1 //===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops
11 // and generates target-independent LLVM-IR.
12 // The vectorizer uses the TargetTransformInfo analysis to estimate the costs
13 // of instructions in order to estimate the profitability of vectorization.
15 // The loop vectorizer combines consecutive loop iterations into a single
16 // 'wide' iteration. After this transformation the index is incremented
17 // by the SIMD vector width, and not by one.
19 // This pass has three parts:
20 // 1. The main loop pass that drives the different parts.
21 // 2. LoopVectorizationLegality - A unit that checks for the legality
22 // of the vectorization.
23 // 3. InnerLoopVectorizer - A unit that performs the actual
24 // widening of instructions.
25 // 4. LoopVectorizationCostModel - A unit that checks for the profitability
26 // of vectorization. It decides on the optimal vector width, which
27 // can be one, if vectorization is not profitable.
29 //===----------------------------------------------------------------------===//
31 // The reduction-variable vectorization is based on the paper:
32 // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
34 // Variable uniformity checks are inspired by:
35 // Karrenberg, R. and Hack, S. Whole Function Vectorization.
37 // Other ideas/concepts are from:
38 // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
40 // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of
41 // Vectorizing Compilers.
43 //===----------------------------------------------------------------------===//
45 #include "llvm/Transforms/Vectorize.h"
46 #include "llvm/ADT/DenseMap.h"
47 #include "llvm/ADT/EquivalenceClasses.h"
48 #include "llvm/ADT/Hashing.h"
49 #include "llvm/ADT/MapVector.h"
50 #include "llvm/ADT/SetVector.h"
51 #include "llvm/ADT/SmallPtrSet.h"
52 #include "llvm/ADT/SmallSet.h"
53 #include "llvm/ADT/SmallVector.h"
54 #include "llvm/ADT/Statistic.h"
55 #include "llvm/ADT/StringExtras.h"
56 #include "llvm/Analysis/AliasAnalysis.h"
57 #include "llvm/Analysis/AliasSetTracker.h"
58 #include "llvm/Analysis/AssumptionTracker.h"
59 #include "llvm/Analysis/BlockFrequencyInfo.h"
60 #include "llvm/Analysis/CodeMetrics.h"
61 #include "llvm/Analysis/LoopInfo.h"
62 #include "llvm/Analysis/LoopIterator.h"
63 #include "llvm/Analysis/LoopPass.h"
64 #include "llvm/Analysis/ScalarEvolution.h"
65 #include "llvm/Analysis/ScalarEvolutionExpander.h"
66 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
67 #include "llvm/Analysis/TargetTransformInfo.h"
68 #include "llvm/Analysis/ValueTracking.h"
69 #include "llvm/IR/Constants.h"
70 #include "llvm/IR/DataLayout.h"
71 #include "llvm/IR/DebugInfo.h"
72 #include "llvm/IR/DerivedTypes.h"
73 #include "llvm/IR/DiagnosticInfo.h"
74 #include "llvm/IR/Dominators.h"
75 #include "llvm/IR/Function.h"
76 #include "llvm/IR/IRBuilder.h"
77 #include "llvm/IR/Instructions.h"
78 #include "llvm/IR/IntrinsicInst.h"
79 #include "llvm/IR/LLVMContext.h"
80 #include "llvm/IR/Module.h"
81 #include "llvm/IR/PatternMatch.h"
82 #include "llvm/IR/Type.h"
83 #include "llvm/IR/Value.h"
84 #include "llvm/IR/ValueHandle.h"
85 #include "llvm/IR/Verifier.h"
86 #include "llvm/Pass.h"
87 #include "llvm/Support/BranchProbability.h"
88 #include "llvm/Support/CommandLine.h"
89 #include "llvm/Support/Debug.h"
90 #include "llvm/Support/raw_ostream.h"
91 #include "llvm/Transforms/Scalar.h"
92 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
93 #include "llvm/Transforms/Utils/Local.h"
94 #include "llvm/Transforms/Utils/VectorUtils.h"
100 using namespace llvm::PatternMatch;
102 #define LV_NAME "loop-vectorize"
103 #define DEBUG_TYPE LV_NAME
105 STATISTIC(LoopsVectorized, "Number of loops vectorized");
106 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
108 static cl::opt<unsigned>
109 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
110 cl::desc("Sets the SIMD width. Zero is autoselect."));
112 static cl::opt<unsigned>
113 VectorizationInterleave("force-vector-interleave", cl::init(0), cl::Hidden,
114 cl::desc("Sets the vectorization interleave count. "
115 "Zero is autoselect."));
118 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
119 cl::desc("Enable if-conversion during vectorization."));
121 /// We don't vectorize loops with a known constant trip count below this number.
122 static cl::opt<unsigned>
123 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
125 cl::desc("Don't vectorize loops with a constant "
126 "trip count that is smaller than this "
129 /// This enables versioning on the strides of symbolically striding memory
130 /// accesses in code like the following.
131 /// for (i = 0; i < N; ++i)
132 /// A[i * Stride1] += B[i * Stride2] ...
134 /// Will be roughly translated to
135 /// if (Stride1 == 1 && Stride2 == 1) {
136 /// for (i = 0; i < N; i+=4)
140 static cl::opt<bool> EnableMemAccessVersioning(
141 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
142 cl::desc("Enable symblic stride memory access versioning"));
144 /// We don't unroll loops with a known constant trip count below this number.
145 static const unsigned TinyTripCountUnrollThreshold = 128;
147 /// When performing memory disambiguation checks at runtime do not make more
148 /// than this number of comparisons.
149 static const unsigned RuntimeMemoryCheckThreshold = 8;
151 /// Maximum simd width.
152 static const unsigned MaxVectorWidth = 64;
154 static cl::opt<unsigned> ForceTargetNumScalarRegs(
155 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
156 cl::desc("A flag that overrides the target's number of scalar registers."));
158 static cl::opt<unsigned> ForceTargetNumVectorRegs(
159 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
160 cl::desc("A flag that overrides the target's number of vector registers."));
162 /// Maximum vectorization interleave count.
