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/AssumptionTracker.h"
59 #include "llvm/Analysis/BlockFrequencyInfo.h"
60 #include "llvm/Analysis/CodeMetrics.h"
61 #include "llvm/Analysis/LoopInfo.h"
62 #include "llvm/Analysis/LoopIterator.h"
63 #include "llvm/Analysis/LoopPass.h"
64 #include "llvm/Analysis/ScalarEvolution.h"
65 #include "llvm/Analysis/ScalarEvolutionExpander.h"
66 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
67 #include "llvm/Analysis/TargetTransformInfo.h"
68 #include "llvm/Analysis/ValueTracking.h"
69 #include "llvm/IR/Constants.h"
70 #include "llvm/IR/DataLayout.h"
71 #include "llvm/IR/DebugInfo.h"
72 #include "llvm/IR/DerivedTypes.h"
73 #include "llvm/IR/DiagnosticInfo.h"
74 #include "llvm/IR/Dominators.h"
75 #include "llvm/IR/Function.h"
76 #include "llvm/IR/IRBuilder.h"
77 #include "llvm/IR/Instructions.h"
78 #include "llvm/IR/IntrinsicInst.h"
79 #include "llvm/IR/LLVMContext.h"
80 #include "llvm/IR/Module.h"
81 #include "llvm/IR/PatternMatch.h"
82 #include "llvm/IR/Type.h"
83 #include "llvm/IR/Value.h"
84 #include "llvm/IR/ValueHandle.h"
85 #include "llvm/IR/Verifier.h"
86 #include "llvm/Pass.h"
87 #include "llvm/Support/BranchProbability.h"
88 #include "llvm/Support/CommandLine.h"
89 #include "llvm/Support/Debug.h"
90 #include "llvm/Support/raw_ostream.h"
91 #include "llvm/Transforms/Scalar.h"
92 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
93 #include "llvm/Transforms/Utils/Local.h"
94 #include "llvm/Transforms/Utils/VectorUtils.h"
100 using namespace llvm::PatternMatch;
102 #define LV_NAME "loop-vectorize"
103 #define DEBUG_TYPE LV_NAME
105 STATISTIC(LoopsVectorized, "Number of loops vectorized");
106 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
108 static cl::opt<unsigned>
109 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
110 cl::desc("Sets the SIMD width. Zero is autoselect."));
112 static cl::opt<unsigned>
113 VectorizationInterleave("force-vector-interleave", cl::init(0), cl::Hidden,
114 cl::desc("Sets the vectorization interleave count. "
115 "Zero is autoselect."));
118 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
119 cl::desc("Enable if-conversion during vectorization."));
121 /// We don't vectorize loops with a known constant trip count below this number.
122 static cl::opt<unsigned>
123 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
125 cl::desc("Don't vectorize loops with a constant "
126 "trip count that is smaller than this "
129 /// This enables versioning on the strides of symbolically striding memory
130 /// accesses in code like the following.
131 /// for (i = 0; i < N; ++i)
132 /// A[i * Stride1] += B[i * Stride2] ...
134 /// Will be roughly translated to
135 /// if (Stride1 == 1 && Stride2 == 1) {
136 /// for (i = 0; i < N; i+=4)
140 static cl::opt<bool> EnableMemAccessVersioning(
141 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
142 cl::desc("Enable symblic stride memory access versioning"));
144 /// We don't unroll loops with a known constant trip count below this number.
145 static const unsigned TinyTripCountUnrollThreshold = 128;
147 /// When performing memory disambiguation checks at runtime do not make more
148 /// than this number of comparisons.
149 static const unsigned RuntimeMemoryCheckThreshold = 8;
151 /// Maximum simd width.
152 static const unsigned MaxVectorWidth = 64;
154 static cl::opt<unsigned> ForceTargetNumScalarRegs(
155 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
156 cl::desc("A flag that overrides the target's number of scalar registers."));
158 static cl::opt<unsigned> ForceTargetNumVectorRegs(
159 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
160 cl::desc("A flag that overrides the target's number of vector registers."));
162 /// Maximum vectorization interleave count.
163 static const unsigned MaxInterleaveFactor = 16;
165 static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
166 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
167 cl::desc("A flag that overrides the target's max interleave factor for "
170 static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
171 "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
172 cl::desc("A flag that overrides the target's max interleave factor for "
173 "vectorized loops."));
175 static cl::opt<unsigned> ForceTargetInstructionCost(
176 "force-target-instruction-cost", cl::init(0), cl::Hidden,
177 cl::desc("A flag that overrides the target's expected cost for "
178 "an instruction to a single constant value. Mostly "
179 "useful for getting consistent testing."));
181 static cl::opt<unsigned> SmallLoopCost(
182 "small-loop-cost", cl::init(20), cl::Hidden,
183 cl::desc("The cost of a loop that is considered 'small' by the unroller."));
185 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
186 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
187 cl::desc("Enable the use of the block frequency analysis to access PGO "
188 "heuristics minimizing code growth in cold regions and being more "
189 "aggressive in hot regions."));
191 // Runtime unroll loops for load/store throughput.
192 static cl::opt<bool> EnableLoadStoreRuntimeUnroll(
193 "enable-loadstore-runtime-unroll", cl::init(true), cl::Hidden,
194 cl::desc("Enable runtime unrolling until load/store ports are saturated"));
196 /// The number of stores in a loop that are allowed to need predication.
197 static cl::opt<unsigned> NumberOfStoresToPredicate(
198 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
199 cl::desc("Max number of stores to be predicated behind an if."));
201 static cl::opt<bool> EnableIndVarRegisterHeur(
202 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
203 cl::desc("Count the induction variable only once when unrolling"));
205 static cl::opt<bool> EnableCondStoresVectorization(
206 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
207 cl::desc("Enable if predication of stores during vectorization."));
209 static cl::opt<unsigned> MaxNestedScalarReductionUF(
210 "max-nested-scalar-reduction-unroll", cl::init(2), cl::Hidden,
211 cl::desc("The maximum unroll factor to use when unrolling a scalar "
212 "reduction in a nested loop."));
216 // Forward declarations.
217 class LoopVectorizationLegality;
218 class LoopVectorizationCostModel;
219 class LoopVectorizeHints;
221 /// Optimization analysis message produced during vectorization. Messages inform
222 /// the user why vectorization did not occur.
225 raw_string_ostream Out;
229 Report(Instruction *I = nullptr) : Out(Message), Instr(I) {
230 Out << "loop not vectorized: ";
233 template <typename A> Report &operator<<(const A &Value) {
238 Instruction *getInstr() { return Instr; }
240 std::string &str() { return Out.str(); }
241 operator Twine() { return Out.str(); }
244 /// InnerLoopVectorizer vectorizes loops which contain only one basic
245 /// block to a specified vectorization factor (VF).
246 /// This class performs the widening of scalars into vectors, or multiple
247 /// scalars. This class also implements the following features:
248 /// * It inserts an epilogue loop for handling loops that don't have iteration
249 /// counts that are known to be a multiple of the vectorization factor.
250 /// * It handles the code generation for reduction variables.
251 /// * Scalarization (implementation using scalars) of un-vectorizable
253 /// InnerLoopVectorizer does not perform any vectorization-legality
254 /// checks, and relies on the caller to check for the different legality
255 /// aspects. The InnerLoopVectorizer relies on the
256 /// LoopVectorizationLegality class to provide information about the induction
257 /// and reduction variables that were found to a given vectorization factor.
258 class InnerLoopVectorizer {
260 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
261 DominatorTree *DT, const DataLayout *DL,
262 const TargetLibraryInfo *TLI, unsigned VecWidth,
263 unsigned UnrollFactor)
264 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
265 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
266 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
269 // Perform the actual loop widening (vectorization).
270 void vectorize(LoopVectorizationLegality *L) {
272 // Create a new empty loop. Unlink the old loop and connect the new one.
274 // Widen each instruction in the old loop to a new one in the new loop.
275 // Use the Legality module to find the induction and reduction variables.
277 // Register the new loop and update the analysis passes.
281 virtual ~InnerLoopVectorizer() {}
284 /// A small list of PHINodes.
285 typedef SmallVector<PHINode*, 4> PhiVector;
286 /// When we unroll loops we have multiple vector values for each scalar.
287 /// This data structure holds the unrolled and vectorized values that
288 /// originated from one scalar instruction.
289 typedef SmallVector<Value*, 2> VectorParts;
291 // When we if-convert we need create edge masks. We have to cache values so
292 // that we don't end up with exponential recursion/IR.
293 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
294 VectorParts> EdgeMaskCache;
296 /// \brief Add code that checks at runtime if the accessed arrays overlap.
298 /// Returns a pair of instructions where the first element is the first
299 /// instruction generated in possibly a sequence of instructions and the
300 /// second value is the final comparator value or NULL if no check is needed.
301 std::pair<Instruction *, Instruction *> addRuntimeCheck(Instruction *Loc);
303 /// \brief Add checks for strides that where assumed to be 1.
305 /// Returns the last check instruction and the first check instruction in the
306 /// pair as (first, last).
307 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
309 /// Create an empty loop, based on the loop ranges of the old loop.
310 void createEmptyLoop();
311 /// Copy and widen the instructions from the old loop.
312 virtual void vectorizeLoop();
314 /// \brief The Loop exit block may have single value PHI nodes where the
315 /// incoming value is 'Undef'. While vectorizing we only handled real values
316 /// that were defined inside the loop. Here we fix the 'undef case'.
320 /// A helper function that computes the predicate of the block BB, assuming
321 /// that the header block of the loop is set to True. It returns the *entry*
322 /// mask for the block BB.
323 VectorParts createBlockInMask(BasicBlock *BB);
324 /// A helper function that computes the predicate of the edge between SRC
326 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
328 /// A helper function to vectorize a single BB within the innermost loop.
329 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
331 /// Vectorize a single PHINode in a block. This method handles the induction
332 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
333 /// arbitrary length vectors.
334 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
335 unsigned UF, unsigned VF, PhiVector *PV);
337 /// Insert the new loop to the loop hierarchy and pass manager
338 /// and update the analysis passes.
339 void updateAnalysis();
341 /// This instruction is un-vectorizable. Implement it as a sequence
342 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
343 /// scalarized instruction behind an if block predicated on the control
344 /// dependence of the instruction.
345 virtual void scalarizeInstruction(Instruction *Instr,
346 bool IfPredicateStore=false);
348 /// Vectorize Load and Store instructions,
349 virtual void vectorizeMemoryInstruction(Instruction *Instr);
351 /// Create a broadcast instruction. This method generates a broadcast
352 /// instruction (shuffle) for loop invariant values and for the induction
353 /// value. If this is the induction variable then we extend it to N, N+1, ...
354 /// this is needed because each iteration in the loop corresponds to a SIMD
356 virtual Value *getBroadcastInstrs(Value *V);
358 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
359 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
360 /// The sequence starts at StartIndex.
361 virtual Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
363 /// When we go over instructions in the basic block we rely on previous
364 /// values within the current basic block or on loop invariant values.
365 /// When we widen (vectorize) values we place them in the map. If the values
366 /// are not within the map, they have to be loop invariant, so we simply
367 /// broadcast them into a vector.
368 VectorParts &getVectorValue(Value *V);
370 /// Generate a shuffle sequence that will reverse the vector Vec.
371 virtual Value *reverseVector(Value *Vec);
373 /// This is a helper class that holds the vectorizer state. It maps scalar
374 /// instructions to vector instructions. When the code is 'unrolled' then
375 /// then a single scalar value is mapped to multiple vector parts. The parts
376 /// are stored in the VectorPart type.
378 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
380 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
382 /// \return True if 'Key' is saved in the Value Map.
383 bool has(Value *Key) const { return MapStorage.count(Key); }
385 /// Initializes a new entry in the map. Sets all of the vector parts to the
386 /// save value in 'Val'.
387 /// \return A reference to a vector with splat values.
388 VectorParts &splat(Value *Key, Value *Val) {
389 VectorParts &Entry = MapStorage[Key];
390 Entry.assign(UF, Val);
394 ///\return A reference to the value that is stored at 'Key'.
395 VectorParts &get(Value *Key) {
396 VectorParts &Entry = MapStorage[Key];
399 assert(Entry.size() == UF);
404 /// The unroll factor. Each entry in the map stores this number of vector
408 /// Map storage. We use std::map and not DenseMap because insertions to a
409 /// dense map invalidates its iterators.
410 std::map<Value *, VectorParts> MapStorage;
413 /// The original loop.
415 /// Scev analysis to use.
424 const DataLayout *DL;
425 /// Target Library Info.
426 const TargetLibraryInfo *TLI;
428 /// The vectorization SIMD factor to use. Each vector will have this many
433 /// The vectorization unroll factor to use. Each scalar is vectorized to this
434 /// many different vector instructions.
437 /// The builder that we use
440 // --- Vectorization state ---
442 /// The vector-loop preheader.
443 BasicBlock *LoopVectorPreHeader;
444 /// The scalar-loop preheader.
445 BasicBlock *LoopScalarPreHeader;
446 /// Middle Block between the vector and the scalar.
447 BasicBlock *LoopMiddleBlock;
448 ///The ExitBlock of the scalar loop.
449 BasicBlock *LoopExitBlock;
450 ///The vector loop body.
451 SmallVector<BasicBlock *, 4> LoopVectorBody;
452 ///The scalar loop body.
453 BasicBlock *LoopScalarBody;
454 /// A list of all bypass blocks. The first block is the entry of the loop.
455 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
457 /// The new Induction variable which was added to the new block.
459 /// The induction variable of the old basic block.
460 PHINode *OldInduction;
461 /// Holds the extended (to the widest induction type) start index.
463 /// Maps scalars to widened vectors.
465 EdgeMaskCache MaskCache;
467 LoopVectorizationLegality *Legal;
470 class InnerLoopUnroller : public InnerLoopVectorizer {
472 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
473 DominatorTree *DT, const DataLayout *DL,
474 const TargetLibraryInfo *TLI, unsigned UnrollFactor) :
475 InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { }
478 void scalarizeInstruction(Instruction *Instr,
479 bool IfPredicateStore = false) override;
480 void vectorizeMemoryInstruction(Instruction *Instr) override;
481 Value *getBroadcastInstrs(Value *V) override;
482 Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate) override;
483 Value *reverseVector(Value *Vec) override;
486 /// \brief Look for a meaningful debug location on the instruction or it's
488 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
493 if (I->getDebugLoc() != Empty)
496 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
497 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
498 if (OpInst->getDebugLoc() != Empty)
505 /// \brief Set the debug location in the builder using the debug location in the
507 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
508 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
509 B.SetCurrentDebugLocation(Inst->getDebugLoc());
511 B.SetCurrentDebugLocation(DebugLoc());
515 /// \return string containing a file name and a line # for the given loop.
516 static std::string getDebugLocString(const Loop *L) {
519 raw_string_ostream OS(Result);
520 const DebugLoc LoopDbgLoc = L->getStartLoc();
521 if (!LoopDbgLoc.isUnknown())
522 LoopDbgLoc.print(L->getHeader()->getContext(), OS);
524 // Just print the module name.
525 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
532 /// \brief Propagate known metadata from one instruction to another.
533 static void propagateMetadata(Instruction *To, const Instruction *From) {
534 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
535 From->getAllMetadataOtherThanDebugLoc(Metadata);
537 for (auto M : Metadata) {
538 unsigned Kind = M.first;
540 // These are safe to transfer (this is safe for TBAA, even when we
541 // if-convert, because should that metadata have had a control dependency
542 // on the condition, and thus actually aliased with some other
543 // non-speculated memory access when the condition was false, this would be
544 // caught by the runtime overlap checks).
545 if (Kind != LLVMContext::MD_tbaa &&
546 Kind != LLVMContext::MD_alias_scope &&
547 Kind != LLVMContext::MD_noalias &&
548 Kind != LLVMContext::MD_fpmath)
551 To->setMetadata(Kind, M.second);
555 /// \brief Propagate known metadata from one instruction to a vector of others.
556 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
558 if (Instruction *I = dyn_cast<Instruction>(V))
559 propagateMetadata(I, From);
562 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
563 /// to what vectorization factor.
564 /// This class does not look at the profitability of vectorization, only the
565 /// legality. This class has two main kinds of checks:
566 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
567 /// will change the order of memory accesses in a way that will change the
568 /// correctness of the program.
569 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
570 /// checks for a number of different conditions, such as the availability of a
571 /// single induction variable, that all types are supported and vectorize-able,
572 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
573 /// This class is also used by InnerLoopVectorizer for identifying
574 /// induction variable and the different reduction variables.
575 class LoopVectorizationLegality {
579 unsigned NumPredStores;
581 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, const DataLayout *DL,
582 DominatorTree *DT, TargetLibraryInfo *TLI,
583 AliasAnalysis *AA, Function *F,
584 const TargetTransformInfo *TTI)
585 : NumLoads(0), NumStores(0), NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
586 DT(DT), TLI(TLI), AA(AA), TheFunction(F), TTI(TTI), Induction(nullptr),
587 WidestIndTy(nullptr), HasFunNoNaNAttr(false), MaxSafeDepDistBytes(-1U) {
590 /// This enum represents the kinds of reductions that we support.
592 RK_NoReduction, ///< Not a reduction.
593 RK_IntegerAdd, ///< Sum of integers.
594 RK_IntegerMult, ///< Product of integers.
595 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
596 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
597 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
598 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
599 RK_FloatAdd, ///< Sum of floats.
600 RK_FloatMult, ///< Product of floats.
601 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
604 /// This enum represents the kinds of inductions that we support.
606 IK_NoInduction, ///< Not an induction variable.
607 IK_IntInduction, ///< Integer induction variable. Step = 1.
608 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
609 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
610 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
613 // This enum represents the kind of minmax reduction.
614 enum MinMaxReductionKind {
624 /// This struct holds information about reduction variables.
625 struct ReductionDescriptor {
626 ReductionDescriptor() : StartValue(nullptr), LoopExitInstr(nullptr),
627 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
629 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
630 MinMaxReductionKind MK)
631 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
633 // The starting value of the reduction.
634 // It does not have to be zero!
635 TrackingVH<Value> StartValue;
636 // The instruction who's value is used outside the loop.
637 Instruction *LoopExitInstr;
638 // The kind of the reduction.
640 // If this a min/max reduction the kind of reduction.
641 MinMaxReductionKind MinMaxKind;
644 /// This POD struct holds information about a potential reduction operation.
645 struct ReductionInstDesc {
646 ReductionInstDesc(bool IsRedux, Instruction *I) :
647 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
649 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
650 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
652 // Is this instruction a reduction candidate.
654 // The last instruction in a min/max pattern (select of the select(icmp())
655 // pattern), or the current reduction instruction otherwise.
656 Instruction *PatternLastInst;
657 // If this is a min/max pattern the comparison predicate.
658 MinMaxReductionKind MinMaxKind;
661 /// This struct holds information about the memory runtime legality
662 /// check that a group of pointers do not overlap.
663 struct RuntimePointerCheck {
664 RuntimePointerCheck() : Need(false) {}
666 /// Reset the state of the pointer runtime information.
673 DependencySetId.clear();
677 /// Insert a pointer and calculate the start and end SCEVs.
678 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
679 unsigned DepSetId, unsigned ASId, ValueToValueMap &Strides);
681 /// This flag indicates if we need to add the runtime check.
683 /// Holds the pointers that we need to check.
684 SmallVector<TrackingVH<Value>, 2> Pointers;
685 /// Holds the pointer value at the beginning of the loop.
686 SmallVector<const SCEV*, 2> Starts;
687 /// Holds the pointer value at the end of the loop.
688 SmallVector<const SCEV*, 2> Ends;
689 /// Holds the information if this pointer is used for writing to memory.
690 SmallVector<bool, 2> IsWritePtr;
691 /// Holds the id of the set of pointers that could be dependent because of a
692 /// shared underlying object.
693 SmallVector<unsigned, 2> DependencySetId;
694 /// Holds the id of the disjoint alias set to which this pointer belongs.
695 SmallVector<unsigned, 2> AliasSetId;
698 /// A struct for saving information about induction variables.
699 struct InductionInfo {
700 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
701 InductionInfo() : StartValue(nullptr), IK(IK_NoInduction) {}
703 TrackingVH<Value> StartValue;
708 /// ReductionList contains the reduction descriptors for all
709 /// of the reductions that were found in the loop.
710 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
712 /// InductionList saves induction variables and maps them to the
713 /// induction descriptor.
714 typedef MapVector<PHINode*, InductionInfo> InductionList;
716 /// Returns true if it is legal to vectorize this loop.
717 /// This does not mean that it is profitable to vectorize this
718 /// loop, only that it is legal to do so.
721 /// Returns the Induction variable.
722 PHINode *getInduction() { return Induction; }
724 /// Returns the reduction variables found in the loop.
725 ReductionList *getReductionVars() { return &Reductions; }
727 /// Returns the induction variables found in the loop.
728 InductionList *getInductionVars() { return &Inductions; }
730 /// Returns the widest induction type.
731 Type *getWidestInductionType() { return WidestIndTy; }
733 /// Returns True if V is an induction variable in this loop.
734 bool isInductionVariable(const Value *V);
736 /// Return true if the block BB needs to be predicated in order for the loop
737 /// to be vectorized.
738 bool blockNeedsPredication(BasicBlock *BB);
740 /// Check if this pointer is consecutive when vectorizing. This happens
741 /// when the last index of the GEP is the induction variable, or that the
742 /// pointer itself is an induction variable.
743 /// This check allows us to vectorize A[idx] into a wide load/store.
745 /// 0 - Stride is unknown or non-consecutive.
746 /// 1 - Address is consecutive.
747 /// -1 - Address is consecutive, and decreasing.
748 int isConsecutivePtr(Value *Ptr);
750 /// Returns true if the value V is uniform within the loop.
751 bool isUniform(Value *V);
753 /// Returns true if this instruction will remain scalar after vectorization.
754 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
756 /// Returns the information that we collected about runtime memory check.
