1 //===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops
11 // and generates target-independent LLVM-IR.
12 // The vectorizer uses the TargetTransformInfo analysis to estimate the costs
13 // of instructions in order to estimate the profitability of vectorization.
15 // The loop vectorizer combines consecutive loop iterations into a single
16 // 'wide' iteration. After this transformation the index is incremented
17 // by the SIMD vector width, and not by one.
19 // This pass has three parts:
20 // 1. The main loop pass that drives the different parts.
21 // 2. LoopVectorizationLegality - A unit that checks for the legality
22 // of the vectorization.
23 // 3. InnerLoopVectorizer - A unit that performs the actual
24 // widening of instructions.
25 // 4. LoopVectorizationCostModel - A unit that checks for the profitability
26 // of vectorization. It decides on the optimal vector width, which
27 // can be one, if vectorization is not profitable.
29 //===----------------------------------------------------------------------===//
31 // The reduction-variable vectorization is based on the paper:
32 // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
34 // Variable uniformity checks are inspired by:
35 // Karrenberg, R. and Hack, S. Whole Function Vectorization.
37 // Other ideas/concepts are from:
38 // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
40 // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of
41 // Vectorizing Compilers.
43 //===----------------------------------------------------------------------===//
45 #include "llvm/Transforms/Vectorize.h"
46 #include "llvm/ADT/DenseMap.h"
47 #include "llvm/ADT/EquivalenceClasses.h"
48 #include "llvm/ADT/Hashing.h"
49 #include "llvm/ADT/MapVector.h"
50 #include "llvm/ADT/SetVector.h"
51 #include "llvm/ADT/SmallPtrSet.h"
52 #include "llvm/ADT/SmallSet.h"
53 #include "llvm/ADT/SmallVector.h"
54 #include "llvm/ADT/Statistic.h"
55 #include "llvm/ADT/StringExtras.h"
56 #include "llvm/Analysis/AliasAnalysis.h"
57 #include "llvm/Analysis/AliasSetTracker.h"
58 #include "llvm/Analysis/AssumptionCache.h"
59 #include "llvm/Analysis/BlockFrequencyInfo.h"
60 #include "llvm/Analysis/CodeMetrics.h"
61 #include "llvm/Analysis/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 collectStridedAccess(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 AssumptionCache *AC, 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, AC, 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 AssumptionCache *AC;
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 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
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, AC, 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<AssumptionCacheTracker>();
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);
1855 VectorParts Mask = createBlockInMask(Instr->getParent());
1858 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1859 "We do not allow storing to uniform addresses");
1860 setDebugLocFromInst(Builder, SI);
1861 // We don't want to update the value in the map as it might be used in
1862 // another expression. So don't use a reference type for "StoredVal".
1863 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1865 for (unsigned Part = 0; Part < UF; ++Part) {
1866 // Calculate the pointer for the specific unroll-part.
1867 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1870 // If we store to reverse consecutive memory locations then we need
1871 // to reverse the order of elements in the stored value.
1872 StoredVal[Part] = reverseVector(StoredVal[Part]);
1873 // If the address is consecutive but reversed, then the
1874 // wide store needs to start at the last vector element.
1875 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1876 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1879 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1880 DataTy->getPointerTo(AddressSpace));
1883 if (Legal->isMaskRequired(SI))
1884 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
1887 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
1888 propagateMetadata(NewSI, SI);
1894 assert(LI && "Must have a load instruction");
1895 setDebugLocFromInst(Builder, LI);
1896 for (unsigned Part = 0; Part < UF; ++Part) {
1897 // Calculate the pointer for the specific unroll-part.
1898 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1901 // If the address is consecutive but reversed, then the
1902 // wide load needs to start at the last vector element.
1903 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1904 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1908 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1909 DataTy->getPointerTo(AddressSpace));
1910 if (Legal->isMaskRequired(LI))
1911 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
1912 UndefValue::get(DataTy),
1913 "wide.masked.load");
1915 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
1916 propagateMetadata(NewLI, LI);
1917 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
1921 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1922 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1923 // Holds vector parameters or scalars, in case of uniform vals.
1924 SmallVector<VectorParts, 4> Params;
1926 setDebugLocFromInst(Builder, Instr);
1928 // Find all of the vectorized parameters.
1929 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1930 Value *SrcOp = Instr->getOperand(op);
1932 // If we are accessing the old induction variable, use the new one.
1933 if (SrcOp == OldInduction) {
1934 Params.push_back(getVectorValue(SrcOp));
1938 // Try using previously calculated values.
1939 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1941 // If the src is an instruction that appeared earlier in the basic block
1942 // then it should already be vectorized.
1943 if (SrcInst && OrigLoop->contains(SrcInst)) {
1944 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1945 // The parameter is a vector value from earlier.
1946 Params.push_back(WidenMap.get(SrcInst));
1948 // The parameter is a scalar from outside the loop. Maybe even a constant.
1949 VectorParts Scalars;
1950 Scalars.append(UF, SrcOp);
1951 Params.push_back(Scalars);
1955 assert(Params.size() == Instr->getNumOperands() &&
1956 "Invalid number of operands");
1958 // Does this instruction return a value ?
1959 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1961 Value *UndefVec = IsVoidRetTy ? nullptr :
1962 UndefValue::get(VectorType::get(Instr->getType(), VF));
1963 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1964 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1966 Instruction *InsertPt = Builder.GetInsertPoint();
1967 BasicBlock *IfBlock = Builder.GetInsertBlock();
1968 BasicBlock *CondBlock = nullptr;
1971 Loop *VectorLp = nullptr;
1972 if (IfPredicateStore) {
1973 assert(Instr->getParent()->getSinglePredecessor() &&
1974 "Only support single predecessor blocks");
1975 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1976 Instr->getParent());
1977 VectorLp = LI->getLoopFor(IfBlock);
1978 assert(VectorLp && "Must have a loop for this block");
1981 // For each vector unroll 'part':
1982 for (unsigned Part = 0; Part < UF; ++Part) {
1983 // For each scalar that we create:
1984 for (unsigned Width = 0; Width < VF; ++Width) {
1987 Value *Cmp = nullptr;
1988 if (IfPredicateStore) {
1989 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1990 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1991 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1992 LoopVectorBody.push_back(CondBlock);
1993 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
1994 // Update Builder with newly created basic block.
1995 Builder.SetInsertPoint(InsertPt);
1998 Instruction *Cloned = Instr->clone();
2000 Cloned->setName(Instr->getName() + ".cloned");
2001 // Replace the operands of the cloned instructions with extracted scalars.
2002 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2003 Value *Op = Params[op][Part];
2004 // Param is a vector. Need to extract the right lane.
2005 if (Op->getType()->isVectorTy())
2006 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
2007 Cloned->setOperand(op, Op);
2010 // Place the cloned scalar in the new loop.
2011 Builder.Insert(Cloned);
2013 // If the original scalar returns a value we need to place it in a vector
2014 // so that future users will be able to use it.
2016 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
2017 Builder.getInt32(Width));
2019 if (IfPredicateStore) {
2020 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2021 LoopVectorBody.push_back(NewIfBlock);
2022 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
2023 Builder.SetInsertPoint(InsertPt);
2024 Instruction *OldBr = IfBlock->getTerminator();
2025 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2026 OldBr->eraseFromParent();
2027 IfBlock = NewIfBlock;
2033 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
2037 if (Instruction *I = dyn_cast<Instruction>(V))
2038 return I->getParent() == Loc->getParent() ? I : nullptr;
2042 std::pair<Instruction *, Instruction *>
2043 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
2044 Instruction *tnullptr = nullptr;
2045 if (!Legal->mustCheckStrides())
2046 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2048 IRBuilder<> ChkBuilder(Loc);
2051 Value *Check = nullptr;
2052 Instruction *FirstInst = nullptr;
2053 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
2054 SE = Legal->strides_end();
2056 Value *Ptr = stripIntegerCast(*SI);
2057 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
2059 // Store the first instruction we create.
2060 FirstInst = getFirstInst(FirstInst, C, Loc);
2062 Check = ChkBuilder.CreateOr(Check, C);
2067 // We have to do this trickery because the IRBuilder might fold the check to a
2068 // constant expression in which case there is no Instruction anchored in a
2070 LLVMContext &Ctx = Loc->getContext();
2071 Instruction *TheCheck =
2072 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2073 ChkBuilder.Insert(TheCheck, "stride.not.one");
2074 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2076 return std::make_pair(FirstInst, TheCheck);
2079 std::pair<Instruction *, Instruction *>
2080 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
2081 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
2082 Legal->getRuntimePointerCheck();
2084 Instruction *tnullptr = nullptr;
2085 if (!PtrRtCheck->Need)
2086 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2088 unsigned NumPointers = PtrRtCheck->Pointers.size();
2089 SmallVector<TrackingVH<Value> , 2> Starts;
2090 SmallVector<TrackingVH<Value> , 2> Ends;
2092 LLVMContext &Ctx = Loc->getContext();
2093 SCEVExpander Exp(*SE, "induction");
2094 Instruction *FirstInst = nullptr;
2096 for (unsigned i = 0; i < NumPointers; ++i) {
2097 Value *Ptr = PtrRtCheck->Pointers[i];
2098 const SCEV *Sc = SE->getSCEV(Ptr);
2100 if (SE->isLoopInvariant(Sc, OrigLoop)) {
2101 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
2103 Starts.push_back(Ptr);
2104 Ends.push_back(Ptr);
2106 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
2107 unsigned AS = Ptr->getType()->getPointerAddressSpace();
2109 // Use this type for pointer arithmetic.
2110 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
2112 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
2113 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
2114 Starts.push_back(Start);
2115 Ends.push_back(End);
2119 IRBuilder<> ChkBuilder(Loc);
2120 // Our instructions might fold to a constant.
2121 Value *MemoryRuntimeCheck = nullptr;
2122 for (unsigned i = 0; i < NumPointers; ++i) {
2123 for (unsigned j = i+1; j < NumPointers; ++j) {
2124 // No need to check if two readonly pointers intersect.
2125 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
2128 // Only need to check pointers between two different dependency sets.
2129 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
2131 // Only need to check pointers in the same alias set.
2132 if (PtrRtCheck->AliasSetId[i] != PtrRtCheck->AliasSetId[j])
2135 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
2136 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
2138 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
2139 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
2140 "Trying to bounds check pointers with different address spaces");
2142 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
2143 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
2145 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
2146 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
2147 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
2148 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
2150 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
2151 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
2152 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
2153 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
2154 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
2155 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2156 if (MemoryRuntimeCheck) {
2157 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
2159 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2161 MemoryRuntimeCheck = IsConflict;
2165 // We have to do this trickery because the IRBuilder might fold the check to a
2166 // constant expression in which case there is no Instruction anchored in a
2168 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
2169 ConstantInt::getTrue(Ctx));
2170 ChkBuilder.Insert(Check, "memcheck.conflict");
2171 FirstInst = getFirstInst(FirstInst, Check, Loc);
2172 return std::make_pair(FirstInst, Check);
2175 void InnerLoopVectorizer::createEmptyLoop() {
2177 In this function we generate a new loop. The new loop will contain
2178 the vectorized instructions while the old loop will continue to run the
2181 [ ] <-- Back-edge taken count overflow check.
2184 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2187 || [ ] <-- vector pre header.
2191 || [ ]_| <-- vector loop.
2194 | >[ ] <--- middle-block.
2197 -|- >[ ] <--- new preheader.
2201 | [ ]_| <-- old scalar loop to handle remainder.
2204 >[ ] <-- exit block.
2208 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2209 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2210 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2211 assert(BypassBlock && "Invalid loop structure");
2212 assert(ExitBlock && "Must have an exit block");
2214 // Some loops have a single integer induction variable, while other loops
2215 // don't. One example is c++ iterators that often have multiple pointer
2216 // induction variables. In the code below we also support a case where we
2217 // don't have a single induction variable.
2218 OldInduction = Legal->getInduction();
2219 Type *IdxTy = Legal->getWidestInductionType();
2221 // Find the loop boundaries.
2222 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2223 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2225 // The exit count might have the type of i64 while the phi is i32. This can
2226 // happen if we have an induction variable that is sign extended before the
2227 // compare. The only way that we get a backedge taken count is that the
2228 // induction variable was signed and as such will not overflow. In such a case
2229 // truncation is legal.
2230 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2231 IdxTy->getPrimitiveSizeInBits())
2232 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2234 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2235 // Get the total trip count from the count by adding 1.
2236 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2237 SE->getConstant(BackedgeTakeCount->getType(), 1));
2239 // Expand the trip count and place the new instructions in the preheader.
2240 // Notice that the pre-header does not change, only the loop body.
2241 SCEVExpander Exp(*SE, "induction");
2243 // We need to test whether the backedge-taken count is uint##_max. Adding one
2244 // to it will cause overflow and an incorrect loop trip count in the vector
2245 // body. In case of overflow we want to directly jump to the scalar remainder
2247 Value *BackedgeCount =
2248 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2249 BypassBlock->getTerminator());
2250 if (BackedgeCount->getType()->isPointerTy())
2251 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2252 "backedge.ptrcnt.to.int",
2253 BypassBlock->getTerminator());
2254 Instruction *CheckBCOverflow =
2255 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2256 Constant::getAllOnesValue(BackedgeCount->getType()),
2257 "backedge.overflow", BypassBlock->getTerminator());
2259 // The loop index does not have to start at Zero. Find the original start
2260 // value from the induction PHI node. If we don't have an induction variable
2261 // then we know that it starts at zero.
2262 Builder.SetInsertPoint(BypassBlock->getTerminator());
2263 Value *StartIdx = ExtendedIdx = OldInduction ?
2264 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2266 ConstantInt::get(IdxTy, 0);
2268 // We need an instruction to anchor the overflow check on. StartIdx needs to
2269 // be defined before the overflow check branch. Because the scalar preheader
2270 // is going to merge the start index and so the overflow branch block needs to
2271 // contain a definition of the start index.
2272 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2273 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2274 BypassBlock->getTerminator());
2276 // Count holds the overall loop count (N).
2277 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2278 BypassBlock->getTerminator());
2280 LoopBypassBlocks.push_back(BypassBlock);
2282 // Split the single block loop into the two loop structure described above.
2283 BasicBlock *VectorPH =
2284 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2285 BasicBlock *VecBody =
2286 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2287 BasicBlock *MiddleBlock =
2288 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2289 BasicBlock *ScalarPH =
2290 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2292 // Create and register the new vector loop.
2293 Loop* Lp = new Loop();
2294 Loop *ParentLoop = OrigLoop->getParentLoop();
2296 // Insert the new loop into the loop nest and register the new basic blocks
2297 // before calling any utilities such as SCEV that require valid LoopInfo.
2299 ParentLoop->addChildLoop(Lp);
2300 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
2301 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
2302 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
2304 LI->addTopLevelLoop(Lp);
2306 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
2308 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2310 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2312 // Generate the induction variable.
2313 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2314 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2315 // The loop step is equal to the vectorization factor (num of SIMD elements)
2316 // times the unroll factor (num of SIMD instructions).
2317 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2319 // This is the IR builder that we use to add all of the logic for bypassing
2320 // the new vector loop.
2321 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2322 setDebugLocFromInst(BypassBuilder,
2323 getDebugLocFromInstOrOperands(OldInduction));
2325 // We may need to extend the index in case there is a type mismatch.
2326 // We know that the count starts at zero and does not overflow.
2327 if (Count->getType() != IdxTy) {
2328 // The exit count can be of pointer type. Convert it to the correct
2330 if (ExitCount->getType()->isPointerTy())
2331 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2333 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2336 // Add the start index to the loop count to get the new end index.
