 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
4542 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
4543 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
4547 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
4548 TargetNumRegisters = ForceTargetNumScalarRegs;
4550 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
4551 TargetNumRegisters = ForceTargetNumVectorRegs;
4554 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4555 // We divide by these constants so assume that we have at least one
4556 // instruction that uses at least one register.
4557 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4558 R.NumInstructions = std::max(R.NumInstructions, 1U);
4560 // We calculate the unroll factor using the following formula.
4561 // Subtract the number of loop invariants from the number of available
4562 // registers. These registers are used by all of the unrolled instances.
4563 // Next, divide the remaining registers by the number of registers that is
4564 // required by the loop, in order to estimate how many parallel instances
4565 // fit without causing spills. All of this is rounded down if necessary to be
4566 // a power of two. We want power of two unroll factors to simplify any
4567 // addressing operations or alignment considerations.
4568 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
4571 // Don't count the induction variable as unrolled.
4572 if (EnableIndVarRegisterHeur)
4573 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
4574 std::max(1U, (R.MaxLocalUsers - 1)));
4576 // Clamp the unroll factor ranges to reasonable factors.
4577 unsigned MaxInterleaveSize = TTI.getMaxInterleaveFactor();
4579 // Check if the user has overridden the unroll max.
4581 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
4582 MaxInterleaveSize = ForceTargetMaxScalarInterleaveFactor;
4584 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
4585 MaxInterleaveSize = ForceTargetMaxVectorInterleaveFactor;
4588 // If we did not calculate the cost for VF (because the user selected the VF)
4589 // then we calculate the cost of VF here.
4591 LoopCost = expectedCost(VF);
4593 // Clamp the calculated UF to be between the 1 and the max unroll factor
4594 // that the target allows.
4595 if (UF > MaxInterleaveSize)
4596 UF = MaxInterleaveSize;
4600 // Unroll if we vectorized this loop and there is a reduction that could
4601 // benefit from unrolling.
4602 if (VF > 1 && Legal->getReductionVars()->size()) {
4603 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
4607 // Note that if we've already vectorized the loop we will have done the
4608 // runtime check and so unrolling won't require further checks.
4609 bool UnrollingRequiresRuntimePointerCheck =
4610 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
4612 // We want to unroll small loops in order to reduce the loop overhead and
4613 // potentially expose ILP opportunities.
4614 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
4615 if (!UnrollingRequiresRuntimePointerCheck &&
4616 LoopCost < SmallLoopCost) {
4617 // We assume that the cost overhead is 1 and we use the cost model
4618 // to estimate the cost of the loop and unroll until the cost of the
4619 // loop overhead is about 5% of the cost of the loop.
4620 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
4622 // Unroll until store/load ports (estimated by max unroll factor) are
4624 unsigned NumStores = Legal->getNumStores();
4625 unsigned NumLoads = Legal->getNumLoads();
4626 unsigned StoresUF = UF / (NumStores ? NumStores : 1);
4627 unsigned LoadsUF = UF / (NumLoads ? NumLoads : 1);
4629 // If we have a scalar reduction (vector reductions are already dealt with
4630 // by this point), we can increase the critical path length if the loop
4631 // we're unrolling is inside another loop. Limit, by default to 2, so the
4632 // critical path only gets increased by one reduction operation.
4633 if (Legal->getReductionVars()->size() &&
4634 TheLoop->getLoopDepth() > 1) {
4635 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionUF);
4636 SmallUF = std::min(SmallUF, F);
4637 StoresUF = std::min(StoresUF, F);
4638 LoadsUF = std::min(LoadsUF, F);
4641 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
4642 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
4643 return std::max(StoresUF, LoadsUF);
4646 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
4650 DEBUG(dbgs() << "LV: Not Unrolling.\n");
4654 LoopVectorizationCostModel::RegisterUsage
4655 LoopVectorizationCostModel::calculateRegisterUsage() {
4656 // This function calculates the register usage by measuring the highest number
4657 // of values that are alive at a single location. Obviously, this is a very
4658 // rough estimation. We scan the loop in a topological order in order and
4659 // assign a number to each instruction. We use RPO to ensure that defs are
4660 // met before their users. We assume that each instruction that has in-loop
4661 // users starts an interval. We record every time that an in-loop value is
4662 // used, so we have a list of the first and last occurrences of each
4663 // instruction. Next, we transpose this data structure into a multi map that
4664 // holds the list of intervals that *end* at a specific location. This multi
4665 // map allows us to perform a linear search. We scan the instructions linearly
4666 // and record each time that a new interval starts, by placing it in a set.