757 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
759 /// This function returns the identity element (or neutral element) for
761 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
763 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
765 bool hasStride(Value *V) { return StrideSet.count(V); }
766 bool mustCheckStrides() { return !StrideSet.empty(); }
767 SmallPtrSet<Value *, 8>::iterator strides_begin() {
768 return StrideSet.begin();
770 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
772 /// Returns true if the target machine supports masked store operation
773 /// for the given \p DataType and kind of access to \p Ptr.
774 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
775 return TTI->isLegalMaskedStore(DataType, isConsecutivePtr(Ptr));
777 /// Returns true if the target machine supports masked load operation
778 /// for the given \p DataType and kind of access to \p Ptr.
779 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
780 return TTI->isLegalMaskedLoad(DataType, isConsecutivePtr(Ptr));
782 /// Returns true if vector representation of the instruction \p I
784 bool isMaskRequired(const Instruction* I) {
785 return (MaskedOp.count(I) != 0);
788 /// Check if a single basic block loop is vectorizable.
789 /// At this point we know that this is a loop with a constant trip count
790 /// and we only need to check individual instructions.
791 bool canVectorizeInstrs();
793 /// When we vectorize loops we may change the order in which
794 /// we read and write from memory. This method checks if it is
795 /// legal to vectorize the code, considering only memory constrains.
796 /// Returns true if the loop is vectorizable
797 bool canVectorizeMemory();
799 /// Return true if we can vectorize this loop using the IF-conversion
801 bool canVectorizeWithIfConvert();
803 /// Collect the variables that need to stay uniform after vectorization.
804 void collectLoopUniforms();
806 /// Return true if all of the instructions in the block can be speculatively
807 /// executed. \p SafePtrs is a list of addresses that are known to be legal
808 /// and we know that we can read from them without segfault.
809 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
811 /// Returns True, if 'Phi' is the kind of reduction variable for type
812 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
813 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
814 /// Returns a struct describing if the instruction 'I' can be a reduction
815 /// variable of type 'Kind'. If the reduction is a min/max pattern of
816 /// select(icmp()) this function advances the instruction pointer 'I' from the
817 /// compare instruction to the select instruction and stores this pointer in
818 /// 'PatternLastInst' member of the returned struct.
819 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
820 ReductionInstDesc &Desc);
821 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
822 /// pattern corresponding to a min(X, Y) or max(X, Y).
823 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
824 ReductionInstDesc &Prev);
825 /// Returns the induction kind of Phi. This function may return NoInduction
826 /// if the PHI is not an induction variable.
827 InductionKind isInductionVariable(PHINode *Phi);
829 /// \brief Collect memory access with loop invariant strides.
831 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
833 void collectStridedAcccess(Value *LoadOrStoreInst);
835 /// Report an analysis message to assist the user in diagnosing loops that are
837 void emitAnalysis(Report &Message) {
838 DebugLoc DL = TheLoop->getStartLoc();
839 if (Instruction *I = Message.getInstr())
840 DL = I->getDebugLoc();
841 emitOptimizationRemarkAnalysis(TheFunction->getContext(), DEBUG_TYPE,
842 *TheFunction, DL, Message.str());
845 /// The loop that we evaluate.
849 /// DataLayout analysis.
850 const DataLayout *DL;
853 /// Target Library Info.
854 TargetLibraryInfo *TLI;
858 Function *TheFunction;
859 /// Target Transform Info
860 const TargetTransformInfo *TTI;
862 // --- vectorization state --- //
864 /// Holds the integer induction variable. This is the counter of the
867 /// Holds the reduction variables.
868 ReductionList Reductions;
869 /// Holds all of the induction variables that we found in the loop.
870 /// Notice that inductions don't need to start at zero and that induction
871 /// variables can be pointers.
872 InductionList Inductions;
873 /// Holds the widest induction type encountered.
876 /// Allowed outside users. This holds the reduction
877 /// vars which can be accessed from outside the loop.
878 SmallPtrSet<Value*, 4> AllowedExit;
879 /// This set holds the variables which are known to be uniform after
881 SmallPtrSet<Instruction*, 4> Uniforms;
882 /// We need to check that all of the pointers in this list are disjoint
884 RuntimePointerCheck PtrRtCheck;
885 /// Can we assume the absence of NaNs.
886 bool HasFunNoNaNAttr;
888 unsigned MaxSafeDepDistBytes;
890 ValueToValueMap Strides;
891 SmallPtrSet<Value *, 8> StrideSet;
893 /// While vectorizing these instructions we have to generate a
894 /// call to the appropriate masked intrinsic
895 SmallPtrSet<const Instruction*, 8> MaskedOp;
898 /// LoopVectorizationCostModel - estimates the expected speedups due to
900 /// In many cases vectorization is not profitable. This can happen because of
901 /// a number of reasons. In this class we mainly attempt to predict the
902 /// expected speedup/slowdowns due to the supported instruction set. We use the
903 /// TargetTransformInfo to query the different backends for the cost of
904 /// different operations.
905 class LoopVectorizationCostModel {
907 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
908 LoopVectorizationLegality *Legal,
909 const TargetTransformInfo &TTI,
910 const DataLayout *DL, const TargetLibraryInfo *TLI,
911 AssumptionTracker *AT, const Function *F,
912 const LoopVectorizeHints *Hints)
913 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI),
914 TheFunction(F), Hints(Hints) {
915 CodeMetrics::collectEphemeralValues(L, AT, EphValues);
918 /// Information about vectorization costs
919 struct VectorizationFactor {
920 unsigned Width; // Vector width with best cost
921 unsigned Cost; // Cost of the loop with that width
923 /// \return The most profitable vectorization factor and the cost of that VF.
924 /// This method checks every power of two up to VF. If UserVF is not ZERO
925 /// then this vectorization factor will be selected if vectorization is
927 VectorizationFactor selectVectorizationFactor(bool OptForSize);
929 /// \return The size (in bits) of the widest type in the code that
930 /// needs to be vectorized. We ignore values that remain scalar such as
931 /// 64 bit loop indices.
932 unsigned getWidestType();
934 /// \return The most profitable unroll factor.
935 /// If UserUF is non-zero then this method finds the best unroll-factor
936 /// based on register pressure and other parameters.
937 /// VF and LoopCost are the selected vectorization factor and the cost of the
939 unsigned selectUnrollFactor(bool OptForSize, unsigned VF, unsigned LoopCost);
941 /// \brief A struct that represents some properties of the register usage
943 struct RegisterUsage {
944 /// Holds the number of loop invariant values that are used in the loop.
945 unsigned LoopInvariantRegs;
946 /// Holds the maximum number of concurrent live intervals in the loop.
947 unsigned MaxLocalUsers;
948 /// Holds the number of instructions in the loop.
949 unsigned NumInstructions;
952 /// \return information about the register usage of the loop.
953 RegisterUsage calculateRegisterUsage();
956 /// Returns the expected execution cost. The unit of the cost does
957 /// not matter because we use the 'cost' units to compare different
958 /// vector widths. The cost that is returned is *not* normalized by
959 /// the factor width.
960 unsigned expectedCost(unsigned VF);
962 /// Returns the execution time cost of an instruction for a given vector
963 /// width. Vector width of one means scalar.
964 unsigned getInstructionCost(Instruction *I, unsigned VF);
966 /// A helper function for converting Scalar types to vector types.
967 /// If the incoming type is void, we return void. If the VF is 1, we return
969 static Type* ToVectorTy(Type *Scalar, unsigned VF);
971 /// Returns whether the instruction is a load or store and will be a emitted
972 /// as a vector operation.
973 bool isConsecutiveLoadOrStore(Instruction *I);
975 /// Report an analysis message to assist the user in diagnosing loops that are
977 void emitAnalysis(Report &Message) {
978 DebugLoc DL = TheLoop->getStartLoc();
979 if (Instruction *I = Message.getInstr())
980 DL = I->getDebugLoc();
981 emitOptimizationRemarkAnalysis(TheFunction->getContext(), DEBUG_TYPE,
982 *TheFunction, DL, Message.str());
985 /// Values used only by @llvm.assume calls.
986 SmallPtrSet<const Value *, 32> EphValues;
988 /// The loop that we evaluate.
992 /// Loop Info analysis.
994 /// Vectorization legality.
995 LoopVectorizationLegality *Legal;
996 /// Vector target information.
997 const TargetTransformInfo &TTI;
998 /// Target data layout information.
999 const DataLayout *DL;
1000 /// Target Library Info.
1001 const TargetLibraryInfo *TLI;
1002 const Function *TheFunction;
1003 // Loop Vectorize Hint.
1004 const LoopVectorizeHints *Hints;
1007 /// Utility class for getting and setting loop vectorizer hints in the form
1008 /// of loop metadata.
1009 /// This class keeps a number of loop annotations locally (as member variables)
1010 /// and can, upon request, write them back as metadata on the loop. It will
1011 /// initially scan the loop for existing metadata, and will update the local
1012 /// values based on information in the loop.
1013 /// We cannot write all values to metadata, as the mere presence of some info,
1014 /// for example 'force', means a decision has been made. So, we need to be
1015 /// careful NOT to add them if the user hasn't specifically asked so.
1016 class LoopVectorizeHints {
1023 /// Hint - associates name and validation with the hint value.
1026 unsigned Value; // This may have to change for non-numeric values.
1029 Hint(const char * Name, unsigned Value, HintKind Kind)
1030 : Name(Name), Value(Value), Kind(Kind) { }
1032 bool validate(unsigned Val) {
1035 return isPowerOf2_32(Val) && Val <= MaxVectorWidth;
1037 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1045 /// Vectorization width.
1047 /// Vectorization interleave factor.
1049 /// Vectorization forced
1052 /// Return the loop metadata prefix.
1053 static StringRef Prefix() { return "llvm.loop."; }
1057 FK_Undefined = -1, ///< Not selected.
1058 FK_Disabled = 0, ///< Forcing disabled.
1059 FK_Enabled = 1, ///< Forcing enabled.
1062 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
1063 : Width("vectorize.width", VectorizationFactor, HK_WIDTH),
1064 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1065 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1067 // Populate values with existing loop metadata.
1068 getHintsFromMetadata();
1070 // force-vector-interleave overrides DisableInterleaving.
1071 if (VectorizationInterleave.getNumOccurrences() > 0)
1072 Interleave.Value = VectorizationInterleave;
1074 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
1075 << "LV: Interleaving disabled by the pass manager\n");
1078 /// Mark the loop L as already vectorized by setting the width to 1.
1079 void setAlreadyVectorized() {
1080 Width.Value = Interleave.Value = 1;
1081 Hint Hints[] = {Width, Interleave};
1082 writeHintsToMetadata(Hints);
1085 /// Dumps all the hint information.
1086 std::string emitRemark() const {
1088 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1089 R << "vectorization is explicitly disabled";
1091 R << "use -Rpass-analysis=loop-vectorize for more info";
1092 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1093 R << " (Force=true";
1094 if (Width.Value != 0)
1095 R << ", Vector Width=" << Width.Value;
1096 if (Interleave.Value != 0)
1097 R << ", Interleave Count=" << Interleave.Value;
1105 unsigned getWidth() const { return Width.Value; }
1106 unsigned getInterleave() const { return Interleave.Value; }
1107 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1110 /// Find hints specified in the loop metadata and update local values.
1111 void getHintsFromMetadata() {
1112 MDNode *LoopID = TheLoop->getLoopID();
1116 // First operand should refer to the loop id itself.
1117 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1118 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1120 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1121 const MDString *S = nullptr;
1122 SmallVector<Metadata *, 4> Args;
1124 // The expected hint is either a MDString or a MDNode with the first
1125 // operand a MDString.
1126 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1127 if (!MD || MD->getNumOperands() == 0)
1129 S = dyn_cast<MDString>(MD->getOperand(0));
1130 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1131 Args.push_back(MD->getOperand(i));
1133 S = dyn_cast<MDString>(LoopID->getOperand(i));
1134 assert(Args.size() == 0 && "too many arguments for MDString");
1140 // Check if the hint starts with the loop metadata prefix.
1141 StringRef Name = S->getString();
1142 if (Args.size() == 1)
1143 setHint(Name, Args[0]);
1147 /// Checks string hint with one operand and set value if valid.
1148 void setHint(StringRef Name, Metadata *Arg) {
1149 if (!Name.startswith(Prefix()))
1151 Name = Name.substr(Prefix().size(), StringRef::npos);
1153 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1155 unsigned Val = C->getZExtValue();
1157 Hint *Hints[] = {&Width, &Interleave, &Force};
1158 for (auto H : Hints) {
1159 if (Name == H->Name) {
1160 if (H->validate(Val))
1163 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1169 /// Create a new hint from name / value pair.
1170 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1171 LLVMContext &Context = TheLoop->getHeader()->getContext();
1172 Metadata *MDs[] = {MDString::get(Context, Name),
1173 ConstantAsMetadata::get(
1174 ConstantInt::get(Type::getInt32Ty(Context), V))};
1175 return MDNode::get(Context, MDs);
1178 /// Matches metadata with hint name.
1179 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1180 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1184 for (auto H : HintTypes)
1185 if (Name->getString().endswith(H.Name))
1190 /// Sets current hints into loop metadata, keeping other values intact.
1191 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1192 if (HintTypes.size() == 0)
1195 // Reserve the first element to LoopID (see below).
1196 SmallVector<Metadata *, 4> MDs(1);
1197 // If the loop already has metadata, then ignore the existing operands.
1198 MDNode *LoopID = TheLoop->getLoopID();
1200 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1201 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1202 // If node in update list, ignore old value.
1203 if (!matchesHintMetadataName(Node, HintTypes))
1204 MDs.push_back(Node);
1208 // Now, add the missing hints.
1209 for (auto H : HintTypes)
1210 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1212 // Replace current metadata node with new one.
1213 LLVMContext &Context = TheLoop->getHeader()->getContext();
1214 MDNode *NewLoopID = MDNode::get(Context, MDs);
1215 // Set operand 0 to refer to the loop id itself.
1216 NewLoopID->replaceOperandWith(0, NewLoopID);
1218 TheLoop->setLoopID(NewLoopID);
1221 /// The loop these hints belong to.
1222 const Loop *TheLoop;
1225 static void emitMissedWarning(Function *F, Loop *L,
1226 const LoopVectorizeHints &LH) {
1227 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1228 L->getStartLoc(), LH.emitRemark());
1230 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1231 if (LH.getWidth() != 1)
1232 emitLoopVectorizeWarning(
1233 F->getContext(), *F, L->getStartLoc(),
1234 "failed explicitly specified loop vectorization");
1235 else if (LH.getInterleave() != 1)
1236 emitLoopInterleaveWarning(
1237 F->getContext(), *F, L->getStartLoc(),
1238 "failed explicitly specified loop interleaving");
1242 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1244 return V.push_back(&L);
1246 for (Loop *InnerL : L)
1247 addInnerLoop(*InnerL, V);
1250 /// The LoopVectorize Pass.
1251 struct LoopVectorize : public FunctionPass {
1252 /// Pass identification, replacement for typeid
1255 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1257 DisableUnrolling(NoUnrolling),
1258 AlwaysVectorize(AlwaysVectorize) {
1259 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1262 ScalarEvolution *SE;
1263 const DataLayout *DL;
1265 TargetTransformInfo *TTI;
1267 BlockFrequencyInfo *BFI;
1268 TargetLibraryInfo *TLI;
1270 AssumptionTracker *AT;
1271 bool DisableUnrolling;
1272 bool AlwaysVectorize;
1274 BlockFrequency ColdEntryFreq;
1276 bool runOnFunction(Function &F) override {
1277 SE = &getAnalysis<ScalarEvolution>();
1278 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1279 DL = DLP ? &DLP->getDataLayout() : nullptr;
1280 LI = &getAnalysis<LoopInfo>();
1281 TTI = &getAnalysis<TargetTransformInfo>();
1282 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1283 BFI = &getAnalysis<BlockFrequencyInfo>();
1284 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
1285 AA = &getAnalysis<AliasAnalysis>();
1286 AT = &getAnalysis<AssumptionTracker>();
1288 // Compute some weights outside of the loop over the loops. Compute this
1289 // using a BranchProbability to re-use its scaling math.
1290 const BranchProbability ColdProb(1, 5); // 20%
1291 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1293 // If the target claims to have no vector registers don't attempt
1295 if (!TTI->getNumberOfRegisters(true))
1299 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1300 << ": Missing data layout\n");
1304 // Build up a worklist of inner-loops to vectorize. This is necessary as
1305 // the act of vectorizing or partially unrolling a loop creates new loops
1306 // and can invalidate iterators across the loops.
1307 SmallVector<Loop *, 8> Worklist;
1310 addInnerLoop(*L, Worklist);
1312 LoopsAnalyzed += Worklist.size();
1314 // Now walk the identified inner loops.
1315 bool Changed = false;
1316 while (!Worklist.empty())
1317 Changed |= processLoop(Worklist.pop_back_val());
1319 // Process each loop nest in the function.
1323 bool processLoop(Loop *L) {
1324 assert(L->empty() && "Only process inner loops.");
1327 const std::string DebugLocStr = getDebugLocString(L);
1330 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1331 << L->getHeader()->getParent()->getName() << "\" from "
1332 << DebugLocStr << "\n");
1334 LoopVectorizeHints Hints(L, DisableUnrolling);
1336 DEBUG(dbgs() << "LV: Loop hints:"
1338 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1340 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1342 : "?")) << " width=" << Hints.getWidth()
1343 << " unroll=" << Hints.getInterleave() << "\n");
1345 // Function containing loop
1346 Function *F = L->getHeader()->getParent();
1348 // Looking at the diagnostic output is the only way to determine if a loop
1349 // was vectorized (other than looking at the IR or machine code), so it
1350 // is important to generate an optimization remark for each loop. Most of
1351 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1352 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1353 // less verbose reporting vectorized loops and unvectorized loops that may
1354 // benefit from vectorization, respectively.
1356 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1357 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1358 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1359 L->getStartLoc(), Hints.emitRemark());
1363 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1364 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1365 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1366 L->getStartLoc(), Hints.emitRemark());
1370 if (Hints.getWidth() == 1 && Hints.getInterleave() == 1) {
1371 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1372 emitOptimizationRemarkAnalysis(
1373 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1374 "loop not vectorized: vector width and interleave count are "
1375 "explicitly set to 1");
1379 // Check the loop for a trip count threshold:
1380 // do not vectorize loops with a tiny trip count.
1381 const unsigned TC = SE->getSmallConstantTripCount(L);
1382 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1383 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1384 << "This loop is not worth vectorizing.");
1385 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1386 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1388 DEBUG(dbgs() << "\n");
1389 emitOptimizationRemarkAnalysis(
1390 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1391 "vectorization is not beneficial and is not explicitly forced");
1396 // Check if it is legal to vectorize the loop.
1397 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, AA, F, TTI);
1398 if (!LVL.canVectorize()) {
1399 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1400 emitMissedWarning(F, L, Hints);
1404 // Use the cost model.
1405 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI, AT, F,
1408 // Check the function attributes to find out if this function should be
1409 // optimized for size.
1410 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1411 F->hasFnAttribute(Attribute::OptimizeForSize);
1413 // Compute the weighted frequency of this loop being executed and see if it
1414 // is less than 20% of the function entry baseline frequency. Note that we
1415 // always have a canonical loop here because we think we *can* vectoriez.
1416 // FIXME: This is hidden behind a flag due to pervasive problems with
1417 // exactly what block frequency models.
1418 if (LoopVectorizeWithBlockFrequency) {
1419 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1420 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1421 LoopEntryFreq < ColdEntryFreq)
1425 // Check the function attributes to see if implicit floats are allowed.a
1426 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1427 // an integer loop and the vector instructions selected are purely integer
1428 // vector instructions?
1429 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1430 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1431 "attribute is used.\n");
1432 emitOptimizationRemarkAnalysis(
1433 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1434 "loop not vectorized due to NoImplicitFloat attribute");
1435 emitMissedWarning(F, L, Hints);
1439 // Select the optimal vectorization factor.
1440 const LoopVectorizationCostModel::VectorizationFactor VF =
1441 CM.selectVectorizationFactor(OptForSize);
1443 // Select the unroll factor.
1445 CM.selectUnrollFactor(OptForSize, VF.Width, VF.Cost);
1447 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1448 << DebugLocStr << '\n');
1449 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1451 if (VF.Width == 1) {
1452 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
1455 emitOptimizationRemarkAnalysis(
1456 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1457 "not beneficial to vectorize and user disabled interleaving");
1460 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1462 // Report the unrolling decision.
1463 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1464 Twine("unrolled with interleaving factor " +
1466 " (vectorization not beneficial)"));
1468 // We decided not to vectorize, but we may want to unroll.
1470 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1471 Unroller.vectorize(&LVL);
1473 // If we decided that it is *legal* to vectorize the loop then do it.
1474 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1478 // Report the vectorization decision.
1479 emitOptimizationRemark(
1480 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1481 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1482 ", unrolling interleave factor: " + Twine(UF) + ")");
1485 // Mark the loop as already vectorized to avoid vectorizing again.
1486 Hints.setAlreadyVectorized();
1488 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1492 void getAnalysisUsage(AnalysisUsage &AU) const override {
1493 AU.addRequired<AssumptionTracker>();
1494 AU.addRequiredID(LoopSimplifyID);
1495 AU.addRequiredID(LCSSAID);
1496 AU.addRequired<BlockFrequencyInfo>();
1497 AU.addRequired<DominatorTreeWrapperPass>();
1498 AU.addRequired<LoopInfo>();
1499 AU.addRequired<ScalarEvolution>();
1500 AU.addRequired<TargetTransformInfo>();
1501 AU.addRequired<AliasAnalysis>();
1502 AU.addPreserved<LoopInfo>();
1503 AU.addPreserved<DominatorTreeWrapperPass>();
1504 AU.addPreserved<AliasAnalysis>();
1509 } // end anonymous namespace
1511 //===----------------------------------------------------------------------===//
1512 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1513 // LoopVectorizationCostModel.