2337 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2339 // Now we need to generate the expression for N - (N % VF), which is
2340 // the part that the vectorized body will execute.
2341 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2342 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2343 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2344 "end.idx.rnd.down");
2346 // Now, compare the new count to zero. If it is zero skip the vector loop and
2347 // jump to the scalar loop.
2349 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2351 BasicBlock *LastBypassBlock = BypassBlock;
2353 // Generate code to check that the loops trip count that we computed by adding
2354 // one to the backedge-taken count will not overflow.
2356 auto PastOverflowCheck =
2357 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2358 BasicBlock *CheckBlock =
2359 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2361 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2362 LoopBypassBlocks.push_back(CheckBlock);
2363 Instruction *OldTerm = LastBypassBlock->getTerminator();
2364 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2365 OldTerm->eraseFromParent();
2366 LastBypassBlock = CheckBlock;
2369 // Generate the code to check that the strides we assumed to be one are really
2370 // one. We want the new basic block to start at the first instruction in a
2371 // sequence of instructions that form a check.
2372 Instruction *StrideCheck;
2373 Instruction *FirstCheckInst;
2374 std::tie(FirstCheckInst, StrideCheck) =
2375 addStrideCheck(LastBypassBlock->getTerminator());
2377 // Create a new block containing the stride check.
2378 BasicBlock *CheckBlock =
2379 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2381 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2382 LoopBypassBlocks.push_back(CheckBlock);
2384 // Replace the branch into the memory check block with a conditional branch
2385 // for the "few elements case".
2386 Instruction *OldTerm = LastBypassBlock->getTerminator();
2387 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2388 OldTerm->eraseFromParent();
2391 LastBypassBlock = CheckBlock;
2394 // Generate the code that checks in runtime if arrays overlap. We put the
2395 // checks into a separate block to make the more common case of few elements
2397 Instruction *MemRuntimeCheck;
2398 std::tie(FirstCheckInst, MemRuntimeCheck) =
2399 addRuntimeCheck(LastBypassBlock->getTerminator());
2400 if (MemRuntimeCheck) {
2401 // Create a new block containing the memory check.
2402 BasicBlock *CheckBlock =
2403 LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
2405 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2406 LoopBypassBlocks.push_back(CheckBlock);
2408 // Replace the branch into the memory check block with a conditional branch
2409 // for the "few elements case".
2410 Instruction *OldTerm = LastBypassBlock->getTerminator();
2411 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2412 OldTerm->eraseFromParent();
2414 Cmp = MemRuntimeCheck;
2415 LastBypassBlock = CheckBlock;
2418 LastBypassBlock->getTerminator()->eraseFromParent();
2419 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2422 // We are going to resume the execution of the scalar loop.
2423 // Go over all of the induction variables that we found and fix the
2424 // PHIs that are left in the scalar version of the loop.
2425 // The starting values of PHI nodes depend on the counter of the last
2426 // iteration in the vectorized loop.
2427 // If we come from a bypass edge then we need to start from the original
2430 // This variable saves the new starting index for the scalar loop.
2431 PHINode *ResumeIndex = nullptr;
2432 LoopVectorizationLegality::InductionList::iterator I, E;
2433 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2434 // Set builder to point to last bypass block.
2435 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2436 for (I = List->begin(), E = List->end(); I != E; ++I) {
2437 PHINode *OrigPhi = I->first;
2438 LoopVectorizationLegality::InductionInfo II = I->second;
2440 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2441 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2442 MiddleBlock->getTerminator());
2443 // We might have extended the type of the induction variable but we need a
2444 // truncated version for the scalar loop.
2445 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2446 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2447 MiddleBlock->getTerminator()) : nullptr;
2449 // Create phi nodes to merge from the backedge-taken check block.
2450 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2451 ScalarPH->getTerminator());
2452 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2454 PHINode *BCTruncResumeVal = nullptr;
2455 if (OrigPhi == OldInduction) {
2457 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2458 ScalarPH->getTerminator());
2459 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2462 Value *EndValue = nullptr;
2464 case LoopVectorizationLegality::IK_NoInduction:
2465 llvm_unreachable("Unknown induction");
2466 case LoopVectorizationLegality::IK_IntInduction: {
2467 // Handle the integer induction counter.
2468 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2470 // We have the canonical induction variable.
2471 if (OrigPhi == OldInduction) {
2472 // Create a truncated version of the resume value for the scalar loop,
2473 // we might have promoted the type to a larger width.
2475 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2476 // The new PHI merges the original incoming value, in case of a bypass,
2477 // or the value at the end of the vectorized loop.
2478 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2479 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2480 TruncResumeVal->addIncoming(EndValue, VecBody);
2482 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2484 // We know what the end value is.
2485 EndValue = IdxEndRoundDown;
2486 // We also know which PHI node holds it.
2487 ResumeIndex = ResumeVal;
2491 // Not the canonical induction variable - add the vector loop count to the
2493 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2494 II.StartValue->getType(),
2496 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
2499 case LoopVectorizationLegality::IK_ReverseIntInduction: {
2500 // Convert the CountRoundDown variable to the PHI size.
2501 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2502 II.StartValue->getType(),
2504 // Handle reverse integer induction counter.
2505 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
2508 case LoopVectorizationLegality::IK_PtrInduction: {
2509 // For pointer induction variables, calculate the offset using
2511 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
2515 case LoopVectorizationLegality::IK_ReversePtrInduction: {
2516 // The value at the end of the loop for the reverse pointer is calculated
2517 // by creating a GEP with a negative index starting from the start value.
2518 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
2519 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
2521 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
2527 // The new PHI merges the original incoming value, in case of a bypass,
2528 // or the value at the end of the vectorized loop.
2529 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2530 if (OrigPhi == OldInduction)
2531 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2533 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2535 ResumeVal->addIncoming(EndValue, VecBody);
2537 // Fix the scalar body counter (PHI node).
2538 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2540 // The old induction's phi node in the scalar body needs the truncated
2542 if (OrigPhi == OldInduction) {
2543 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2544 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2546 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2547 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2551 // If we are generating a new induction variable then we also need to
2552 // generate the code that calculates the exit value. This value is not
2553 // simply the end of the counter because we may skip the vectorized body
2554 // in case of a runtime check.
2556 assert(!ResumeIndex && "Unexpected resume value found");
2557 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2558 MiddleBlock->getTerminator());
2559 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2560 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2561 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2564 // Make sure that we found the index where scalar loop needs to continue.
2565 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2566 "Invalid resume Index");
2568 // Add a check in the middle block to see if we have completed
2569 // all of the iterations in the first vector loop.
2570 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2571 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2572 ResumeIndex, "cmp.n",
2573 MiddleBlock->getTerminator());
2575 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2576 // Remove the old terminator.
2577 MiddleBlock->getTerminator()->eraseFromParent();
2579 // Create i+1 and fill the PHINode.
2580 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2581 Induction->addIncoming(StartIdx, VectorPH);
2582 Induction->addIncoming(NextIdx, VecBody);
2583 // Create the compare.
2584 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2585 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2587 // Now we have two terminators. Remove the old one from the block.
2588 VecBody->getTerminator()->eraseFromParent();
2590 // Get ready to start creating new instructions into the vectorized body.
2591 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2594 LoopVectorPreHeader = VectorPH;
2595 LoopScalarPreHeader = ScalarPH;
2596 LoopMiddleBlock = MiddleBlock;
2597 LoopExitBlock = ExitBlock;
2598 LoopVectorBody.push_back(VecBody);
2599 LoopScalarBody = OldBasicBlock;
2601 LoopVectorizeHints Hints(Lp, true);
2602 Hints.setAlreadyVectorized();
2605 /// This function returns the identity element (or neutral element) for
2606 /// the operation K.
2608 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2613 // Adding, Xoring, Oring zero to a number does not change it.
2614 return ConstantInt::get(Tp, 0);
2615 case RK_IntegerMult:
2616 // Multiplying a number by 1 does not change it.
2617 return ConstantInt::get(Tp, 1);
2619 // AND-ing a number with an all-1 value does not change it.
2620 return ConstantInt::get(Tp, -1, true);
2622 // Multiplying a number by 1 does not change it.
2623 return ConstantFP::get(Tp, 1.0L);
2625 // Adding zero to a number does not change it.
2626 return ConstantFP::get(Tp, 0.0L);
2628 llvm_unreachable("Unknown reduction kind");
2632 /// This function translates the reduction kind to an LLVM binary operator.
2634 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2636 case LoopVectorizationLegality::RK_IntegerAdd:
2637 return Instruction::Add;
2638 case LoopVectorizationLegality::RK_IntegerMult:
2639 return Instruction::Mul;
2640 case LoopVectorizationLegality::RK_IntegerOr:
2641 return Instruction::Or;
2642 case LoopVectorizationLegality::RK_IntegerAnd:
2643 return Instruction::And;
2644 case LoopVectorizationLegality::RK_IntegerXor:
2645 return Instruction::Xor;
2646 case LoopVectorizationLegality::RK_FloatMult:
2647 return Instruction::FMul;
2648 case LoopVectorizationLegality::RK_FloatAdd:
2649 return Instruction::FAdd;
2650 case LoopVectorizationLegality::RK_IntegerMinMax:
2651 return Instruction::ICmp;
2652 case LoopVectorizationLegality::RK_FloatMinMax:
2653 return Instruction::FCmp;
2655 llvm_unreachable("Unknown reduction operation");
2659 Value *createMinMaxOp(IRBuilder<> &Builder,
2660 LoopVectorizationLegality::MinMaxReductionKind RK,
2663 CmpInst::Predicate P = CmpInst::ICMP_NE;
2666 llvm_unreachable("Unknown min/max reduction kind");
2667 case LoopVectorizationLegality::MRK_UIntMin:
2668 P = CmpInst::ICMP_ULT;
2670 case LoopVectorizationLegality::MRK_UIntMax:
2671 P = CmpInst::ICMP_UGT;
2673 case LoopVectorizationLegality::MRK_SIntMin:
2674 P = CmpInst::ICMP_SLT;
2676 case LoopVectorizationLegality::MRK_SIntMax:
2677 P = CmpInst::ICMP_SGT;
2679 case LoopVectorizationLegality::MRK_FloatMin:
2680 P = CmpInst::FCMP_OLT;
2682 case LoopVectorizationLegality::MRK_FloatMax:
2683 P = CmpInst::FCMP_OGT;
2688 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2689 RK == LoopVectorizationLegality::MRK_FloatMax)
2690 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2692 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2694 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2699 struct CSEDenseMapInfo {
2700 static bool canHandle(Instruction *I) {
2701 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2702 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2704 static inline Instruction *getEmptyKey() {
2705 return DenseMapInfo<Instruction *>::getEmptyKey();
2707 static inline Instruction *getTombstoneKey() {
2708 return DenseMapInfo<Instruction *>::getTombstoneKey();
2710 static unsigned getHashValue(Instruction *I) {
2711 assert(canHandle(I) && "Unknown instruction!");
2712 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2713 I->value_op_end()));
2715 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2716 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2717 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2719 return LHS->isIdenticalTo(RHS);
2724 /// \brief Check whether this block is a predicated block.
2725 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2726 /// = ...; " blocks. We start with one vectorized basic block. For every
2727 /// conditional block we split this vectorized block. Therefore, every second
2728 /// block will be a predicated one.
2729 static bool isPredicatedBlock(unsigned BlockNum) {
2730 return BlockNum % 2;
2733 ///\brief Perform cse of induction variable instructions.
2734 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2735 // Perform simple cse.
2736 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2737 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2738 BasicBlock *BB = BBs[i];
2739 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2740 Instruction *In = I++;
2742 if (!CSEDenseMapInfo::canHandle(In))
2745 // Check if we can replace this instruction with any of the
2746 // visited instructions.
2747 if (Instruction *V = CSEMap.lookup(In)) {
2748 In->replaceAllUsesWith(V);
2749 In->eraseFromParent();
2752 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2753 // ...;" blocks for predicated stores. Every second block is a predicated
2755 if (isPredicatedBlock(i))
2763 /// \brief Adds a 'fast' flag to floating point operations.
2764 static Value *addFastMathFlag(Value *V) {
2765 if (isa<FPMathOperator>(V)){
2766 FastMathFlags Flags;
2767 Flags.setUnsafeAlgebra();
2768 cast<Instruction>(V)->setFastMathFlags(Flags);
2773 void InnerLoopVectorizer::vectorizeLoop() {
2774 //===------------------------------------------------===//
2776 // Notice: any optimization or new instruction that go
2777 // into the code below should be also be implemented in
2780 //===------------------------------------------------===//
2781 Constant *Zero = Builder.getInt32(0);
2783 // In order to support reduction variables we need to be able to vectorize
2784 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2785 // stages. First, we create a new vector PHI node with no incoming edges.
2786 // We use this value when we vectorize all of the instructions that use the
2787 // PHI. Next, after all of the instructions in the block are complete we
2788 // add the new incoming edges to the PHI. At this point all of the
2789 // instructions in the basic block are vectorized, so we can use them to
2790 // construct the PHI.
2791 PhiVector RdxPHIsToFix;
2793 // Scan the loop in a topological order to ensure that defs are vectorized
2795 LoopBlocksDFS DFS(OrigLoop);
2798 // Vectorize all of the blocks in the original loop.
2799 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2800 be = DFS.endRPO(); bb != be; ++bb)
2801 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2803 // At this point every instruction in the original loop is widened to
2804 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2805 // that we vectorized. The PHI nodes are currently empty because we did
2806 // not want to introduce cycles. Notice that the remaining PHI nodes
2807 // that we need to fix are reduction variables.
2809 // Create the 'reduced' values for each of the induction vars.
2810 // The reduced values are the vector values that we scalarize and combine
2811 // after the loop is finished.
2812 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2814 PHINode *RdxPhi = *it;
2815 assert(RdxPhi && "Unable to recover vectorized PHI");
2817 // Find the reduction variable descriptor.
2818 assert(Legal->getReductionVars()->count(RdxPhi) &&
2819 "Unable to find the reduction variable");
2820 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2821 (*Legal->getReductionVars())[RdxPhi];
2823 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2825 // We need to generate a reduction vector from the incoming scalar.
2826 // To do so, we need to generate the 'identity' vector and override
2827 // one of the elements with the incoming scalar reduction. We need
2828 // to do it in the vector-loop preheader.
2829 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
2831 // This is the vector-clone of the value that leaves the loop.
2832 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2833 Type *VecTy = VectorExit[0]->getType();
2835 // Find the reduction identity variable. Zero for addition, or, xor,
2836 // one for multiplication, -1 for And.
2839 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2840 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2841 // MinMax reduction have the start value as their identify.
2843 VectorStart = Identity = RdxDesc.StartValue;
2845 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2850 // Handle other reduction kinds:
2852 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2853 VecTy->getScalarType());
2856 // This vector is the Identity vector where the first element is the
2857 // incoming scalar reduction.
2858 VectorStart = RdxDesc.StartValue;
2860 Identity = ConstantVector::getSplat(VF, Iden);
2862 // This vector is the Identity vector where the first element is the
2863 // incoming scalar reduction.
2864 VectorStart = Builder.CreateInsertElement(Identity,
2865 RdxDesc.StartValue, Zero);
2869 // Fix the vector-loop phi.
2871 // Reductions do not have to start at zero. They can start with
2872 // any loop invariant values.