4667 // If we find this value in the multi-map then we remove it from the set.
4668 // The max register usage is the maximum size of the set.
4669 // We also search for instructions that are defined outside the loop, but are
4670 // used inside the loop. We need this number separately from the max-interval
4671 // usage number because when we unroll, loop-invariant values do not take
4673 LoopBlocksDFS DFS(TheLoop);
4677 R.NumInstructions = 0;
4679 // Each 'key' in the map opens a new interval. The values
4680 // of the map are the index of the 'last seen' usage of the
4681 // instruction that is the key.
4682 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4683 // Maps instruction to its index.
4684 DenseMap<unsigned, Instruction*> IdxToInstr;
4685 // Marks the end of each interval.
4686 IntervalMap EndPoint;
4687 // Saves the list of instruction indices that are used in the loop.
4688 SmallSet<Instruction*, 8> Ends;
4689 // Saves the list of values that are used in the loop but are
4690 // defined outside the loop, such as arguments and constants.
4691 SmallPtrSet<Value*, 8> LoopInvariants;
4694 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4695 be = DFS.endRPO(); bb != be; ++bb) {
4696 R.NumInstructions += (*bb)->size();
4697 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4699 Instruction *I = it;
4700 IdxToInstr[Index++] = I;
4702 // Save the end location of each USE.
4703 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4704 Value *U = I->getOperand(i);
4705 Instruction *Instr = dyn_cast<Instruction>(U);
4707 // Ignore non-instruction values such as arguments, constants, etc.
4708 if (!Instr) continue;
4710 // If this instruction is outside the loop then record it and continue.
4711 if (!TheLoop->contains(Instr)) {
4712 LoopInvariants.insert(Instr);
4716 // Overwrite previous end points.
4717 EndPoint[Instr] = Index;
4723 // Saves the list of intervals that end with the index in 'key'.
4724 typedef SmallVector<Instruction*, 2> InstrList;
4725 DenseMap<unsigned, InstrList> TransposeEnds;
4727 // Transpose the EndPoints to a list of values that end at each index.
4728 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4730 TransposeEnds[it->second].push_back(it->first);
4732 SmallSet<Instruction*, 8> OpenIntervals;
4733 unsigned MaxUsage = 0;
4736 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4737 for (unsigned int i = 0; i < Index; ++i) {
4738 Instruction *I = IdxToInstr[i];
4739 // Ignore instructions that are never used within the loop.
4740 if (!Ends.count(I)) continue;
4742 // Ignore ephemeral values.
4743 if (EphValues.count(I))
4746 // Remove all of the instructions that end at this location.
4747 InstrList &List = TransposeEnds[i];
4748 for (unsigned int j=0, e = List.size(); j < e; ++j)
4749 OpenIntervals.erase(List[j]);
4751 // Count the number of live interals.
4752 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4754 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4755 OpenIntervals.size() << '\n');
4757 // Add the current instruction to the list of open intervals.
4758 OpenIntervals.insert(I);
4761 unsigned Invariant = LoopInvariants.size();
4762 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
4763 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
4764 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
4766 R.LoopInvariantRegs = Invariant;
4767 R.MaxLocalUsers = MaxUsage;
4771 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4775 for (Loop::block_iterator bb = TheLoop->block_begin(),
4776 be = TheLoop->block_end(); bb != be; ++bb) {
4777 unsigned BlockCost = 0;
4778 BasicBlock *BB = *bb;
4780 // For each instruction in the old loop.
4781 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4782 // Skip dbg intrinsics.
4783 if (isa<DbgInfoIntrinsic>(it))
4786 // Ignore ephemeral values.
4787 if (EphValues.count(it))
4790 unsigned C = getInstructionCost(it, VF);
4792 // Check if we should override the cost.