1514 //===----------------------------------------------------------------------===//
1516 static Value *stripIntegerCast(Value *V) {
1517 if (CastInst *CI = dyn_cast<CastInst>(V))
1518 if (CI->getOperand(0)->getType()->isIntegerTy())
1519 return CI->getOperand(0);
1523 ///\brief Replaces the symbolic stride in a pointer SCEV expression by one.
1525 /// If \p OrigPtr is not null, use it to look up the stride value instead of
1527 static const SCEV *replaceSymbolicStrideSCEV(ScalarEvolution *SE,
1528 ValueToValueMap &PtrToStride,
1529 Value *Ptr, Value *OrigPtr = nullptr) {
1531 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
1533 // If there is an entry in the map return the SCEV of the pointer with the
1534 // symbolic stride replaced by one.
1535 ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
1536 if (SI != PtrToStride.end()) {
1537 Value *StrideVal = SI->second;
1540 StrideVal = stripIntegerCast(StrideVal);
1542 // Replace symbolic stride by one.
1543 Value *One = ConstantInt::get(StrideVal->getType(), 1);
1544 ValueToValueMap RewriteMap;
1545 RewriteMap[StrideVal] = One;
1548 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
1549 DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
1554 // Otherwise, just return the SCEV of the original pointer.
1555 return SE->getSCEV(Ptr);
1558 void LoopVectorizationLegality::RuntimePointerCheck::insert(
1559 ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
1560 unsigned ASId, ValueToValueMap &Strides) {
1561 // Get the stride replaced scev.
1562 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
1563 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1564 assert(AR && "Invalid addrec expression");
1565 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1566 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1567 Pointers.push_back(Ptr);
1568 Starts.push_back(AR->getStart());
1569 Ends.push_back(ScEnd);
1570 IsWritePtr.push_back(WritePtr);
1571 DependencySetId.push_back(DepSetId);
1572 AliasSetId.push_back(ASId);
1575 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1576 // We need to place the broadcast of invariant variables outside the loop.
1577 Instruction *Instr = dyn_cast<Instruction>(V);
1579 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1580 Instr->getParent()) != LoopVectorBody.end());
1581 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1583 // Place the code for broadcasting invariant variables in the new preheader.
1584 IRBuilder<>::InsertPointGuard Guard(Builder);
1586 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1588 // Broadcast the scalar into all locations in the vector.
1589 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1594 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1596 assert(Val->getType()->isVectorTy() && "Must be a vector");
1597 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1598 "Elem must be an integer");
1599 // Create the types.
1600 Type *ITy = Val->getType()->getScalarType();
1601 VectorType *Ty = cast<VectorType>(Val->getType());
1602 int VLen = Ty->getNumElements();
1603 SmallVector<Constant*, 8> Indices;
1605 // Create a vector of consecutive numbers from zero to VF.
1606 for (int i = 0; i < VLen; ++i) {
1607 int64_t Idx = Negate ? (-i) : i;
1608 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1611 // Add the consecutive indices to the vector value.
1612 Constant *Cv = ConstantVector::get(Indices);
1613 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1614 return Builder.CreateAdd(Val, Cv, "induction");
1617 /// \brief Find the operand of the GEP that should be checked for consecutive
1618 /// stores. This ignores trailing indices that have no effect on the final
1620 static unsigned getGEPInductionOperand(const DataLayout *DL,
1621 const GetElementPtrInst *Gep) {
1622 unsigned LastOperand = Gep->getNumOperands() - 1;
1623 unsigned GEPAllocSize = DL->getTypeAllocSize(
1624 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1626 // Walk backwards and try to peel off zeros.
1627 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1628 // Find the type we're currently indexing into.
1629 gep_type_iterator GEPTI = gep_type_begin(Gep);
1630 std::advance(GEPTI, LastOperand - 1);
1632 // If it's a type with the same allocation size as the result of the GEP we
1633 // can peel off the zero index.
1634 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1642 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1643 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1644 // Make sure that the pointer does not point to structs.
1645 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1648 // If this value is a pointer induction variable we know it is consecutive.
1649 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1650 if (Phi && Inductions.count(Phi)) {
1651 InductionInfo II = Inductions[Phi];
1652 if (IK_PtrInduction == II.IK)
1654 else if (IK_ReversePtrInduction == II.IK)
1658 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1662 unsigned NumOperands = Gep->getNumOperands();
1663 Value *GpPtr = Gep->getPointerOperand();
1664 // If this GEP value is a consecutive pointer induction variable and all of
1665 // the indices are constant then we know it is consecutive. We can
1666 Phi = dyn_cast<PHINode>(GpPtr);
1667 if (Phi && Inductions.count(Phi)) {
1669 // Make sure that the pointer does not point to structs.
1670 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1671 if (GepPtrType->getElementType()->isAggregateType())
1674 // Make sure that all of the index operands are loop invariant.
1675 for (unsigned i = 1; i < NumOperands; ++i)
1676 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1679 InductionInfo II = Inductions[Phi];
1680 if (IK_PtrInduction == II.IK)
1682 else if (IK_ReversePtrInduction == II.IK)
1686 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1688 // Check that all of the gep indices are uniform except for our induction
1690 for (unsigned i = 0; i != NumOperands; ++i)
1691 if (i != InductionOperand &&
1692 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1695 // We can emit wide load/stores only if the last non-zero index is the
1696 // induction variable.
1697 const SCEV *Last = nullptr;
1698 if (!Strides.count(Gep))
1699 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1701 // Because of the multiplication by a stride we can have a s/zext cast.
1702 // We are going to replace this stride by 1 so the cast is safe to ignore.
1704 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1705 // %0 = trunc i64 %indvars.iv to i32
1706 // %mul = mul i32 %0, %Stride1
1707 // %idxprom = zext i32 %mul to i64 << Safe cast.
1708 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1710 Last = replaceSymbolicStrideSCEV(SE, Strides,
1711 Gep->getOperand(InductionOperand), Gep);
1712 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1714 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1718 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1719 const SCEV *Step = AR->getStepRecurrence(*SE);
1721 // The memory is consecutive because the last index is consecutive
1722 // and all other indices are loop invariant.
1725 if (Step->isAllOnesValue())
1732 bool LoopVectorizationLegality::isUniform(Value *V) {
1733 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1736 InnerLoopVectorizer::VectorParts&
1737 InnerLoopVectorizer::getVectorValue(Value *V) {
1738 assert(V != Induction && "The new induction variable should not be used.");
1739 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1741 // If we have a stride that is replaced by one, do it here.
1742 if (Legal->hasStride(V))
1743 V = ConstantInt::get(V->getType(), 1);
1745 // If we have this scalar in the map, return it.
1746 if (WidenMap.has(V))
1747 return WidenMap.get(V);
1749 // If this scalar is unknown, assume that it is a constant or that it is
1750 // loop invariant. Broadcast V and save the value for future uses.
1751 Value *B = getBroadcastInstrs(V);
1752 return WidenMap.splat(V, B);
1755 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1756 assert(Vec->getType()->isVectorTy() && "Invalid type");
1757 SmallVector<Constant*, 8> ShuffleMask;
1758 for (unsigned i = 0; i < VF; ++i)
1759 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1761 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1762 ConstantVector::get(ShuffleMask),
1766 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1767 // Attempt to issue a wide load.
1768 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1769 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1771 assert((LI || SI) && "Invalid Load/Store instruction");
1773 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1774 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1775 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1776 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1777 // An alignment of 0 means target abi alignment. We need to use the scalar's
1778 // target abi alignment in such a case.
1780 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1781 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1782 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1783 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1785 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
1786 !Legal->isMaskRequired(SI))
1787 return scalarizeInstruction(Instr, true);
1789 if (ScalarAllocatedSize != VectorElementSize)
1790 return scalarizeInstruction(Instr);
1792 // If the pointer is loop invariant or if it is non-consecutive,
1793 // scalarize the load.
1794 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1795 bool Reverse = ConsecutiveStride < 0;
1796 bool UniformLoad = LI && Legal->isUniform(Ptr);
1797 if (!ConsecutiveStride || UniformLoad)
1798 return scalarizeInstruction(Instr);
1800 Constant *Zero = Builder.getInt32(0);
1801 VectorParts &Entry = WidenMap.get(Instr);
1803 // Handle consecutive loads/stores.
1804 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1805 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1806 setDebugLocFromInst(Builder, Gep);
1807 Value *PtrOperand = Gep->getPointerOperand();
1808 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1809 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1811 // Create the new GEP with the new induction variable.
1812 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1813 Gep2->setOperand(0, FirstBasePtr);
1814 Gep2->setName("gep.indvar.base");
1815 Ptr = Builder.Insert(Gep2);
1817 setDebugLocFromInst(Builder, Gep);
1818 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1819 OrigLoop) && "Base ptr must be invariant");
1821 // The last index does not have to be the induction. It can be
1822 // consecutive and be a function of the index. For example A[I+1];
1823 unsigned NumOperands = Gep->getNumOperands();
1824 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1825 // Create the new GEP with the new induction variable.
1826 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1828 for (unsigned i = 0; i < NumOperands; ++i) {
1829 Value *GepOperand = Gep->getOperand(i);
1830 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1832 // Update last index or loop invariant instruction anchored in loop.
1833 if (i == InductionOperand ||
1834 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1835 assert((i == InductionOperand ||
1836 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1837 "Must be last index or loop invariant");
1839 VectorParts &GEPParts = getVectorValue(GepOperand);
1840 Value *Index = GEPParts[0];
1841 Index = Builder.CreateExtractElement(Index, Zero);
1842 Gep2->setOperand(i, Index);
1843 Gep2->setName("gep.indvar.idx");
1846 Ptr = Builder.Insert(Gep2);
1848 // Use the induction element ptr.
1849 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1850 setDebugLocFromInst(Builder, Ptr);
1851 VectorParts &PtrVal = getVectorValue(Ptr);
1852 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1857 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1858 "We do not allow storing to uniform addresses");
1859 setDebugLocFromInst(Builder, SI);
1860 // We don't want to update the value in the map as it might be used in
1861 // another expression. So don't use a reference type for "StoredVal".
1862 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1864 for (unsigned Part = 0; Part < UF; ++Part) {
1865 // Calculate the pointer for the specific unroll-part.
1866 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1869 // If we store to reverse consecutive memory locations then we need
1870 // to reverse the order of elements in the stored value.
1871 StoredVal[Part] = reverseVector(StoredVal[Part]);
1872 // If the address is consecutive but reversed, then the
1873 // wide store needs to start at the last vector element.
1874 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1875 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1878 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1879 DataTy->getPointerTo(AddressSpace));
1882 if (Legal->isMaskRequired(SI)) {
1884 Builder.getInt8PtrTy(PartPtr->getType()->getPointerAddressSpace());
1886 Value *I8Ptr = Builder.CreateBitCast(PartPtr, I8PtrTy);
1888 VectorParts Cond = createBlockInMask(SI->getParent());
1889 SmallVector <Value *, 8> Ops;
1890 Ops.push_back(I8Ptr);
1891 Ops.push_back(StoredVal[Part]);
1892 Ops.push_back(Builder.getInt32(Alignment));
1893 Ops.push_back(Cond[Part]);
1894 NewSI = Builder.CreateMaskedStore(Ops);
1897 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
1898 propagateMetadata(NewSI, SI);
1904 assert(LI && "Must have a load instruction");
1905 setDebugLocFromInst(Builder, LI);
1906 for (unsigned Part = 0; Part < UF; ++Part) {
1907 // Calculate the pointer for the specific unroll-part.
1908 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1911 // If the address is consecutive but reversed, then the
1912 // wide load needs to start at the last vector element.
1913 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1914 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1918 if (Legal->isMaskRequired(LI)) {
1920 Builder.getInt8PtrTy(PartPtr->getType()->getPointerAddressSpace());
1922 Value *I8Ptr = Builder.CreateBitCast(PartPtr, I8PtrTy);
1924 VectorParts SrcMask = createBlockInMask(LI->getParent());
1925 SmallVector <Value *, 8> Ops;
1926 Ops.push_back(I8Ptr);
1927 Ops.push_back(UndefValue::get(DataTy));
1928 Ops.push_back(Builder.getInt32(Alignment));
1929 Ops.push_back(SrcMask[Part]);
1930 NewLI = Builder.CreateMaskedLoad(Ops);
1933 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1934 DataTy->getPointerTo(AddressSpace));
1935 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
1937 propagateMetadata(NewLI, LI);
1938 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
1942 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1943 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1944 // Holds vector parameters or scalars, in case of uniform vals.
1945 SmallVector<VectorParts, 4> Params;
1947 setDebugLocFromInst(Builder, Instr);
1949 // Find all of the vectorized parameters.
1950 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1951 Value *SrcOp = Instr->getOperand(op);
1953 // If we are accessing the old induction variable, use the new one.
1954 if (SrcOp == OldInduction) {
1955 Params.push_back(getVectorValue(SrcOp));
1959 // Try using previously calculated values.
1960 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1962 // If the src is an instruction that appeared earlier in the basic block
1963 // then it should already be vectorized.
1964 if (SrcInst && OrigLoop->contains(SrcInst)) {
1965 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1966 // The parameter is a vector value from earlier.
1967 Params.push_back(WidenMap.get(SrcInst));
1969 // The parameter is a scalar from outside the loop. Maybe even a constant.
1970 VectorParts Scalars;
1971 Scalars.append(UF, SrcOp);
1972 Params.push_back(Scalars);
1976 assert(Params.size() == Instr->getNumOperands() &&
1977 "Invalid number of operands");
1979 // Does this instruction return a value ?
1980 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1982 Value *UndefVec = IsVoidRetTy ? nullptr :
1983 UndefValue::get(VectorType::get(Instr->getType(), VF));
1984 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1985 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1987 Instruction *InsertPt = Builder.GetInsertPoint();
1988 BasicBlock *IfBlock = Builder.GetInsertBlock();
1989 BasicBlock *CondBlock = nullptr;
1992 Loop *VectorLp = nullptr;
1993 if (IfPredicateStore) {
1994 assert(Instr->getParent()->getSinglePredecessor() &&
1995 "Only support single predecessor blocks");
1996 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1997 Instr->getParent());
1998 VectorLp = LI->getLoopFor(IfBlock);
1999 assert(VectorLp && "Must have a loop for this block");
2002 // For each vector unroll 'part':
2003 for (unsigned Part = 0; Part < UF; ++Part) {
2004 // For each scalar that we create:
2005 for (unsigned Width = 0; Width < VF; ++Width) {
2008 Value *Cmp = nullptr;
2009 if (IfPredicateStore) {
2010 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
2011 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
2012 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
2013 LoopVectorBody.push_back(CondBlock);
2014 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
2015 // Update Builder with newly created basic block.
2016 Builder.SetInsertPoint(InsertPt);
2019 Instruction *Cloned = Instr->clone();
2021 Cloned->setName(Instr->getName() + ".cloned");
2022 // Replace the operands of the cloned instructions with extracted scalars.
2023 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2024 Value *Op = Params[op][Part];
2025 // Param is a vector. Need to extract the right lane.
2026 if (Op->getType()->isVectorTy())
2027 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
2028 Cloned->setOperand(op, Op);
2031 // Place the cloned scalar in the new loop.
2032 Builder.Insert(Cloned);
2034 // If the original scalar returns a value we need to place it in a vector
2035 // so that future users will be able to use it.
2037 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
2038 Builder.getInt32(Width));
2040 if (IfPredicateStore) {
2041 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2042 LoopVectorBody.push_back(NewIfBlock);
2043 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
2044 Builder.SetInsertPoint(InsertPt);
2045 Instruction *OldBr = IfBlock->getTerminator();
2046 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2047 OldBr->eraseFromParent();
2048 IfBlock = NewIfBlock;
2054 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
2058 if (Instruction *I = dyn_cast<Instruction>(V))
2059 return I->getParent() == Loc->getParent() ? I : nullptr;
2063 std::pair<Instruction *, Instruction *>
2064 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
2065 Instruction *tnullptr = nullptr;
2066 if (!Legal->mustCheckStrides())
2067 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2069 IRBuilder<> ChkBuilder(Loc);
2072 Value *Check = nullptr;
2073 Instruction *FirstInst = nullptr;
2074 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
2075 SE = Legal->strides_end();
2077 Value *Ptr = stripIntegerCast(*SI);
2078 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
2080 // Store the first instruction we create.
2081 FirstInst = getFirstInst(FirstInst, C, Loc);
2083 Check = ChkBuilder.CreateOr(Check, C);
2088 // We have to do this trickery because the IRBuilder might fold the check to a
2089 // constant expression in which case there is no Instruction anchored in a
2091 LLVMContext &Ctx = Loc->getContext();
2092 Instruction *TheCheck =
2093 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2094 ChkBuilder.Insert(TheCheck, "stride.not.one");
2095 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2097 return std::make_pair(FirstInst, TheCheck);
2100 std::pair<Instruction *, Instruction *>
2101 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
2102 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
2103 Legal->getRuntimePointerCheck();
2105 Instruction *tnullptr = nullptr;
2106 if (!PtrRtCheck->Need)
2107 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2109 unsigned NumPointers = PtrRtCheck->Pointers.size();
2110 SmallVector<TrackingVH<Value> , 2> Starts;
2111 SmallVector<TrackingVH<Value> , 2> Ends;
2113 LLVMContext &Ctx = Loc->getContext();
2114 SCEVExpander Exp(*SE, "induction");
2115 Instruction *FirstInst = nullptr;
2117 for (unsigned i = 0; i < NumPointers; ++i) {
2118 Value *Ptr = PtrRtCheck->Pointers[i];
2119 const SCEV *Sc = SE->getSCEV(Ptr);
2121 if (SE->isLoopInvariant(Sc, OrigLoop)) {
2122 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
2124 Starts.push_back(Ptr);
2125 Ends.push_back(Ptr);
2127 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
2128 unsigned AS = Ptr->getType()->getPointerAddressSpace();
2130 // Use this type for pointer arithmetic.
2131 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
2133 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
2134 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
2135 Starts.push_back(Start);
2136 Ends.push_back(End);
2140 IRBuilder<> ChkBuilder(Loc);
2141 // Our instructions might fold to a constant.
2142 Value *MemoryRuntimeCheck = nullptr;
2143 for (unsigned i = 0; i < NumPointers; ++i) {
2144 for (unsigned j = i+1; j < NumPointers; ++j) {
2145 // No need to check if two readonly pointers intersect.
2146 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
2149 // Only need to check pointers between two different dependency sets.
2150 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
2152 // Only need to check pointers in the same alias set.
2153 if (PtrRtCheck->AliasSetId[i] != PtrRtCheck->AliasSetId[j])
2156 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
2157 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
2159 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
2160 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
2161 "Trying to bounds check pointers with different address spaces");
2163 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
2164 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
2166 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
2167 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
2168 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
2169 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
2171 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
2172 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
2173 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
2174 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
2175 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
2176 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2177 if (MemoryRuntimeCheck) {
2178 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
2180 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2182 MemoryRuntimeCheck = IsConflict;
2186 // We have to do this trickery because the IRBuilder might fold the check to a
2187 // constant expression in which case there is no Instruction anchored in a
2189 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
2190 ConstantInt::getTrue(Ctx));
2191 ChkBuilder.Insert(Check, "memcheck.conflict");
2192 FirstInst = getFirstInst(FirstInst, Check, Loc);
2193 return std::make_pair(FirstInst, Check);
2196 void InnerLoopVectorizer::createEmptyLoop() {
2198 In this function we generate a new loop. The new loop will contain
2199 the vectorized instructions while the old loop will continue to run the
2202 [ ] <-- Back-edge taken count overflow check.
2205 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2208 || [ ] <-- vector pre header.
2212 || [ ]_| <-- vector loop.
2215 | >[ ] <--- middle-block.
2218 -|- >[ ] <--- new preheader.
2222 | [ ]_| <-- old scalar loop to handle remainder.
2225 >[ ] <-- exit block.
2229 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2230 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2231 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2232 assert(BypassBlock && "Invalid loop structure");
2233 assert(ExitBlock && "Must have an exit block");
2235 // Some loops have a single integer induction variable, while other loops
2236 // don't. One example is c++ iterators that often have multiple pointer
2237 // induction variables. In the code below we also support a case where we
2238 // don't have a single induction variable.
2239 OldInduction = Legal->getInduction();
2240 Type *IdxTy = Legal->getWidestInductionType();
2242 // Find the loop boundaries.
2243 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2244 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2246 // The exit count might have the type of i64 while the phi is i32. This can
2247 // happen if we have an induction variable that is sign extended before the
2248 // compare. The only way that we get a backedge taken count is that the
2249 // induction variable was signed and as such will not overflow. In such a case
2250 // truncation is legal.
2251 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2252 IdxTy->getPrimitiveSizeInBits())
2253 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2255 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2256 // Get the total trip count from the count by adding 1.
2257 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2258 SE->getConstant(BackedgeTakeCount->getType(), 1));
2260 // Expand the trip count and place the new instructions in the preheader.
2261 // Notice that the pre-header does not change, only the loop body.
2262 SCEVExpander Exp(*SE, "induction");
2264 // We need to test whether the backedge-taken count is uint##_max. Adding one
2265 // to it will cause overflow and an incorrect loop trip count in the vector
2266 // body. In case of overflow we want to directly jump to the scalar remainder
2268 Value *BackedgeCount =
2269 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2270 BypassBlock->getTerminator());
2271 if (BackedgeCount->getType()->isPointerTy())
2272 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2273 "backedge.ptrcnt.to.int",
2274 BypassBlock->getTerminator());
2275 Instruction *CheckBCOverflow =
2276 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2277 Constant::getAllOnesValue(BackedgeCount->getType()),
2278 "backedge.overflow", BypassBlock->getTerminator());
2280 // The loop index does not have to start at Zero. Find the original start
2281 // value from the induction PHI node. If we don't have an induction variable
2282 // then we know that it starts at zero.