2873 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2874 BasicBlock *Latch = OrigLoop->getLoopLatch();
2875 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2876 VectorParts &Val = getVectorValue(LoopVal);
2877 for (unsigned part = 0; part < UF; ++part) {
2878 // Make sure to add the reduction stat value only to the
2879 // first unroll part.
2880 Value *StartVal = (part == 0) ? VectorStart : Identity;
2881 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
2882 LoopVectorPreHeader);
2883 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2884 LoopVectorBody.back());
2887 // Before each round, move the insertion point right between
2888 // the PHIs and the values we are going to write.
2889 // This allows us to write both PHINodes and the extractelement
2891 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2893 VectorParts RdxParts;
2894 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2895 for (unsigned part = 0; part < UF; ++part) {
2896 // This PHINode contains the vectorized reduction variable, or
2897 // the initial value vector, if we bypass the vector loop.
2898 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2899 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2900 Value *StartVal = (part == 0) ? VectorStart : Identity;
2901 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2902 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2903 NewPhi->addIncoming(RdxExitVal[part],
2904 LoopVectorBody.back());
2905 RdxParts.push_back(NewPhi);
2908 // Reduce all of the unrolled parts into a single vector.
2909 Value *ReducedPartRdx = RdxParts[0];
2910 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2911 setDebugLocFromInst(Builder, ReducedPartRdx);
2912 for (unsigned part = 1; part < UF; ++part) {
2913 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2914 // Floating point operations had to be 'fast' to enable the reduction.
2915 ReducedPartRdx = addFastMathFlag(
2916 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2917 ReducedPartRdx, "bin.rdx"));
2919 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2920 ReducedPartRdx, RdxParts[part]);
2924 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2925 // and vector ops, reducing the set of values being computed by half each
2927 assert(isPowerOf2_32(VF) &&
2928 "Reduction emission only supported for pow2 vectors!");
2929 Value *TmpVec = ReducedPartRdx;
2930 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2931 for (unsigned i = VF; i != 1; i >>= 1) {
2932 // Move the upper half of the vector to the lower half.
2933 for (unsigned j = 0; j != i/2; ++j)
2934 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2936 // Fill the rest of the mask with undef.
2937 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2938 UndefValue::get(Builder.getInt32Ty()));
2941 Builder.CreateShuffleVector(TmpVec,
2942 UndefValue::get(TmpVec->getType()),
2943 ConstantVector::get(ShuffleMask),
2946 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2947 // Floating point operations had to be 'fast' to enable the reduction.
2948 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2949 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2951 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2954 // The result is in the first element of the vector.
2955 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2956 Builder.getInt32(0));
2959 // Create a phi node that merges control-flow from the backedge-taken check
2960 // block and the middle block.
2961 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
2962 LoopScalarPreHeader->getTerminator());
2963 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
2964 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2966 // Now, we need to fix the users of the reduction variable
2967 // inside and outside of the scalar remainder loop.
2968 // We know that the loop is in LCSSA form. We need to update the
2969 // PHI nodes in the exit blocks.
2970 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2971 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2972 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2973 if (!LCSSAPhi) break;
2975 // All PHINodes need to have a single entry edge, or two if
2976 // we already fixed them.
2977 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2979 // We found our reduction value exit-PHI. Update it with the
2980 // incoming bypass edge.
2981 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2982 // Add an edge coming from the bypass.
2983 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2986 }// end of the LCSSA phi scan.
2988 // Fix the scalar loop reduction variable with the incoming reduction sum
2989 // from the vector body and from the backedge value.
2990 int IncomingEdgeBlockIdx =
2991 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2992 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2993 // Pick the other block.
2994 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2995 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
2996 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2997 }// end of for each redux variable.
3001 // Remove redundant induction instructions.
3002 cse(LoopVectorBody);
3005 void InnerLoopVectorizer::fixLCSSAPHIs() {
3006 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3007 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3008 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3009 if (!LCSSAPhi) break;
3010 if (LCSSAPhi->getNumIncomingValues() == 1)
3011 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
3016 InnerLoopVectorizer::VectorParts
3017 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
3018 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
3021 // Look for cached value.
3022 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
3023 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
3024 if (ECEntryIt != MaskCache.end())
3025 return ECEntryIt->second;
3027 VectorParts SrcMask = createBlockInMask(Src);
3029 // The terminator has to be a branch inst!
3030 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
3031 assert(BI && "Unexpected terminator found");
3033 if (BI->isConditional()) {
3034 VectorParts EdgeMask = getVectorValue(BI->getCondition());
3036 if (BI->getSuccessor(0) != Dst)
3037 for (unsigned part = 0; part < UF; ++part)
3038 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
3040 for (unsigned part = 0; part < UF; ++part)
3041 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
3043 MaskCache[Edge] = EdgeMask;
3047 MaskCache[Edge] = SrcMask;
3051 InnerLoopVectorizer::VectorParts
3052 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
3053 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
3055 // Loop incoming mask is all-one.
3056 if (OrigLoop->getHeader() == BB) {
3057 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
3058 return getVectorValue(C);
3061 // This is the block mask. We OR all incoming edges, and with zero.
3062 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
3063 VectorParts BlockMask = getVectorValue(Zero);
3066 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
3067 VectorParts EM = createEdgeMask(*it, BB);
3068 for (unsigned part = 0; part < UF; ++part)
3069 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
3075 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
3076 InnerLoopVectorizer::VectorParts &Entry,
3077 unsigned UF, unsigned VF, PhiVector *PV) {
3078 PHINode* P = cast<PHINode>(PN);
3079 // Handle reduction variables:
3080 if (Legal->getReductionVars()->count(P)) {
3081 for (unsigned part = 0; part < UF; ++part) {
3082 // This is phase one of vectorizing PHIs.
3083 Type *VecTy = (VF == 1) ? PN->getType() :
3084 VectorType::get(PN->getType(), VF);
3085 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
3086 LoopVectorBody.back()-> getFirstInsertionPt());
3092 setDebugLocFromInst(Builder, P);
3093 // Check for PHI nodes that are lowered to vector selects.
3094 if (P->getParent() != OrigLoop->getHeader()) {
3095 // We know that all PHIs in non-header blocks are converted into
3096 // selects, so we don't have to worry about the insertion order and we
3097 // can just use the builder.
3098 // At this point we generate the predication tree. There may be
3099 // duplications since this is a simple recursive scan, but future
3100 // optimizations will clean it up.
3102 unsigned NumIncoming = P->getNumIncomingValues();
3104 // Generate a sequence of selects of the form:
3105 // SELECT(Mask3, In3,
3106 // SELECT(Mask2, In2,
3108 for (unsigned In = 0; In < NumIncoming; In++) {
3109 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
3111 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3113 for (unsigned part = 0; part < UF; ++part) {
3114 // We might have single edge PHIs (blocks) - use an identity
3115 // 'select' for the first PHI operand.
3117 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3120 // Select between the current value and the previous incoming edge
3121 // based on the incoming mask.
3122 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3123 Entry[part], "predphi");
3129 // This PHINode must be an induction variable.
3130 // Make sure that we know about it.
3131 assert(Legal->getInductionVars()->count(P) &&
3132 "Not an induction variable");
3134 LoopVectorizationLegality::InductionInfo II =
3135 Legal->getInductionVars()->lookup(P);
3138 case LoopVectorizationLegality::IK_NoInduction:
3139 llvm_unreachable("Unknown induction");
3140 case LoopVectorizationLegality::IK_IntInduction: {
3141 assert(P->getType() == II.StartValue->getType() && "Types must match");
3142 Type *PhiTy = P->getType();
3144 if (P == OldInduction) {
3145 // Handle the canonical induction variable. We might have had to
3147 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
3149 // Handle other induction variables that are now based on the
3151 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
3153 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
3154 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
3157 Broadcasted = getBroadcastInstrs(Broadcasted);
3158 // After broadcasting the induction variable we need to make the vector
3159 // consecutive by adding 0, 1, 2, etc.
3160 for (unsigned part = 0; part < UF; ++part)
3161 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
3164 case LoopVectorizationLegality::IK_ReverseIntInduction:
3165 case LoopVectorizationLegality::IK_PtrInduction:
3166 case LoopVectorizationLegality::IK_ReversePtrInduction:
3167 // Handle reverse integer and pointer inductions.
3168 Value *StartIdx = ExtendedIdx;
3169 // This is the normalized GEP that starts counting at zero.
3170 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
3173 // Handle the reverse integer induction variable case.
3174 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
3175 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
3176 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
3178 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
3181 // This is a new value so do not hoist it out.
3182 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
3183 // After broadcasting the induction variable we need to make the
3184 // vector consecutive by adding ... -3, -2, -1, 0.
3185 for (unsigned part = 0; part < UF; ++part)
3186 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
3191 // Handle the pointer induction variable case.
3192 assert(P->getType()->isPointerTy() && "Unexpected type.");
3194 // Is this a reverse induction ptr or a consecutive induction ptr.
3195 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
3198 // This is the vector of results. Notice that we don't generate
3199 // vector geps because scalar geps result in better code.
3200 for (unsigned part = 0; part < UF; ++part) {
3202 int EltIndex = (part) * (Reverse ? -1 : 1);
3203 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3206 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3208 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3210 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3212 Entry[part] = SclrGep;
3216 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3217 for (unsigned int i = 0; i < VF; ++i) {
3218 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
3219 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3222 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3224 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3226 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3228 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3229 Builder.getInt32(i),
3232 Entry[part] = VecVal;
3238 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3239 // For each instruction in the old loop.
3240 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3241 VectorParts &Entry = WidenMap.get(it);
3242 switch (it->getOpcode()) {
3243 case Instruction::Br:
3244 // Nothing to do for PHIs and BR, since we already took care of the
3245 // loop control flow instructions.
3247 case Instruction::PHI:{
3248 // Vectorize PHINodes.
3249 widenPHIInstruction(it, Entry, UF, VF, PV);
3253 case Instruction::Add:
3254 case Instruction::FAdd:
3255 case Instruction::Sub:
3256 case Instruction::FSub:
3257 case Instruction::Mul:
3258 case Instruction::FMul:
3259 case Instruction::UDiv:
3260 case Instruction::SDiv:
3261 case Instruction::FDiv:
3262 case Instruction::URem:
3263 case Instruction::SRem:
3264 case Instruction::FRem:
3265 case Instruction::Shl:
3266 case Instruction::LShr:
3267 case Instruction::AShr:
3268 case Instruction::And:
3269 case Instruction::Or:
3270 case Instruction::Xor: {
3271 // Just widen binops.
3272 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3273 setDebugLocFromInst(Builder, BinOp);
3274 VectorParts &A = getVectorValue(it->getOperand(0));
3275 VectorParts &B = getVectorValue(it->getOperand(1));
3277 // Use this vector value for all users of the original instruction.
3278 for (unsigned Part = 0; Part < UF; ++Part) {
3279 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3281 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3282 VecOp->copyIRFlags(BinOp);
3287 propagateMetadata(Entry, it);
3290 case Instruction::Select: {
3292 // If the selector is loop invariant we can create a select
3293 // instruction with a scalar condition. Otherwise, use vector-select.
3294 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3296 setDebugLocFromInst(Builder, it);
3298 // The condition can be loop invariant but still defined inside the
3299 // loop. This means that we can't just use the original 'cond' value.
3300 // We have to take the 'vectorized' value and pick the first lane.
3301 // Instcombine will make this a no-op.
3302 VectorParts &Cond = getVectorValue(it->getOperand(0));
3303 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3304 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3306 Value *ScalarCond = (VF == 1) ? Cond[0] :
3307 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3309 for (unsigned Part = 0; Part < UF; ++Part) {
3310 Entry[Part] = Builder.CreateSelect(
3311 InvariantCond ? ScalarCond : Cond[Part],
3316 propagateMetadata(Entry, it);
3320 case Instruction::ICmp:
3321 case Instruction::FCmp: {
3322 // Widen compares. Generate vector compares.
3323 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3324 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3325 setDebugLocFromInst(Builder, it);
3326 VectorParts &A = getVectorValue(it->getOperand(0));
3327 VectorParts &B = getVectorValue(it->getOperand(1));
3328 for (unsigned Part = 0; Part < UF; ++Part) {
3331 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3333 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3337 propagateMetadata(Entry, it);
3341 case Instruction::Store:
3342 case Instruction::Load:
3343 vectorizeMemoryInstruction(it);
3345 case Instruction::ZExt:
3346 case Instruction::SExt:
3347 case Instruction::FPToUI:
3348 case Instruction::FPToSI:
3349 case Instruction::FPExt:
3350 case Instruction::PtrToInt:
3351 case Instruction::IntToPtr:
3352 case Instruction::SIToFP:
3353 case Instruction::UIToFP:
3354 case Instruction::Trunc:
3355 case Instruction::FPTrunc:
3356 case Instruction::BitCast: {
3357 CastInst *CI = dyn_cast<CastInst>(it);
3358 setDebugLocFromInst(Builder, it);
3359 /// Optimize the special case where the source is the induction
3360 /// variable. Notice that we can only optimize the 'trunc' case
3361 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3362 /// c. other casts depend on pointer size.
3363 if (CI->getOperand(0) == OldInduction &&
3364 it->getOpcode() == Instruction::Trunc) {
3365 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3367 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3368 for (unsigned Part = 0; Part < UF; ++Part)
3369 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
3370 propagateMetadata(Entry, it);
3373 /// Vectorize casts.
3374 Type *DestTy = (VF == 1) ? CI->getType() :
3375 VectorType::get(CI->getType(), VF);
3377 VectorParts &A = getVectorValue(it->getOperand(0));
3378 for (unsigned Part = 0; Part < UF; ++Part)
3379 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3380 propagateMetadata(Entry, it);
3384 case Instruction::Call: {
3385 // Ignore dbg intrinsics.
3386 if (isa<DbgInfoIntrinsic>(it))
3388 setDebugLocFromInst(Builder, it);
3390 Module *M = BB->getParent()->getParent();
3391 CallInst *CI = cast<CallInst>(it);
3392 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3393 assert(ID && "Not an intrinsic call!");
3395 case Intrinsic::assume:
3396 case Intrinsic::lifetime_end:
3397 case Intrinsic::lifetime_start:
3398 scalarizeInstruction(it);
3401 bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
3402 for (unsigned Part = 0; Part < UF; ++Part) {
3403 SmallVector<Value *, 4> Args;
3404 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3405 if (HasScalarOpd && i == 1) {
3406 Args.push_back(CI->getArgOperand(i));
3409 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3410 Args.push_back(Arg[Part]);
3412 Type *Tys[] = {CI->getType()};
3414 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3416 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3417 Entry[Part] = Builder.CreateCall(F, Args);
3420 propagateMetadata(Entry, it);
3427 // All other instructions are unsupported. Scalarize them.
3428 scalarizeInstruction(it);
3431 }// end of for_each instr.
3434 void InnerLoopVectorizer::updateAnalysis() {
3435 // Forget the original basic block.
3436 SE->forgetLoop(OrigLoop);
3438 // Update the dominator tree information.
3439 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3440 "Entry does not dominate exit.");
3442 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3443 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3444 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3446 // Due to if predication of stores we might create a sequence of "if(pred)
3447 // a[i] = ...; " blocks.
3448 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3450 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3451 else if (isPredicatedBlock(i)) {
3452 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3454 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3458 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3459 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3460 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3461 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3463 DEBUG(DT->verifyDomTree());
3466 /// \brief Check whether it is safe to if-convert this phi node.