4793 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
4794 C = ForceTargetInstructionCost;
4797 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
4798 VF << " For instruction: " << *it << '\n');
4801 // We assume that if-converted blocks have a 50% chance of being executed.
4802 // When the code is scalar then some of the blocks are avoided due to CF.
4803 // When the code is vectorized we execute all code paths.
4804 if (VF == 1 && Legal->blockNeedsPredication(*bb))
4813 /// \brief Check whether the address computation for a non-consecutive memory
4814 /// access looks like an unlikely candidate for being merged into the indexing
4817 /// We look for a GEP which has one index that is an induction variable and all
4818 /// other indices are loop invariant. If the stride of this access is also
4819 /// within a small bound we decide that this address computation can likely be
4820 /// merged into the addressing mode.
4821 /// In all other cases, we identify the address computation as complex.
4822 static bool isLikelyComplexAddressComputation(Value *Ptr,
4823 LoopVectorizationLegality *Legal,
4824 ScalarEvolution *SE,
4825 const Loop *TheLoop) {
4826 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
4830 // We are looking for a gep with all loop invariant indices except for one
4831 // which should be an induction variable.
4832 unsigned NumOperands = Gep->getNumOperands();
4833 for (unsigned i = 1; i < NumOperands; ++i) {
4834 Value *Opd = Gep->getOperand(i);
4835 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
4836 !Legal->isInductionVariable(Opd))
4840 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
4841 // can likely be merged into the address computation.
4842 unsigned MaxMergeDistance = 64;
4844 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
4848 // Check the step is constant.
4849 const SCEV *Step = AddRec->getStepRecurrence(*SE);
4850 // Calculate the pointer stride and check if it is consecutive.
4851 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4855 const APInt &APStepVal = C->getValue()->getValue();
4857 // Huge step value - give up.
4858 if (APStepVal.getBitWidth() > 64)
4861 int64_t StepVal = APStepVal.getSExtValue();
4863 return StepVal > MaxMergeDistance;
4866 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
4867 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
4873 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
4874 // If we know that this instruction will remain uniform, check the cost of
4875 // the scalar version.
4876 if (Legal->isUniformAfterVectorization(I))
4879 Type *RetTy = I->getType();
4880 Type *VectorTy = ToVectorTy(RetTy, VF);
4882 // TODO: We need to estimate the cost of intrinsic calls.
4883 switch (I->getOpcode()) {
4884 case Instruction::GetElementPtr:
4885 // We mark this instruction as zero-cost because the cost of GEPs in
4886 // vectorized code depends on whether the corresponding memory instruction
4887 // is scalarized or not. Therefore, we handle GEPs with the memory
4888 // instruction cost.
4890 case Instruction::Br: {
4891 return TTI.getCFInstrCost(I->getOpcode());
4893 case Instruction::PHI:
4894 //TODO: IF-converted IFs become selects.
4896 case Instruction::Add:
4897 case Instruction::FAdd:
4898 case Instruction::Sub:
4899 case Instruction::FSub:
4900 case Instruction::Mul:
4901 case Instruction::FMul:
4902 case Instruction::UDiv:
4903 case Instruction::SDiv:
4904 case Instruction::FDiv:
4905 case Instruction::URem:
4906 case Instruction::SRem:
4907 case Instruction::FRem:
4908 case Instruction::Shl:
4909 case Instruction::LShr:
4910 case Instruction::AShr:
4911 case Instruction::And:
4912 case Instruction::Or:
4913 case Instruction::Xor: {
4914 // Since we will replace the stride by 1 the multiplication should go away.
4915 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
4917 // Certain instructions can be cheaper to vectorize if they have a constant
4918 // second vector operand. One example of this are shifts on x86.
4919 TargetTransformInfo::OperandValueKind Op1VK =
4920 TargetTransformInfo::OK_AnyValue;
4921 TargetTransformInfo::OperandValueKind Op2VK =
4922 TargetTransformInfo::OK_AnyValue;
4923 TargetTransformInfo::OperandValueProperties Op1VP =
4924 TargetTransformInfo::OP_None;
4925 TargetTransformInfo::OperandValueProperties Op2VP =
4926 TargetTransformInfo::OP_None;
4927 Value *Op2 = I->getOperand(1);
4929 // Check for a splat of a constant or for a non uniform vector of constants.