2283 Builder.SetInsertPoint(BypassBlock->getTerminator());
2284 Value *StartIdx = ExtendedIdx = OldInduction ?
2285 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2287 ConstantInt::get(IdxTy, 0);
2289 // We need an instruction to anchor the overflow check on. StartIdx needs to
2290 // be defined before the overflow check branch. Because the scalar preheader
2291 // is going to merge the start index and so the overflow branch block needs to
2292 // contain a definition of the start index.
2293 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2294 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2295 BypassBlock->getTerminator());
2297 // Count holds the overall loop count (N).
2298 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2299 BypassBlock->getTerminator());
2301 LoopBypassBlocks.push_back(BypassBlock);
2303 // Split the single block loop into the two loop structure described above.
2304 BasicBlock *VectorPH =
2305 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2306 BasicBlock *VecBody =
2307 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2308 BasicBlock *MiddleBlock =
2309 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2310 BasicBlock *ScalarPH =
2311 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2313 // Create and register the new vector loop.
2314 Loop* Lp = new Loop();
2315 Loop *ParentLoop = OrigLoop->getParentLoop();
2317 // Insert the new loop into the loop nest and register the new basic blocks
2318 // before calling any utilities such as SCEV that require valid LoopInfo.
2320 ParentLoop->addChildLoop(Lp);
2321 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
2322 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
2323 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
2325 LI->addTopLevelLoop(Lp);
2327 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
2329 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2331 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2333 // Generate the induction variable.
2334 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2335 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2336 // The loop step is equal to the vectorization factor (num of SIMD elements)
2337 // times the unroll factor (num of SIMD instructions).
2338 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2340 // This is the IR builder that we use to add all of the logic for bypassing
2341 // the new vector loop.
2342 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2343 setDebugLocFromInst(BypassBuilder,
2344 getDebugLocFromInstOrOperands(OldInduction));
2346 // We may need to extend the index in case there is a type mismatch.
2347 // We know that the count starts at zero and does not overflow.
2348 if (Count->getType() != IdxTy) {
2349 // The exit count can be of pointer type. Convert it to the correct
2351 if (ExitCount->getType()->isPointerTy())
2352 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2354 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2357 // Add the start index to the loop count to get the new end index.
2358 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2360 // Now we need to generate the expression for N - (N % VF), which is
2361 // the part that the vectorized body will execute.
2362 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2363 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2364 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2365 "end.idx.rnd.down");
2367 // Now, compare the new count to zero. If it is zero skip the vector loop and
2368 // jump to the scalar loop.
2370 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2372 BasicBlock *LastBypassBlock = BypassBlock;
2374 // Generate code to check that the loops trip count that we computed by adding
2375 // one to the backedge-taken count will not overflow.
2377 auto PastOverflowCheck =
2378 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2379 BasicBlock *CheckBlock =
2380 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2382 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2383 LoopBypassBlocks.push_back(CheckBlock);
2384 Instruction *OldTerm = LastBypassBlock->getTerminator();
2385 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2386 OldTerm->eraseFromParent();
2387 LastBypassBlock = CheckBlock;
2390 // Generate the code to check that the strides we assumed to be one are really
2391 // one. We want the new basic block to start at the first instruction in a
2392 // sequence of instructions that form a check.
2393 Instruction *StrideCheck;
2394 Instruction *FirstCheckInst;
2395 std::tie(FirstCheckInst, StrideCheck) =
2396 addStrideCheck(LastBypassBlock->getTerminator());
2398 // Create a new block containing the stride check.
2399 BasicBlock *CheckBlock =
2400 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2402 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2403 LoopBypassBlocks.push_back(CheckBlock);
2405 // Replace the branch into the memory check block with a conditional branch
2406 // for the "few elements case".
2407 Instruction *OldTerm = LastBypassBlock->getTerminator();
2408 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2409 OldTerm->eraseFromParent();
2412 LastBypassBlock = CheckBlock;
2415 // Generate the code that checks in runtime if arrays overlap. We put the
2416 // checks into a separate block to make the more common case of few elements
2418 Instruction *MemRuntimeCheck;
2419 std::tie(FirstCheckInst, MemRuntimeCheck) =
2420 addRuntimeCheck(LastBypassBlock->getTerminator());
2421 if (MemRuntimeCheck) {
2422 // Create a new block containing the memory check.
2423 BasicBlock *CheckBlock =
2424 LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
2426 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2427 LoopBypassBlocks.push_back(CheckBlock);
2429 // Replace the branch into the memory check block with a conditional branch
2430 // for the "few elements case".
2431 Instruction *OldTerm = LastBypassBlock->getTerminator();
2432 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2433 OldTerm->eraseFromParent();
2435 Cmp = MemRuntimeCheck;
2436 LastBypassBlock = CheckBlock;
2439 LastBypassBlock->getTerminator()->eraseFromParent();
2440 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2443 // We are going to resume the execution of the scalar loop.
2444 // Go over all of the induction variables that we found and fix the
2445 // PHIs that are left in the scalar version of the loop.
2446 // The starting values of PHI nodes depend on the counter of the last
2447 // iteration in the vectorized loop.
2448 // If we come from a bypass edge then we need to start from the original
2451 // This variable saves the new starting index for the scalar loop.
2452 PHINode *ResumeIndex = nullptr;
2453 LoopVectorizationLegality::InductionList::iterator I, E;
2454 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2455 // Set builder to point to last bypass block.
2456 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2457 for (I = List->begin(), E = List->end(); I != E; ++I) {
2458 PHINode *OrigPhi = I->first;
2459 LoopVectorizationLegality::InductionInfo II = I->second;
2461 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2462 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2463 MiddleBlock->getTerminator());
2464 // We might have extended the type of the induction variable but we need a
2465 // truncated version for the scalar loop.
2466 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2467 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2468 MiddleBlock->getTerminator()) : nullptr;
2470 // Create phi nodes to merge from the backedge-taken check block.
2471 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2472 ScalarPH->getTerminator());
2473 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2475 PHINode *BCTruncResumeVal = nullptr;
2476 if (OrigPhi == OldInduction) {
2478 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2479 ScalarPH->getTerminator());
2480 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2483 Value *EndValue = nullptr;
2485 case LoopVectorizationLegality::IK_NoInduction:
2486 llvm_unreachable("Unknown induction");
2487 case LoopVectorizationLegality::IK_IntInduction: {
2488 // Handle the integer induction counter.
2489 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2491 // We have the canonical induction variable.
2492 if (OrigPhi == OldInduction) {
2493 // Create a truncated version of the resume value for the scalar loop,
2494 // we might have promoted the type to a larger width.
2496 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2497 // The new PHI merges the original incoming value, in case of a bypass,
2498 // or the value at the end of the vectorized loop.
2499 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2500 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2501 TruncResumeVal->addIncoming(EndValue, VecBody);
2503 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2505 // We know what the end value is.
2506 EndValue = IdxEndRoundDown;
2507 // We also know which PHI node holds it.
2508 ResumeIndex = ResumeVal;
2512 // Not the canonical induction variable - add the vector loop count to the
2514 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2515 II.StartValue->getType(),
2517 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
2520 case LoopVectorizationLegality::IK_ReverseIntInduction: {
2521 // Convert the CountRoundDown variable to the PHI size.
2522 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2523 II.StartValue->getType(),
2525 // Handle reverse integer induction counter.
2526 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
2529 case LoopVectorizationLegality::IK_PtrInduction: {
2530 // For pointer induction variables, calculate the offset using
2532 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
2536 case LoopVectorizationLegality::IK_ReversePtrInduction: {
2537 // The value at the end of the loop for the reverse pointer is calculated
2538 // by creating a GEP with a negative index starting from the start value.
2539 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
2540 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
2542 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
2548 // The new PHI merges the original incoming value, in case of a bypass,
2549 // or the value at the end of the vectorized loop.
2550 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2551 if (OrigPhi == OldInduction)
2552 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2554 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2556 ResumeVal->addIncoming(EndValue, VecBody);
2558 // Fix the scalar body counter (PHI node).
2559 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2561 // The old induction's phi node in the scalar body needs the truncated
2563 if (OrigPhi == OldInduction) {
2564 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2565 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2567 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2568 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2572 // If we are generating a new induction variable then we also need to
2573 // generate the code that calculates the exit value. This value is not
2574 // simply the end of the counter because we may skip the vectorized body
2575 // in case of a runtime check.
2577 assert(!ResumeIndex && "Unexpected resume value found");
2578 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2579 MiddleBlock->getTerminator());
2580 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2581 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2582 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2585 // Make sure that we found the index where scalar loop needs to continue.
2586 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2587 "Invalid resume Index");
2589 // Add a check in the middle block to see if we have completed
2590 // all of the iterations in the first vector loop.
2591 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2592 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2593 ResumeIndex, "cmp.n",
2594 MiddleBlock->getTerminator());
2596 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2597 // Remove the old terminator.
2598 MiddleBlock->getTerminator()->eraseFromParent();
2600 // Create i+1 and fill the PHINode.
2601 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2602 Induction->addIncoming(StartIdx, VectorPH);
2603 Induction->addIncoming(NextIdx, VecBody);
2604 // Create the compare.
2605 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2606 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2608 // Now we have two terminators. Remove the old one from the block.
2609 VecBody->getTerminator()->eraseFromParent();
2611 // Get ready to start creating new instructions into the vectorized body.
2612 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2615 LoopVectorPreHeader = VectorPH;
2616 LoopScalarPreHeader = ScalarPH;
2617 LoopMiddleBlock = MiddleBlock;
2618 LoopExitBlock = ExitBlock;
2619 LoopVectorBody.push_back(VecBody);
2620 LoopScalarBody = OldBasicBlock;
2622 LoopVectorizeHints Hints(Lp, true);
2623 Hints.setAlreadyVectorized();
2626 /// This function returns the identity element (or neutral element) for
2627 /// the operation K.
2629 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2634 // Adding, Xoring, Oring zero to a number does not change it.
2635 return ConstantInt::get(Tp, 0);
2636 case RK_IntegerMult:
2637 // Multiplying a number by 1 does not change it.
2638 return ConstantInt::get(Tp, 1);
2640 // AND-ing a number with an all-1 value does not change it.
2641 return ConstantInt::get(Tp, -1, true);
2643 // Multiplying a number by 1 does not change it.
2644 return ConstantFP::get(Tp, 1.0L);
2646 // Adding zero to a number does not change it.
2647 return ConstantFP::get(Tp, 0.0L);
2649 llvm_unreachable("Unknown reduction kind");
2653 /// This function translates the reduction kind to an LLVM binary operator.
2655 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2657 case LoopVectorizationLegality::RK_IntegerAdd:
2658 return Instruction::Add;
2659 case LoopVectorizationLegality::RK_IntegerMult:
2660 return Instruction::Mul;
2661 case LoopVectorizationLegality::RK_IntegerOr:
2662 return Instruction::Or;
2663 case LoopVectorizationLegality::RK_IntegerAnd:
2664 return Instruction::And;
2665 case LoopVectorizationLegality::RK_IntegerXor:
2666 return Instruction::Xor;
2667 case LoopVectorizationLegality::RK_FloatMult:
2668 return Instruction::FMul;
2669 case LoopVectorizationLegality::RK_FloatAdd:
2670 return Instruction::FAdd;
2671 case LoopVectorizationLegality::RK_IntegerMinMax:
2672 return Instruction::ICmp;
2673 case LoopVectorizationLegality::RK_FloatMinMax:
2674 return Instruction::FCmp;
2676 llvm_unreachable("Unknown reduction operation");
2680 Value *createMinMaxOp(IRBuilder<> &Builder,
2681 LoopVectorizationLegality::MinMaxReductionKind RK,
2684 CmpInst::Predicate P = CmpInst::ICMP_NE;
2687 llvm_unreachable("Unknown min/max reduction kind");
2688 case LoopVectorizationLegality::MRK_UIntMin:
2689 P = CmpInst::ICMP_ULT;
2691 case LoopVectorizationLegality::MRK_UIntMax:
2692 P = CmpInst::ICMP_UGT;
2694 case LoopVectorizationLegality::MRK_SIntMin:
2695 P = CmpInst::ICMP_SLT;
2697 case LoopVectorizationLegality::MRK_SIntMax:
2698 P = CmpInst::ICMP_SGT;
2700 case LoopVectorizationLegality::MRK_FloatMin:
2701 P = CmpInst::FCMP_OLT;
2703 case LoopVectorizationLegality::MRK_FloatMax:
2704 P = CmpInst::FCMP_OGT;
2709 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2710 RK == LoopVectorizationLegality::MRK_FloatMax)
2711 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2713 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2715 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2720 struct CSEDenseMapInfo {
2721 static bool canHandle(Instruction *I) {
2722 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2723 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2725 static inline Instruction *getEmptyKey() {
2726 return DenseMapInfo<Instruction *>::getEmptyKey();
2728 static inline Instruction *getTombstoneKey() {
2729 return DenseMapInfo<Instruction *>::getTombstoneKey();
2731 static unsigned getHashValue(Instruction *I) {
2732 assert(canHandle(I) && "Unknown instruction!");
2733 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2734 I->value_op_end()));
2736 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2737 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2738 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2740 return LHS->isIdenticalTo(RHS);
2745 /// \brief Check whether this block is a predicated block.
2746 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2747 /// = ...; " blocks. We start with one vectorized basic block. For every
2748 /// conditional block we split this vectorized block. Therefore, every second
2749 /// block will be a predicated one.
2750 static bool isPredicatedBlock(unsigned BlockNum) {
2751 return BlockNum % 2;
2754 ///\brief Perform cse of induction variable instructions.
2755 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2756 // Perform simple cse.
2757 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2758 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2759 BasicBlock *BB = BBs[i];
2760 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2761 Instruction *In = I++;
2763 if (!CSEDenseMapInfo::canHandle(In))
2766 // Check if we can replace this instruction with any of the
2767 // visited instructions.
2768 if (Instruction *V = CSEMap.lookup(In)) {
2769 In->replaceAllUsesWith(V);
2770 In->eraseFromParent();
2773 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2774 // ...;" blocks for predicated stores. Every second block is a predicated
2776 if (isPredicatedBlock(i))
2784 /// \brief Adds a 'fast' flag to floating point operations.
2785 static Value *addFastMathFlag(Value *V) {
2786 if (isa<FPMathOperator>(V)){
2787 FastMathFlags Flags;
2788 Flags.setUnsafeAlgebra();
2789 cast<Instruction>(V)->setFastMathFlags(Flags);
2794 void InnerLoopVectorizer::vectorizeLoop() {
2795 //===------------------------------------------------===//
2797 // Notice: any optimization or new instruction that go
2798 // into the code below should be also be implemented in
2801 //===------------------------------------------------===//
2802 Constant *Zero = Builder.getInt32(0);
2804 // In order to support reduction variables we need to be able to vectorize
2805 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2806 // stages. First, we create a new vector PHI node with no incoming edges.
2807 // We use this value when we vectorize all of the instructions that use the
2808 // PHI. Next, after all of the instructions in the block are complete we
2809 // add the new incoming edges to the PHI. At this point all of the
2810 // instructions in the basic block are vectorized, so we can use them to
2811 // construct the PHI.
2812 PhiVector RdxPHIsToFix;
2814 // Scan the loop in a topological order to ensure that defs are vectorized
2816 LoopBlocksDFS DFS(OrigLoop);
2819 // Vectorize all of the blocks in the original loop.
2820 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2821 be = DFS.endRPO(); bb != be; ++bb)
2822 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2824 // At this point every instruction in the original loop is widened to
2825 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2826 // that we vectorized. The PHI nodes are currently empty because we did
2827 // not want to introduce cycles. Notice that the remaining PHI nodes
2828 // that we need to fix are reduction variables.
2830 // Create the 'reduced' values for each of the induction vars.
2831 // The reduced values are the vector values that we scalarize and combine
2832 // after the loop is finished.
2833 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2835 PHINode *RdxPhi = *it;
2836 assert(RdxPhi && "Unable to recover vectorized PHI");
2838 // Find the reduction variable descriptor.
2839 assert(Legal->getReductionVars()->count(RdxPhi) &&
2840 "Unable to find the reduction variable");
2841 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2842 (*Legal->getReductionVars())[RdxPhi];
2844 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2846 // We need to generate a reduction vector from the incoming scalar.
2847 // To do so, we need to generate the 'identity' vector and override
2848 // one of the elements with the incoming scalar reduction. We need
2849 // to do it in the vector-loop preheader.
2850 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
2852 // This is the vector-clone of the value that leaves the loop.
2853 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2854 Type *VecTy = VectorExit[0]->getType();
2856 // Find the reduction identity variable. Zero for addition, or, xor,
2857 // one for multiplication, -1 for And.
2860 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2861 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2862 // MinMax reduction have the start value as their identify.
2864 VectorStart = Identity = RdxDesc.StartValue;
2866 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2871 // Handle other reduction kinds:
2873 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2874 VecTy->getScalarType());
2877 // This vector is the Identity vector where the first element is the
2878 // incoming scalar reduction.
2879 VectorStart = RdxDesc.StartValue;
2881 Identity = ConstantVector::getSplat(VF, Iden);
2883 // This vector is the Identity vector where the first element is the
2884 // incoming scalar reduction.
2885 VectorStart = Builder.CreateInsertElement(Identity,
2886 RdxDesc.StartValue, Zero);
2890 // Fix the vector-loop phi.
2892 // Reductions do not have to start at zero. They can start with
2893 // any loop invariant values.
2894 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2895 BasicBlock *Latch = OrigLoop->getLoopLatch();
2896 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2897 VectorParts &Val = getVectorValue(LoopVal);
2898 for (unsigned part = 0; part < UF; ++part) {
2899 // Make sure to add the reduction stat value only to the
2900 // first unroll part.
2901 Value *StartVal = (part == 0) ? VectorStart : Identity;
2902 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
2903 LoopVectorPreHeader);
2904 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2905 LoopVectorBody.back());
2908 // Before each round, move the insertion point right between
2909 // the PHIs and the values we are going to write.
2910 // This allows us to write both PHINodes and the extractelement
2912 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2914 VectorParts RdxParts;
2915 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2916 for (unsigned part = 0; part < UF; ++part) {
2917 // This PHINode contains the vectorized reduction variable, or
2918 // the initial value vector, if we bypass the vector loop.
2919 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2920 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2921 Value *StartVal = (part == 0) ? VectorStart : Identity;
2922 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2923 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2924 NewPhi->addIncoming(RdxExitVal[part],
2925 LoopVectorBody.back());
2926 RdxParts.push_back(NewPhi);
2929 // Reduce all of the unrolled parts into a single vector.
2930 Value *ReducedPartRdx = RdxParts[0];
2931 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2932 setDebugLocFromInst(Builder, ReducedPartRdx);
2933 for (unsigned part = 1; part < UF; ++part) {
2934 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2935 // Floating point operations had to be 'fast' to enable the reduction.
2936 ReducedPartRdx = addFastMathFlag(
2937 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2938 ReducedPartRdx, "bin.rdx"));
2940 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2941 ReducedPartRdx, RdxParts[part]);
2945 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2946 // and vector ops, reducing the set of values being computed by half each
2948 assert(isPowerOf2_32(VF) &&
2949 "Reduction emission only supported for pow2 vectors!");
2950 Value *TmpVec = ReducedPartRdx;
2951 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2952 for (unsigned i = VF; i != 1; i >>= 1) {
2953 // Move the upper half of the vector to the lower half.
2954 for (unsigned j = 0; j != i/2; ++j)
2955 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2957 // Fill the rest of the mask with undef.
2958 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2959 UndefValue::get(Builder.getInt32Ty()));
2962 Builder.CreateShuffleVector(TmpVec,
2963 UndefValue::get(TmpVec->getType()),
2964 ConstantVector::get(ShuffleMask),
2967 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2968 // Floating point operations had to be 'fast' to enable the reduction.
2969 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2970 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2972 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2975 // The result is in the first element of the vector.
2976 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2977 Builder.getInt32(0));
2980 // Create a phi node that merges control-flow from the backedge-taken check
2981 // block and the middle block.
2982 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
2983 LoopScalarPreHeader->getTerminator());
2984 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
2985 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2987 // Now, we need to fix the users of the reduction variable
2988 // inside and outside of the scalar remainder loop.
2989 // We know that the loop is in LCSSA form. We need to update the
2990 // PHI nodes in the exit blocks.
2991 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2992 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2993 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2994 if (!LCSSAPhi) break;
2996 // All PHINodes need to have a single entry edge, or two if
2997 // we already fixed them.
2998 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
3000 // We found our reduction value exit-PHI. Update it with the
3001 // incoming bypass edge.
3002 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
3003 // Add an edge coming from the bypass.
3004 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3007 }// end of the LCSSA phi scan.
3009 // Fix the scalar loop reduction variable with the incoming reduction sum
3010 // from the vector body and from the backedge value.
3011 int IncomingEdgeBlockIdx =
3012 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
3013 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
3014 // Pick the other block.
3015 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
3016 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
3017 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
3018 }// end of for each redux variable.
3022 // Remove redundant induction instructions.
3023 cse(LoopVectorBody);
3026 void InnerLoopVectorizer::fixLCSSAPHIs() {
3027 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3028 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3029 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3030 if (!LCSSAPhi) break;
3031 if (LCSSAPhi->getNumIncomingValues() == 1)
3032 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
3037 InnerLoopVectorizer::VectorParts
3038 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
3039 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
3042 // Look for cached value.
3043 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
3044 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
3045 if (ECEntryIt != MaskCache.end())
3046 return ECEntryIt->second;
3048 VectorParts SrcMask = createBlockInMask(Src);
3050 // The terminator has to be a branch inst!
3051 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
3052 assert(BI && "Unexpected terminator found");
3054 if (BI->isConditional()) {
3055 VectorParts EdgeMask = getVectorValue(BI->getCondition());
3057 if (BI->getSuccessor(0) != Dst)
3058 for (unsigned part = 0; part < UF; ++part)
3059 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
3061 for (unsigned part = 0; part < UF; ++part)
3062 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
3064 MaskCache[Edge] = EdgeMask;
3068 MaskCache[Edge] = SrcMask;
3072 InnerLoopVectorizer::VectorParts
3073 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
3074 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
3076 // Loop incoming mask is all-one.