163 static const unsigned MaxInterleaveFactor = 16;
165 static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
166 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
167 cl::desc("A flag that overrides the target's max interleave factor for "
170 static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
171 "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
172 cl::desc("A flag that overrides the target's max interleave factor for "
173 "vectorized loops."));
175 static cl::opt<unsigned> ForceTargetInstructionCost(
176 "force-target-instruction-cost", cl::init(0), cl::Hidden,
177 cl::desc("A flag that overrides the target's expected cost for "
178 "an instruction to a single constant value. Mostly "
179 "useful for getting consistent testing."));
181 static cl::opt<unsigned> SmallLoopCost(
182 "small-loop-cost", cl::init(20), cl::Hidden,
183 cl::desc("The cost of a loop that is considered 'small' by the unroller."));
185 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
186 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
187 cl::desc("Enable the use of the block frequency analysis to access PGO "
188 "heuristics minimizing code growth in cold regions and being more "
189 "aggressive in hot regions."));
191 // Runtime unroll loops for load/store throughput.
192 static cl::opt<bool> EnableLoadStoreRuntimeUnroll(
193 "enable-loadstore-runtime-unroll", cl::init(true), cl::Hidden,
194 cl::desc("Enable runtime unrolling until load/store ports are saturated"));
196 /// The number of stores in a loop that are allowed to need predication.
197 static cl::opt<unsigned> NumberOfStoresToPredicate(
198 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
199 cl::desc("Max number of stores to be predicated behind an if."));
201 static cl::opt<bool> EnableIndVarRegisterHeur(
202 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
203 cl::desc("Count the induction variable only once when unrolling"));
205 static cl::opt<bool> EnableCondStoresVectorization(
206 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
207 cl::desc("Enable if predication of stores during vectorization."));
209 static cl::opt<unsigned> MaxNestedScalarReductionUF(
210 "max-nested-scalar-reduction-unroll", cl::init(2), cl::Hidden,
211 cl::desc("The maximum unroll factor to use when unrolling a scalar "
212 "reduction in a nested loop."));
216 // Forward declarations.
217 class LoopVectorizationLegality;
218 class LoopVectorizationCostModel;
219 class LoopVectorizeHints;
221 /// Optimization analysis message produced during vectorization. Messages inform
222 /// the user why vectorization did not occur.
225 raw_string_ostream Out;
229 Report(Instruction *I = nullptr) : Out(Message), Instr(I) {
230 Out << "loop not vectorized: ";
233 template <typename A> Report &operator<<(const A &Value) {
238 Instruction *getInstr() { return Instr; }
240 std::string &str() { return Out.str(); }
241 operator Twine() { return Out.str(); }
244 /// InnerLoopVectorizer vectorizes loops which contain only one basic
245 /// block to a specified vectorization factor (VF).
246 /// This class performs the widening of scalars into vectors, or multiple
247 /// scalars. This class also implements the following features:
248 /// * It inserts an epilogue loop for handling loops that don't have iteration
249 /// counts that are known to be a multiple of the vectorization factor.
250 /// * It handles the code generation for reduction variables.
251 /// * Scalarization (implementation using scalars) of un-vectorizable
253 /// InnerLoopVectorizer does not perform any vectorization-legality
254 /// checks, and relies on the caller to check for the different legality
255 /// aspects. The InnerLoopVectorizer relies on the
256 /// LoopVectorizationLegality class to provide information about the induction
257 /// and reduction variables that were found to a given vectorization factor.
258 class InnerLoopVectorizer {
260 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
261 DominatorTree *DT, const DataLayout *DL,
262 const TargetLibraryInfo *TLI, unsigned VecWidth,
263 unsigned UnrollFactor)
264 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
265 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
266 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
269 // Perform the actual loop widening (vectorization).
270 void vectorize(LoopVectorizationLegality *L) {
272 // Create a new empty loop. Unlink the old loop and connect the new one.
274 // Widen each instruction in the old loop to a new one in the new loop.
275 // Use the Legality module to find the induction and reduction variables.
277 // Register the new loop and update the analysis passes.
281 virtual ~InnerLoopVectorizer() {}
284 /// A small list of PHINodes.
285 typedef SmallVector<PHINode*, 4> PhiVector;
286 /// When we unroll loops we have multiple vector values for each scalar.
287 /// This data structure holds the unrolled and vectorized values that
288 /// originated from one scalar instruction.
289 typedef SmallVector<Value*, 2> VectorParts;
291 // When we if-convert we need create edge masks. We have to cache values so
292 // that we don't end up with exponential recursion/IR.
293 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
294 VectorParts> EdgeMaskCache;
296 /// \brief Add code that checks at runtime if the accessed arrays overlap.
298 /// Returns a pair of instructions where the first element is the first
299 /// instruction generated in possibly a sequence of instructions and the
300 /// second value is the final comparator value or NULL if no check is needed.
301 std::pair<Instruction *, Instruction *> addRuntimeCheck(Instruction *Loc);
303 /// \brief Add checks for strides that where assumed to be 1.
305 /// Returns the last check instruction and the first check instruction in the
306 /// pair as (first, last).
307 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
309 /// Create an empty loop, based on the loop ranges of the old loop.
310 void createEmptyLoop();
311 /// Copy and widen the instructions from the old loop.
312 virtual void vectorizeLoop();
314 /// \brief The Loop exit block may have single value PHI nodes where the
315 /// incoming value is 'Undef'. While vectorizing we only handled real values
316 /// that were defined inside the loop. Here we fix the 'undef case'.
320 /// A helper function that computes the predicate of the block BB, assuming
321 /// that the header block of the loop is set to True. It returns the *entry*
322 /// mask for the block BB.
323 VectorParts createBlockInMask(BasicBlock *BB);
324 /// A helper function that computes the predicate of the edge between SRC
326 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
328 /// A helper function to vectorize a single BB within the innermost loop.
329 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
331 /// Vectorize a single PHINode in a block. This method handles the induction
332 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
333 /// arbitrary length vectors.
334 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
335 unsigned UF, unsigned VF, PhiVector *PV);
337 /// Insert the new loop to the loop hierarchy and pass manager
338 /// and update the analysis passes.
339 void updateAnalysis();
341 /// This instruction is un-vectorizable. Implement it as a sequence
342 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
343 /// scalarized instruction behind an if block predicated on the control
344 /// dependence of the instruction.
345 virtual void scalarizeInstruction(Instruction *Instr,
346 bool IfPredicateStore=false);
348 /// Vectorize Load and Store instructions,
349 virtual void vectorizeMemoryInstruction(Instruction *Instr);
351 /// Create a broadcast instruction. This method generates a broadcast
352 /// instruction (shuffle) for loop invariant values and for the induction
353 /// value. If this is the induction variable then we extend it to N, N+1, ...