4930 if (isa<ConstantInt>(Op2)) {
4931 ConstantInt *CInt = cast<ConstantInt>(Op2);
4932 if (CInt && CInt->getValue().isPowerOf2())
4933 Op2VP = TargetTransformInfo::OP_PowerOf2;
4934 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4935 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
4936 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
4937 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
4939 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
4940 if (CInt && CInt->getValue().isPowerOf2())
4941 Op2VP = TargetTransformInfo::OP_PowerOf2;
4942 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4946 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
4949 case Instruction::Select: {
4950 SelectInst *SI = cast<SelectInst>(I);
4951 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
4952 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
4953 Type *CondTy = SI->getCondition()->getType();
4955 CondTy = VectorType::get(CondTy, VF);
4957 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
4959 case Instruction::ICmp:
4960 case Instruction::FCmp: {
4961 Type *ValTy = I->getOperand(0)->getType();
4962 VectorTy = ToVectorTy(ValTy, VF);
4963 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
4965 case Instruction::Store:
4966 case Instruction::Load: {
4967 StoreInst *SI = dyn_cast<StoreInst>(I);
4968 LoadInst *LI = dyn_cast<LoadInst>(I);
4969 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
4971 VectorTy = ToVectorTy(ValTy, VF);
4973 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
4974 unsigned AS = SI ? SI->getPointerAddressSpace() :
4975 LI->getPointerAddressSpace();
4976 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
4977 // We add the cost of address computation here instead of with the gep
4978 // instruction because only here we know whether the operation is
4981 return TTI.getAddressComputationCost(VectorTy) +
4982 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4984 // Scalarized loads/stores.
4985 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
4986 bool Reverse = ConsecutiveStride < 0;
4987 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
4988 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
4989 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
4990 bool IsComplexComputation =
4991 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
4993 // The cost of extracting from the value vector and pointer vector.
4994 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
4995 for (unsigned i = 0; i < VF; ++i) {
4996 // The cost of extracting the pointer operand.
4997 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
4998 // In case of STORE, the cost of ExtractElement from the vector.
4999 // In case of LOAD, the cost of InsertElement into the returned
5001 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5002 Instruction::InsertElement,
5006 // The cost of the scalar loads/stores.
5007 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5008 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5013 // Wide load/stores.
5014 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5015 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5018 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5022 case Instruction::ZExt:
5023 case Instruction::SExt:
5024 case Instruction::FPToUI:
5025 case Instruction::FPToSI:
5026 case Instruction::FPExt:
5027 case Instruction::PtrToInt:
5028 case Instruction::IntToPtr:
5029 case Instruction::SIToFP:
5030 case Instruction::UIToFP:
5031 case Instruction::Trunc:
5032 case Instruction::FPTrunc:
5033 case Instruction::BitCast: {
5034 // We optimize the truncation of induction variable.
5035 // The cost of these is the same as the scalar operation.
5036 if (I->getOpcode() == Instruction::Trunc &&
5037 Legal->isInductionVariable(I->getOperand(0)))
5038 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5039 I->getOperand(0)->getType());
5041 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5042 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5044 case Instruction::Call: {
5045 CallInst *CI = cast<CallInst>(I);
5046 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
5047 assert(ID && "Not an intrinsic call!");
5048 Type *RetTy = ToVectorTy(CI->getType(), VF);
5049 SmallVector<Type*, 4> Tys;
5050 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
5051 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
5052 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
5055 // We are scalarizing the instruction. Return the cost of the scalar
5056 // instruction, plus the cost of insert and extract into vector
5057 // elements, times the vector width.
5060 if (!RetTy->isVoidTy() && VF != 1) {
5061 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5063 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5066 // The cost of inserting the results plus extracting each one of the
5068 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5071 // The cost of executing VF copies of the scalar instruction. This opcode
5072 // is unknown. Assume that it is the same as 'mul'.