3077 if (OrigLoop->getHeader() == BB) {
3078 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
3079 return getVectorValue(C);
3082 // This is the block mask. We OR all incoming edges, and with zero.
3083 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
3084 VectorParts BlockMask = getVectorValue(Zero);
3087 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
3088 VectorParts EM = createEdgeMask(*it, BB);
3089 for (unsigned part = 0; part < UF; ++part)
3090 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
3096 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
3097 InnerLoopVectorizer::VectorParts &Entry,
3098 unsigned UF, unsigned VF, PhiVector *PV) {
3099 PHINode* P = cast<PHINode>(PN);
3100 // Handle reduction variables:
3101 if (Legal->getReductionVars()->count(P)) {
3102 for (unsigned part = 0; part < UF; ++part) {
3103 // This is phase one of vectorizing PHIs.
3104 Type *VecTy = (VF == 1) ? PN->getType() :
3105 VectorType::get(PN->getType(), VF);
3106 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
3107 LoopVectorBody.back()-> getFirstInsertionPt());
3113 setDebugLocFromInst(Builder, P);
3114 // Check for PHI nodes that are lowered to vector selects.
3115 if (P->getParent() != OrigLoop->getHeader()) {
3116 // We know that all PHIs in non-header blocks are converted into
3117 // selects, so we don't have to worry about the insertion order and we
3118 // can just use the builder.
3119 // At this point we generate the predication tree. There may be
3120 // duplications since this is a simple recursive scan, but future
3121 // optimizations will clean it up.
3123 unsigned NumIncoming = P->getNumIncomingValues();
3125 // Generate a sequence of selects of the form:
3126 // SELECT(Mask3, In3,
3127 // SELECT(Mask2, In2,
3129 for (unsigned In = 0; In < NumIncoming; In++) {
3130 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
3132 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3134 for (unsigned part = 0; part < UF; ++part) {
3135 // We might have single edge PHIs (blocks) - use an identity
3136 // 'select' for the first PHI operand.
3138 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3141 // Select between the current value and the previous incoming edge
3142 // based on the incoming mask.
3143 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3144 Entry[part], "predphi");
3150 // This PHINode must be an induction variable.
3151 // Make sure that we know about it.
3152 assert(Legal->getInductionVars()->count(P) &&
3153 "Not an induction variable");
3155 LoopVectorizationLegality::InductionInfo II =
3156 Legal->getInductionVars()->lookup(P);
3159 case LoopVectorizationLegality::IK_NoInduction:
3160 llvm_unreachable("Unknown induction");
3161 case LoopVectorizationLegality::IK_IntInduction: {
3162 assert(P->getType() == II.StartValue->getType() && "Types must match");
3163 Type *PhiTy = P->getType();
3165 if (P == OldInduction) {
3166 // Handle the canonical induction variable. We might have had to
3168 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
3170 // Handle other induction variables that are now based on the
3172 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
3174 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
3175 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
3178 Broadcasted = getBroadcastInstrs(Broadcasted);
3179 // After broadcasting the induction variable we need to make the vector
3180 // consecutive by adding 0, 1, 2, etc.
3181 for (unsigned part = 0; part < UF; ++part)
3182 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
3185 case LoopVectorizationLegality::IK_ReverseIntInduction:
3186 case LoopVectorizationLegality::IK_PtrInduction:
3187 case LoopVectorizationLegality::IK_ReversePtrInduction:
3188 // Handle reverse integer and pointer inductions.
3189 Value *StartIdx = ExtendedIdx;
3190 // This is the normalized GEP that starts counting at zero.
3191 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
3194 // Handle the reverse integer induction variable case.
3195 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
3196 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
3197 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
3199 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
3202 // This is a new value so do not hoist it out.
3203 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
3204 // After broadcasting the induction variable we need to make the
3205 // vector consecutive by adding ... -3, -2, -1, 0.
3206 for (unsigned part = 0; part < UF; ++part)
3207 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
3212 // Handle the pointer induction variable case.
3213 assert(P->getType()->isPointerTy() && "Unexpected type.");
3215 // Is this a reverse induction ptr or a consecutive induction ptr.
3216 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
3219 // This is the vector of results. Notice that we don't generate
3220 // vector geps because scalar geps result in better code.
3221 for (unsigned part = 0; part < UF; ++part) {
3223 int EltIndex = (part) * (Reverse ? -1 : 1);
3224 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3227 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3229 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3231 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3233 Entry[part] = SclrGep;
3237 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3238 for (unsigned int i = 0; i < VF; ++i) {
3239 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
3240 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3243 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3245 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3247 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3249 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3250 Builder.getInt32(i),
3253 Entry[part] = VecVal;
3259 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3260 // For each instruction in the old loop.
3261 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3262 VectorParts &Entry = WidenMap.get(it);
3263 switch (it->getOpcode()) {
3264 case Instruction::Br:
3265 // Nothing to do for PHIs and BR, since we already took care of the
3266 // loop control flow instructions.
3268 case Instruction::PHI:{
3269 // Vectorize PHINodes.
3270 widenPHIInstruction(it, Entry, UF, VF, PV);
3274 case Instruction::Add:
3275 case Instruction::FAdd:
3276 case Instruction::Sub:
3277 case Instruction::FSub:
3278 case Instruction::Mul:
3279 case Instruction::FMul:
3280 case Instruction::UDiv:
3281 case Instruction::SDiv:
3282 case Instruction::FDiv:
3283 case Instruction::URem:
3284 case Instruction::SRem:
3285 case Instruction::FRem:
3286 case Instruction::Shl:
3287 case Instruction::LShr:
3288 case Instruction::AShr:
3289 case Instruction::And:
3290 case Instruction::Or:
3291 case Instruction::Xor: {
3292 // Just widen binops.
3293 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3294 setDebugLocFromInst(Builder, BinOp);
3295 VectorParts &A = getVectorValue(it->getOperand(0));
3296 VectorParts &B = getVectorValue(it->getOperand(1));
3298 // Use this vector value for all users of the original instruction.
3299 for (unsigned Part = 0; Part < UF; ++Part) {
3300 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3302 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3303 VecOp->copyIRFlags(BinOp);
3308 propagateMetadata(Entry, it);
3311 case Instruction::Select: {
3313 // If the selector is loop invariant we can create a select
3314 // instruction with a scalar condition. Otherwise, use vector-select.
3315 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3317 setDebugLocFromInst(Builder, it);
3319 // The condition can be loop invariant but still defined inside the
3320 // loop. This means that we can't just use the original 'cond' value.
3321 // We have to take the 'vectorized' value and pick the first lane.
3322 // Instcombine will make this a no-op.
3323 VectorParts &Cond = getVectorValue(it->getOperand(0));
3324 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3325 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3327 Value *ScalarCond = (VF == 1) ? Cond[0] :
3328 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3330 for (unsigned Part = 0; Part < UF; ++Part) {
3331 Entry[Part] = Builder.CreateSelect(
3332 InvariantCond ? ScalarCond : Cond[Part],
3337 propagateMetadata(Entry, it);
3341 case Instruction::ICmp:
3342 case Instruction::FCmp: {
3343 // Widen compares. Generate vector compares.
3344 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3345 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3346 setDebugLocFromInst(Builder, it);
3347 VectorParts &A = getVectorValue(it->getOperand(0));
3348 VectorParts &B = getVectorValue(it->getOperand(1));
3349 for (unsigned Part = 0; Part < UF; ++Part) {
3352 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3354 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3358 propagateMetadata(Entry, it);
3362 case Instruction::Store:
3363 case Instruction::Load:
3364 vectorizeMemoryInstruction(it);
3366 case Instruction::ZExt:
3367 case Instruction::SExt:
3368 case Instruction::FPToUI:
3369 case Instruction::FPToSI:
3370 case Instruction::FPExt:
3371 case Instruction::PtrToInt:
3372 case Instruction::IntToPtr:
3373 case Instruction::SIToFP:
3374 case Instruction::UIToFP:
3375 case Instruction::Trunc:
3376 case Instruction::FPTrunc:
3377 case Instruction::BitCast: {
3378 CastInst *CI = dyn_cast<CastInst>(it);
3379 setDebugLocFromInst(Builder, it);
3380 /// Optimize the special case where the source is the induction
3381 /// variable. Notice that we can only optimize the 'trunc' case
3382 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3383 /// c. other casts depend on pointer size.
3384 if (CI->getOperand(0) == OldInduction &&
3385 it->getOpcode() == Instruction::Trunc) {
3386 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3388 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3389 for (unsigned Part = 0; Part < UF; ++Part)
3390 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
3391 propagateMetadata(Entry, it);
3394 /// Vectorize casts.
3395 Type *DestTy = (VF == 1) ? CI->getType() :
3396 VectorType::get(CI->getType(), VF);
3398 VectorParts &A = getVectorValue(it->getOperand(0));
3399 for (unsigned Part = 0; Part < UF; ++Part)
3400 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3401 propagateMetadata(Entry, it);
3405 case Instruction::Call: {
3406 // Ignore dbg intrinsics.
3407 if (isa<DbgInfoIntrinsic>(it))
3409 setDebugLocFromInst(Builder, it);
3411 Module *M = BB->getParent()->getParent();
3412 CallInst *CI = cast<CallInst>(it);
3413 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3414 assert(ID && "Not an intrinsic call!");
3416 case Intrinsic::assume:
3417 case Intrinsic::lifetime_end:
3418 case Intrinsic::lifetime_start:
3419 scalarizeInstruction(it);
3422 bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
3423 for (unsigned Part = 0; Part < UF; ++Part) {
3424 SmallVector<Value *, 4> Args;
3425 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3426 if (HasScalarOpd && i == 1) {
3427 Args.push_back(CI->getArgOperand(i));
3430 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3431 Args.push_back(Arg[Part]);
3433 Type *Tys[] = {CI->getType()};
3435 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3437 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3438 Entry[Part] = Builder.CreateCall(F, Args);
3441 propagateMetadata(Entry, it);
3448 // All other instructions are unsupported. Scalarize them.
3449 scalarizeInstruction(it);
3452 }// end of for_each instr.
3455 void InnerLoopVectorizer::updateAnalysis() {
3456 // Forget the original basic block.
3457 SE->forgetLoop(OrigLoop);
3459 // Update the dominator tree information.
3460 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3461 "Entry does not dominate exit.");
3463 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3464 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3465 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3467 // Due to if predication of stores we might create a sequence of "if(pred)
3468 // a[i] = ...; " blocks.
3469 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3471 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3472 else if (isPredicatedBlock(i)) {
3473 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3475 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3479 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3480 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3481 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3482 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3484 DEBUG(DT->verifyDomTree());
3487 /// \brief Check whether it is safe to if-convert this phi node.
3489 /// Phi nodes with constant expressions that can trap are not safe to if
3491 static bool canIfConvertPHINodes(BasicBlock *BB) {
3492 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3493 PHINode *Phi = dyn_cast<PHINode>(I);
3496 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3497 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3504 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3505 if (!EnableIfConversion) {
3506 emitAnalysis(Report() << "if-conversion is disabled");
3510 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3512 // A list of pointers that we can safely read and write to.
3513 SmallPtrSet<Value *, 8> SafePointes;
3515 // Collect safe addresses.
3516 for (Loop::block_iterator BI = TheLoop->block_begin(),
3517 BE = TheLoop->block_end(); BI != BE; ++BI) {
3518 BasicBlock *BB = *BI;
3520 if (blockNeedsPredication(BB))
3523 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3524 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3525 SafePointes.insert(LI->getPointerOperand());
3526 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3527 SafePointes.insert(SI->getPointerOperand());
3531 // Collect the blocks that need predication.
3532 BasicBlock *Header = TheLoop->getHeader();
3533 for (Loop::block_iterator BI = TheLoop->block_begin(),
3534 BE = TheLoop->block_end(); BI != BE; ++BI) {
3535 BasicBlock *BB = *BI;
3537 // We don't support switch statements inside loops.
3538 if (!isa<BranchInst>(BB->getTerminator())) {
3539 emitAnalysis(Report(BB->getTerminator())
3540 << "loop contains a switch statement");
3544 // We must be able to predicate all blocks that need to be predicated.
3545 if (blockNeedsPredication(BB)) {
3546 if (!blockCanBePredicated(BB, SafePointes)) {
3547 emitAnalysis(Report(BB->getTerminator())
3548 << "control flow cannot be substituted for a select");
3551 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3552 emitAnalysis(Report(BB->getTerminator())
3553 << "control flow cannot be substituted for a select");
3558 // We can if-convert this loop.
3562 bool LoopVectorizationLegality::canVectorize() {
3563 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3564 // be canonicalized.
3565 if (!TheLoop->getLoopPreheader()) {
3567 Report() << "loop control flow is not understood by vectorizer");
3571 // We can only vectorize innermost loops.
3572 if (TheLoop->getSubLoopsVector().size()) {
3573 emitAnalysis(Report() << "loop is not the innermost loop");
3577 // We must have a single backedge.
3578 if (TheLoop->getNumBackEdges() != 1) {
3580 Report() << "loop control flow is not understood by vectorizer");
3584 // We must have a single exiting block.
3585 if (!TheLoop->getExitingBlock()) {
3587 Report() << "loop control flow is not understood by vectorizer");
3591 // We only handle bottom-tested loops, i.e. loop in which the condition is
3592 // checked at the end of each iteration. With that we can assume that all
3593 // instructions in the loop are executed the same number of times.
3594 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
3596 Report() << "loop control flow is not understood by vectorizer");
3600 // We need to have a loop header.
3601 DEBUG(dbgs() << "LV: Found a loop: " <<
3602 TheLoop->getHeader()->getName() << '\n');
3604 // Check if we can if-convert non-single-bb loops.
3605 unsigned NumBlocks = TheLoop->getNumBlocks();
3606 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3607 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3611 // ScalarEvolution needs to be able to find the exit count.
3612 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3613 if (ExitCount == SE->getCouldNotCompute()) {
3614 emitAnalysis(Report() << "could not determine number of loop iterations");
3615 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3619 // Check if we can vectorize the instructions and CFG in this loop.
3620 if (!canVectorizeInstrs()) {
3621 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3625 // Go over each instruction and look at memory deps.
3626 if (!canVectorizeMemory()) {
3627 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3631 // Collect all of the variables that remain uniform after vectorization.
3632 collectLoopUniforms();
3634 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3635 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
3638 // Okay! We can vectorize. At this point we don't have any other mem analysis
3639 // which may limit our maximum vectorization factor, so just return true with
3644 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3645 if (Ty->isPointerTy())
3646 return DL.getIntPtrType(Ty);
3648 // It is possible that char's or short's overflow when we ask for the loop's
3649 // trip count, work around this by changing the type size.
3650 if (Ty->getScalarSizeInBits() < 32)
3651 return Type::getInt32Ty(Ty->getContext());
3656 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3657 Ty0 = convertPointerToIntegerType(DL, Ty0);
3658 Ty1 = convertPointerToIntegerType(DL, Ty1);
3659 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3664 /// \brief Check that the instruction has outside loop users and is not an
3665 /// identified reduction variable.
3666 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3667 SmallPtrSetImpl<Value *> &Reductions) {
3668 // Reduction instructions are allowed to have exit users. All other
3669 // instructions must not have external users.
3670 if (!Reductions.count(Inst))
3671 //Check that all of the users of the loop are inside the BB.
3672 for (User *U : Inst->users()) {
3673 Instruction *UI = cast<Instruction>(U);
3674 // This user may be a reduction exit value.
3675 if (!TheLoop->contains(UI)) {
3676 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3683 bool LoopVectorizationLegality::canVectorizeInstrs() {
3684 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3685 BasicBlock *Header = TheLoop->getHeader();
3687 // Look for the attribute signaling the absence of NaNs.
3688 Function &F = *Header->getParent();
3689 if (F.hasFnAttribute("no-nans-fp-math"))
3690 HasFunNoNaNAttr = F.getAttributes().getAttribute(
3691 AttributeSet::FunctionIndex,
3692 "no-nans-fp-math").getValueAsString() == "true";
3694 // For each block in the loop.
3695 for (Loop::block_iterator bb = TheLoop->block_begin(),
3696 be = TheLoop->block_end(); bb != be; ++bb) {
3698 // Scan the instructions in the block and look for hazards.
3699 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3702 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3703 Type *PhiTy = Phi->getType();
3704 // Check that this PHI type is allowed.
3705 if (!PhiTy->isIntegerTy() &&
3706 !PhiTy->isFloatingPointTy() &&
3707 !PhiTy->isPointerTy()) {
3708 emitAnalysis(Report(it)
3709 << "loop control flow is not understood by vectorizer");
3710 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3714 // If this PHINode is not in the header block, then we know that we
3715 // can convert it to select during if-conversion. No need to check if
3716 // the PHIs in this block are induction or reduction variables.
3717 if (*bb != Header) {
3718 // Check that this instruction has no outside users or is an
3719 // identified reduction value with an outside user.
3720 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3722 emitAnalysis(Report(it) << "value could not be identified as "
3723 "an induction or reduction variable");
3727 // We only allow if-converted PHIs with more than two incoming values.
3728 if (Phi->getNumIncomingValues() != 2) {
3729 emitAnalysis(Report(it)
3730 << "control flow not understood by vectorizer");
3731 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3735 // This is the value coming from the preheader.
3736 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3737 // Check if this is an induction variable.
3738 InductionKind IK = isInductionVariable(Phi);
3740 if (IK_NoInduction != IK) {
3741 // Get the widest type.
3743 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3745 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3747 // Int inductions are special because we only allow one IV.
3748 if (IK == IK_IntInduction) {
3749 // Use the phi node with the widest type as induction. Use the last
3750 // one if there are multiple (no good reason for doing this other
3751 // than it is expedient).
3752 if (!Induction || PhiTy == WidestIndTy)
3756 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3757 Inductions[Phi] = InductionInfo(StartValue, IK);
3759 // Until we explicitly handle the case of an induction variable with
3760 // an outside loop user we have to give up vectorizing this loop.
3761 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3762 emitAnalysis(Report(it) << "use of induction value outside of the "
3763 "loop is not handled by vectorizer");
3770 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3771 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3774 if (AddReductionVar(Phi, RK_IntegerMult)) {
3775 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3778 if (AddReductionVar(Phi, RK_IntegerOr)) {
3779 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3782 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3783 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3786 if (AddReductionVar(Phi, RK_IntegerXor)) {
3787 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3790 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3791 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3794 if (AddReductionVar(Phi, RK_FloatMult)) {
3795 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3798 if (AddReductionVar(Phi, RK_FloatAdd)) {
3799 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3802 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3803 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3808 emitAnalysis(Report(it) << "value that could not be identified as "
3809 "reduction is used outside the loop");
3810 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3812 }// end of PHI handling
3814 // We still don't handle functions. However, we can ignore dbg intrinsic
3815 // calls and we do handle certain intrinsic and libm functions.
3816 CallInst *CI = dyn_cast<CallInst>(it);
3817 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3818 emitAnalysis(Report(it) << "call instruction cannot be vectorized");
3819 DEBUG(dbgs() << "LV: Found a call site.\n");
3823 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
3824 // second argument is the same (i.e. loop invariant)
3826 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
3827 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
3828 emitAnalysis(Report(it)
3829 << "intrinsic instruction cannot be vectorized");
3830 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
3835 // Check that the instruction return type is vectorizable.
3836 // Also, we can't vectorize extractelement instructions.
3837 if ((!VectorType::isValidElementType(it->getType()) &&
3838 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3839 emitAnalysis(Report(it)
3840 << "instruction return type cannot be vectorized");
3841 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3845 // Check that the stored type is vectorizable.
3846 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3847 Type *T = ST->getValueOperand()->getType();
3848 if (!VectorType::isValidElementType(T)) {
3849 emitAnalysis(Report(ST) << "store instruction cannot be vectorized");
3852 if (EnableMemAccessVersioning)
3853 collectStridedAcccess(ST);
3856 if (EnableMemAccessVersioning)
3857 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3858 collectStridedAcccess(LI);
3860 // Reduction instructions are allowed to have exit users.
3861 // All other instructions must not have external users.
3862 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3863 emitAnalysis(Report(it) << "value cannot be used outside the loop");
3872 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3873 if (Inductions.empty()) {
3874 emitAnalysis(Report()
3875 << "loop induction variable could not be identified");
3883 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3884 /// return the induction operand of the gep pointer.
3885 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3886 const DataLayout *DL, Loop *Lp) {
3887 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3891 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3893 // Check that all of the gep indices are uniform except for our induction
3895 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3896 if (i != InductionOperand &&
3897 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3899 return GEP->getOperand(InductionOperand);
3902 ///\brief Look for a cast use of the passed value.
3903 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3904 Value *UniqueCast = nullptr;
3905 for (User *U : Ptr->users()) {
3906 CastInst *CI = dyn_cast<CastInst>(U);
3907 if (CI && CI->getType() == Ty) {
3917 ///\brief Get the stride of a pointer access in a loop.
3918 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3919 /// pointer to the Value, or null otherwise.
3920 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3921 const DataLayout *DL, Loop *Lp) {
3922 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3923 if (!PtrTy || PtrTy->isAggregateType())
3926 // Try to remove a gep instruction to make the pointer (actually index at this
3927 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3928 // pointer, otherwise, we are analyzing the index.
3929 Value *OrigPtr = Ptr;
3931 // The size of the pointer access.
3932 int64_t PtrAccessSize = 1;
3934 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3935 const SCEV *V = SE->getSCEV(Ptr);
3939 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3940 V = C->getOperand();
3942 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3946 V = S->getStepRecurrence(*SE);
3950 // Strip off the size of access multiplication if we are still analyzing the
3952 if (OrigPtr == Ptr) {
3953 DL->getTypeAllocSize(PtrTy->getElementType());
3954 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3955 if (M->getOperand(0)->getSCEVType() != scConstant)
3958 const APInt &APStepVal =
3959 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3961 // Huge step value - give up.