354 /// this is needed because each iteration in the loop corresponds to a SIMD
356 virtual Value *getBroadcastInstrs(Value *V);
358 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
359 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
360 /// The sequence starts at StartIndex.
361 virtual Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
363 /// When we go over instructions in the basic block we rely on previous
364 /// values within the current basic block or on loop invariant values.
365 /// When we widen (vectorize) values we place them in the map. If the values
366 /// are not within the map, they have to be loop invariant, so we simply
367 /// broadcast them into a vector.
368 VectorParts &getVectorValue(Value *V);
370 /// Generate a shuffle sequence that will reverse the vector Vec.
371 virtual Value *reverseVector(Value *Vec);
373 /// This is a helper class that holds the vectorizer state. It maps scalar
374 /// instructions to vector instructions. When the code is 'unrolled' then
375 /// then a single scalar value is mapped to multiple vector parts. The parts
376 /// are stored in the VectorPart type.
378 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
380 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
382 /// \return True if 'Key' is saved in the Value Map.
383 bool has(Value *Key) const { return MapStorage.count(Key); }
385 /// Initializes a new entry in the map. Sets all of the vector parts to the
386 /// save value in 'Val'.
387 /// \return A reference to a vector with splat values.
388 VectorParts &splat(Value *Key, Value *Val) {
389 VectorParts &Entry = MapStorage[Key];
390 Entry.assign(UF, Val);
394 ///\return A reference to the value that is stored at 'Key'.
395 VectorParts &get(Value *Key) {
396 VectorParts &Entry = MapStorage[Key];
399 assert(Entry.size() == UF);
404 /// The unroll factor. Each entry in the map stores this number of vector
408 /// Map storage. We use std::map and not DenseMap because insertions to a
409 /// dense map invalidates its iterators.
410 std::map<Value *, VectorParts> MapStorage;
413 /// The original loop.
415 /// Scev analysis to use.
424 const DataLayout *DL;
425 /// Target Library Info.
426 const TargetLibraryInfo *TLI;
428 /// The vectorization SIMD factor to use. Each vector will have this many
433 /// The vectorization unroll factor to use. Each scalar is vectorized to this
434 /// many different vector instructions.
437 /// The builder that we use
440 // --- Vectorization state ---
442 /// The vector-loop preheader.
443 BasicBlock *LoopVectorPreHeader;
444 /// The scalar-loop preheader.
445 BasicBlock *LoopScalarPreHeader;
446 /// Middle Block between the vector and the scalar.
447 BasicBlock *LoopMiddleBlock;
448 ///The ExitBlock of the scalar loop.
449 BasicBlock *LoopExitBlock;
450 ///The vector loop body.
451 SmallVector<BasicBlock *, 4> LoopVectorBody;
452 ///The scalar loop body.
453 BasicBlock *LoopScalarBody;
454 /// A list of all bypass blocks. The first block is the entry of the loop.
455 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
457 /// The new Induction variable which was added to the new block.
459 /// The induction variable of the old basic block.
460 PHINode *OldInduction;
461 /// Holds the extended (to the widest induction type) start index.
463 /// Maps scalars to widened vectors.
465 EdgeMaskCache MaskCache;
467 LoopVectorizationLegality *Legal;
470 class InnerLoopUnroller : public InnerLoopVectorizer {
472 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
473 DominatorTree *DT, const DataLayout *DL,
474 const TargetLibraryInfo *TLI, unsigned UnrollFactor) :
475 InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { }
478 void scalarizeInstruction(Instruction *Instr,
479 bool IfPredicateStore = false) override;
480 void vectorizeMemoryInstruction(Instruction *Instr) override;
481 Value *getBroadcastInstrs(Value *V) override;
482 Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate) override;
483 Value *reverseVector(Value *Vec) override;
486 /// \brief Look for a meaningful debug location on the instruction or it's
488 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
493 if (I->getDebugLoc() != Empty)
496 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
497 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
498 if (OpInst->getDebugLoc() != Empty)
505 /// \brief Set the debug location in the builder using the debug location in the
507 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
508 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
509 B.SetCurrentDebugLocation(Inst->getDebugLoc());
511 B.SetCurrentDebugLocation(DebugLoc());
515 /// \return string containing a file name and a line # for the given loop.
516 static std::string getDebugLocString(const Loop *L) {
519 raw_string_ostream OS(Result);
520 const DebugLoc LoopDbgLoc = L->getStartLoc();
521 if (!LoopDbgLoc.isUnknown())
522 LoopDbgLoc.print(L->getHeader()->getContext(), OS);
524 // Just print the module name.
525 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
532 /// \brief Propagate known metadata from one instruction to another.
533 static void propagateMetadata(Instruction *To, const Instruction *From) {
534 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
535 From->getAllMetadataOtherThanDebugLoc(Metadata);
537 for (auto M : Metadata) {
538 unsigned Kind = M.first;
540 // These are safe to transfer (this is safe for TBAA, even when we
541 // if-convert, because should that metadata have had a control dependency
542 // on the condition, and thus actually aliased with some other
543 // non-speculated memory access when the condition was false, this would be
544 // caught by the runtime overlap checks).
545 if (Kind != LLVMContext::MD_tbaa &&
546 Kind != LLVMContext::MD_alias_scope &&
547 Kind != LLVMContext::MD_noalias &&
548 Kind != LLVMContext::MD_fpmath)
551 To->setMetadata(Kind, M.second);
555 /// \brief Propagate known metadata from one instruction to a vector of others.
556 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
558 if (Instruction *I = dyn_cast<Instruction>(V))
559 propagateMetadata(I, From);
562 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
563 /// to what vectorization factor.
564 /// This class does not look at the profitability of vectorization, only the
565 /// legality. This class has two main kinds of checks:
566 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
567 /// will change the order of memory accesses in a way that will change the
568 /// correctness of the program.
569 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
570 /// checks for a number of different conditions, such as the availability of a
571 /// single induction variable, that all types are supported and vectorize-able,
572 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
573 /// This class is also used by InnerLoopVectorizer for identifying
574 /// induction variable and the different reduction variables.