3468 /// Phi nodes with constant expressions that can trap are not safe to if
3470 static bool canIfConvertPHINodes(BasicBlock *BB) {
3471 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3472 PHINode *Phi = dyn_cast<PHINode>(I);
3475 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3476 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3483 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3484 if (!EnableIfConversion) {
3485 emitAnalysis(Report() << "if-conversion is disabled");
3489 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3491 // A list of pointers that we can safely read and write to.
3492 SmallPtrSet<Value *, 8> SafePointes;
3494 // Collect safe addresses.
3495 for (Loop::block_iterator BI = TheLoop->block_begin(),
3496 BE = TheLoop->block_end(); BI != BE; ++BI) {
3497 BasicBlock *BB = *BI;
3499 if (blockNeedsPredication(BB))
3502 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3503 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3504 SafePointes.insert(LI->getPointerOperand());
3505 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3506 SafePointes.insert(SI->getPointerOperand());
3510 // Collect the blocks that need predication.
3511 BasicBlock *Header = TheLoop->getHeader();
3512 for (Loop::block_iterator BI = TheLoop->block_begin(),
3513 BE = TheLoop->block_end(); BI != BE; ++BI) {
3514 BasicBlock *BB = *BI;
3516 // We don't support switch statements inside loops.
3517 if (!isa<BranchInst>(BB->getTerminator())) {
3518 emitAnalysis(Report(BB->getTerminator())
3519 << "loop contains a switch statement");
3523 // We must be able to predicate all blocks that need to be predicated.
3524 if (blockNeedsPredication(BB)) {
3525 if (!blockCanBePredicated(BB, SafePointes)) {
3526 emitAnalysis(Report(BB->getTerminator())
3527 << "control flow cannot be substituted for a select");
3530 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3531 emitAnalysis(Report(BB->getTerminator())
3532 << "control flow cannot be substituted for a select");
3537 // We can if-convert this loop.
3541 bool LoopVectorizationLegality::canVectorize() {
3542 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3543 // be canonicalized.
3544 if (!TheLoop->getLoopPreheader()) {
3546 Report() << "loop control flow is not understood by vectorizer");
3550 // We can only vectorize innermost loops.
3551 if (TheLoop->getSubLoopsVector().size()) {
3552 emitAnalysis(Report() << "loop is not the innermost loop");
3556 // We must have a single backedge.
3557 if (TheLoop->getNumBackEdges() != 1) {
3559 Report() << "loop control flow is not understood by vectorizer");
3563 // We must have a single exiting block.
3564 if (!TheLoop->getExitingBlock()) {
3566 Report() << "loop control flow is not understood by vectorizer");
3570 // We only handle bottom-tested loops, i.e. loop in which the condition is
3571 // checked at the end of each iteration. With that we can assume that all
3572 // instructions in the loop are executed the same number of times.
3573 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
3575 Report() << "loop control flow is not understood by vectorizer");
3579 // We need to have a loop header.
3580 DEBUG(dbgs() << "LV: Found a loop: " <<
3581 TheLoop->getHeader()->getName() << '\n');
3583 // Check if we can if-convert non-single-bb loops.
3584 unsigned NumBlocks = TheLoop->getNumBlocks();
3585 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3586 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3590 // ScalarEvolution needs to be able to find the exit count.
3591 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3592 if (ExitCount == SE->getCouldNotCompute()) {
3593 emitAnalysis(Report() << "could not determine number of loop iterations");
3594 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3598 // Check if we can vectorize the instructions and CFG in this loop.
3599 if (!canVectorizeInstrs()) {
3600 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3604 // Go over each instruction and look at memory deps.
3605 if (!canVectorizeMemory()) {
3606 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3610 // Collect all of the variables that remain uniform after vectorization.
3611 collectLoopUniforms();
3613 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3614 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
3617 // Okay! We can vectorize. At this point we don't have any other mem analysis
3618 // which may limit our maximum vectorization factor, so just return true with
3623 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3624 if (Ty->isPointerTy())
3625 return DL.getIntPtrType(Ty);
3627 // It is possible that char's or short's overflow when we ask for the loop's
3628 // trip count, work around this by changing the type size.
3629 if (Ty->getScalarSizeInBits() < 32)
3630 return Type::getInt32Ty(Ty->getContext());
3635 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3636 Ty0 = convertPointerToIntegerType(DL, Ty0);
3637 Ty1 = convertPointerToIntegerType(DL, Ty1);
3638 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3643 /// \brief Check that the instruction has outside loop users and is not an
3644 /// identified reduction variable.
3645 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3646 SmallPtrSetImpl<Value *> &Reductions) {
3647 // Reduction instructions are allowed to have exit users. All other
3648 // instructions must not have external users.
3649 if (!Reductions.count(Inst))
3650 //Check that all of the users of the loop are inside the BB.
3651 for (User *U : Inst->users()) {
3652 Instruction *UI = cast<Instruction>(U);
3653 // This user may be a reduction exit value.
3654 if (!TheLoop->contains(UI)) {
3655 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3662 bool LoopVectorizationLegality::canVectorizeInstrs() {
3663 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3664 BasicBlock *Header = TheLoop->getHeader();
3666 // Look for the attribute signaling the absence of NaNs.
3667 Function &F = *Header->getParent();
3668 if (F.hasFnAttribute("no-nans-fp-math"))
3669 HasFunNoNaNAttr = F.getAttributes().getAttribute(
3670 AttributeSet::FunctionIndex,
3671 "no-nans-fp-math").getValueAsString() == "true";
3673 // For each block in the loop.
3674 for (Loop::block_iterator bb = TheLoop->block_begin(),
3675 be = TheLoop->block_end(); bb != be; ++bb) {
3677 // Scan the instructions in the block and look for hazards.
3678 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3681 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3682 Type *PhiTy = Phi->getType();
3683 // Check that this PHI type is allowed.
3684 if (!PhiTy->isIntegerTy() &&
3685 !PhiTy->isFloatingPointTy() &&
3686 !PhiTy->isPointerTy()) {
3687 emitAnalysis(Report(it)
3688 << "loop control flow is not understood by vectorizer");
3689 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3693 // If this PHINode is not in the header block, then we know that we
3694 // can convert it to select during if-conversion. No need to check if
3695 // the PHIs in this block are induction or reduction variables.
3696 if (*bb != Header) {
3697 // Check that this instruction has no outside users or is an
3698 // identified reduction value with an outside user.
3699 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3701 emitAnalysis(Report(it) << "value could not be identified as "
3702 "an induction or reduction variable");
3706 // We only allow if-converted PHIs with exactly two incoming values.
3707 if (Phi->getNumIncomingValues() != 2) {
3708 emitAnalysis(Report(it)
3709 << "control flow not understood by vectorizer");
3710 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3714 // This is the value coming from the preheader.
3715 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3716 // Check if this is an induction variable.
3717 InductionKind IK = isInductionVariable(Phi);
3719 if (IK_NoInduction != IK) {
3720 // Get the widest type.
3722 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3724 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3726 // Int inductions are special because we only allow one IV.
3727 if (IK == IK_IntInduction) {
3728 // Use the phi node with the widest type as induction. Use the last
3729 // one if there are multiple (no good reason for doing this other
3730 // than it is expedient).
3731 if (!Induction || PhiTy == WidestIndTy)
3735 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3736 Inductions[Phi] = InductionInfo(StartValue, IK);
3738 // Until we explicitly handle the case of an induction variable with
3739 // an outside loop user we have to give up vectorizing this loop.
3740 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3741 emitAnalysis(Report(it) << "use of induction value outside of the "
3742 "loop is not handled by vectorizer");
3749 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3750 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3753 if (AddReductionVar(Phi, RK_IntegerMult)) {
3754 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3757 if (AddReductionVar(Phi, RK_IntegerOr)) {
3758 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3761 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3762 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3765 if (AddReductionVar(Phi, RK_IntegerXor)) {
3766 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3769 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3770 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3773 if (AddReductionVar(Phi, RK_FloatMult)) {
3774 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3777 if (AddReductionVar(Phi, RK_FloatAdd)) {
3778 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3781 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3782 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3787 emitAnalysis(Report(it) << "value that could not be identified as "
3788 "reduction is used outside the loop");
3789 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3791 }// end of PHI handling
3793 // We still don't handle functions. However, we can ignore dbg intrinsic
3794 // calls and we do handle certain intrinsic and libm functions.
3795 CallInst *CI = dyn_cast<CallInst>(it);
3796 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3797 emitAnalysis(Report(it) << "call instruction cannot be vectorized");
3798 DEBUG(dbgs() << "LV: Found a call site.\n");
3802 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
3803 // second argument is the same (i.e. loop invariant)
3805 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
3806 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
3807 emitAnalysis(Report(it)
3808 << "intrinsic instruction cannot be vectorized");
3809 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
3814 // Check that the instruction return type is vectorizable.
3815 // Also, we can't vectorize extractelement instructions.
3816 if ((!VectorType::isValidElementType(it->getType()) &&
3817 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3818 emitAnalysis(Report(it)
3819 << "instruction return type cannot be vectorized");
3820 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3824 // Check that the stored type is vectorizable.
3825 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3826 Type *T = ST->getValueOperand()->getType();
3827 if (!VectorType::isValidElementType(T)) {
3828 emitAnalysis(Report(ST) << "store instruction cannot be vectorized");
3831 if (EnableMemAccessVersioning)
3832 collectStridedAccess(ST);
3835 if (EnableMemAccessVersioning)
3836 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3837 collectStridedAccess(LI);
3839 // Reduction instructions are allowed to have exit users.
3840 // All other instructions must not have external users.
3841 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3842 emitAnalysis(Report(it) << "value cannot be used outside the loop");
3851 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3852 if (Inductions.empty()) {
3853 emitAnalysis(Report()
3854 << "loop induction variable could not be identified");
3862 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3863 /// return the induction operand of the gep pointer.
3864 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3865 const DataLayout *DL, Loop *Lp) {
3866 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3870 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3872 // Check that all of the gep indices are uniform except for our induction
3874 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3875 if (i != InductionOperand &&
3876 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3878 return GEP->getOperand(InductionOperand);
3881 ///\brief Look for a cast use of the passed value.
3882 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3883 Value *UniqueCast = nullptr;
3884 for (User *U : Ptr->users()) {
3885 CastInst *CI = dyn_cast<CastInst>(U);
3886 if (CI && CI->getType() == Ty) {
3896 ///\brief Get the stride of a pointer access in a loop.
3897 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3898 /// pointer to the Value, or null otherwise.
3899 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3900 const DataLayout *DL, Loop *Lp) {
3901 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3902 if (!PtrTy || PtrTy->isAggregateType())
3905 // Try to remove a gep instruction to make the pointer (actually index at this
3906 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3907 // pointer, otherwise, we are analyzing the index.
3908 Value *OrigPtr = Ptr;
3910 // The size of the pointer access.
3911 int64_t PtrAccessSize = 1;
3913 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3914 const SCEV *V = SE->getSCEV(Ptr);
3918 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3919 V = C->getOperand();
3921 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3925 V = S->getStepRecurrence(*SE);
3929 // Strip off the size of access multiplication if we are still analyzing the
3931 if (OrigPtr == Ptr) {
3932 DL->getTypeAllocSize(PtrTy->getElementType());
3933 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3934 if (M->getOperand(0)->getSCEVType() != scConstant)
3937 const APInt &APStepVal =
3938 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3940 // Huge step value - give up.
3941 if (APStepVal.getBitWidth() > 64)
3944 int64_t StepVal = APStepVal.getSExtValue();
3945 if (PtrAccessSize != StepVal)
3947 V = M->getOperand(1);
3952 Type *StripedOffRecurrenceCast = nullptr;
3953 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3954 StripedOffRecurrenceCast = C->getType();
3955 V = C->getOperand();
3958 // Look for the loop invariant symbolic value.
3959 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3963 Value *Stride = U->getValue();
3964 if (!Lp->isLoopInvariant(Stride))
3967 // If we have stripped off the recurrence cast we have to make sure that we
3968 // return the value that is used in this loop so that we can replace it later.
3969 if (StripedOffRecurrenceCast)
3970 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3975 void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) {
3976 Value *Ptr = nullptr;
3977 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3978 Ptr = LI->getPointerOperand();
3979 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3980 Ptr = SI->getPointerOperand();
3984 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3988 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3989 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3990 Strides[Ptr] = Stride;
3991 StrideSet.insert(Stride);
3994 void LoopVectorizationLegality::collectLoopUniforms() {
3995 // We now know that the loop is vectorizable!
3996 // Collect variables that will remain uniform after vectorization.
3997 std::vector<Value*> Worklist;
3998 BasicBlock *Latch = TheLoop->getLoopLatch();
4000 // Start with the conditional branch and walk up the block.
4001 Worklist.push_back(Latch->getTerminator()->getOperand(0));
4003 // Also add all consecutive pointer values; these values will be uniform
4004 // after vectorization (and subsequent cleanup) and, until revectorization is
4005 // supported, all dependencies must also be uniform.
4006 for (Loop::block_iterator B = TheLoop->block_begin(),
4007 BE = TheLoop->block_end(); B != BE; ++B)
4008 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
4010 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
4011 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4013 while (Worklist.size()) {
4014 Instruction *I = dyn_cast<Instruction>(Worklist.back());
4015 Worklist.pop_back();
4017 // Look at instructions inside this loop.
4018 // Stop when reaching PHI nodes.
4019 // TODO: we need to follow values all over the loop, not only in this block.
4020 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
4023 // This is a known uniform.
4026 // Insert all operands.
4027 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4032 /// \brief Analyses memory accesses in a loop.
4034 /// Checks whether run time pointer checks are needed and builds sets for data
4035 /// dependence checking.
4036 class AccessAnalysis {
4038 /// \brief Read or write access location.
4039 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
4040 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
4042 /// \brief Set of potential dependent memory accesses.
4043 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
4045 AccessAnalysis(const DataLayout *Dl, AliasAnalysis *AA, DepCandidates &DA) :
4046 DL(Dl), AST(*AA), DepCands(DA), IsRTCheckNeeded(false) {}
4048 /// \brief Register a load and whether it is only read from.
4049 void addLoad(AliasAnalysis::Location &Loc, bool IsReadOnly) {
4050 Value *Ptr = const_cast<Value*>(Loc.Ptr);
4051 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
4052 Accesses.insert(MemAccessInfo(Ptr, false));
4054 ReadOnlyPtr.insert(Ptr);
4057 /// \brief Register a store.
4058 void addStore(AliasAnalysis::Location &Loc) {
4059 Value *Ptr = const_cast<Value*>(Loc.Ptr);
4060 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
4061 Accesses.insert(MemAccessInfo(Ptr, true));
4064 /// \brief Check whether we can check the pointers at runtime for
4065 /// non-intersection.
4066 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
4067 unsigned &NumComparisons, ScalarEvolution *SE,
4068 Loop *TheLoop, ValueToValueMap &Strides,
4069 bool ShouldCheckStride = false);
4071 /// \brief Goes over all memory accesses, checks whether a RT check is needed
4072 /// and builds sets of dependent accesses.
4073 void buildDependenceSets() {
4074 processMemAccesses();
4077 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
4079 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
4080 void resetDepChecks() { CheckDeps.clear(); }
4082 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
4085 typedef SetVector<MemAccessInfo> PtrAccessSet;
4087 /// \brief Go over all memory access and check whether runtime pointer checks
4088 /// are needed /// and build sets of dependency check candidates.
4089 void processMemAccesses();
4091 /// Set of all accesses.
4092 PtrAccessSet Accesses;
4094 /// Set of accesses that need a further dependence check.
4095 MemAccessInfoSet CheckDeps;
4097 /// Set of pointers that are read only.