3962 if (APStepVal.getBitWidth() > 64)
3965 int64_t StepVal = APStepVal.getSExtValue();
3966 if (PtrAccessSize != StepVal)
3968 V = M->getOperand(1);
3973 Type *StripedOffRecurrenceCast = nullptr;
3974 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3975 StripedOffRecurrenceCast = C->getType();
3976 V = C->getOperand();
3979 // Look for the loop invariant symbolic value.
3980 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3984 Value *Stride = U->getValue();
3985 if (!Lp->isLoopInvariant(Stride))
3988 // If we have stripped off the recurrence cast we have to make sure that we
3989 // return the value that is used in this loop so that we can replace it later.
3990 if (StripedOffRecurrenceCast)
3991 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3996 void LoopVectorizationLegality::collectStridedAcccess(Value *MemAccess) {
3997 Value *Ptr = nullptr;
3998 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3999 Ptr = LI->getPointerOperand();
4000 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
4001 Ptr = SI->getPointerOperand();
4005 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
4009 DEBUG(dbgs() << "LV: Found a strided access that we can version");
4010 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
4011 Strides[Ptr] = Stride;
4012 StrideSet.insert(Stride);
4015 void LoopVectorizationLegality::collectLoopUniforms() {
4016 // We now know that the loop is vectorizable!
4017 // Collect variables that will remain uniform after vectorization.
4018 std::vector<Value*> Worklist;
4019 BasicBlock *Latch = TheLoop->getLoopLatch();
4021 // Start with the conditional branch and walk up the block.
4022 Worklist.push_back(Latch->getTerminator()->getOperand(0));
4024 // Also add all consecutive pointer values; these values will be uniform
4025 // after vectorization (and subsequent cleanup) and, until revectorization is
4026 // supported, all dependencies must also be uniform.
4027 for (Loop::block_iterator B = TheLoop->block_begin(),
4028 BE = TheLoop->block_end(); B != BE; ++B)
4029 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
4031 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
4032 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4034 while (Worklist.size()) {
4035 Instruction *I = dyn_cast<Instruction>(Worklist.back());
4036 Worklist.pop_back();
4038 // Look at instructions inside this loop.
4039 // Stop when reaching PHI nodes.
4040 // TODO: we need to follow values all over the loop, not only in this block.
4041 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
4044 // This is a known uniform.
4047 // Insert all operands.
4048 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4053 /// \brief Analyses memory accesses in a loop.
4055 /// Checks whether run time pointer checks are needed and builds sets for data
4056 /// dependence checking.
4057 class AccessAnalysis {
4059 /// \brief Read or write access location.
4060 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
4061 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
4063 /// \brief Set of potential dependent memory accesses.
4064 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
4066 AccessAnalysis(const DataLayout *Dl, AliasAnalysis *AA, DepCandidates &DA) :
4067 DL(Dl), AST(*AA), DepCands(DA), IsRTCheckNeeded(false) {}
4069 /// \brief Register a load and whether it is only read from.
4070 void addLoad(AliasAnalysis::Location &Loc, bool IsReadOnly) {
4071 Value *Ptr = const_cast<Value*>(Loc.Ptr);
4072 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
4073 Accesses.insert(MemAccessInfo(Ptr, false));
4075 ReadOnlyPtr.insert(Ptr);
4078 /// \brief Register a store.
4079 void addStore(AliasAnalysis::Location &Loc) {
4080 Value *Ptr = const_cast<Value*>(Loc.Ptr);
4081 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
4082 Accesses.insert(MemAccessInfo(Ptr, true));
4085 /// \brief Check whether we can check the pointers at runtime for
4086 /// non-intersection.
4087 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
4088 unsigned &NumComparisons, ScalarEvolution *SE,
4089 Loop *TheLoop, ValueToValueMap &Strides,
4090 bool ShouldCheckStride = false);
4092 /// \brief Goes over all memory accesses, checks whether a RT check is needed
4093 /// and builds sets of dependent accesses.
4094 void buildDependenceSets() {
4095 processMemAccesses();
4098 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
4100 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
4101 void resetDepChecks() { CheckDeps.clear(); }
4103 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
4106 typedef SetVector<MemAccessInfo> PtrAccessSet;
4108 /// \brief Go over all memory access and check whether runtime pointer checks
4109 /// are needed /// and build sets of dependency check candidates.
4110 void processMemAccesses();
4112 /// Set of all accesses.
4113 PtrAccessSet Accesses;
4115 /// Set of accesses that need a further dependence check.
4116 MemAccessInfoSet CheckDeps;
4118 /// Set of pointers that are read only.
4119 SmallPtrSet<Value*, 16> ReadOnlyPtr;
4121 const DataLayout *DL;
4123 /// An alias set tracker to partition the access set by underlying object and
4124 //intrinsic property (such as TBAA metadata).
4125 AliasSetTracker AST;
4127 /// Sets of potentially dependent accesses - members of one set share an
4128 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
4129 /// dependence check.
4130 DepCandidates &DepCands;
4132 bool IsRTCheckNeeded;
4135 } // end anonymous namespace
4137 /// \brief Check whether a pointer can participate in a runtime bounds check.
4138 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
4140 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
4141 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4145 return AR->isAffine();
4148 /// \brief Check the stride of the pointer and ensure that it does not wrap in
4149 /// the address space.
4150 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4151 const Loop *Lp, ValueToValueMap &StridesMap);
4153 bool AccessAnalysis::canCheckPtrAtRT(
4154 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
4155 unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
4156 ValueToValueMap &StridesMap, bool ShouldCheckStride) {
4157 // Find pointers with computable bounds. We are going to use this information
4158 // to place a runtime bound check.
4159 bool CanDoRT = true;
4161 bool IsDepCheckNeeded = isDependencyCheckNeeded();
4164 // We assign a consecutive id to access from different alias sets.
4165 // Accesses between different groups doesn't need to be checked.
4167 for (auto &AS : AST) {
4168 unsigned NumReadPtrChecks = 0;
4169 unsigned NumWritePtrChecks = 0;
4171 // We assign consecutive id to access from different dependence sets.
4172 // Accesses within the same set don't need a runtime check.
4173 unsigned RunningDepId = 1;
4174 DenseMap<Value *, unsigned> DepSetId;
4177 Value *Ptr = A.getValue();
4178 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
4179 MemAccessInfo Access(Ptr, IsWrite);
4182 ++NumWritePtrChecks;
4186 if (hasComputableBounds(SE, StridesMap, Ptr) &&
4187 // When we run after a failing dependency check we have to make sure we
4188 // don't have wrapping pointers.
4189 (!ShouldCheckStride ||
4190 isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
4191 // The id of the dependence set.
4194 if (IsDepCheckNeeded) {
4195 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
4196 unsigned &LeaderId = DepSetId[Leader];
4198 LeaderId = RunningDepId++;
4201 // Each access has its own dependence set.
4202 DepId = RunningDepId++;
4204 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
4206 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
4212 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
4213 NumComparisons += 0; // Only one dependence set.
4215 NumComparisons += (NumWritePtrChecks * (NumReadPtrChecks +
4216 NumWritePtrChecks - 1));
4222 // If the pointers that we would use for the bounds comparison have different
4223 // address spaces, assume the values aren't directly comparable, so we can't
4224 // use them for the runtime check. We also have to assume they could
4225 // overlap. In the future there should be metadata for whether address spaces
4227 unsigned NumPointers = RtCheck.Pointers.size();
4228 for (unsigned i = 0; i < NumPointers; ++i) {
4229 for (unsigned j = i + 1; j < NumPointers; ++j) {
4230 // Only need to check pointers between two different dependency sets.
4231 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
4233 // Only need to check pointers in the same alias set.
4234 if (RtCheck.AliasSetId[i] != RtCheck.AliasSetId[j])
4237 Value *PtrI = RtCheck.Pointers[i];
4238 Value *PtrJ = RtCheck.Pointers[j];
4240 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
4241 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
4243 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
4244 " different address spaces\n");
4253 void AccessAnalysis::processMemAccesses() {
4254 // We process the set twice: first we process read-write pointers, last we
4255 // process read-only pointers. This allows us to skip dependence tests for
4256 // read-only pointers.
4258 DEBUG(dbgs() << "LV: Processing memory accesses...\n");
4259 DEBUG(dbgs() << " AST: "; AST.dump());
4260 DEBUG(dbgs() << "LV: Accesses:\n");
4262 for (auto A : Accesses)
4263 dbgs() << "\t" << *A.getPointer() << " (" <<
4264 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
4265 "read-only" : "read")) << ")\n";
4268 // The AliasSetTracker has nicely partitioned our pointers by metadata
4269 // compatibility and potential for underlying-object overlap. As a result, we
4270 // only need to check for potential pointer dependencies within each alias
4272 for (auto &AS : AST) {
4273 // Note that both the alias-set tracker and the alias sets themselves used
4274 // linked lists internally and so the iteration order here is deterministic
4275 // (matching the original instruction order within each set).
4277 bool SetHasWrite = false;
4279 // Map of pointers to last access encountered.
4280 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
4281 UnderlyingObjToAccessMap ObjToLastAccess;
4283 // Set of access to check after all writes have been processed.
4284 PtrAccessSet DeferredAccesses;
4286 // Iterate over each alias set twice, once to process read/write pointers,
4287 // and then to process read-only pointers.
4288 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
4289 bool UseDeferred = SetIteration > 0;
4290 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
4293 Value *Ptr = A.getValue();
4294 bool IsWrite = S.count(MemAccessInfo(Ptr, true));
4296 // If we're using the deferred access set, then it contains only reads.
4297 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
4298 if (UseDeferred && !IsReadOnlyPtr)
4300 // Otherwise, the pointer must be in the PtrAccessSet, either as a read
4302 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
4303 S.count(MemAccessInfo(Ptr, false))) &&
4304 "Alias-set pointer not in the access set?");
4306 MemAccessInfo Access(Ptr, IsWrite);
4307 DepCands.insert(Access);
4309 // Memorize read-only pointers for later processing and skip them in the
4310 // first round (they need to be checked after we have seen all write
4311 // pointers). Note: we also mark pointer that are not consecutive as
4312 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need
4313 // the second check for "!IsWrite".
4314 if (!UseDeferred && IsReadOnlyPtr) {
4315 DeferredAccesses.insert(Access);
4319 // If this is a write - check other reads and writes for conflicts. If
4320 // this is a read only check other writes for conflicts (but only if
4321 // there is no other write to the ptr - this is an optimization to
4322 // catch "a[i] = a[i] + " without having to do a dependence check).
4323 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
4324 CheckDeps.insert(Access);
4325 IsRTCheckNeeded = true;
4331 // Create sets of pointers connected by a shared alias set and
4332 // underlying object.
4333 typedef SmallVector<Value *, 16> ValueVector;
4334 ValueVector TempObjects;
4335 GetUnderlyingObjects(Ptr, TempObjects, DL);
4336 for (Value *UnderlyingObj : TempObjects) {
4337 UnderlyingObjToAccessMap::iterator Prev =
4338 ObjToLastAccess.find(UnderlyingObj);
4339 if (Prev != ObjToLastAccess.end())
4340 DepCands.unionSets(Access, Prev->second);
4342 ObjToLastAccess[UnderlyingObj] = Access;
4350 /// \brief Checks memory dependences among accesses to the same underlying
4351 /// object to determine whether there vectorization is legal or not (and at
4352 /// which vectorization factor).
4354 /// This class works under the assumption that we already checked that memory
4355 /// locations with different underlying pointers are "must-not alias".
4356 /// We use the ScalarEvolution framework to symbolically evalutate access
4357 /// functions pairs. Since we currently don't restructure the loop we can rely
4358 /// on the program order of memory accesses to determine their safety.
4359 /// At the moment we will only deem accesses as safe for:
4360 /// * A negative constant distance assuming program order.
4362 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
4363 /// a[i] = tmp; y = a[i];
4365 /// The latter case is safe because later checks guarantuee that there can't
4366 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
4367 /// the same variable: a header phi can only be an induction or a reduction, a
4368 /// reduction can't have a memory sink, an induction can't have a memory
4369 /// source). This is important and must not be violated (or we have to
4370 /// resort to checking for cycles through memory).
4372 /// * A positive constant distance assuming program order that is bigger
4373 /// than the biggest memory access.
4375 /// tmp = a[i] OR b[i] = x
4376 /// a[i+2] = tmp y = b[i+2];
4378 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
4380 /// * Zero distances and all accesses have the same size.
4382 class MemoryDepChecker {
4384 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
4385 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
4387 MemoryDepChecker(ScalarEvolution *Se, const DataLayout *Dl, const Loop *L)
4388 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
4389 ShouldRetryWithRuntimeCheck(false) {}
4391 /// \brief Register the location (instructions are given increasing numbers)
4392 /// of a write access.
4393 void addAccess(StoreInst *SI) {
4394 Value *Ptr = SI->getPointerOperand();
4395 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
4396 InstMap.push_back(SI);
4400 /// \brief Register the location (instructions are given increasing numbers)
4401 /// of a write access.
4402 void addAccess(LoadInst *LI) {
4403 Value *Ptr = LI->getPointerOperand();
4404 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
4405 InstMap.push_back(LI);
4409 /// \brief Check whether the dependencies between the accesses are safe.
4411 /// Only checks sets with elements in \p CheckDeps.
4412 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4413 MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
4415 /// \brief The maximum number of bytes of a vector register we can vectorize
4416 /// the accesses safely with.
4417 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
4419 /// \brief In same cases when the dependency check fails we can still
4420 /// vectorize the loop with a dynamic array access check.
4421 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
4424 ScalarEvolution *SE;
4425 const DataLayout *DL;
4426 const Loop *InnermostLoop;
4428 /// \brief Maps access locations (ptr, read/write) to program order.
4429 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
4431 /// \brief Memory access instructions in program order.
4432 SmallVector<Instruction *, 16> InstMap;
4434 /// \brief The program order index to be used for the next instruction.
4437 // We can access this many bytes in parallel safely.
4438 unsigned MaxSafeDepDistBytes;
4440 /// \brief If we see a non-constant dependence distance we can still try to
4441 /// vectorize this loop with runtime checks.
4442 bool ShouldRetryWithRuntimeCheck;
4444 /// \brief Check whether there is a plausible dependence between the two
4447 /// Access \p A must happen before \p B in program order. The two indices
4448 /// identify the index into the program order map.
4450 /// This function checks whether there is a plausible dependence (or the
4451 /// absence of such can't be proved) between the two accesses. If there is a
4452 /// plausible dependence but the dependence distance is bigger than one
4453 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
4454 /// distance is smaller than any other distance encountered so far).
4455 /// Otherwise, this function returns true signaling a possible dependence.
4456 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
4457 const MemAccessInfo &B, unsigned BIdx,
4458 ValueToValueMap &Strides);
4460 /// \brief Check whether the data dependence could prevent store-load
4462 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
4465 } // end anonymous namespace
4467 static bool isInBoundsGep(Value *Ptr) {
4468 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
4469 return GEP->isInBounds();
4473 /// \brief Check whether the access through \p Ptr has a constant stride.
4474 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4475 const Loop *Lp, ValueToValueMap &StridesMap) {
4476 const Type *Ty = Ptr->getType();
4477 assert(Ty->isPointerTy() && "Unexpected non-ptr");
4479 // Make sure that the pointer does not point to aggregate types.
4480 const PointerType *PtrTy = cast<PointerType>(Ty);
4481 if (PtrTy->getElementType()->isAggregateType()) {
4482 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
4487 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
4489 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4491 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
4492 << *Ptr << " SCEV: " << *PtrScev << "\n");
4496 // The accesss function must stride over the innermost loop.
4497 if (Lp != AR->getLoop()) {
4498 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
4499 *Ptr << " SCEV: " << *PtrScev << "\n");
4502 // The address calculation must not wrap. Otherwise, a dependence could be
4504 // An inbounds getelementptr that is a AddRec with a unit stride
4505 // cannot wrap per definition. The unit stride requirement is checked later.
4506 // An getelementptr without an inbounds attribute and unit stride would have
4507 // to access the pointer value "0" which is undefined behavior in address
4508 // space 0, therefore we can also vectorize this case.
4509 bool IsInBoundsGEP = isInBoundsGep(Ptr);
4510 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
4511 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
4512 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
4513 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
4514 << *Ptr << " SCEV: " << *PtrScev << "\n");
4518 // Check the step is constant.
4519 const SCEV *Step = AR->getStepRecurrence(*SE);
4521 // Calculate the pointer stride and check if it is consecutive.
4522 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4524 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
4525 " SCEV: " << *PtrScev << "\n");
4529 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
4530 const APInt &APStepVal = C->getValue()->getValue();
4532 // Huge step value - give up.
4533 if (APStepVal.getBitWidth() > 64)
4536 int64_t StepVal = APStepVal.getSExtValue();
4539 int64_t Stride = StepVal / Size;
4540 int64_t Rem = StepVal % Size;
4544 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
4545 // know we can't "wrap around the address space". In case of address space
4546 // zero we know that this won't happen without triggering undefined behavior.
4547 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
4548 Stride != 1 && Stride != -1)
4554 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
4555 unsigned TypeByteSize) {
4556 // If loads occur at a distance that is not a multiple of a feasible vector
4557 // factor store-load forwarding does not take place.
4558 // Positive dependences might cause troubles because vectorizing them might
4559 // prevent store-load forwarding making vectorized code run a lot slower.
4560 // a[i] = a[i-3] ^ a[i-8];
4561 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
4562 // hence on your typical architecture store-load forwarding does not take
4563 // place. Vectorizing in such cases does not make sense.
4564 // Store-load forwarding distance.
4565 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
4566 // Maximum vector factor.
4567 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
4568 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
4569 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
4571 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
4573 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
4574 MaxVFWithoutSLForwardIssues = (vf >>=1);
4579 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
4580 DEBUG(dbgs() << "LV: Distance " << Distance <<
4581 " that could cause a store-load forwarding conflict\n");
4585 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
4586 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
4587 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
4591 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
4592 const MemAccessInfo &B, unsigned BIdx,
4593 ValueToValueMap &Strides) {
4594 assert (AIdx < BIdx && "Must pass arguments in program order");
4596 Value *APtr = A.getPointer();
4597 Value *BPtr = B.getPointer();
4598 bool AIsWrite = A.getInt();
4599 bool BIsWrite = B.getInt();
4601 // Two reads are independent.
4602 if (!AIsWrite && !BIsWrite)
4605 // We cannot check pointers in different address spaces.
4606 if (APtr->getType()->getPointerAddressSpace() !=
4607 BPtr->getType()->getPointerAddressSpace())
4610 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
4611 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
4613 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
4614 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
4616 const SCEV *Src = AScev;
4617 const SCEV *Sink = BScev;
4619 // If the induction step is negative we have to invert source and sink of the
4621 if (StrideAPtr < 0) {
4624 std::swap(APtr, BPtr);
4625 std::swap(Src, Sink);
4626 std::swap(AIsWrite, BIsWrite);
4627 std::swap(AIdx, BIdx);
4628 std::swap(StrideAPtr, StrideBPtr);
4631 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
4633 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
4634 << "(Induction step: " << StrideAPtr << ")\n");
4635 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
4636 << *InstMap[BIdx] << ": " << *Dist << "\n");
4638 // Need consecutive accesses. We don't want to vectorize
4639 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
4640 // the address space.
4641 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
4642 DEBUG(dbgs() << "Non-consecutive pointer access\n");
4646 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
4648 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
4649 ShouldRetryWithRuntimeCheck = true;
4653 Type *ATy = APtr->getType()->getPointerElementType();
4654 Type *BTy = BPtr->getType()->getPointerElementType();
4655 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
4657 // Negative distances are not plausible dependencies.
4658 const APInt &Val = C->getValue()->getValue();
4659 if (Val.isNegative()) {
4660 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
4661 if (IsTrueDataDependence &&
4662 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
4666 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
4670 // Write to the same location with the same size.
4671 // Could be improved to assert type sizes are the same (i32 == float, etc).
4675 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
4679 assert(Val.isStrictlyPositive() && "Expect a positive value");
4681 // Positive distance bigger than max vectorization factor.
4684 "LV: ReadWrite-Write positive dependency with different types\n");
4688 unsigned Distance = (unsigned) Val.getZExtValue();
4690 // Bail out early if passed-in parameters make vectorization not feasible.
4691 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
4692 unsigned ForcedUnroll = VectorizationInterleave ? VectorizationInterleave : 1;
4694 // The distance must be bigger than the size needed for a vectorized version
4695 // of the operation and the size of the vectorized operation must not be
4696 // bigger than the currrent maximum size.
4697 if (Distance < 2*TypeByteSize ||
4698 2*TypeByteSize > MaxSafeDepDistBytes ||
4699 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
4700 DEBUG(dbgs() << "LV: Failure because of Positive distance "
4701 << Val.getSExtValue() << '\n');
4705 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
4706 Distance : MaxSafeDepDistBytes;
4708 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
4709 if (IsTrueDataDependence &&
4710 couldPreventStoreLoadForward(Distance, TypeByteSize))
4713 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
4714 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
4719 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4720 MemAccessInfoSet &CheckDeps,
4721 ValueToValueMap &Strides) {
4723 MaxSafeDepDistBytes = -1U;
4724 while (!CheckDeps.empty()) {
4725 MemAccessInfo CurAccess = *CheckDeps.begin();
4727 // Get the relevant memory access set.
4728 EquivalenceClasses<MemAccessInfo>::iterator I =
4729 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
4731 // Check accesses within this set.
4732 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
4733 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
4735 // Check every access pair.
4737 CheckDeps.erase(*AI);
4738 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
4740 // Check every accessing instruction pair in program order.
4741 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
4742 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
4743 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
4744 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
4745 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
4747 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
4758 bool LoopVectorizationLegality::canVectorizeMemory() {
4760 typedef SmallVector<Value*, 16> ValueVector;
4761 typedef SmallPtrSet<Value*, 16> ValueSet;
4763 // Holds the Load and Store *instructions*.
4767 // Holds all the different accesses in the loop.