575 class LoopVectorizationLegality {
579 unsigned NumPredStores;
581 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, const DataLayout *DL,
582 DominatorTree *DT, TargetLibraryInfo *TLI,
583 AliasAnalysis *AA, Function *F,
584 const TargetTransformInfo *TTI)
585 : NumLoads(0), NumStores(0), NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
586 DT(DT), TLI(TLI), AA(AA), TheFunction(F), TTI(TTI), Induction(nullptr),
587 WidestIndTy(nullptr), HasFunNoNaNAttr(false), MaxSafeDepDistBytes(-1U) {
590 /// This enum represents the kinds of reductions that we support.
592 RK_NoReduction, ///< Not a reduction.
593 RK_IntegerAdd, ///< Sum of integers.
594 RK_IntegerMult, ///< Product of integers.
595 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
596 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
597 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
598 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
599 RK_FloatAdd, ///< Sum of floats.
600 RK_FloatMult, ///< Product of floats.
601 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
604 /// This enum represents the kinds of inductions that we support.
606 IK_NoInduction, ///< Not an induction variable.
607 IK_IntInduction, ///< Integer induction variable. Step = 1.
608 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
609 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
610 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
613 // This enum represents the kind of minmax reduction.
614 enum MinMaxReductionKind {
624 /// This struct holds information about reduction variables.
625 struct ReductionDescriptor {
626 ReductionDescriptor() : StartValue(nullptr), LoopExitInstr(nullptr),
627 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
629 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
630 MinMaxReductionKind MK)
631 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
633 // The starting value of the reduction.
634 // It does not have to be zero!
635 TrackingVH<Value> StartValue;
636 // The instruction who's value is used outside the loop.
637 Instruction *LoopExitInstr;
638 // The kind of the reduction.
640 // If this a min/max reduction the kind of reduction.
641 MinMaxReductionKind MinMaxKind;
644 /// This POD struct holds information about a potential reduction operation.
645 struct ReductionInstDesc {
646 ReductionInstDesc(bool IsRedux, Instruction *I) :
647 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
649 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
650 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
652 // Is this instruction a reduction candidate.
654 // The last instruction in a min/max pattern (select of the select(icmp())
655 // pattern), or the current reduction instruction otherwise.
656 Instruction *PatternLastInst;
657 // If this is a min/max pattern the comparison predicate.
658 MinMaxReductionKind MinMaxKind;
661 /// This struct holds information about the memory runtime legality
662 /// check that a group of pointers do not overlap.
663 struct RuntimePointerCheck {
664 RuntimePointerCheck() : Need(false) {}
666 /// Reset the state of the pointer runtime information.
673 DependencySetId.clear();
677 /// Insert a pointer and calculate the start and end SCEVs.
678 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
679 unsigned DepSetId, unsigned ASId, ValueToValueMap &Strides);
681 /// This flag indicates if we need to add the runtime check.
683 /// Holds the pointers that we need to check.
684 SmallVector<TrackingVH<Value>, 2> Pointers;
685 /// Holds the pointer value at the beginning of the loop.
686 SmallVector<const SCEV*, 2> Starts;
687 /// Holds the pointer value at the end of the loop.
688 SmallVector<const SCEV*, 2> Ends;
689 /// Holds the information if this pointer is used for writing to memory.
690 SmallVector<bool, 2> IsWritePtr;
691 /// Holds the id of the set of pointers that could be dependent because of a
692 /// shared underlying object.
693 SmallVector<unsigned, 2> DependencySetId;
694 /// Holds the id of the disjoint alias set to which this pointer belongs.
695 SmallVector<unsigned, 2> AliasSetId;
698 /// A struct for saving information about induction variables.
699 struct InductionInfo {
700 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
701 InductionInfo() : StartValue(nullptr), IK(IK_NoInduction) {}
703 TrackingVH<Value> StartValue;
708 /// ReductionList contains the reduction descriptors for all
709 /// of the reductions that were found in the loop.
710 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
712 /// InductionList saves induction variables and maps them to the
713 /// induction descriptor.
714 typedef MapVector<PHINode*, InductionInfo> InductionList;
716 /// Returns true if it is legal to vectorize this loop.
717 /// This does not mean that it is profitable to vectorize this
718 /// loop, only that it is legal to do so.
721 /// Returns the Induction variable.
722 PHINode *getInduction() { return Induction; }
724 /// Returns the reduction variables found in the loop.
725 ReductionList *getReductionVars() { return &Reductions; }
727 /// Returns the induction variables found in the loop.
728 InductionList *getInductionVars() { return &Inductions; }
730 /// Returns the widest induction type.
731 Type *getWidestInductionType() { return WidestIndTy; }
733 /// Returns True if V is an induction variable in this loop.
734 bool isInductionVariable(const Value *V);
736 /// Return true if the block BB needs to be predicated in order for the loop
737 /// to be vectorized.
738 bool blockNeedsPredication(BasicBlock *BB);
740 /// Check if this pointer is consecutive when vectorizing. This happens
741 /// when the last index of the GEP is the induction variable, or that the
742 /// pointer itself is an induction variable.
743 /// This check allows us to vectorize A[idx] into a wide load/store.
745 /// 0 - Stride is unknown or non-consecutive.
746 /// 1 - Address is consecutive.
747 /// -1 - Address is consecutive, and decreasing.
748 int isConsecutivePtr(Value *Ptr);
750 /// Returns true if the value V is uniform within the loop.
751 bool isUniform(Value *V);
753 /// Returns true if this instruction will remain scalar after vectorization.
754 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
756 /// Returns the information that we collected about runtime memory check.
757 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
759 /// This function returns the identity element (or neutral element) for
761 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
763 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
765 bool hasStride(Value *V) { return StrideSet.count(V); }
766 bool mustCheckStrides() { return !StrideSet.empty(); }
767 SmallPtrSet<Value *, 8>::iterator strides_begin() {
768 return StrideSet.begin();
770 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
772 bool canPredicateStore(Type *DataType, Value *Ptr) {
773 return TTI->isLegalPredicatedStore(DataType, isConsecutivePtr(Ptr));
775 bool canPredicateLoad(Type *DataType, Value *Ptr) {
776 return TTI->isLegalPredicatedLoad(DataType, isConsecutivePtr(Ptr));
778 bool setMaskedOp(const Instruction* I) {
779 return (MaskedOp.find(I) != MaskedOp.end());
782 /// Check if a single basic block loop is vectorizable.
783 /// At this point we know that this is a loop with a constant trip count
784 /// and we only need to check individual instructions.