4098 SmallPtrSet<Value*, 16> ReadOnlyPtr;
4100 const DataLayout *DL;
4102 /// An alias set tracker to partition the access set by underlying object and
4103 //intrinsic property (such as TBAA metadata).
4104 AliasSetTracker AST;
4106 /// Sets of potentially dependent accesses - members of one set share an
4107 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
4108 /// dependence check.
4109 DepCandidates &DepCands;
4111 bool IsRTCheckNeeded;
4114 } // end anonymous namespace
4116 /// \brief Check whether a pointer can participate in a runtime bounds check.
4117 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
4119 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
4120 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4124 return AR->isAffine();
4127 /// \brief Check the stride of the pointer and ensure that it does not wrap in
4128 /// the address space.
4129 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4130 const Loop *Lp, ValueToValueMap &StridesMap);
4132 bool AccessAnalysis::canCheckPtrAtRT(
4133 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
4134 unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
4135 ValueToValueMap &StridesMap, bool ShouldCheckStride) {
4136 // Find pointers with computable bounds. We are going to use this information
4137 // to place a runtime bound check.
4138 bool CanDoRT = true;
4140 bool IsDepCheckNeeded = isDependencyCheckNeeded();
4143 // We assign a consecutive id to access from different alias sets.
4144 // Accesses between different groups doesn't need to be checked.
4146 for (auto &AS : AST) {
4147 unsigned NumReadPtrChecks = 0;
4148 unsigned NumWritePtrChecks = 0;
4150 // We assign consecutive id to access from different dependence sets.
4151 // Accesses within the same set don't need a runtime check.
4152 unsigned RunningDepId = 1;
4153 DenseMap<Value *, unsigned> DepSetId;
4156 Value *Ptr = A.getValue();
4157 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
4158 MemAccessInfo Access(Ptr, IsWrite);
4161 ++NumWritePtrChecks;
4165 if (hasComputableBounds(SE, StridesMap, Ptr) &&
4166 // When we run after a failing dependency check we have to make sure we
4167 // don't have wrapping pointers.
4168 (!ShouldCheckStride ||
4169 isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
4170 // The id of the dependence set.
4173 if (IsDepCheckNeeded) {
4174 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
4175 unsigned &LeaderId = DepSetId[Leader];
4177 LeaderId = RunningDepId++;
4180 // Each access has its own dependence set.
4181 DepId = RunningDepId++;
4183 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
4185 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
4191 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
4192 NumComparisons += 0; // Only one dependence set.
4194 NumComparisons += (NumWritePtrChecks * (NumReadPtrChecks +
4195 NumWritePtrChecks - 1));
4201 // If the pointers that we would use for the bounds comparison have different
4202 // address spaces, assume the values aren't directly comparable, so we can't
4203 // use them for the runtime check. We also have to assume they could
4204 // overlap. In the future there should be metadata for whether address spaces
4206 unsigned NumPointers = RtCheck.Pointers.size();
4207 for (unsigned i = 0; i < NumPointers; ++i) {
4208 for (unsigned j = i + 1; j < NumPointers; ++j) {
4209 // Only need to check pointers between two different dependency sets.
4210 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
4212 // Only need to check pointers in the same alias set.
4213 if (RtCheck.AliasSetId[i] != RtCheck.AliasSetId[j])
4216 Value *PtrI = RtCheck.Pointers[i];
4217 Value *PtrJ = RtCheck.Pointers[j];
4219 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
4220 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
4222 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
4223 " different address spaces\n");
4232 void AccessAnalysis::processMemAccesses() {
4233 // We process the set twice: first we process read-write pointers, last we
4234 // process read-only pointers. This allows us to skip dependence tests for
4235 // read-only pointers.
4237 DEBUG(dbgs() << "LV: Processing memory accesses...\n");
4238 DEBUG(dbgs() << " AST: "; AST.dump());
4239 DEBUG(dbgs() << "LV: Accesses:\n");
4241 for (auto A : Accesses)
4242 dbgs() << "\t" << *A.getPointer() << " (" <<
4243 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
4244 "read-only" : "read")) << ")\n";
4247 // The AliasSetTracker has nicely partitioned our pointers by metadata
4248 // compatibility and potential for underlying-object overlap. As a result, we
4249 // only need to check for potential pointer dependencies within each alias
4251 for (auto &AS : AST) {
4252 // Note that both the alias-set tracker and the alias sets themselves used
4253 // linked lists internally and so the iteration order here is deterministic
4254 // (matching the original instruction order within each set).
4256 bool SetHasWrite = false;
4258 // Map of pointers to last access encountered.
4259 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
4260 UnderlyingObjToAccessMap ObjToLastAccess;
4262 // Set of access to check after all writes have been processed.
4263 PtrAccessSet DeferredAccesses;
4265 // Iterate over each alias set twice, once to process read/write pointers,
4266 // and then to process read-only pointers.
4267 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
4268 bool UseDeferred = SetIteration > 0;
4269 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
4271 for (auto AV : AS) {
4272 Value *Ptr = AV.getValue();
4274 // For a single memory access in AliasSetTracker, Accesses may contain
4275 // both read and write, and they both need to be handled for CheckDeps.
4277 if (AC.getPointer() != Ptr)
4280 bool IsWrite = AC.getInt();
4282 // If we're using the deferred access set, then it contains only
4284 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
4285 if (UseDeferred && !IsReadOnlyPtr)
4287 // Otherwise, the pointer must be in the PtrAccessSet, either as a
4289 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
4290 S.count(MemAccessInfo(Ptr, false))) &&
4291 "Alias-set pointer not in the access set?");
4293 MemAccessInfo Access(Ptr, IsWrite);
4294 DepCands.insert(Access);
4296 // Memorize read-only pointers for later processing and skip them in
4297 // the first round (they need to be checked after we have seen all
4298 // write pointers). Note: we also mark pointer that are not
4299 // consecutive as "read-only" pointers (so that we check
4300 // "a[b[i]] +="). Hence, we need the second check for "!IsWrite".
4301 if (!UseDeferred && IsReadOnlyPtr) {
4302 DeferredAccesses.insert(Access);
4306 // If this is a write - check other reads and writes for conflicts. If
4307 // this is a read only check other writes for conflicts (but only if
4308 // there is no other write to the ptr - this is an optimization to
4309 // catch "a[i] = a[i] + " without having to do a dependence check).
4310 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
4311 CheckDeps.insert(Access);
4312 IsRTCheckNeeded = true;
4318 // Create sets of pointers connected by a shared alias set and
4319 // underlying object.
4320 typedef SmallVector<Value *, 16> ValueVector;
4321 ValueVector TempObjects;
4322 GetUnderlyingObjects(Ptr, TempObjects, DL);
4323 for (Value *UnderlyingObj : TempObjects) {
4324 UnderlyingObjToAccessMap::iterator Prev =
4325 ObjToLastAccess.find(UnderlyingObj);
4326 if (Prev != ObjToLastAccess.end())
4327 DepCands.unionSets(Access, Prev->second);
4329 ObjToLastAccess[UnderlyingObj] = Access;
4338 /// \brief Checks memory dependences among accesses to the same underlying
4339 /// object to determine whether there vectorization is legal or not (and at
4340 /// which vectorization factor).
4342 /// This class works under the assumption that we already checked that memory
4343 /// locations with different underlying pointers are "must-not alias".
4344 /// We use the ScalarEvolution framework to symbolically evalutate access
4345 /// functions pairs. Since we currently don't restructure the loop we can rely
4346 /// on the program order of memory accesses to determine their safety.
4347 /// At the moment we will only deem accesses as safe for:
4348 /// * A negative constant distance assuming program order.
4350 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
4351 /// a[i] = tmp; y = a[i];
4353 /// The latter case is safe because later checks guarantuee that there can't
4354 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
4355 /// the same variable: a header phi can only be an induction or a reduction, a
4356 /// reduction can't have a memory sink, an induction can't have a memory
4357 /// source). This is important and must not be violated (or we have to
4358 /// resort to checking for cycles through memory).
4360 /// * A positive constant distance assuming program order that is bigger
4361 /// than the biggest memory access.
4363 /// tmp = a[i] OR b[i] = x
4364 /// a[i+2] = tmp y = b[i+2];
4366 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
4368 /// * Zero distances and all accesses have the same size.
4370 class MemoryDepChecker {
4372 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
4373 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
4375 MemoryDepChecker(ScalarEvolution *Se, const DataLayout *Dl, const Loop *L)
4376 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
4377 ShouldRetryWithRuntimeCheck(false) {}
4379 /// \brief Register the location (instructions are given increasing numbers)
4380 /// of a write access.
4381 void addAccess(StoreInst *SI) {
4382 Value *Ptr = SI->getPointerOperand();
4383 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
4384 InstMap.push_back(SI);
4388 /// \brief Register the location (instructions are given increasing numbers)
4389 /// of a write access.
4390 void addAccess(LoadInst *LI) {
4391 Value *Ptr = LI->getPointerOperand();
4392 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
4393 InstMap.push_back(LI);
4397 /// \brief Check whether the dependencies between the accesses are safe.
4399 /// Only checks sets with elements in \p CheckDeps.
4400 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4401 MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
4403 /// \brief The maximum number of bytes of a vector register we can vectorize
4404 /// the accesses safely with.
4405 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
4407 /// \brief In same cases when the dependency check fails we can still
4408 /// vectorize the loop with a dynamic array access check.
4409 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
4412 ScalarEvolution *SE;
4413 const DataLayout *DL;
4414 const Loop *InnermostLoop;
4416 /// \brief Maps access locations (ptr, read/write) to program order.
4417 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
4419 /// \brief Memory access instructions in program order.
4420 SmallVector<Instruction *, 16> InstMap;
4422 /// \brief The program order index to be used for the next instruction.
4425 // We can access this many bytes in parallel safely.
4426 unsigned MaxSafeDepDistBytes;
4428 /// \brief If we see a non-constant dependence distance we can still try to
4429 /// vectorize this loop with runtime checks.
4430 bool ShouldRetryWithRuntimeCheck;
4432 /// \brief Check whether there is a plausible dependence between the two
4435 /// Access \p A must happen before \p B in program order. The two indices
4436 /// identify the index into the program order map.
4438 /// This function checks whether there is a plausible dependence (or the
4439 /// absence of such can't be proved) between the two accesses. If there is a
4440 /// plausible dependence but the dependence distance is bigger than one
4441 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
4442 /// distance is smaller than any other distance encountered so far).
4443 /// Otherwise, this function returns true signaling a possible dependence.
4444 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
4445 const MemAccessInfo &B, unsigned BIdx,
4446 ValueToValueMap &Strides);
4448 /// \brief Check whether the data dependence could prevent store-load
4450 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
4453 } // end anonymous namespace
4455 static bool isInBoundsGep(Value *Ptr) {
4456 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
4457 return GEP->isInBounds();
4461 /// \brief Check whether the access through \p Ptr has a constant stride.
4462 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4463 const Loop *Lp, ValueToValueMap &StridesMap) {
4464 const Type *Ty = Ptr->getType();
4465 assert(Ty->isPointerTy() && "Unexpected non-ptr");
4467 // Make sure that the pointer does not point to aggregate types.
4468 const PointerType *PtrTy = cast<PointerType>(Ty);
4469 if (PtrTy->getElementType()->isAggregateType()) {
4470 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
4475 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
4477 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4479 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
4480 << *Ptr << " SCEV: " << *PtrScev << "\n");
4484 // The accesss function must stride over the innermost loop.
4485 if (Lp != AR->getLoop()) {
4486 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
4487 *Ptr << " SCEV: " << *PtrScev << "\n");
4490 // The address calculation must not wrap. Otherwise, a dependence could be
4492 // An inbounds getelementptr that is a AddRec with a unit stride
4493 // cannot wrap per definition. The unit stride requirement is checked later.
4494 // An getelementptr without an inbounds attribute and unit stride would have
4495 // to access the pointer value "0" which is undefined behavior in address
4496 // space 0, therefore we can also vectorize this case.
4497 bool IsInBoundsGEP = isInBoundsGep(Ptr);
4498 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
4499 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
4500 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
4501 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
4502 << *Ptr << " SCEV: " << *PtrScev << "\n");
4506 // Check the step is constant.
4507 const SCEV *Step = AR->getStepRecurrence(*SE);
4509 // Calculate the pointer stride and check if it is consecutive.
4510 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4512 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
4513 " SCEV: " << *PtrScev << "\n");
4517 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
4518 const APInt &APStepVal = C->getValue()->getValue();
4520 // Huge step value - give up.
4521 if (APStepVal.getBitWidth() > 64)
4524 int64_t StepVal = APStepVal.getSExtValue();
4527 int64_t Stride = StepVal / Size;
4528 int64_t Rem = StepVal % Size;
4532 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
4533 // know we can't "wrap around the address space". In case of address space
4534 // zero we know that this won't happen without triggering undefined behavior.
4535 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
4536 Stride != 1 && Stride != -1)
4542 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
4543 unsigned TypeByteSize) {
4544 // If loads occur at a distance that is not a multiple of a feasible vector
4545 // factor store-load forwarding does not take place.
4546 // Positive dependences might cause troubles because vectorizing them might
4547 // prevent store-load forwarding making vectorized code run a lot slower.
4548 // a[i] = a[i-3] ^ a[i-8];
4549 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
4550 // hence on your typical architecture store-load forwarding does not take
4551 // place. Vectorizing in such cases does not make sense.
4552 // Store-load forwarding distance.
4553 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
4554 // Maximum vector factor.
4555 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
4556 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
4557 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
4559 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
4561 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
4562 MaxVFWithoutSLForwardIssues = (vf >>=1);
4567 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
4568 DEBUG(dbgs() << "LV: Distance " << Distance <<
4569 " that could cause a store-load forwarding conflict\n");
4573 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
4574 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
4575 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
4579 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
4580 const MemAccessInfo &B, unsigned BIdx,
4581 ValueToValueMap &Strides) {
4582 assert (AIdx < BIdx && "Must pass arguments in program order");
4584 Value *APtr = A.getPointer();
4585 Value *BPtr = B.getPointer();
4586 bool AIsWrite = A.getInt();
4587 bool BIsWrite = B.getInt();
4589 // Two reads are independent.
4590 if (!AIsWrite && !BIsWrite)
4593 // We cannot check pointers in different address spaces.
4594 if (APtr->getType()->getPointerAddressSpace() !=
4595 BPtr->getType()->getPointerAddressSpace())
4598 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
4599 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
4601 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
4602 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
4604 const SCEV *Src = AScev;
4605 const SCEV *Sink = BScev;
4607 // If the induction step is negative we have to invert source and sink of the
4609 if (StrideAPtr < 0) {
4612 std::swap(APtr, BPtr);
4613 std::swap(Src, Sink);
4614 std::swap(AIsWrite, BIsWrite);
4615 std::swap(AIdx, BIdx);
4616 std::swap(StrideAPtr, StrideBPtr);
4619 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
4621 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
4622 << "(Induction step: " << StrideAPtr << ")\n");
4623 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
4624 << *InstMap[BIdx] << ": " << *Dist << "\n");
4626 // Need consecutive accesses. We don't want to vectorize
4627 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
4628 // the address space.
4629 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
4630 DEBUG(dbgs() << "Non-consecutive pointer access\n");
4634 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
4636 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
4637 ShouldRetryWithRuntimeCheck = true;
4641 Type *ATy = APtr->getType()->getPointerElementType();
4642 Type *BTy = BPtr->getType()->getPointerElementType();
4643 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
4645 // Negative distances are not plausible dependencies.