4768 unsigned NumReads = 0;
4769 unsigned NumReadWrites = 0;
4771 PtrRtCheck.Pointers.clear();
4772 PtrRtCheck.Need = false;
4774 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
4775 MemoryDepChecker DepChecker(SE, DL, TheLoop);
4778 for (Loop::block_iterator bb = TheLoop->block_begin(),
4779 be = TheLoop->block_end(); bb != be; ++bb) {
4781 // Scan the BB and collect legal loads and stores.
4782 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4785 // If this is a load, save it. If this instruction can read from memory
4786 // but is not a load, then we quit. Notice that we don't handle function
4787 // calls that read or write.
4788 if (it->mayReadFromMemory()) {
4789 // Many math library functions read the rounding mode. We will only
4790 // vectorize a loop if it contains known function calls that don't set
4791 // the flag. Therefore, it is safe to ignore this read from memory.
4792 CallInst *Call = dyn_cast<CallInst>(it);
4793 if (Call && getIntrinsicIDForCall(Call, TLI))
4796 LoadInst *Ld = dyn_cast<LoadInst>(it);
4797 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
4798 emitAnalysis(Report(Ld)
4799 << "read with atomic ordering or volatile read");
4800 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
4804 Loads.push_back(Ld);
4805 DepChecker.addAccess(Ld);
4809 // Save 'store' instructions. Abort if other instructions write to memory.
4810 if (it->mayWriteToMemory()) {
4811 StoreInst *St = dyn_cast<StoreInst>(it);
4813 emitAnalysis(Report(it) << "instruction cannot be vectorized");
4816 if (!St->isSimple() && !IsAnnotatedParallel) {
4817 emitAnalysis(Report(St)
4818 << "write with atomic ordering or volatile write");
4819 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
4823 Stores.push_back(St);
4824 DepChecker.addAccess(St);
4829 // Now we have two lists that hold the loads and the stores.
4830 // Next, we find the pointers that they use.
4832 // Check if we see any stores. If there are no stores, then we don't
4833 // care if the pointers are *restrict*.
4834 if (!Stores.size()) {
4835 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
4839 AccessAnalysis::DepCandidates DependentAccesses;
4840 AccessAnalysis Accesses(DL, AA, DependentAccesses);
4842 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
4843 // multiple times on the same object. If the ptr is accessed twice, once
4844 // for read and once for write, it will only appear once (on the write
4845 // list). This is okay, since we are going to check for conflicts between
4846 // writes and between reads and writes, but not between reads and reads.
4849 ValueVector::iterator I, IE;
4850 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
4851 StoreInst *ST = cast<StoreInst>(*I);
4852 Value* Ptr = ST->getPointerOperand();
4854 if (isUniform(Ptr)) {
4857 << "write to a loop invariant address could not be vectorized");
4858 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4862 // If we did *not* see this pointer before, insert it to the read-write
4863 // list. At this phase it is only a 'write' list.
4864 if (Seen.insert(Ptr).second) {
4867 AliasAnalysis::Location Loc = AA->getLocation(ST);
4868 // The TBAA metadata could have a control dependency on the predication
4869 // condition, so we cannot rely on it when determining whether or not we
4870 // need runtime pointer checks.
4871 if (blockNeedsPredication(ST->getParent()))
4872 Loc.AATags.TBAA = nullptr;
4874 Accesses.addStore(Loc);
4878 if (IsAnnotatedParallel) {
4880 << "LV: A loop annotated parallel, ignore memory dependency "
4885 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
4886 LoadInst *LD = cast<LoadInst>(*I);
4887 Value* Ptr = LD->getPointerOperand();
4888 // If we did *not* see this pointer before, insert it to the
4889 // read list. If we *did* see it before, then it is already in
4890 // the read-write list. This allows us to vectorize expressions
4891 // such as A[i] += x; Because the address of A[i] is a read-write
4892 // pointer. This only works if the index of A[i] is consecutive.
4893 // If the address of i is unknown (for example A[B[i]]) then we may
4894 // read a few words, modify, and write a few words, and some of the
4895 // words may be written to the same address.
4896 bool IsReadOnlyPtr = false;
4897 if (Seen.insert(Ptr).second ||
4898 !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
4900 IsReadOnlyPtr = true;
4903 AliasAnalysis::Location Loc = AA->getLocation(LD);
4904 // The TBAA metadata could have a control dependency on the predication
4905 // condition, so we cannot rely on it when determining whether or not we
4906 // need runtime pointer checks.
4907 if (blockNeedsPredication(LD->getParent()))
4908 Loc.AATags.TBAA = nullptr;
4910 Accesses.addLoad(Loc, IsReadOnlyPtr);
4913 // If we write (or read-write) to a single destination and there are no
4914 // other reads in this loop then is it safe to vectorize.
4915 if (NumReadWrites == 1 && NumReads == 0) {
4916 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
4920 // Build dependence sets and check whether we need a runtime pointer bounds
4922 Accesses.buildDependenceSets();
4923 bool NeedRTCheck = Accesses.isRTCheckNeeded();
4925 // Find pointers with computable bounds. We are going to use this information
4926 // to place a runtime bound check.
4927 unsigned NumComparisons = 0;
4928 bool CanDoRT = false;
4930 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
4933 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
4934 " pointer comparisons.\n");
4936 // If we only have one set of dependences to check pointers among we don't
4937 // need a runtime check.
4938 if (NumComparisons == 0 && NeedRTCheck)
4939 NeedRTCheck = false;
4941 // Check that we did not collect too many pointers or found an unsizeable
4943 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4949 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
4952 if (NeedRTCheck && !CanDoRT) {
4953 emitAnalysis(Report() << "cannot identify array bounds");
4954 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
4955 "the array bounds.\n");
4960 PtrRtCheck.Need = NeedRTCheck;
4962 bool CanVecMem = true;
4963 if (Accesses.isDependencyCheckNeeded()) {
4964 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
4965 CanVecMem = DepChecker.areDepsSafe(
4966 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
4967 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
4969 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
4970 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
4973 // Clear the dependency checks. We assume they are not needed.
4974 Accesses.resetDepChecks();
4977 PtrRtCheck.Need = true;
4979 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
4980 TheLoop, Strides, true);
4981 // Check that we did not collect too many pointers or found an unsizeable
4983 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4984 if (!CanDoRT && NumComparisons > 0)
4985 emitAnalysis(Report()
4986 << "cannot check memory dependencies at runtime");
4988 emitAnalysis(Report()
4989 << NumComparisons << " exceeds limit of "
4990 << RuntimeMemoryCheckThreshold
4991 << " dependent memory operations checked at runtime");
4992 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
5002 emitAnalysis(Report() << "unsafe dependent memory operations in loop");
5004 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
5005 " need a runtime memory check.\n");
5010 static bool hasMultipleUsesOf(Instruction *I,
5011 SmallPtrSetImpl<Instruction *> &Insts) {
5012 unsigned NumUses = 0;
5013 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
5014 if (Insts.count(dyn_cast<Instruction>(*Use)))
5023 static bool areAllUsesIn(Instruction *I, SmallPtrSetImpl<Instruction *> &Set) {
5024 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
5025 if (!Set.count(dyn_cast<Instruction>(*Use)))
5030 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
5031 ReductionKind Kind) {
5032 if (Phi->getNumIncomingValues() != 2)
5035 // Reduction variables are only found in the loop header block.
5036 if (Phi->getParent() != TheLoop->getHeader())
5039 // Obtain the reduction start value from the value that comes from the loop
5041 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
5043 // ExitInstruction is the single value which is used outside the loop.
5044 // We only allow for a single reduction value to be used outside the loop.
5045 // This includes users of the reduction, variables (which form a cycle
5046 // which ends in the phi node).
5047 Instruction *ExitInstruction = nullptr;
5048 // Indicates that we found a reduction operation in our scan.
5049 bool FoundReduxOp = false;
5051 // We start with the PHI node and scan for all of the users of this
5052 // instruction. All users must be instructions that can be used as reduction
5053 // variables (such as ADD). We must have a single out-of-block user. The cycle
5054 // must include the original PHI.
5055 bool FoundStartPHI = false;
5057 // To recognize min/max patterns formed by a icmp select sequence, we store
5058 // the number of instruction we saw from the recognized min/max pattern,
5059 // to make sure we only see exactly the two instructions.
5060 unsigned NumCmpSelectPatternInst = 0;
5061 ReductionInstDesc ReduxDesc(false, nullptr);
5063 SmallPtrSet<Instruction *, 8> VisitedInsts;
5064 SmallVector<Instruction *, 8> Worklist;
5065 Worklist.push_back(Phi);
5066 VisitedInsts.insert(Phi);
5068 // A value in the reduction can be used:
5069 // - By the reduction:
5070 // - Reduction operation:
5071 // - One use of reduction value (safe).
5072 // - Multiple use of reduction value (not safe).
5074 // - All uses of the PHI must be the reduction (safe).
5075 // - Otherwise, not safe.
5076 // - By one instruction outside of the loop (safe).
5077 // - By further instructions outside of the loop (not safe).
5078 // - By an instruction that is not part of the reduction (not safe).
5080 // * An instruction type other than PHI or the reduction operation.
5081 // * A PHI in the header other than the initial PHI.
5082 while (!Worklist.empty()) {
5083 Instruction *Cur = Worklist.back();
5084 Worklist.pop_back();
5087 // If the instruction has no users then this is a broken chain and can't be
5088 // a reduction variable.
5089 if (Cur->use_empty())
5092 bool IsAPhi = isa<PHINode>(Cur);
5094 // A header PHI use other than the original PHI.
5095 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
5098 // Reductions of instructions such as Div, and Sub is only possible if the
5099 // LHS is the reduction variable.
5100 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
5101 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
5102 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
5105 // Any reduction instruction must be of one of the allowed kinds.
5106 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
5107 if (!ReduxDesc.IsReduction)
5110 // A reduction operation must only have one use of the reduction value.
5111 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
5112 hasMultipleUsesOf(Cur, VisitedInsts))
5115 // All inputs to a PHI node must be a reduction value.
5116 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
5119 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
5120 isa<SelectInst>(Cur)))
5121 ++NumCmpSelectPatternInst;
5122 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
5123 isa<SelectInst>(Cur)))
5124 ++NumCmpSelectPatternInst;
5126 // Check whether we found a reduction operator.
5127 FoundReduxOp |= !IsAPhi;
5129 // Process users of current instruction. Push non-PHI nodes after PHI nodes
5130 // onto the stack. This way we are going to have seen all inputs to PHI
5131 // nodes once we get to them.
5132 SmallVector<Instruction *, 8> NonPHIs;
5133 SmallVector<Instruction *, 8> PHIs;
5134 for (User *U : Cur->users()) {
5135 Instruction *UI = cast<Instruction>(U);
5137 // Check if we found the exit user.
5138 BasicBlock *Parent = UI->getParent();
5139 if (!TheLoop->contains(Parent)) {
5140 // Exit if you find multiple outside users or if the header phi node is
5141 // being used. In this case the user uses the value of the previous
5142 // iteration, in which case we would loose "VF-1" iterations of the
5143 // reduction operation if we vectorize.
5144 if (ExitInstruction != nullptr || Cur == Phi)
5147 // The instruction used by an outside user must be the last instruction
5148 // before we feed back to the reduction phi. Otherwise, we loose VF-1
5149 // operations on the value.
5150 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
5153 ExitInstruction = Cur;
5157 // Process instructions only once (termination). Each reduction cycle
5158 // value must only be used once, except by phi nodes and min/max
5159 // reductions which are represented as a cmp followed by a select.
5160 ReductionInstDesc IgnoredVal(false, nullptr);
5161 if (VisitedInsts.insert(UI).second) {
5162 if (isa<PHINode>(UI))
5165 NonPHIs.push_back(UI);
5166 } else if (!isa<PHINode>(UI) &&
5167 ((!isa<FCmpInst>(UI) &&
5168 !isa<ICmpInst>(UI) &&
5169 !isa<SelectInst>(UI)) ||
5170 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
5173 // Remember that we completed the cycle.
5175 FoundStartPHI = true;
5177 Worklist.append(PHIs.begin(), PHIs.end());
5178 Worklist.append(NonPHIs.begin(), NonPHIs.end());
5181 // This means we have seen one but not the other instruction of the
5182 // pattern or more than just a select and cmp.
5183 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
5184 NumCmpSelectPatternInst != 2)
5187 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
5190 // We found a reduction var if we have reached the original phi node and we
5191 // only have a single instruction with out-of-loop users.
5193 // This instruction is allowed to have out-of-loop users.
5194 AllowedExit.insert(ExitInstruction);
5196 // Save the description of this reduction variable.
5197 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
5198 ReduxDesc.MinMaxKind);
5199 Reductions[Phi] = RD;
5200 // We've ended the cycle. This is a reduction variable if we have an
5201 // outside user and it has a binary op.
5206 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
5207 /// pattern corresponding to a min(X, Y) or max(X, Y).
5208 LoopVectorizationLegality::ReductionInstDesc
5209 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
5210 ReductionInstDesc &Prev) {
5212 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
5213 "Expect a select instruction");
5214 Instruction *Cmp = nullptr;
5215 SelectInst *Select = nullptr;
5217 // We must handle the select(cmp()) as a single instruction. Advance to the
5219 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
5220 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
5221 return ReductionInstDesc(false, I);
5222 return ReductionInstDesc(Select, Prev.MinMaxKind);
5225 // Only handle single use cases for now.
5226 if (!(Select = dyn_cast<SelectInst>(I)))
5227 return ReductionInstDesc(false, I);
5228 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
5229 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
5230 return ReductionInstDesc(false, I);
5231 if (!Cmp->hasOneUse())
5232 return ReductionInstDesc(false, I);
5237 // Look for a min/max pattern.
5238 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5239 return ReductionInstDesc(Select, MRK_UIntMin);
5240 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5241 return ReductionInstDesc(Select, MRK_UIntMax);
5242 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5243 return ReductionInstDesc(Select, MRK_SIntMax);
5244 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5245 return ReductionInstDesc(Select, MRK_SIntMin);
5246 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5247 return ReductionInstDesc(Select, MRK_FloatMin);
5248 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5249 return ReductionInstDesc(Select, MRK_FloatMax);
5250 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5251 return ReductionInstDesc(Select, MRK_FloatMin);
5252 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5253 return ReductionInstDesc(Select, MRK_FloatMax);
5255 return ReductionInstDesc(false, I);
5258 LoopVectorizationLegality::ReductionInstDesc
5259 LoopVectorizationLegality::isReductionInstr(Instruction *I,
5261 ReductionInstDesc &Prev) {
5262 bool FP = I->getType()->isFloatingPointTy();
5263 bool FastMath = FP && I->hasUnsafeAlgebra();
5264 switch (I->getOpcode()) {
5266 return ReductionInstDesc(false, I);
5267 case Instruction::PHI:
5268 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
5269 Kind != RK_FloatMinMax))
5270 return ReductionInstDesc(false, I);
5271 return ReductionInstDesc(I, Prev.MinMaxKind);
5272 case Instruction::Sub:
5273 case Instruction::Add:
5274 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
5275 case Instruction::Mul:
5276 return ReductionInstDesc(Kind == RK_IntegerMult, I);
5277 case Instruction::And:
5278 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
5279 case Instruction::Or:
5280 return ReductionInstDesc(Kind == RK_IntegerOr, I);
5281 case Instruction::Xor:
5282 return ReductionInstDesc(Kind == RK_IntegerXor, I);
5283 case Instruction::FMul:
5284 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
5285 case Instruction::FSub:
5286 case Instruction::FAdd:
5287 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
5288 case Instruction::FCmp:
5289 case Instruction::ICmp:
5290 case Instruction::Select:
5291 if (Kind != RK_IntegerMinMax &&
5292 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
5293 return ReductionInstDesc(false, I);
5294 return isMinMaxSelectCmpPattern(I, Prev);
5298 LoopVectorizationLegality::InductionKind
5299 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
5300 Type *PhiTy = Phi->getType();
5301 // We only handle integer and pointer inductions variables.
5302 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
5303 return IK_NoInduction;
5305 // Check that the PHI is consecutive.
5306 const SCEV *PhiScev = SE->getSCEV(Phi);
5307 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
5309 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
5310 return IK_NoInduction;
5312 const SCEV *Step = AR->getStepRecurrence(*SE);
5314 // Integer inductions need to have a stride of one.
5315 if (PhiTy->isIntegerTy()) {
5317 return IK_IntInduction;
5318 if (Step->isAllOnesValue())
5319 return IK_ReverseIntInduction;
5320 return IK_NoInduction;
5323 // Calculate the pointer stride and check if it is consecutive.
5324 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5326 return IK_NoInduction;
5328 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
5329 Type *PointerElementType = PhiTy->getPointerElementType();
5330 // The pointer stride cannot be determined if the pointer element type is not
5332 if (!PointerElementType->isSized())
5333 return IK_NoInduction;
5335 uint64_t Size = DL->getTypeAllocSize(PointerElementType);
5336 if (C->getValue()->equalsInt(Size))
5337 return IK_PtrInduction;
5338 else if (C->getValue()->equalsInt(0 - Size))
5339 return IK_ReversePtrInduction;
5341 return IK_NoInduction;
5344 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
5345 Value *In0 = const_cast<Value*>(V);
5346 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
5350 return Inductions.count(PN);
5353 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
5354 assert(TheLoop->contains(BB) && "Unknown block used");
5356 // Blocks that do not dominate the latch need predication.
5357 BasicBlock* Latch = TheLoop->getLoopLatch();
5358 return !DT->dominates(BB, Latch);
5361 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
5362 SmallPtrSetImpl<Value *> &SafePtrs) {
5364 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5365 // Check that we don't have a constant expression that can trap as operand.
5366 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
5368 if (Constant *C = dyn_cast<Constant>(*OI))
5372 // We might be able to hoist the load.
5373 if (it->mayReadFromMemory()) {
5374 LoadInst *LI = dyn_cast<LoadInst>(it);
5377 if (!SafePtrs.count(LI->getPointerOperand())) {
5378 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
5379 MaskedOp.insert(LI);
5386 // We don't predicate stores at the moment.
5387 if (it->mayWriteToMemory()) {
5388 StoreInst *SI = dyn_cast<StoreInst>(it);
5389 // We only support predication of stores in basic blocks with one
5394 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
5395 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
5397 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
5398 !isSinglePredecessor) {
5399 // Build a masked store if it is legal for the target, otherwise scalarize
5401 bool isLegalMaskedOp =
5402 isLegalMaskedStore(SI->getValueOperand()->getType(),
5403 SI->getPointerOperand());
5404 if (isLegalMaskedOp) {
5406 MaskedOp.insert(SI);
5415 // The instructions below can trap.
5416 switch (it->getOpcode()) {
5418 case Instruction::UDiv:
5419 case Instruction::SDiv:
5420 case Instruction::URem:
5421 case Instruction::SRem:
5429 LoopVectorizationCostModel::VectorizationFactor
5430 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
5431 // Width 1 means no vectorize
5432 VectorizationFactor Factor = { 1U, 0U };
5433 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
5434 emitAnalysis(Report() << "runtime pointer checks needed. Enable vectorization of this loop with '#pragma clang loop vectorize(enable)' when compiling with -Os");
5435 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
5439 if (!EnableCondStoresVectorization && Legal->NumPredStores) {
5440 emitAnalysis(Report() << "store that is conditionally executed prevents vectorization");
5441 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
5445 // Find the trip count.
5446 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
5447 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
5449 unsigned WidestType = getWidestType();
5450 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
5451 unsigned MaxSafeDepDist = -1U;
5452 if (Legal->getMaxSafeDepDistBytes() != -1U)
5453 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
5454 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
5455 WidestRegister : MaxSafeDepDist);
5456 unsigned MaxVectorSize = WidestRegister / WidestType;
5457 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
5458 DEBUG(dbgs() << "LV: The Widest register is: "
5459 << WidestRegister << " bits.\n");
5461 if (MaxVectorSize == 0) {
5462 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
5466 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
5467 " into one vector!");
5469 unsigned VF = MaxVectorSize;
5471 // If we optimize the program for size, avoid creating the tail loop.
5473 // If we are unable to calculate the trip count then don't try to vectorize.
5475 emitAnalysis(Report() << "unable to calculate the loop count due to complex control flow");
5476 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5480 // Find the maximum SIMD width that can fit within the trip count.
5481 VF = TC % MaxVectorSize;
5486 // If the trip count that we found modulo the vectorization factor is not
5487 // zero then we require a tail.
5489 emitAnalysis(Report() << "cannot optimize for size and vectorize at the "
5490 "same time. Enable vectorization of this loop "
5491 "with '#pragma clang loop vectorize(enable)' "
5492 "when compiling with -Os");
5493 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5498 int UserVF = Hints->getWidth();
5500 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
5501 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
5503 Factor.Width = UserVF;
5507 float Cost = expectedCost(1);
5509 const float ScalarCost = Cost;
5512 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
5514 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
5515 // Ignore scalar width, because the user explicitly wants vectorization.
5516 if (ForceVectorization && VF > 1) {
5518 Cost = expectedCost(Width) / (float)Width;
5521 for (unsigned i=2; i <= VF; i*=2) {
5522 // Notice that the vector loop needs to be executed less times, so
5523 // we need to divide the cost of the vector loops by the width of
5524 // the vector elements.
5525 float VectorCost = expectedCost(i) / (float)i;
5526 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
5527 (int)VectorCost << ".\n");
5528 if (VectorCost < Cost) {
5534 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
5535 << "LV: Vectorization seems to be not beneficial, "
5536 << "but was forced by a user.\n");
5537 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
5538 Factor.Width = Width;
5539 Factor.Cost = Width * Cost;
5543 unsigned LoopVectorizationCostModel::getWidestType() {
5544 unsigned MaxWidth = 8;
5547 for (Loop::block_iterator bb = TheLoop->block_begin(),
5548 be = TheLoop->block_end(); bb != be; ++bb) {
5549 BasicBlock *BB = *bb;
5551 // For each instruction in the loop.
5552 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5553 Type *T = it->getType();
5555 // Ignore ephemeral values.
5556 if (EphValues.count(it))
5559 // Only examine Loads, Stores and PHINodes.