785 bool canVectorizeInstrs();
787 /// When we vectorize loops we may change the order in which
788 /// we read and write from memory. This method checks if it is
789 /// legal to vectorize the code, considering only memory constrains.
790 /// Returns true if the loop is vectorizable
791 bool canVectorizeMemory();
793 /// Return true if we can vectorize this loop using the IF-conversion
795 bool canVectorizeWithIfConvert();
797 /// Collect the variables that need to stay uniform after vectorization.
798 void collectLoopUniforms();
800 /// Return true if all of the instructions in the block can be speculatively
801 /// executed. \p SafePtrs is a list of addresses that are known to be legal
802 /// and we know that we can read from them without segfault.
803 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
805 /// Returns True, if 'Phi' is the kind of reduction variable for type
806 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
807 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
808 /// Returns a struct describing if the instruction 'I' can be a reduction
809 /// variable of type 'Kind'. If the reduction is a min/max pattern of
810 /// select(icmp()) this function advances the instruction pointer 'I' from the
811 /// compare instruction to the select instruction and stores this pointer in
812 /// 'PatternLastInst' member of the returned struct.
813 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
814 ReductionInstDesc &Desc);
815 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
816 /// pattern corresponding to a min(X, Y) or max(X, Y).
817 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
818 ReductionInstDesc &Prev);
819 /// Returns the induction kind of Phi. This function may return NoInduction
820 /// if the PHI is not an induction variable.
821 InductionKind isInductionVariable(PHINode *Phi);
823 /// \brief Collect memory access with loop invariant strides.
825 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
827 void collectStridedAcccess(Value *LoadOrStoreInst);
829 /// Report an analysis message to assist the user in diagnosing loops that are
831 void emitAnalysis(Report &Message) {
832 DebugLoc DL = TheLoop->getStartLoc();
833 if (Instruction *I = Message.getInstr())
834 DL = I->getDebugLoc();
835 emitOptimizationRemarkAnalysis(TheFunction->getContext(), DEBUG_TYPE,
836 *TheFunction, DL, Message.str());
839 /// The loop that we evaluate.
843 /// DataLayout analysis.
844 const DataLayout *DL;
847 /// Target Library Info.
848 TargetLibraryInfo *TLI;
852 Function *TheFunction;
853 /// Target Transform Info
854 const TargetTransformInfo *TTI;
856 // --- vectorization state --- //
858 /// Holds the integer induction variable. This is the counter of the
861 /// Holds the reduction variables.
862 ReductionList Reductions;
863 /// Holds all of the induction variables that we found in the loop.
864 /// Notice that inductions don't need to start at zero and that induction
865 /// variables can be pointers.
866 InductionList Inductions;
867 /// Holds the widest induction type encountered.
870 /// Allowed outside users. This holds the reduction
871 /// vars which can be accessed from outside the loop.
872 SmallPtrSet<Value*, 4> AllowedExit;
873 /// This set holds the variables which are known to be uniform after
875 SmallPtrSet<Instruction*, 4> Uniforms;
876 /// We need to check that all of the pointers in this list are disjoint
878 RuntimePointerCheck PtrRtCheck;
879 /// Can we assume the absence of NaNs.
880 bool HasFunNoNaNAttr;
882 unsigned MaxSafeDepDistBytes;
884 ValueToValueMap Strides;
885 SmallPtrSet<Value *, 8> StrideSet;
887 /// While vectorizing these instructions we have to generate a
888 /// call to an appropriate masked intrinsic
889 std::set<const Instruction*> MaskedOp;
892 /// LoopVectorizationCostModel - estimates the expected speedups due to
894 /// In many cases vectorization is not profitable. This can happen because of
895 /// a number of reasons. In this class we mainly attempt to predict the
896 /// expected speedup/slowdowns due to the supported instruction set. We use the
897 /// TargetTransformInfo to query the different backends for the cost of
898 /// different operations.
899 class LoopVectorizationCostModel {
901 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
902 LoopVectorizationLegality *Legal,
903 const TargetTransformInfo &TTI,
904 const DataLayout *DL, const TargetLibraryInfo *TLI,
905 AssumptionTracker *AT, const Function *F,
906 const LoopVectorizeHints *Hints)
907 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI),
908 TheFunction(F), Hints(Hints) {
909 CodeMetrics::collectEphemeralValues(L, AT, EphValues);
912 /// Information about vectorization costs
913 struct VectorizationFactor {
914 unsigned Width; // Vector width with best cost
915 unsigned Cost; // Cost of the loop with that width
917 /// \return The most profitable vectorization factor and the cost of that VF.
918 /// This method checks every power of two up to VF. If UserVF is not ZERO
919 /// then this vectorization factor will be selected if vectorization is
921 VectorizationFactor selectVectorizationFactor(bool OptForSize);
923 /// \return The size (in bits) of the widest type in the code that
924 /// needs to be vectorized. We ignore values that remain scalar such as
925 /// 64 bit loop indices.
926 unsigned getWidestType();
928 /// \return The most profitable unroll factor.
929 /// If UserUF is non-zero then this method finds the best unroll-factor
930 /// based on register pressure and other parameters.
931 /// VF and LoopCost are the selected vectorization factor and the cost of the
933 unsigned selectUnrollFactor(bool OptForSize, unsigned VF, unsigned LoopCost);
935 /// \brief A struct that represents some properties of the register usage
937 struct RegisterUsage {
938 /// Holds the number of loop invariant values that are used in the loop.
939 unsigned LoopInvariantRegs;
940 /// Holds the maximum number of concurrent live intervals in the loop.
941 unsigned MaxLocalUsers;
942 /// Holds the number of instructions in the loop.
943 unsigned NumInstructions;
946 /// \return information about the register usage of the loop.
947 RegisterUsage calculateRegisterUsage();
950 /// Returns the expected execution cost. The unit of the cost does
951 /// not matter because we use the 'cost' units to compare different
952 /// vector widths. The cost that is returned is *not* normalized by
953 /// the factor width.
954 unsigned expectedCost(unsigned VF);
956 /// Returns the execution time cost of an instruction for a given vector
957 /// width. Vector width of one means scalar.
958 unsigned getInstructionCost(Instruction *I, unsigned VF);
960 /// A helper function for converting Scalar types to vector types.
961 /// If the incoming type is void, we return void. If the VF is 1, we return
963 static Type* ToVectorTy(Type *Scalar, unsigned VF);
965 /// Returns whether the instruction is a load or store and will be a emitted
966 /// as a vector operation.