4646 const APInt &Val = C->getValue()->getValue();
4647 if (Val.isNegative()) {
4648 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
4649 if (IsTrueDataDependence &&
4650 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
4654 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
4658 // Write to the same location with the same size.
4659 // Could be improved to assert type sizes are the same (i32 == float, etc).
4663 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
4667 assert(Val.isStrictlyPositive() && "Expect a positive value");
4669 // Positive distance bigger than max vectorization factor.
4672 "LV: ReadWrite-Write positive dependency with different types\n");
4676 unsigned Distance = (unsigned) Val.getZExtValue();
4678 // Bail out early if passed-in parameters make vectorization not feasible.
4679 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
4680 unsigned ForcedUnroll = VectorizationInterleave ? VectorizationInterleave : 1;
4682 // The distance must be bigger than the size needed for a vectorized version
4683 // of the operation and the size of the vectorized operation must not be
4684 // bigger than the currrent maximum size.
4685 if (Distance < 2*TypeByteSize ||
4686 2*TypeByteSize > MaxSafeDepDistBytes ||
4687 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
4688 DEBUG(dbgs() << "LV: Failure because of Positive distance "
4689 << Val.getSExtValue() << '\n');
4693 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
4694 Distance : MaxSafeDepDistBytes;
4696 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
4697 if (IsTrueDataDependence &&
4698 couldPreventStoreLoadForward(Distance, TypeByteSize))
4701 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
4702 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
4707 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4708 MemAccessInfoSet &CheckDeps,
4709 ValueToValueMap &Strides) {
4711 MaxSafeDepDistBytes = -1U;
4712 while (!CheckDeps.empty()) {
4713 MemAccessInfo CurAccess = *CheckDeps.begin();
4715 // Get the relevant memory access set.
4716 EquivalenceClasses<MemAccessInfo>::iterator I =
4717 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
4719 // Check accesses within this set.
4720 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
4721 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
4723 // Check every access pair.
4725 CheckDeps.erase(*AI);
4726 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
4728 // Check every accessing instruction pair in program order.
4729 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
4730 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
4731 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
4732 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
4733 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
4735 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
4746 bool LoopVectorizationLegality::canVectorizeMemory() {
4748 typedef SmallVector<Value*, 16> ValueVector;
4749 typedef SmallPtrSet<Value*, 16> ValueSet;
4751 // Holds the Load and Store *instructions*.
4755 // Holds all the different accesses in the loop.
4756 unsigned NumReads = 0;
4757 unsigned NumReadWrites = 0;
4759 PtrRtCheck.Pointers.clear();
4760 PtrRtCheck.Need = false;
4762 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
4763 MemoryDepChecker DepChecker(SE, DL, TheLoop);
4766 for (Loop::block_iterator bb = TheLoop->block_begin(),
4767 be = TheLoop->block_end(); bb != be; ++bb) {
4769 // Scan the BB and collect legal loads and stores.
4770 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4773 // If this is a load, save it. If this instruction can read from memory
4774 // but is not a load, then we quit. Notice that we don't handle function
4775 // calls that read or write.
4776 if (it->mayReadFromMemory()) {
4777 // Many math library functions read the rounding mode. We will only
4778 // vectorize a loop if it contains known function calls that don't set
4779 // the flag. Therefore, it is safe to ignore this read from memory.
4780 CallInst *Call = dyn_cast<CallInst>(it);
4781 if (Call && getIntrinsicIDForCall(Call, TLI))
4784 LoadInst *Ld = dyn_cast<LoadInst>(it);
4785 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
4786 emitAnalysis(Report(Ld)
4787 << "read with atomic ordering or volatile read");
4788 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
4792 Loads.push_back(Ld);
4793 DepChecker.addAccess(Ld);
4797 // Save 'store' instructions. Abort if other instructions write to memory.
4798 if (it->mayWriteToMemory()) {
4799 StoreInst *St = dyn_cast<StoreInst>(it);
4801 emitAnalysis(Report(it) << "instruction cannot be vectorized");
4804 if (!St->isSimple() && !IsAnnotatedParallel) {
4805 emitAnalysis(Report(St)
4806 << "write with atomic ordering or volatile write");
4807 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
4811 Stores.push_back(St);
4812 DepChecker.addAccess(St);
4817 // Now we have two lists that hold the loads and the stores.
4818 // Next, we find the pointers that they use.
4820 // Check if we see any stores. If there are no stores, then we don't
4821 // care if the pointers are *restrict*.
4822 if (!Stores.size()) {
4823 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
4827 AccessAnalysis::DepCandidates DependentAccesses;
4828 AccessAnalysis Accesses(DL, AA, DependentAccesses);
4830 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
4831 // multiple times on the same object. If the ptr is accessed twice, once
4832 // for read and once for write, it will only appear once (on the write
4833 // list). This is okay, since we are going to check for conflicts between
4834 // writes and between reads and writes, but not between reads and reads.
4837 ValueVector::iterator I, IE;
4838 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
4839 StoreInst *ST = cast<StoreInst>(*I);
4840 Value* Ptr = ST->getPointerOperand();
4842 if (isUniform(Ptr)) {
4845 << "write to a loop invariant address could not be vectorized");
4846 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4850 // If we did *not* see this pointer before, insert it to the read-write
4851 // list. At this phase it is only a 'write' list.
4852 if (Seen.insert(Ptr).second) {
4855 AliasAnalysis::Location Loc = AA->getLocation(ST);
4856 // The TBAA metadata could have a control dependency on the predication
4857 // condition, so we cannot rely on it when determining whether or not we
4858 // need runtime pointer checks.
4859 if (blockNeedsPredication(ST->getParent()))
4860 Loc.AATags.TBAA = nullptr;
4862 Accesses.addStore(Loc);
4866 if (IsAnnotatedParallel) {
4868 << "LV: A loop annotated parallel, ignore memory dependency "
4873 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
4874 LoadInst *LD = cast<LoadInst>(*I);
4875 Value* Ptr = LD->getPointerOperand();
4876 // If we did *not* see this pointer before, insert it to the
4877 // read list. If we *did* see it before, then it is already in
4878 // the read-write list. This allows us to vectorize expressions
4879 // such as A[i] += x; Because the address of A[i] is a read-write
4880 // pointer. This only works if the index of A[i] is consecutive.
4881 // If the address of i is unknown (for example A[B[i]]) then we may
4882 // read a few words, modify, and write a few words, and some of the
4883 // words may be written to the same address.
4884 bool IsReadOnlyPtr = false;
4885 if (Seen.insert(Ptr).second ||
4886 !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
4888 IsReadOnlyPtr = true;
4891 AliasAnalysis::Location Loc = AA->getLocation(LD);
4892 // The TBAA metadata could have a control dependency on the predication
4893 // condition, so we cannot rely on it when determining whether or not we
4894 // need runtime pointer checks.
4895 if (blockNeedsPredication(LD->getParent()))
4896 Loc.AATags.TBAA = nullptr;
4898 Accesses.addLoad(Loc, IsReadOnlyPtr);
4901 // If we write (or read-write) to a single destination and there are no
4902 // other reads in this loop then is it safe to vectorize.
4903 if (NumReadWrites == 1 && NumReads == 0) {
4904 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
4908 // Build dependence sets and check whether we need a runtime pointer bounds
4910 Accesses.buildDependenceSets();
4911 bool NeedRTCheck = Accesses.isRTCheckNeeded();
4913 // Find pointers with computable bounds. We are going to use this information
4914 // to place a runtime bound check.
4915 unsigned NumComparisons = 0;
4916 bool CanDoRT = false;
4918 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
4921 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
4922 " pointer comparisons.\n");
4924 // If we only have one set of dependences to check pointers among we don't
4925 // need a runtime check.
4926 if (NumComparisons == 0 && NeedRTCheck)
4927 NeedRTCheck = false;
4929 // Check that we did not collect too many pointers or found an unsizeable
4931 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4937 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
4940 if (NeedRTCheck && !CanDoRT) {
4941 emitAnalysis(Report() << "cannot identify array bounds");
4942 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
4943 "the array bounds.\n");
4948 PtrRtCheck.Need = NeedRTCheck;
4950 bool CanVecMem = true;
4951 if (Accesses.isDependencyCheckNeeded()) {
4952 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
4953 CanVecMem = DepChecker.areDepsSafe(
4954 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
4955 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
4957 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
4958 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
4961 // Clear the dependency checks. We assume they are not needed.
4962 Accesses.resetDepChecks();
4965 PtrRtCheck.Need = true;
4967 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
4968 TheLoop, Strides, true);
4969 // Check that we did not collect too many pointers or found an unsizeable
4971 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4972 if (!CanDoRT && NumComparisons > 0)
4973 emitAnalysis(Report()
4974 << "cannot check memory dependencies at runtime");
4976 emitAnalysis(Report()
4977 << NumComparisons << " exceeds limit of "
4978 << RuntimeMemoryCheckThreshold
4979 << " dependent memory operations checked at runtime");
4980 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
4990 emitAnalysis(Report() << "unsafe dependent memory operations in loop");
4992 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
4993 " need a runtime memory check.\n");
4998 static bool hasMultipleUsesOf(Instruction *I,
4999 SmallPtrSetImpl<Instruction *> &Insts) {
5000 unsigned NumUses = 0;
5001 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
5002 if (Insts.count(dyn_cast<Instruction>(*Use)))
5011 static bool areAllUsesIn(Instruction *I, SmallPtrSetImpl<Instruction *> &Set) {
5012 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
5013 if (!Set.count(dyn_cast<Instruction>(*Use)))
5018 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
5019 ReductionKind Kind) {
5020 if (Phi->getNumIncomingValues() != 2)
5023 // Reduction variables are only found in the loop header block.
5024 if (Phi->getParent() != TheLoop->getHeader())
5027 // Obtain the reduction start value from the value that comes from the loop
5029 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
5031 // ExitInstruction is the single value which is used outside the loop.
5032 // We only allow for a single reduction value to be used outside the loop.
5033 // This includes users of the reduction, variables (which form a cycle
5034 // which ends in the phi node).
5035 Instruction *ExitInstruction = nullptr;
5036 // Indicates that we found a reduction operation in our scan.
5037 bool FoundReduxOp = false;
5039 // We start with the PHI node and scan for all of the users of this
5040 // instruction. All users must be instructions that can be used as reduction
5041 // variables (such as ADD). We must have a single out-of-block user. The cycle
5042 // must include the original PHI.
5043 bool FoundStartPHI = false;
5045 // To recognize min/max patterns formed by a icmp select sequence, we store
5046 // the number of instruction we saw from the recognized min/max pattern,
5047 // to make sure we only see exactly the two instructions.
5048 unsigned NumCmpSelectPatternInst = 0;
5049 ReductionInstDesc ReduxDesc(false, nullptr);
5051 SmallPtrSet<Instruction *, 8> VisitedInsts;
5052 SmallVector<Instruction *, 8> Worklist;
5053 Worklist.push_back(Phi);
5054 VisitedInsts.insert(Phi);
5056 // A value in the reduction can be used:
5057 // - By the reduction:
5058 // - Reduction operation:
5059 // - One use of reduction value (safe).
5060 // - Multiple use of reduction value (not safe).
5062 // - All uses of the PHI must be the reduction (safe).
5063 // - Otherwise, not safe.
5064 // - By one instruction outside of the loop (safe).
5065 // - By further instructions outside of the loop (not safe).
5066 // - By an instruction that is not part of the reduction (not safe).
5068 // * An instruction type other than PHI or the reduction operation.
5069 // * A PHI in the header other than the initial PHI.
5070 while (!Worklist.empty()) {
5071 Instruction *Cur = Worklist.back();
5072 Worklist.pop_back();
5075 // If the instruction has no users then this is a broken chain and can't be
5076 // a reduction variable.
5077 if (Cur->use_empty())
5080 bool IsAPhi = isa<PHINode>(Cur);
5082 // A header PHI use other than the original PHI.
5083 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
5086 // Reductions of instructions such as Div, and Sub is only possible if the
5087 // LHS is the reduction variable.
5088 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
5089 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
5090 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
5093 // Any reduction instruction must be of one of the allowed kinds.
5094 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
5095 if (!ReduxDesc.IsReduction)
5098 // A reduction operation must only have one use of the reduction value.
5099 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
5100 hasMultipleUsesOf(Cur, VisitedInsts))
5103 // All inputs to a PHI node must be a reduction value.
5104 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
5107 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
5108 isa<SelectInst>(Cur)))
5109 ++NumCmpSelectPatternInst;
5110 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
5111 isa<SelectInst>(Cur)))
5112 ++NumCmpSelectPatternInst;
5114 // Check whether we found a reduction operator.
5115 FoundReduxOp |= !IsAPhi;
5117 // Process users of current instruction. Push non-PHI nodes after PHI nodes
5118 // onto the stack. This way we are going to have seen all inputs to PHI
5119 // nodes once we get to them.
5120 SmallVector<Instruction *, 8> NonPHIs;
5121 SmallVector<Instruction *, 8> PHIs;
5122 for (User *U : Cur->users()) {
5123 Instruction *UI = cast<Instruction>(U);
5125 // Check if we found the exit user.
5126 BasicBlock *Parent = UI->getParent();
5127 if (!TheLoop->contains(Parent)) {
5128 // Exit if you find multiple outside users or if the header phi node is
5129 // being used. In this case the user uses the value of the previous
5130 // iteration, in which case we would loose "VF-1" iterations of the
5131 // reduction operation if we vectorize.
5132 if (ExitInstruction != nullptr || Cur == Phi)
5135 // The instruction used by an outside user must be the last instruction
5136 // before we feed back to the reduction phi. Otherwise, we loose VF-1
5137 // operations on the value.
5138 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
5141 ExitInstruction = Cur;
5145 // Process instructions only once (termination). Each reduction cycle
5146 // value must only be used once, except by phi nodes and min/max
5147 // reductions which are represented as a cmp followed by a select.
5148 ReductionInstDesc IgnoredVal(false, nullptr);
5149 if (VisitedInsts.insert(UI).second) {
5150 if (isa<PHINode>(UI))
5153 NonPHIs.push_back(UI);
5154 } else if (!isa<PHINode>(UI) &&
5155 ((!isa<FCmpInst>(UI) &&
5156 !isa<ICmpInst>(UI) &&
5157 !isa<SelectInst>(UI)) ||
5158 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
5161 // Remember that we completed the cycle.
5163 FoundStartPHI = true;
5165 Worklist.append(PHIs.begin(), PHIs.end());
5166 Worklist.append(NonPHIs.begin(), NonPHIs.end());
5169 // This means we have seen one but not the other instruction of the
5170 // pattern or more than just a select and cmp.
5171 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
5172 NumCmpSelectPatternInst != 2)
5175 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
5178 // We found a reduction var if we have reached the original phi node and we
5179 // only have a single instruction with out-of-loop users.
5181 // This instruction is allowed to have out-of-loop users.
5182 AllowedExit.insert(ExitInstruction);
5184 // Save the description of this reduction variable.
5185 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
5186 ReduxDesc.MinMaxKind);
5187 Reductions[Phi] = RD;
5188 // We've ended the cycle. This is a reduction variable if we have an
5189 // outside user and it has a binary op.
5194 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
5195 /// pattern corresponding to a min(X, Y) or max(X, Y).
5196 LoopVectorizationLegality::ReductionInstDesc
5197 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
5198 ReductionInstDesc &Prev) {
5200 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
5201 "Expect a select instruction");
5202 Instruction *Cmp = nullptr;
5203 SelectInst *Select = nullptr;
5205 // We must handle the select(cmp()) as a single instruction. Advance to the
5207 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
5208 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
5209 return ReductionInstDesc(false, I);
5210 return ReductionInstDesc(Select, Prev.MinMaxKind);
5213 // Only handle single use cases for now.