5560 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
5563 // Examine PHI nodes that are reduction variables.
5564 if (PHINode *PN = dyn_cast<PHINode>(it))
5565 if (!Legal->getReductionVars()->count(PN))
5568 // Examine the stored values.
5569 if (StoreInst *ST = dyn_cast<StoreInst>(it))
5570 T = ST->getValueOperand()->getType();
5572 // Ignore loaded pointer types and stored pointer types that are not
5573 // consecutive. However, we do want to take consecutive stores/loads of
5574 // pointer vectors into account.
5575 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
5578 MaxWidth = std::max(MaxWidth,
5579 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
5587 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
5589 unsigned LoopCost) {
5591 // -- The unroll heuristics --
5592 // We unroll the loop in order to expose ILP and reduce the loop overhead.
5593 // There are many micro-architectural considerations that we can't predict
5594 // at this level. For example, frontend pressure (on decode or fetch) due to
5595 // code size, or the number and capabilities of the execution ports.
5597 // We use the following heuristics to select the unroll factor:
5598 // 1. If the code has reductions, then we unroll in order to break the cross
5599 // iteration dependency.
5600 // 2. If the loop is really small, then we unroll in order to reduce the loop
5602 // 3. We don't unroll if we think that we will spill registers to memory due
5603 // to the increased register pressure.
5605 // Use the user preference, unless 'auto' is selected.
5606 int UserUF = Hints->getInterleave();
5610 // When we optimize for size, we don't unroll.
5614 // We used the distance for the unroll factor.
5615 if (Legal->getMaxSafeDepDistBytes() != -1U)
5618 // Do not unroll loops with a relatively small trip count.
5619 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
5620 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
5623 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
5624 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
5628 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
5629 TargetNumRegisters = ForceTargetNumScalarRegs;
5631 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
5632 TargetNumRegisters = ForceTargetNumVectorRegs;
5635 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
5636 // We divide by these constants so assume that we have at least one
5637 // instruction that uses at least one register.
5638 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
5639 R.NumInstructions = std::max(R.NumInstructions, 1U);
5641 // We calculate the unroll factor using the following formula.
5642 // Subtract the number of loop invariants from the number of available
5643 // registers. These registers are used by all of the unrolled instances.
5644 // Next, divide the remaining registers by the number of registers that is
5645 // required by the loop, in order to estimate how many parallel instances
5646 // fit without causing spills. All of this is rounded down if necessary to be
5647 // a power of two. We want power of two unroll factors to simplify any
5648 // addressing operations or alignment considerations.
5649 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
5652 // Don't count the induction variable as unrolled.
5653 if (EnableIndVarRegisterHeur)
5654 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
5655 std::max(1U, (R.MaxLocalUsers - 1)));
5657 // Clamp the unroll factor ranges to reasonable factors.
5658 unsigned MaxInterleaveSize = TTI.getMaxInterleaveFactor();
5660 // Check if the user has overridden the unroll max.
5662 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
5663 MaxInterleaveSize = ForceTargetMaxScalarInterleaveFactor;
5665 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
5666 MaxInterleaveSize = ForceTargetMaxVectorInterleaveFactor;
5669 // If we did not calculate the cost for VF (because the user selected the VF)
5670 // then we calculate the cost of VF here.
5672 LoopCost = expectedCost(VF);
5674 // Clamp the calculated UF to be between the 1 and the max unroll factor
5675 // that the target allows.
5676 if (UF > MaxInterleaveSize)
5677 UF = MaxInterleaveSize;
5681 // Unroll if we vectorized this loop and there is a reduction that could
5682 // benefit from unrolling.
5683 if (VF > 1 && Legal->getReductionVars()->size()) {
5684 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
5688 // Note that if we've already vectorized the loop we will have done the
5689 // runtime check and so unrolling won't require further checks.
5690 bool UnrollingRequiresRuntimePointerCheck =
5691 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
5693 // We want to unroll small loops in order to reduce the loop overhead and
5694 // potentially expose ILP opportunities.
5695 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
5696 if (!UnrollingRequiresRuntimePointerCheck &&
5697 LoopCost < SmallLoopCost) {
5698 // We assume that the cost overhead is 1 and we use the cost model
5699 // to estimate the cost of the loop and unroll until the cost of the
5700 // loop overhead is about 5% of the cost of the loop.
5701 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
5703 // Unroll until store/load ports (estimated by max unroll factor) are
5705 unsigned StoresUF = UF / (Legal->NumStores ? Legal->NumStores : 1);
5706 unsigned LoadsUF = UF / (Legal->NumLoads ? Legal->NumLoads : 1);
5708 // If we have a scalar reduction (vector reductions are already dealt with
5709 // by this point), we can increase the critical path length if the loop
5710 // we're unrolling is inside another loop. Limit, by default to 2, so the
5711 // critical path only gets increased by one reduction operation.
5712 if (Legal->getReductionVars()->size() &&
5713 TheLoop->getLoopDepth() > 1) {
5714 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionUF);
5715 SmallUF = std::min(SmallUF, F);
5716 StoresUF = std::min(StoresUF, F);
5717 LoadsUF = std::min(LoadsUF, F);
5720 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
5721 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
5722 return std::max(StoresUF, LoadsUF);
5725 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
5729 DEBUG(dbgs() << "LV: Not Unrolling.\n");
5733 LoopVectorizationCostModel::RegisterUsage
5734 LoopVectorizationCostModel::calculateRegisterUsage() {
5735 // This function calculates the register usage by measuring the highest number
5736 // of values that are alive at a single location. Obviously, this is a very
5737 // rough estimation. We scan the loop in a topological order in order and
5738 // assign a number to each instruction. We use RPO to ensure that defs are
5739 // met before their users. We assume that each instruction that has in-loop
5740 // users starts an interval. We record every time that an in-loop value is
5741 // used, so we have a list of the first and last occurrences of each
5742 // instruction. Next, we transpose this data structure into a multi map that
5743 // holds the list of intervals that *end* at a specific location. This multi
5744 // map allows us to perform a linear search. We scan the instructions linearly
5745 // and record each time that a new interval starts, by placing it in a set.
5746 // If we find this value in the multi-map then we remove it from the set.
5747 // The max register usage is the maximum size of the set.
5748 // We also search for instructions that are defined outside the loop, but are
5749 // used inside the loop. We need this number separately from the max-interval
5750 // usage number because when we unroll, loop-invariant values do not take
5752 LoopBlocksDFS DFS(TheLoop);
5756 R.NumInstructions = 0;
5758 // Each 'key' in the map opens a new interval. The values
5759 // of the map are the index of the 'last seen' usage of the
5760 // instruction that is the key.
5761 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5762 // Maps instruction to its index.
5763 DenseMap<unsigned, Instruction*> IdxToInstr;
5764 // Marks the end of each interval.
5765 IntervalMap EndPoint;
5766 // Saves the list of instruction indices that are used in the loop.
5767 SmallSet<Instruction*, 8> Ends;
5768 // Saves the list of values that are used in the loop but are
5769 // defined outside the loop, such as arguments and constants.
5770 SmallPtrSet<Value*, 8> LoopInvariants;
5773 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5774 be = DFS.endRPO(); bb != be; ++bb) {
5775 R.NumInstructions += (*bb)->size();
5776 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
5778 Instruction *I = it;
5779 IdxToInstr[Index++] = I;
5781 // Save the end location of each USE.
5782 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
5783 Value *U = I->getOperand(i);
5784 Instruction *Instr = dyn_cast<Instruction>(U);
5786 // Ignore non-instruction values such as arguments, constants, etc.
5787 if (!Instr) continue;
5789 // If this instruction is outside the loop then record it and continue.
5790 if (!TheLoop->contains(Instr)) {
5791 LoopInvariants.insert(Instr);
5795 // Overwrite previous end points.
5796 EndPoint[Instr] = Index;
5802 // Saves the list of intervals that end with the index in 'key'.
5803 typedef SmallVector<Instruction*, 2> InstrList;
5804 DenseMap<unsigned, InstrList> TransposeEnds;
5806 // Transpose the EndPoints to a list of values that end at each index.
5807 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5809 TransposeEnds[it->second].push_back(it->first);
5811 SmallSet<Instruction*, 8> OpenIntervals;
5812 unsigned MaxUsage = 0;
5815 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5816 for (unsigned int i = 0; i < Index; ++i) {
5817 Instruction *I = IdxToInstr[i];
5818 // Ignore instructions that are never used within the loop.
5819 if (!Ends.count(I)) continue;
5821 // Ignore ephemeral values.
5822 if (EphValues.count(I))
5825 // Remove all of the instructions that end at this location.
5826 InstrList &List = TransposeEnds[i];
5827 for (unsigned int j=0, e = List.size(); j < e; ++j)
5828 OpenIntervals.erase(List[j]);
5830 // Count the number of live interals.
5831 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5833 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5834 OpenIntervals.size() << '\n');
5836 // Add the current instruction to the list of open intervals.
5837 OpenIntervals.insert(I);
5840 unsigned Invariant = LoopInvariants.size();
5841 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5842 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5843 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5845 R.LoopInvariantRegs = Invariant;
5846 R.MaxLocalUsers = MaxUsage;
5850 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5854 for (Loop::block_iterator bb = TheLoop->block_begin(),
5855 be = TheLoop->block_end(); bb != be; ++bb) {
5856 unsigned BlockCost = 0;
5857 BasicBlock *BB = *bb;
5859 // For each instruction in the old loop.
5860 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5861 // Skip dbg intrinsics.
5862 if (isa<DbgInfoIntrinsic>(it))
5865 // Ignore ephemeral values.
5866 if (EphValues.count(it))
5869 unsigned C = getInstructionCost(it, VF);
5871 // Check if we should override the cost.
5872 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5873 C = ForceTargetInstructionCost;
5876 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5877 VF << " For instruction: " << *it << '\n');
5880 // We assume that if-converted blocks have a 50% chance of being executed.
5881 // When the code is scalar then some of the blocks are avoided due to CF.
5882 // When the code is vectorized we execute all code paths.
5883 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5892 /// \brief Check whether the address computation for a non-consecutive memory
5893 /// access looks like an unlikely candidate for being merged into the indexing
5896 /// We look for a GEP which has one index that is an induction variable and all
5897 /// other indices are loop invariant. If the stride of this access is also
5898 /// within a small bound we decide that this address computation can likely be
5899 /// merged into the addressing mode.
5900 /// In all other cases, we identify the address computation as complex.
5901 static bool isLikelyComplexAddressComputation(Value *Ptr,
5902 LoopVectorizationLegality *Legal,
5903 ScalarEvolution *SE,
5904 const Loop *TheLoop) {
5905 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5909 // We are looking for a gep with all loop invariant indices except for one
5910 // which should be an induction variable.
5911 unsigned NumOperands = Gep->getNumOperands();
5912 for (unsigned i = 1; i < NumOperands; ++i) {
5913 Value *Opd = Gep->getOperand(i);
5914 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5915 !Legal->isInductionVariable(Opd))
5919 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5920 // can likely be merged into the address computation.
5921 unsigned MaxMergeDistance = 64;
5923 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5927 // Check the step is constant.
5928 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5929 // Calculate the pointer stride and check if it is consecutive.
5930 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5934 const APInt &APStepVal = C->getValue()->getValue();
5936 // Huge step value - give up.
5937 if (APStepVal.getBitWidth() > 64)
5940 int64_t StepVal = APStepVal.getSExtValue();
5942 return StepVal > MaxMergeDistance;
5945 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5946 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5952 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5953 // If we know that this instruction will remain uniform, check the cost of
5954 // the scalar version.
5955 if (Legal->isUniformAfterVectorization(I))
5958 Type *RetTy = I->getType();
5959 Type *VectorTy = ToVectorTy(RetTy, VF);
5961 // TODO: We need to estimate the cost of intrinsic calls.
5962 switch (I->getOpcode()) {
5963 case Instruction::GetElementPtr:
5964 // We mark this instruction as zero-cost because the cost of GEPs in
5965 // vectorized code depends on whether the corresponding memory instruction
5966 // is scalarized or not. Therefore, we handle GEPs with the memory
5967 // instruction cost.
5969 case Instruction::Br: {
5970 return TTI.getCFInstrCost(I->getOpcode());
5972 case Instruction::PHI:
5973 //TODO: IF-converted IFs become selects.
5975 case Instruction::Add:
5976 case Instruction::FAdd:
5977 case Instruction::Sub:
5978 case Instruction::FSub:
5979 case Instruction::Mul:
5980 case Instruction::FMul:
5981 case Instruction::UDiv:
5982 case Instruction::SDiv:
5983 case Instruction::FDiv:
5984 case Instruction::URem:
5985 case Instruction::SRem:
5986 case Instruction::FRem:
5987 case Instruction::Shl:
5988 case Instruction::LShr:
5989 case Instruction::AShr:
5990 case Instruction::And:
5991 case Instruction::Or:
5992 case Instruction::Xor: {
5993 // Since we will replace the stride by 1 the multiplication should go away.
5994 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5996 // Certain instructions can be cheaper to vectorize if they have a constant
5997 // second vector operand. One example of this are shifts on x86.
5998 TargetTransformInfo::OperandValueKind Op1VK =
5999 TargetTransformInfo::OK_AnyValue;
6000 TargetTransformInfo::OperandValueKind Op2VK =
6001 TargetTransformInfo::OK_AnyValue;
6002 TargetTransformInfo::OperandValueProperties Op1VP =
6003 TargetTransformInfo::OP_None;
6004 TargetTransformInfo::OperandValueProperties Op2VP =
6005 TargetTransformInfo::OP_None;
6006 Value *Op2 = I->getOperand(1);
6008 // Check for a splat of a constant or for a non uniform vector of constants.
6009 if (isa<ConstantInt>(Op2)) {
6010 ConstantInt *CInt = cast<ConstantInt>(Op2);
6011 if (CInt && CInt->getValue().isPowerOf2())
6012 Op2VP = TargetTransformInfo::OP_PowerOf2;
6013 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
6014 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
6015 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
6016 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
6018 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
6019 if (CInt && CInt->getValue().isPowerOf2())
6020 Op2VP = TargetTransformInfo::OP_PowerOf2;
6021 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
6025 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
6028 case Instruction::Select: {
6029 SelectInst *SI = cast<SelectInst>(I);
6030 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
6031 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
6032 Type *CondTy = SI->getCondition()->getType();
6034 CondTy = VectorType::get(CondTy, VF);
6036 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
6038 case Instruction::ICmp:
6039 case Instruction::FCmp: {
6040 Type *ValTy = I->getOperand(0)->getType();
6041 VectorTy = ToVectorTy(ValTy, VF);
6042 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
6044 case Instruction::Store:
6045 case Instruction::Load: {
6046 StoreInst *SI = dyn_cast<StoreInst>(I);
6047 LoadInst *LI = dyn_cast<LoadInst>(I);
6048 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
6050 VectorTy = ToVectorTy(ValTy, VF);
6052 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
6053 unsigned AS = SI ? SI->getPointerAddressSpace() :
6054 LI->getPointerAddressSpace();
6055 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
6056 // We add the cost of address computation here instead of with the gep
6057 // instruction because only here we know whether the operation is
6060 return TTI.getAddressComputationCost(VectorTy) +
6061 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
6063 // Scalarized loads/stores.
6064 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
6065 bool Reverse = ConsecutiveStride < 0;
6066 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
6067 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
6068 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
6069 bool IsComplexComputation =
6070 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
6072 // The cost of extracting from the value vector and pointer vector.
6073 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
6074 for (unsigned i = 0; i < VF; ++i) {
6075 // The cost of extracting the pointer operand.
6076 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
6077 // In case of STORE, the cost of ExtractElement from the vector.
6078 // In case of LOAD, the cost of InsertElement into the returned
6080 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
6081 Instruction::InsertElement,
6085 // The cost of the scalar loads/stores.
6086 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
6087 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
6092 // Wide load/stores.
6093 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
6094 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
6097 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
6101 case Instruction::ZExt:
6102 case Instruction::SExt:
6103 case Instruction::FPToUI:
6104 case Instruction::FPToSI:
6105 case Instruction::FPExt:
6106 case Instruction::PtrToInt:
6107 case Instruction::IntToPtr:
6108 case Instruction::SIToFP:
6109 case Instruction::UIToFP:
6110 case Instruction::Trunc:
6111 case Instruction::FPTrunc:
6112 case Instruction::BitCast: {
6113 // We optimize the truncation of induction variable.
6114 // The cost of these is the same as the scalar operation.
6115 if (I->getOpcode() == Instruction::Trunc &&
6116 Legal->isInductionVariable(I->getOperand(0)))
6117 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
6118 I->getOperand(0)->getType());
6120 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
6121 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
6123 case Instruction::Call: {
6124 CallInst *CI = cast<CallInst>(I);
6125 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
6126 assert(ID && "Not an intrinsic call!");
6127 Type *RetTy = ToVectorTy(CI->getType(), VF);
6128 SmallVector<Type*, 4> Tys;
6129 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
6130 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
6131 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
6134 // We are scalarizing the instruction. Return the cost of the scalar
6135 // instruction, plus the cost of insert and extract into vector
6136 // elements, times the vector width.
6139 if (!RetTy->isVoidTy() && VF != 1) {
6140 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
6142 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
6145 // The cost of inserting the results plus extracting each one of the
6147 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
6150 // The cost of executing VF copies of the scalar instruction. This opcode
6151 // is unknown. Assume that it is the same as 'mul'.
6152 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
6158 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
6159 if (Scalar->isVoidTy() || VF == 1)
6161 return VectorType::get(Scalar, VF);
6164 char LoopVectorize::ID = 0;
6165 static const char lv_name[] = "Loop Vectorization";
6166 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
6167 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
6168 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
6169 INITIALIZE_PASS_DEPENDENCY(AssumptionTracker)
6170 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
6171 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
6172 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
6173 INITIALIZE_PASS_DEPENDENCY(LCSSA)
6174 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
6175 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
6176 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
6179 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
6180 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
6184 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
6185 // Check for a store.
6186 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
6187 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
6189 // Check for a load.
6190 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
6191 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
6197 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
6198 bool IfPredicateStore) {
6199 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
6200 // Holds vector parameters or scalars, in case of uniform vals.
6201 SmallVector<VectorParts, 4> Params;
6203 setDebugLocFromInst(Builder, Instr);
6205 // Find all of the vectorized parameters.
6206 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
6207 Value *SrcOp = Instr->getOperand(op);
6209 // If we are accessing the old induction variable, use the new one.
6210 if (SrcOp == OldInduction) {
6211 Params.push_back(getVectorValue(SrcOp));
6215 // Try using previously calculated values.
6216 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
6218 // If the src is an instruction that appeared earlier in the basic block
6219 // then it should already be vectorized.
6220 if (SrcInst && OrigLoop->contains(SrcInst)) {
6221 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
6222 // The parameter is a vector value from earlier.
6223 Params.push_back(WidenMap.get(SrcInst));
6225 // The parameter is a scalar from outside the loop. Maybe even a constant.
6226 VectorParts Scalars;
6227 Scalars.append(UF, SrcOp);
6228 Params.push_back(Scalars);
6232 assert(Params.size() == Instr->getNumOperands() &&
6233 "Invalid number of operands");
6235 // Does this instruction return a value ?
6236 bool IsVoidRetTy = Instr->getType()->isVoidTy();
6238 Value *UndefVec = IsVoidRetTy ? nullptr :
6239 UndefValue::get(Instr->getType());
6240 // Create a new entry in the WidenMap and initialize it to Undef or Null.
6241 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
6243 Instruction *InsertPt = Builder.GetInsertPoint();
6244 BasicBlock *IfBlock = Builder.GetInsertBlock();
6245 BasicBlock *CondBlock = nullptr;
6248 Loop *VectorLp = nullptr;
6249 if (IfPredicateStore) {
6250 assert(Instr->getParent()->getSinglePredecessor() &&
6251 "Only support single predecessor blocks");
6252 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
6253 Instr->getParent());
6254 VectorLp = LI->getLoopFor(IfBlock);
6255 assert(VectorLp && "Must have a loop for this block");
6258 // For each vector unroll 'part':
6259 for (unsigned Part = 0; Part < UF; ++Part) {
6260 // For each scalar that we create:
6262 // Start an "if (pred) a[i] = ..." block.
6263 Value *Cmp = nullptr;
6264 if (IfPredicateStore) {
6265 if (Cond[Part]->getType()->isVectorTy())
6267 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
6268 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
6269 ConstantInt::get(Cond[Part]->getType(), 1));
6270 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
6271 LoopVectorBody.push_back(CondBlock);
6272 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
6273 // Update Builder with newly created basic block.
6274 Builder.SetInsertPoint(InsertPt);
6277 Instruction *Cloned = Instr->clone();
6279 Cloned->setName(Instr->getName() + ".cloned");
6280 // Replace the operands of the cloned instructions with extracted scalars.
6281 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
6282 Value *Op = Params[op][Part];
6283 Cloned->setOperand(op, Op);
6286 // Place the cloned scalar in the new loop.
6287 Builder.Insert(Cloned);
6289 // If the original scalar returns a value we need to place it in a vector
6290 // so that future users will be able to use it.
6292 VecResults[Part] = Cloned;
6295 if (IfPredicateStore) {
6296 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
6297 LoopVectorBody.push_back(NewIfBlock);
6298 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
6299 Builder.SetInsertPoint(InsertPt);
6300 Instruction *OldBr = IfBlock->getTerminator();
6301 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
6302 OldBr->eraseFromParent();
6303 IfBlock = NewIfBlock;
6308 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
6309 StoreInst *SI = dyn_cast<StoreInst>(Instr);
6310 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
6312 return scalarizeInstruction(Instr, IfPredicateStore);
6315 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
6319 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
6323 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
6325 // When unrolling and the VF is 1, we only need to add a simple scalar.
6326 Type *ITy = Val->getType();
6327 assert(!ITy->isVectorTy() && "Val must be a scalar");
6328 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
6329 return Builder.CreateAdd(Val, C, "induction");