967 bool isConsecutiveLoadOrStore(Instruction *I);
969 /// Report an analysis message to assist the user in diagnosing loops that are
971 void emitAnalysis(Report &Message) {
972 DebugLoc DL = TheLoop->getStartLoc();
973 if (Instruction *I = Message.getInstr())
974 DL = I->getDebugLoc();
975 emitOptimizationRemarkAnalysis(TheFunction->getContext(), DEBUG_TYPE,
976 *TheFunction, DL, Message.str());
979 /// Values used only by @llvm.assume calls.
980 SmallPtrSet<const Value *, 32> EphValues;
982 /// The loop that we evaluate.
986 /// Loop Info analysis.
988 /// Vectorization legality.
989 LoopVectorizationLegality *Legal;
990 /// Vector target information.
991 const TargetTransformInfo &TTI;
992 /// Target data layout information.
993 const DataLayout *DL;
994 /// Target Library Info.
995 const TargetLibraryInfo *TLI;
996 const Function *TheFunction;
997 // Loop Vectorize Hint.
998 const LoopVectorizeHints *Hints;
1001 /// Utility class for getting and setting loop vectorizer hints in the form
1002 /// of loop metadata.
1003 /// This class keeps a number of loop annotations locally (as member variables)
1004 /// and can, upon request, write them back as metadata on the loop. It will
1005 /// initially scan the loop for existing metadata, and will update the local
1006 /// values based on information in the loop.
1007 /// We cannot write all values to metadata, as the mere presence of some info,
1008 /// for example 'force', means a decision has been made. So, we need to be
1009 /// careful NOT to add them if the user hasn't specifically asked so.
1010 class LoopVectorizeHints {
1017 /// Hint - associates name and validation with the hint value.
1020 unsigned Value; // This may have to change for non-numeric values.
1023 Hint(const char * Name, unsigned Value, HintKind Kind)
1024 : Name(Name), Value(Value), Kind(Kind) { }
1026 bool validate(unsigned Val) {
1029 return isPowerOf2_32(Val) && Val <= MaxVectorWidth;
1031 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1039 /// Vectorization width.
1041 /// Vectorization interleave factor.
1043 /// Vectorization forced
1046 /// Return the loop metadata prefix.
1047 static StringRef Prefix() { return "llvm.loop."; }
1051 FK_Undefined = -1, ///< Not selected.
1052 FK_Disabled = 0, ///< Forcing disabled.
1053 FK_Enabled = 1, ///< Forcing enabled.
1056 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
1057 : Width("vectorize.width", VectorizationFactor, HK_WIDTH),
1058 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1059 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1061 // Populate values with existing loop metadata.
1062 getHintsFromMetadata();
1064 // force-vector-interleave overrides DisableInterleaving.
1065 if (VectorizationInterleave.getNumOccurrences() > 0)
1066 Interleave.Value = VectorizationInterleave;
1068 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
1069 << "LV: Interleaving disabled by the pass manager\n");
1072 /// Mark the loop L as already vectorized by setting the width to 1.
1073 void setAlreadyVectorized() {
1074 Width.Value = Interleave.Value = 1;
1075 Hint Hints[] = {Width, Interleave};
1076 writeHintsToMetadata(Hints);
1079 /// Dumps all the hint information.
1080 std::string emitRemark() const {
1082 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1083 R << "vectorization is explicitly disabled";
1085 R << "use -Rpass-analysis=loop-vectorize for more info";
1086 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1087 R << " (Force=true";
1088 if (Width.Value != 0)
1089 R << ", Vector Width=" << Width.Value;
1090 if (Interleave.Value != 0)
1091 R << ", Interleave Count=" << Interleave.Value;
1099 unsigned getWidth() const { return Width.Value; }
1100 unsigned getInterleave() const { return Interleave.Value; }
1101 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1104 /// Find hints specified in the loop metadata and update local values.
1105 void getHintsFromMetadata() {
1106 MDNode *LoopID = TheLoop->getLoopID();
1110 // First operand should refer to the loop id itself.
1111 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1112 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1114 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1115 const MDString *S = nullptr;
1116 SmallVector<Value*, 4> Args;
1118 // The expected hint is either a MDString or a MDNode with the first
1119 // operand a MDString.
1120 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1121 if (!MD || MD->getNumOperands() == 0)
1123 S = dyn_cast<MDString>(MD->getOperand(0));
1124 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1125 Args.push_back(MD->getOperand(i));
1127 S = dyn_cast<MDString>(LoopID->getOperand(i));
1128 assert(Args.size() == 0 && "too many arguments for MDString");
1134 // Check if the hint starts with the loop metadata prefix.
1135 StringRef Name = S->getString();
1136 if (Args.size() == 1)
1137 setHint(Name, Args[0]);
1141 /// Checks string hint with one operand and set value if valid.
1142 void setHint(StringRef Name, Value *Arg) {
1143 if (!Name.startswith(Prefix()))
1145 Name = Name.substr(Prefix().size(), StringRef::npos);
1147 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
1149 unsigned Val = C->getZExtValue();
1151 Hint *Hints[] = {&Width, &Interleave, &Force};
1152 for (auto H : Hints) {
1153 if (Name == H->Name) {
1154 if (H->validate(Val))
1157 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1163 /// Create a new hint from name / value pair.
1164 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1165 LLVMContext &Context = TheLoop->getHeader()->getContext();
1166 Value *Vals[] = {MDString::get(Context, Name),
1167 ConstantInt::get(Type::getInt32Ty(Context), V)};
1168 return MDNode::get(Context, Vals);
1171 /// Matches metadata with hint name.
1172 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1173 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1177 for (auto H : HintTypes)
1178 if (Name->getString().endswith(H.Name))
1183 /// Sets current hints into loop metadata, keeping other values intact.
1184 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1185 if (HintTypes.size() == 0)
1188 // Reserve the first element to LoopID (see below).
1189 SmallVector<Value*, 4> Vals(1);
1190 // If the loop already has metadata, then ignore the existing operands.
1191 MDNode *LoopID = TheLoop->getLoopID();
1193 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1194 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1195 // If node in update list, ignore old value.
1196 if (!matchesHintMetadataName(Node, HintTypes))
1197 Vals.push_back(Node);
1201 // Now, add the missing hints.
1202 for (auto H : HintTypes)
1204 createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1206 // Replace current metadata node with new one.