5214 if (!(Select = dyn_cast<SelectInst>(I)))
5215 return ReductionInstDesc(false, I);
5216 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
5217 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
5218 return ReductionInstDesc(false, I);
5219 if (!Cmp->hasOneUse())
5220 return ReductionInstDesc(false, I);
5225 // Look for a min/max pattern.
5226 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5227 return ReductionInstDesc(Select, MRK_UIntMin);
5228 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5229 return ReductionInstDesc(Select, MRK_UIntMax);
5230 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5231 return ReductionInstDesc(Select, MRK_SIntMax);
5232 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5233 return ReductionInstDesc(Select, MRK_SIntMin);
5234 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5235 return ReductionInstDesc(Select, MRK_FloatMin);
5236 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5237 return ReductionInstDesc(Select, MRK_FloatMax);
5238 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5239 return ReductionInstDesc(Select, MRK_FloatMin);
5240 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5241 return ReductionInstDesc(Select, MRK_FloatMax);
5243 return ReductionInstDesc(false, I);
5246 LoopVectorizationLegality::ReductionInstDesc
5247 LoopVectorizationLegality::isReductionInstr(Instruction *I,
5249 ReductionInstDesc &Prev) {
5250 bool FP = I->getType()->isFloatingPointTy();
5251 bool FastMath = FP && I->hasUnsafeAlgebra();
5252 switch (I->getOpcode()) {
5254 return ReductionInstDesc(false, I);
5255 case Instruction::PHI:
5256 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
5257 Kind != RK_FloatMinMax))
5258 return ReductionInstDesc(false, I);
5259 return ReductionInstDesc(I, Prev.MinMaxKind);
5260 case Instruction::Sub:
5261 case Instruction::Add:
5262 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
5263 case Instruction::Mul:
5264 return ReductionInstDesc(Kind == RK_IntegerMult, I);
5265 case Instruction::And:
5266 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
5267 case Instruction::Or:
5268 return ReductionInstDesc(Kind == RK_IntegerOr, I);
5269 case Instruction::Xor:
5270 return ReductionInstDesc(Kind == RK_IntegerXor, I);
5271 case Instruction::FMul:
5272 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
5273 case Instruction::FSub:
5274 case Instruction::FAdd:
5275 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
5276 case Instruction::FCmp:
5277 case Instruction::ICmp:
5278 case Instruction::Select:
5279 if (Kind != RK_IntegerMinMax &&
5280 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
5281 return ReductionInstDesc(false, I);
5282 return isMinMaxSelectCmpPattern(I, Prev);
5286 LoopVectorizationLegality::InductionKind
5287 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
5288 Type *PhiTy = Phi->getType();
5289 // We only handle integer and pointer inductions variables.
5290 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
5291 return IK_NoInduction;
5293 // Check that the PHI is consecutive.
5294 const SCEV *PhiScev = SE->getSCEV(Phi);
5295 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
5297 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
5298 return IK_NoInduction;
5300 const SCEV *Step = AR->getStepRecurrence(*SE);
5302 // Integer inductions need to have a stride of one.
5303 if (PhiTy->isIntegerTy()) {
5305 return IK_IntInduction;
5306 if (Step->isAllOnesValue())
5307 return IK_ReverseIntInduction;
5308 return IK_NoInduction;
5311 // Calculate the pointer stride and check if it is consecutive.
5312 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5314 return IK_NoInduction;
5316 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
5317 Type *PointerElementType = PhiTy->getPointerElementType();
5318 // The pointer stride cannot be determined if the pointer element type is not
5320 if (!PointerElementType->isSized())
5321 return IK_NoInduction;
5323 uint64_t Size = DL->getTypeAllocSize(PointerElementType);
5324 if (C->getValue()->equalsInt(Size))
5325 return IK_PtrInduction;
5326 else if (C->getValue()->equalsInt(0 - Size))
5327 return IK_ReversePtrInduction;
5329 return IK_NoInduction;
5332 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
5333 Value *In0 = const_cast<Value*>(V);
5334 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
5338 return Inductions.count(PN);
5341 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
5342 assert(TheLoop->contains(BB) && "Unknown block used");
5344 // Blocks that do not dominate the latch need predication.
5345 BasicBlock* Latch = TheLoop->getLoopLatch();
5346 return !DT->dominates(BB, Latch);
5349 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
5350 SmallPtrSetImpl<Value *> &SafePtrs) {
5352 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5353 // Check that we don't have a constant expression that can trap as operand.
5354 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
5356 if (Constant *C = dyn_cast<Constant>(*OI))
5360 // We might be able to hoist the load.
5361 if (it->mayReadFromMemory()) {
5362 LoadInst *LI = dyn_cast<LoadInst>(it);
5365 if (!SafePtrs.count(LI->getPointerOperand())) {
5366 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
5367 MaskedOp.insert(LI);
5374 // We don't predicate stores at the moment.
5375 if (it->mayWriteToMemory()) {
5376 StoreInst *SI = dyn_cast<StoreInst>(it);
5377 // We only support predication of stores in basic blocks with one
5382 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
5383 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
5385 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
5386 !isSinglePredecessor) {
5387 // Build a masked store if it is legal for the target, otherwise scalarize
5389 bool isLegalMaskedOp =
5390 isLegalMaskedStore(SI->getValueOperand()->getType(),
5391 SI->getPointerOperand());
5392 if (isLegalMaskedOp) {
5394 MaskedOp.insert(SI);
5403 // The instructions below can trap.
5404 switch (it->getOpcode()) {
5406 case Instruction::UDiv:
5407 case Instruction::SDiv:
5408 case Instruction::URem:
5409 case Instruction::SRem:
5417 LoopVectorizationCostModel::VectorizationFactor
5418 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
5419 // Width 1 means no vectorize
5420 VectorizationFactor Factor = { 1U, 0U };
5421 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
5422 emitAnalysis(Report() << "runtime pointer checks needed. Enable vectorization of this loop with '#pragma clang loop vectorize(enable)' when compiling with -Os");
5423 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
5427 if (!EnableCondStoresVectorization && Legal->NumPredStores) {
5428 emitAnalysis(Report() << "store that is conditionally executed prevents vectorization");
5429 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
5433 // Find the trip count.
5434 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
5435 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
5437 unsigned WidestType = getWidestType();
5438 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
5439 unsigned MaxSafeDepDist = -1U;
5440 if (Legal->getMaxSafeDepDistBytes() != -1U)
5441 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
5442 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
5443 WidestRegister : MaxSafeDepDist);
5444 unsigned MaxVectorSize = WidestRegister / WidestType;
5445 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
5446 DEBUG(dbgs() << "LV: The Widest register is: "
5447 << WidestRegister << " bits.\n");
5449 if (MaxVectorSize == 0) {
5450 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
5454 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
5455 " into one vector!");
5457 unsigned VF = MaxVectorSize;
5459 // If we optimize the program for size, avoid creating the tail loop.
5461 // If we are unable to calculate the trip count then don't try to vectorize.
5463 emitAnalysis(Report() << "unable to calculate the loop count due to complex control flow");
5464 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5468 // Find the maximum SIMD width that can fit within the trip count.
5469 VF = TC % MaxVectorSize;
5474 // If the trip count that we found modulo the vectorization factor is not
5475 // zero then we require a tail.
5477 emitAnalysis(Report() << "cannot optimize for size and vectorize at the "
5478 "same time. Enable vectorization of this loop "
5479 "with '#pragma clang loop vectorize(enable)' "
5480 "when compiling with -Os");
5481 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5486 int UserVF = Hints->getWidth();
5488 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
5489 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
5491 Factor.Width = UserVF;
5495 float Cost = expectedCost(1);
5497 const float ScalarCost = Cost;
5500 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
5502 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
5503 // Ignore scalar width, because the user explicitly wants vectorization.
5504 if (ForceVectorization && VF > 1) {
5506 Cost = expectedCost(Width) / (float)Width;
5509 for (unsigned i=2; i <= VF; i*=2) {
5510 // Notice that the vector loop needs to be executed less times, so
5511 // we need to divide the cost of the vector loops by the width of
5512 // the vector elements.
5513 float VectorCost = expectedCost(i) / (float)i;
5514 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
5515 (int)VectorCost << ".\n");
5516 if (VectorCost < Cost) {
5522 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
5523 << "LV: Vectorization seems to be not beneficial, "
5524 << "but was forced by a user.\n");
5525 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
5526 Factor.Width = Width;
5527 Factor.Cost = Width * Cost;
5531 unsigned LoopVectorizationCostModel::getWidestType() {
5532 unsigned MaxWidth = 8;
5535 for (Loop::block_iterator bb = TheLoop->block_begin(),
5536 be = TheLoop->block_end(); bb != be; ++bb) {
5537 BasicBlock *BB = *bb;
5539 // For each instruction in the loop.
5540 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5541 Type *T = it->getType();
5543 // Ignore ephemeral values.
5544 if (EphValues.count(it))
5547 // Only examine Loads, Stores and PHINodes.
5548 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
5551 // Examine PHI nodes that are reduction variables.
5552 if (PHINode *PN = dyn_cast<PHINode>(it))
5553 if (!Legal->getReductionVars()->count(PN))
5556 // Examine the stored values.
5557 if (StoreInst *ST = dyn_cast<StoreInst>(it))
5558 T = ST->getValueOperand()->getType();
5560 // Ignore loaded pointer types and stored pointer types that are not
5561 // consecutive. However, we do want to take consecutive stores/loads of
5562 // pointer vectors into account.
5563 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
5566 MaxWidth = std::max(MaxWidth,
5567 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
5575 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
5577 unsigned LoopCost) {
5579 // -- The unroll heuristics --
5580 // We unroll the loop in order to expose ILP and reduce the loop overhead.
5581 // There are many micro-architectural considerations that we can't predict
5582 // at this level. For example, frontend pressure (on decode or fetch) due to
5583 // code size, or the number and capabilities of the execution ports.
5585 // We use the following heuristics to select the unroll factor:
5586 // 1. If the code has reductions, then we unroll in order to break the cross
5587 // iteration dependency.
5588 // 2. If the loop is really small, then we unroll in order to reduce the loop
5590 // 3. We don't unroll if we think that we will spill registers to memory due
5591 // to the increased register pressure.
5593 // Use the user preference, unless 'auto' is selected.
5594 int UserUF = Hints->getInterleave();
5598 // When we optimize for size, we don't unroll.
5602 // We used the distance for the unroll factor.
5603 if (Legal->getMaxSafeDepDistBytes() != -1U)
5606 // Do not unroll loops with a relatively small trip count.
5607 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
5608 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
5611 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
5612 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
5616 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
5617 TargetNumRegisters = ForceTargetNumScalarRegs;
5619 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
5620 TargetNumRegisters = ForceTargetNumVectorRegs;
5623 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
5624 // We divide by these constants so assume that we have at least one
5625 // instruction that uses at least one register.
5626 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
5627 R.NumInstructions = std::max(R.NumInstructions, 1U);
5629 // We calculate the unroll factor using the following formula.
5630 // Subtract the number of loop invariants from the number of available
5631 // registers. These registers are used by all of the unrolled instances.
5632 // Next, divide the remaining registers by the number of registers that is
5633 // required by the loop, in order to estimate how many parallel instances
5634 // fit without causing spills. All of this is rounded down if necessary to be
5635 // a power of two. We want power of two unroll factors to simplify any
5636 // addressing operations or alignment considerations.
5637 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
5640 // Don't count the induction variable as unrolled.
5641 if (EnableIndVarRegisterHeur)
5642 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
5643 std::max(1U, (R.MaxLocalUsers - 1)));
5645 // Clamp the unroll factor ranges to reasonable factors.
5646 unsigned MaxInterleaveSize = TTI.getMaxInterleaveFactor();
5648 // Check if the user has overridden the unroll max.
5650 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
5651 MaxInterleaveSize = ForceTargetMaxScalarInterleaveFactor;
5653 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
5654 MaxInterleaveSize = ForceTargetMaxVectorInterleaveFactor;
5657 // If we did not calculate the cost for VF (because the user selected the VF)
5658 // then we calculate the cost of VF here.
5660 LoopCost = expectedCost(VF);
5662 // Clamp the calculated UF to be between the 1 and the max unroll factor
5663 // that the target allows.
5664 if (UF > MaxInterleaveSize)
5665 UF = MaxInterleaveSize;
5669 // Unroll if we vectorized this loop and there is a reduction that could
5670 // benefit from unrolling.
5671 if (VF > 1 && Legal->getReductionVars()->size()) {
5672 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
5676 // Note that if we've already vectorized the loop we will have done the
5677 // runtime check and so unrolling won't require further checks.
5678 bool UnrollingRequiresRuntimePointerCheck =
5679 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
5681 // We want to unroll small loops in order to reduce the loop overhead and
5682 // potentially expose ILP opportunities.
5683 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
5684 if (!UnrollingRequiresRuntimePointerCheck &&
5685 LoopCost < SmallLoopCost) {
5686 // We assume that the cost overhead is 1 and we use the cost model
5687 // to estimate the cost of the loop and unroll until the cost of the
5688 // loop overhead is about 5% of the cost of the loop.
5689 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
5691 // Unroll until store/load ports (estimated by max unroll factor) are
5693 unsigned StoresUF = UF / (Legal->NumStores ? Legal->NumStores : 1);
5694 unsigned LoadsUF = UF / (Legal->NumLoads ? Legal->NumLoads : 1);
5696 // If we have a scalar reduction (vector reductions are already dealt with
5697 // by this point), we can increase the critical path length if the loop
5698 // we're unrolling is inside another loop. Limit, by default to 2, so the
5699 // critical path only gets increased by one reduction operation.
5700 if (Legal->getReductionVars()->size() &&
5701 TheLoop->getLoopDepth() > 1) {
5702 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionUF);
5703 SmallUF = std::min(SmallUF, F);
5704 StoresUF = std::min(StoresUF, F);
5705 LoadsUF = std::min(LoadsUF, F);
5708 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
5709 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
5710 return std::max(StoresUF, LoadsUF);
5713 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
5717 DEBUG(dbgs() << "LV: Not Unrolling.\n");
5721 LoopVectorizationCostModel::RegisterUsage
5722 LoopVectorizationCostModel::calculateRegisterUsage() {
5723 // This function calculates the register usage by measuring the highest number
5724 // of values that are alive at a single location. Obviously, this is a very
5725 // rough estimation. We scan the loop in a topological order in order and
5726 // assign a number to each instruction. We use RPO to ensure that defs are
5727 // met before their users. We assume that each instruction that has in-loop
5728 // users starts an interval. We record every time that an in-loop value is
5729 // used, so we have a list of the first and last occurrences of each
5730 // instruction. Next, we transpose this data structure into a multi map that
5731 // holds the list of intervals that *end* at a specific location. This multi
5732 // map allows us to perform a linear search. We scan the instructions linearly
5733 // and record each time that a new interval starts, by placing it in a set.
5734 // If we find this value in the multi-map then we remove it from the set.
5735 // The max register usage is the maximum size of the set.
5736 // We also search for instructions that are defined outside the loop, but are
5737 // used inside the loop. We need this number separately from the max-interval
5738 // usage number because when we unroll, loop-invariant values do not take
5740 LoopBlocksDFS DFS(TheLoop);
5744 R.NumInstructions = 0;
5746 // Each 'key' in the map opens a new interval. The values
5747 // of the map are the index of the 'last seen' usage of the
5748 // instruction that is the key.