5073 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5079 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
5080 if (Scalar->isVoidTy() || VF == 1)
5082 return VectorType::get(Scalar, VF);
5085 char LoopVectorize::ID = 0;
5086 static const char lv_name[] = "Loop Vectorization";
5087 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5088 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
5089 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
5090 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
5091 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
5092 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5093 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5094 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5095 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
5096 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5097 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5100 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5101 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5105 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5106 // Check for a store.
5107 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5108 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5110 // Check for a load.
5111 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5112 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5118 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5119 bool IfPredicateStore) {
5120 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5121 // Holds vector parameters or scalars, in case of uniform vals.
5122 SmallVector<VectorParts, 4> Params;
5124 setDebugLocFromInst(Builder, Instr);
5126 // Find all of the vectorized parameters.
5127 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5128 Value *SrcOp = Instr->getOperand(op);
5130 // If we are accessing the old induction variable, use the new one.
5131 if (SrcOp == OldInduction) {
5132 Params.push_back(getVectorValue(SrcOp));
5136 // Try using previously calculated values.
5137 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5139 // If the src is an instruction that appeared earlier in the basic block
5140 // then it should already be vectorized.
5141 if (SrcInst && OrigLoop->contains(SrcInst)) {
5142 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5143 // The parameter is a vector value from earlier.
5144 Params.push_back(WidenMap.get(SrcInst));
5146 // The parameter is a scalar from outside the loop. Maybe even a constant.
5147 VectorParts Scalars;
5148 Scalars.append(UF, SrcOp);
5149 Params.push_back(Scalars);
5153 assert(Params.size() == Instr->getNumOperands() &&
5154 "Invalid number of operands");
5156 // Does this instruction return a value ?
5157 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5159 Value *UndefVec = IsVoidRetTy ? nullptr :
5160 UndefValue::get(Instr->getType());
5161 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5162 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5164 Instruction *InsertPt = Builder.GetInsertPoint();
5165 BasicBlock *IfBlock = Builder.GetInsertBlock();
5166 BasicBlock *CondBlock = nullptr;
5169 Loop *VectorLp = nullptr;
5170 if (IfPredicateStore) {
5171 assert(Instr->getParent()->getSinglePredecessor() &&
5172 "Only support single predecessor blocks");
5173 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5174 Instr->getParent());
5175 VectorLp = LI->getLoopFor(IfBlock);
5176 assert(VectorLp && "Must have a loop for this block");
5179 // For each vector unroll 'part':
5180 for (unsigned Part = 0; Part < UF; ++Part) {
5181 // For each scalar that we create:
5183 // Start an "if (pred) a[i] = ..." block.
5184 Value *Cmp = nullptr;
5185 if (IfPredicateStore) {
5186 if (Cond[Part]->getType()->isVectorTy())
5188 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5189 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5190 ConstantInt::get(Cond[Part]->getType(), 1));
5191 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5192 LoopVectorBody.push_back(CondBlock);
5193 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
5194 // Update Builder with newly created basic block.
5195 Builder.SetInsertPoint(InsertPt);
5198 Instruction *Cloned = Instr->clone();
5200 Cloned->setName(Instr->getName() + ".cloned");
5201 // Replace the operands of the cloned instructions with extracted scalars.
5202 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5203 Value *Op = Params[op][Part];
5204 Cloned->setOperand(op, Op);
5207 // Place the cloned scalar in the new loop.
5208 Builder.Insert(Cloned);
5210 // If the original scalar returns a value we need to place it in a vector
5211 // so that future users will be able to use it.
5213 VecResults[Part] = Cloned;
5216 if (IfPredicateStore) {
5217 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
5218 LoopVectorBody.push_back(NewIfBlock);
5219 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
5220 Builder.SetInsertPoint(InsertPt);
5221 Instruction *OldBr = IfBlock->getTerminator();
5222 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
5223 OldBr->eraseFromParent();
5224 IfBlock = NewIfBlock;
5229 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5230 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5231 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5233 return scalarizeInstruction(Instr, IfPredicateStore);
5236 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5240 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5244 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) {
5245 // When unrolling and the VF is 1, we only need to add a simple scalar.
5246 Type *ITy = Val->getType();
5247 assert(!ITy->isVectorTy() && "Val must be a scalar");
5248 Constant *C = ConstantInt::get(ITy, StartIdx);
5249 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");