1207 LLVMContext &Context = TheLoop->getHeader()->getContext();
1208 MDNode *NewLoopID = MDNode::get(Context, Vals);
1209 // Set operand 0 to refer to the loop id itself.
1210 NewLoopID->replaceOperandWith(0, NewLoopID);
1212 TheLoop->setLoopID(NewLoopID);
1214 LoopID->replaceAllUsesWith(NewLoopID);
1218 /// The loop these hints belong to.
1219 const Loop *TheLoop;
1222 static void emitMissedWarning(Function *F, Loop *L,
1223 const LoopVectorizeHints &LH) {
1224 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1225 L->getStartLoc(), LH.emitRemark());
1227 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1228 if (LH.getWidth() != 1)
1229 emitLoopVectorizeWarning(
1230 F->getContext(), *F, L->getStartLoc(),
1231 "failed explicitly specified loop vectorization");
1232 else if (LH.getInterleave() != 1)
1233 emitLoopInterleaveWarning(
1234 F->getContext(), *F, L->getStartLoc(),
1235 "failed explicitly specified loop interleaving");
1239 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1241 return V.push_back(&L);
1243 for (Loop *InnerL : L)
1244 addInnerLoop(*InnerL, V);
1247 /// The LoopVectorize Pass.
1248 struct LoopVectorize : public FunctionPass {
1249 /// Pass identification, replacement for typeid
1252 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1254 DisableUnrolling(NoUnrolling),
1255 AlwaysVectorize(AlwaysVectorize) {
1256 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1259 ScalarEvolution *SE;
1260 const DataLayout *DL;
1262 TargetTransformInfo *TTI;
1264 BlockFrequencyInfo *BFI;
1265 TargetLibraryInfo *TLI;
1267 AssumptionTracker *AT;
1268 bool DisableUnrolling;
1269 bool AlwaysVectorize;
1271 BlockFrequency ColdEntryFreq;
1273 bool runOnFunction(Function &F) override {
1274 SE = &getAnalysis<ScalarEvolution>();
1275 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1276 DL = DLP ? &DLP->getDataLayout() : nullptr;
1277 LI = &getAnalysis<LoopInfo>();
1278 TTI = &getAnalysis<TargetTransformInfo>();
1279 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1280 BFI = &getAnalysis<BlockFrequencyInfo>();
1281 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
1282 AA = &getAnalysis<AliasAnalysis>();
1283 AT = &getAnalysis<AssumptionTracker>();
1285 // Compute some weights outside of the loop over the loops. Compute this
1286 // using a BranchProbability to re-use its scaling math.
1287 const BranchProbability ColdProb(1, 5); // 20%
1288 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1290 // If the target claims to have no vector registers don't attempt
1292 if (!TTI->getNumberOfRegisters(true))
1296 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1297 << ": Missing data layout\n");
1301 // Build up a worklist of inner-loops to vectorize. This is necessary as
1302 // the act of vectorizing or partially unrolling a loop creates new loops
1303 // and can invalidate iterators across the loops.
1304 SmallVector<Loop *, 8> Worklist;
1307 addInnerLoop(*L, Worklist);
1309 LoopsAnalyzed += Worklist.size();
1311 // Now walk the identified inner loops.
1312 bool Changed = false;
1313 while (!Worklist.empty())
1314 Changed |= processLoop(Worklist.pop_back_val());
1316 // Process each loop nest in the function.
1320 bool processLoop(Loop *L) {
1321 assert(L->empty() && "Only process inner loops.");
1324 const std::string DebugLocStr = getDebugLocString(L);
1327 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1328 << L->getHeader()->getParent()->getName() << "\" from "
1329 << DebugLocStr << "\n");
1331 LoopVectorizeHints Hints(L, DisableUnrolling);
1333 DEBUG(dbgs() << "LV: Loop hints:"
1335 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1337 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1339 : "?")) << " width=" << Hints.getWidth()
1340 << " unroll=" << Hints.getInterleave() << "\n");
1342 // Function containing loop
1343 Function *F = L->getHeader()->getParent();
1345 // Looking at the diagnostic output is the only way to determine if a loop
1346 // was vectorized (other than looking at the IR or machine code), so it
1347 // is important to generate an optimization remark for each loop. Most of
1348 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1349 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1350 // less verbose reporting vectorized loops and unvectorized loops that may
1351 // benefit from vectorization, respectively.
1353 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1354 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1355 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1356 L->getStartLoc(), Hints.emitRemark());
1360 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1361 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1362 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1363 L->getStartLoc(), Hints.emitRemark());
1367 if (Hints.getWidth() == 1 && Hints.getInterleave() == 1) {
1368 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1369 emitOptimizationRemarkAnalysis(
1370 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1371 "loop not vectorized: vector width and interleave count are "
1372 "explicitly set to 1");
1376 // Check the loop for a trip count threshold:
1377 // do not vectorize loops with a tiny trip count.
1378 const unsigned TC = SE->getSmallConstantTripCount(L);
1379 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1380 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1381 << "This loop is not worth vectorizing.");
1382 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1383 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1385 DEBUG(dbgs() << "\n");
1386 emitOptimizationRemarkAnalysis(
1387 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1388 "vectorization is not beneficial and is not explicitly forced");
1393 // Check if it is legal to vectorize the loop.
1394 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, AA, F, TTI);
1395 if (!LVL.canVectorize()) {
1396 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1397 emitMissedWarning(F, L, Hints);
1401 // Use the cost model.
1402 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI, AT, F,
1405 // Check the function attributes to find out if this function should be
1406 // optimized for size.
1407 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1408 F->hasFnAttribute(Attribute::OptimizeForSize);
1410 // Compute the weighted frequency of this loop being executed and see if it
1411 // is less than 20% of the function entry baseline frequency. Note that we
1412 // always have a canonical loop here because we think we *can* vectoriez.
1413 // FIXME: This is hidden behind a flag due to pervasive problems with
1414 // exactly what block frequency models.
1415 if (LoopVectorizeWithBlockFrequency) {
1416 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1417 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1418 LoopEntryFreq < ColdEntryFreq)
1422 // Check the function attributes to see if implicit floats are allowed.a
1423 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1424 // an integer loop and the vector instructions selected are purely integer
1425 // vector instructions?