5749 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5750 // Maps instruction to its index.
5751 DenseMap<unsigned, Instruction*> IdxToInstr;
5752 // Marks the end of each interval.
5753 IntervalMap EndPoint;
5754 // Saves the list of instruction indices that are used in the loop.
5755 SmallSet<Instruction*, 8> Ends;
5756 // Saves the list of values that are used in the loop but are
5757 // defined outside the loop, such as arguments and constants.
5758 SmallPtrSet<Value*, 8> LoopInvariants;
5761 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5762 be = DFS.endRPO(); bb != be; ++bb) {
5763 R.NumInstructions += (*bb)->size();
5764 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
5766 Instruction *I = it;
5767 IdxToInstr[Index++] = I;
5769 // Save the end location of each USE.
5770 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
5771 Value *U = I->getOperand(i);
5772 Instruction *Instr = dyn_cast<Instruction>(U);
5774 // Ignore non-instruction values such as arguments, constants, etc.
5775 if (!Instr) continue;
5777 // If this instruction is outside the loop then record it and continue.
5778 if (!TheLoop->contains(Instr)) {
5779 LoopInvariants.insert(Instr);
5783 // Overwrite previous end points.
5784 EndPoint[Instr] = Index;
5790 // Saves the list of intervals that end with the index in 'key'.
5791 typedef SmallVector<Instruction*, 2> InstrList;
5792 DenseMap<unsigned, InstrList> TransposeEnds;
5794 // Transpose the EndPoints to a list of values that end at each index.
5795 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5797 TransposeEnds[it->second].push_back(it->first);
5799 SmallSet<Instruction*, 8> OpenIntervals;
5800 unsigned MaxUsage = 0;
5803 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5804 for (unsigned int i = 0; i < Index; ++i) {
5805 Instruction *I = IdxToInstr[i];
5806 // Ignore instructions that are never used within the loop.
5807 if (!Ends.count(I)) continue;
5809 // Ignore ephemeral values.
5810 if (EphValues.count(I))
5813 // Remove all of the instructions that end at this location.
5814 InstrList &List = TransposeEnds[i];
5815 for (unsigned int j=0, e = List.size(); j < e; ++j)
5816 OpenIntervals.erase(List[j]);
5818 // Count the number of live interals.
5819 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5821 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5822 OpenIntervals.size() << '\n');
5824 // Add the current instruction to the list of open intervals.
5825 OpenIntervals.insert(I);
5828 unsigned Invariant = LoopInvariants.size();
5829 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5830 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5831 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5833 R.LoopInvariantRegs = Invariant;
5834 R.MaxLocalUsers = MaxUsage;
5838 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5842 for (Loop::block_iterator bb = TheLoop->block_begin(),
5843 be = TheLoop->block_end(); bb != be; ++bb) {
5844 unsigned BlockCost = 0;
5845 BasicBlock *BB = *bb;
5847 // For each instruction in the old loop.
5848 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5849 // Skip dbg intrinsics.
5850 if (isa<DbgInfoIntrinsic>(it))
5853 // Ignore ephemeral values.
5854 if (EphValues.count(it))
5857 unsigned C = getInstructionCost(it, VF);
5859 // Check if we should override the cost.
5860 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5861 C = ForceTargetInstructionCost;
5864 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5865 VF << " For instruction: " << *it << '\n');
5868 // We assume that if-converted blocks have a 50% chance of being executed.
5869 // When the code is scalar then some of the blocks are avoided due to CF.
5870 // When the code is vectorized we execute all code paths.
5871 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5880 /// \brief Check whether the address computation for a non-consecutive memory
5881 /// access looks like an unlikely candidate for being merged into the indexing
5884 /// We look for a GEP which has one index that is an induction variable and all
5885 /// other indices are loop invariant. If the stride of this access is also
5886 /// within a small bound we decide that this address computation can likely be
5887 /// merged into the addressing mode.
5888 /// In all other cases, we identify the address computation as complex.
5889 static bool isLikelyComplexAddressComputation(Value *Ptr,
5890 LoopVectorizationLegality *Legal,
5891 ScalarEvolution *SE,
5892 const Loop *TheLoop) {
5893 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5897 // We are looking for a gep with all loop invariant indices except for one
5898 // which should be an induction variable.
5899 unsigned NumOperands = Gep->getNumOperands();
5900 for (unsigned i = 1; i < NumOperands; ++i) {
5901 Value *Opd = Gep->getOperand(i);
5902 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5903 !Legal->isInductionVariable(Opd))
5907 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5908 // can likely be merged into the address computation.
5909 unsigned MaxMergeDistance = 64;
5911 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5915 // Check the step is constant.
5916 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5917 // Calculate the pointer stride and check if it is consecutive.
5918 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5922 const APInt &APStepVal = C->getValue()->getValue();
5924 // Huge step value - give up.
5925 if (APStepVal.getBitWidth() > 64)
5928 int64_t StepVal = APStepVal.getSExtValue();
5930 return StepVal > MaxMergeDistance;
5933 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5934 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5940 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5941 // If we know that this instruction will remain uniform, check the cost of
5942 // the scalar version.
5943 if (Legal->isUniformAfterVectorization(I))
5946 Type *RetTy = I->getType();
5947 Type *VectorTy = ToVectorTy(RetTy, VF);
5949 // TODO: We need to estimate the cost of intrinsic calls.
5950 switch (I->getOpcode()) {
5951 case Instruction::GetElementPtr:
5952 // We mark this instruction as zero-cost because the cost of GEPs in
5953 // vectorized code depends on whether the corresponding memory instruction
5954 // is scalarized or not. Therefore, we handle GEPs with the memory
5955 // instruction cost.
5957 case Instruction::Br: {
5958 return TTI.getCFInstrCost(I->getOpcode());
5960 case Instruction::PHI:
5961 //TODO: IF-converted IFs become selects.
5963 case Instruction::Add:
5964 case Instruction::FAdd:
5965 case Instruction::Sub:
5966 case Instruction::FSub:
5967 case Instruction::Mul:
5968 case Instruction::FMul:
5969 case Instruction::UDiv:
5970 case Instruction::SDiv:
5971 case Instruction::FDiv:
5972 case Instruction::URem:
5973 case Instruction::SRem:
5974 case Instruction::FRem:
5975 case Instruction::Shl:
5976 case Instruction::LShr:
5977 case Instruction::AShr:
5978 case Instruction::And:
5979 case Instruction::Or:
5980 case Instruction::Xor: {
5981 // Since we will replace the stride by 1 the multiplication should go away.
5982 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5984 // Certain instructions can be cheaper to vectorize if they have a constant
5985 // second vector operand. One example of this are shifts on x86.
5986 TargetTransformInfo::OperandValueKind Op1VK =
5987 TargetTransformInfo::OK_AnyValue;
5988 TargetTransformInfo::OperandValueKind Op2VK =
5989 TargetTransformInfo::OK_AnyValue;
5990 TargetTransformInfo::OperandValueProperties Op1VP =
5991 TargetTransformInfo::OP_None;
5992 TargetTransformInfo::OperandValueProperties Op2VP =
5993 TargetTransformInfo::OP_None;
5994 Value *Op2 = I->getOperand(1);
5996 // Check for a splat of a constant or for a non uniform vector of constants.
5997 if (isa<ConstantInt>(Op2)) {
5998 ConstantInt *CInt = cast<ConstantInt>(Op2);
5999 if (CInt && CInt->getValue().isPowerOf2())
6000 Op2VP = TargetTransformInfo::OP_PowerOf2;
6001 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
6002 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
6003 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
6004 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
6006 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
6007 if (CInt && CInt->getValue().isPowerOf2())
6008 Op2VP = TargetTransformInfo::OP_PowerOf2;
6009 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
6013 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
6016 case Instruction::Select: {
6017 SelectInst *SI = cast<SelectInst>(I);
6018 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
6019 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
6020 Type *CondTy = SI->getCondition()->getType();
6022 CondTy = VectorType::get(CondTy, VF);
6024 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
6026 case Instruction::ICmp:
6027 case Instruction::FCmp: {
6028 Type *ValTy = I->getOperand(0)->getType();
6029 VectorTy = ToVectorTy(ValTy, VF);
6030 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
6032 case Instruction::Store:
6033 case Instruction::Load: {
6034 StoreInst *SI = dyn_cast<StoreInst>(I);
6035 LoadInst *LI = dyn_cast<LoadInst>(I);
6036 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
6038 VectorTy = ToVectorTy(ValTy, VF);
6040 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
6041 unsigned AS = SI ? SI->getPointerAddressSpace() :
6042 LI->getPointerAddressSpace();
6043 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
6044 // We add the cost of address computation here instead of with the gep
6045 // instruction because only here we know whether the operation is
6048 return TTI.getAddressComputationCost(VectorTy) +
6049 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
6051 // Scalarized loads/stores.
6052 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
6053 bool Reverse = ConsecutiveStride < 0;
6054 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
6055 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
6056 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
6057 bool IsComplexComputation =
6058 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
6060 // The cost of extracting from the value vector and pointer vector.
6061 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
6062 for (unsigned i = 0; i < VF; ++i) {
6063 // The cost of extracting the pointer operand.
6064 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
6065 // In case of STORE, the cost of ExtractElement from the vector.
6066 // In case of LOAD, the cost of InsertElement into the returned
6068 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
6069 Instruction::InsertElement,
6073 // The cost of the scalar loads/stores.
6074 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
6075 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
6080 // Wide load/stores.
6081 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
6082 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
6085 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
6089 case Instruction::ZExt:
6090 case Instruction::SExt:
6091 case Instruction::FPToUI:
6092 case Instruction::FPToSI:
6093 case Instruction::FPExt:
6094 case Instruction::PtrToInt:
6095 case Instruction::IntToPtr:
6096 case Instruction::SIToFP:
6097 case Instruction::UIToFP:
6098 case Instruction::Trunc:
6099 case Instruction::FPTrunc:
6100 case Instruction::BitCast: {
6101 // We optimize the truncation of induction variable.
6102 // The cost of these is the same as the scalar operation.
6103 if (I->getOpcode() == Instruction::Trunc &&
6104 Legal->isInductionVariable(I->getOperand(0)))
6105 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
6106 I->getOperand(0)->getType());
6108 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
6109 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
6111 case Instruction::Call: {
6112 CallInst *CI = cast<CallInst>(I);
6113 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
6114 assert(ID && "Not an intrinsic call!");
6115 Type *RetTy = ToVectorTy(CI->getType(), VF);
6116 SmallVector<Type*, 4> Tys;
6117 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
6118 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
6119 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
6122 // We are scalarizing the instruction. Return the cost of the scalar
6123 // instruction, plus the cost of insert and extract into vector
6124 // elements, times the vector width.
6127 if (!RetTy->isVoidTy() && VF != 1) {
6128 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
6130 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
6133 // The cost of inserting the results plus extracting each one of the
6135 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
6138 // The cost of executing VF copies of the scalar instruction. This opcode
6139 // is unknown. Assume that it is the same as 'mul'.
6140 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
6146 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
6147 if (Scalar->isVoidTy() || VF == 1)
6149 return VectorType::get(Scalar, VF);
6152 char LoopVectorize::ID = 0;
6153 static const char lv_name[] = "Loop Vectorization";
6154 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
6155 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
6156 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
6157 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
6158 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
6159 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
6160 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
6161 INITIALIZE_PASS_DEPENDENCY(LCSSA)
6162 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
6163 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
6164 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
6167 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
6168 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
6172 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
6173 // Check for a store.
6174 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
6175 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
6177 // Check for a load.
6178 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
6179 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
6185 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
6186 bool IfPredicateStore) {
6187 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
6188 // Holds vector parameters or scalars, in case of uniform vals.
6189 SmallVector<VectorParts, 4> Params;
6191 setDebugLocFromInst(Builder, Instr);
6193 // Find all of the vectorized parameters.
6194 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
6195 Value *SrcOp = Instr->getOperand(op);
6197 // If we are accessing the old induction variable, use the new one.
6198 if (SrcOp == OldInduction) {
6199 Params.push_back(getVectorValue(SrcOp));
6203 // Try using previously calculated values.
6204 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
6206 // If the src is an instruction that appeared earlier in the basic block
6207 // then it should already be vectorized.
6208 if (SrcInst && OrigLoop->contains(SrcInst)) {
6209 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
6210 // The parameter is a vector value from earlier.
6211 Params.push_back(WidenMap.get(SrcInst));
6213 // The parameter is a scalar from outside the loop. Maybe even a constant.
6214 VectorParts Scalars;
6215 Scalars.append(UF, SrcOp);
6216 Params.push_back(Scalars);
6220 assert(Params.size() == Instr->getNumOperands() &&
6221 "Invalid number of operands");
6223 // Does this instruction return a value ?
6224 bool IsVoidRetTy = Instr->getType()->isVoidTy();
6226 Value *UndefVec = IsVoidRetTy ? nullptr :
6227 UndefValue::get(Instr->getType());
6228 // Create a new entry in the WidenMap and initialize it to Undef or Null.
6229 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
6231 Instruction *InsertPt = Builder.GetInsertPoint();
6232 BasicBlock *IfBlock = Builder.GetInsertBlock();
6233 BasicBlock *CondBlock = nullptr;
6236 Loop *VectorLp = nullptr;
6237 if (IfPredicateStore) {
6238 assert(Instr->getParent()->getSinglePredecessor() &&
6239 "Only support single predecessor blocks");
6240 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
6241 Instr->getParent());
6242 VectorLp = LI->getLoopFor(IfBlock);
6243 assert(VectorLp && "Must have a loop for this block");
6246 // For each vector unroll 'part':
6247 for (unsigned Part = 0; Part < UF; ++Part) {
6248 // For each scalar that we create:
6250 // Start an "if (pred) a[i] = ..." block.
6251 Value *Cmp = nullptr;
6252 if (IfPredicateStore) {
6253 if (Cond[Part]->getType()->isVectorTy())
6255 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
6256 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
6257 ConstantInt::get(Cond[Part]->getType(), 1));
6258 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
6259 LoopVectorBody.push_back(CondBlock);
6260 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
6261 // Update Builder with newly created basic block.
6262 Builder.SetInsertPoint(InsertPt);
6265 Instruction *Cloned = Instr->clone();
6267 Cloned->setName(Instr->getName() + ".cloned");
6268 // Replace the operands of the cloned instructions with extracted scalars.
6269 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
6270 Value *Op = Params[op][Part];
6271 Cloned->setOperand(op, Op);
6274 // Place the cloned scalar in the new loop.
6275 Builder.Insert(Cloned);
6277 // If the original scalar returns a value we need to place it in a vector
6278 // so that future users will be able to use it.
6280 VecResults[Part] = Cloned;
6283 if (IfPredicateStore) {
6284 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
6285 LoopVectorBody.push_back(NewIfBlock);
6286 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
6287 Builder.SetInsertPoint(InsertPt);
6288 Instruction *OldBr = IfBlock->getTerminator();
6289 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
6290 OldBr->eraseFromParent();
6291 IfBlock = NewIfBlock;
6296 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
6297 StoreInst *SI = dyn_cast<StoreInst>(Instr);
6298 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
6300 return scalarizeInstruction(Instr, IfPredicateStore);
6303 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
6307 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
6311 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
6313 // When unrolling and the VF is 1, we only need to add a simple scalar.
6314 Type *ITy = Val->getType();
6315 assert(!ITy->isVectorTy() && "Val must be a scalar");
6316 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
6317 return Builder.CreateAdd(Val, C, "induction");