1426 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1427 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1428 "attribute is used.\n");
1429 emitOptimizationRemarkAnalysis(
1430 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1431 "loop not vectorized due to NoImplicitFloat attribute");
1432 emitMissedWarning(F, L, Hints);
1436 // Select the optimal vectorization factor.
1437 const LoopVectorizationCostModel::VectorizationFactor VF =
1438 CM.selectVectorizationFactor(OptForSize);
1440 // Select the unroll factor.
1442 CM.selectUnrollFactor(OptForSize, VF.Width, VF.Cost);
1444 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1445 << DebugLocStr << '\n');
1446 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1448 if (VF.Width == 1) {
1449 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
1452 emitOptimizationRemarkAnalysis(
1453 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1454 "not beneficial to vectorize and user disabled interleaving");
1457 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1459 // Report the unrolling decision.
1460 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1461 Twine("unrolled with interleaving factor " +
1463 " (vectorization not beneficial)"));
1465 // We decided not to vectorize, but we may want to unroll.
1467 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1468 Unroller.vectorize(&LVL);
1470 // If we decided that it is *legal* to vectorize the loop then do it.
1471 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1475 // Report the vectorization decision.
1476 emitOptimizationRemark(
1477 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1478 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1479 ", unrolling interleave factor: " + Twine(UF) + ")");
1482 // Mark the loop as already vectorized to avoid vectorizing again.
1483 Hints.setAlreadyVectorized();
1485 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1489 void getAnalysisUsage(AnalysisUsage &AU) const override {
1490 AU.addRequired<AssumptionTracker>();
1491 AU.addRequiredID(LoopSimplifyID);
1492 AU.addRequiredID(LCSSAID);
1493 AU.addRequired<BlockFrequencyInfo>();
1494 AU.addRequired<DominatorTreeWrapperPass>();
1495 AU.addRequired<LoopInfo>();
1496 AU.addRequired<ScalarEvolution>();
1497 AU.addRequired<TargetTransformInfo>();
1498 AU.addRequired<AliasAnalysis>();
1499 AU.addPreserved<LoopInfo>();
1500 AU.addPreserved<DominatorTreeWrapperPass>();
1501 AU.addPreserved<AliasAnalysis>();
1506 } // end anonymous namespace
1508 //===----------------------------------------------------------------------===//
1509 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1510 // LoopVectorizationCostModel.
1511 //===----------------------------------------------------------------------===//
1513 static Value *stripIntegerCast(Value *V) {
1514 if (CastInst *CI = dyn_cast<CastInst>(V))
1515 if (CI->getOperand(0)->getType()->isIntegerTy())
1516 return CI->getOperand(0);
1520 ///\brief Replaces the symbolic stride in a pointer SCEV expression by one.
1522 /// If \p OrigPtr is not null, use it to look up the stride value instead of
1524 static const SCEV *replaceSymbolicStrideSCEV(ScalarEvolution *SE,
1525 ValueToValueMap &PtrToStride,
1526 Value *Ptr, Value *OrigPtr = nullptr) {
1528 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
1530 // If there is an entry in the map return the SCEV of the pointer with the
1531 // symbolic stride replaced by one.
1532 ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
1533 if (SI != PtrToStride.end()) {
1534 Value *StrideVal = SI->second;
1537 StrideVal = stripIntegerCast(StrideVal);
1539 // Replace symbolic stride by one.
1540 Value *One = ConstantInt::get(StrideVal->getType(), 1);
1541 ValueToValueMap RewriteMap;
1542 RewriteMap[StrideVal] = One;
1545 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
1546 DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
1551 // Otherwise, just return the SCEV of the original pointer.
1552 return SE->getSCEV(Ptr);
1555 void LoopVectorizationLegality::RuntimePointerCheck::insert(
1556 ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
1557 unsigned ASId, ValueToValueMap &Strides) {
1558 // Get the stride replaced scev.
1559 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
1560 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1561 assert(AR && "Invalid addrec expression");
1562 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1563 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1564 Pointers.push_back(Ptr);
1565 Starts.push_back(AR->getStart());
1566 Ends.push_back(ScEnd);
1567 IsWritePtr.push_back(WritePtr);
1568 DependencySetId.push_back(DepSetId);
1569 AliasSetId.push_back(ASId);
1572 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1573 // We need to place the broadcast of invariant variables outside the loop.
1574 Instruction *Instr = dyn_cast<Instruction>(V);
1576 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1577 Instr->getParent()) != LoopVectorBody.end());
1578 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1580 // Place the code for broadcasting invariant variables in the new preheader.
1581 IRBuilder<>::InsertPointGuard Guard(Builder);
1583 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1585 // Broadcast the scalar into all locations in the vector.
1586 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1591 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1593 assert(Val->getType()->isVectorTy() && "Must be a vector");
1594 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1595 "Elem must be an integer");
1596 // Create the types.
1597 Type *ITy = Val->getType()->getScalarType();
1598 VectorType *Ty = cast<VectorType>(Val->getType());
1599 int VLen = Ty->getNumElements();
1600 SmallVector<Constant*, 8> Indices;
1602 // Create a vector of consecutive numbers from zero to VF.
1603 for (int i = 0; i < VLen; ++i) {
1604 int64_t Idx = Negate ? (-i) : i;
1605 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1608 // Add the consecutive indices to the vector value.
1609 Constant *Cv = ConstantVector::get(Indices);
1610 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1611 return Builder.CreateAdd(Val, Cv, "induction");
1614 /// \brief Find the operand of the GEP that should be checked for consecutive
1615 /// stores. This ignores trailing indices that have no effect on the final
1617 static unsigned getGEPInductionOperand(const DataLayout *DL,
1618 const GetElementPtrInst *Gep) {
1619 unsigned LastOperand = Gep->getNumOperands() - 1;
1620 unsigned GEPAllocSize = DL->getTypeAllocSize(
1621 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1623 // Walk backwards and try to peel off zeros.
1624 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1625 // Find the type we're currently indexing into.
1626 gep_type_iterator GEPTI = gep_type_begin(Gep);
1627 std::advance(GEPTI, LastOperand - 1);
1629 // If it's a type with the same allocation size as the result of the GEP we
1630 // can peel off the zero index.
1631 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1639 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1640 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1641 // Make sure that the pointer does not point to structs.
1642 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1645 // If this value is a pointer induction variable we know it is consecutive.
1646 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1647 if (Phi && Inductions.count(Phi)) {
1648 InductionInfo II = Inductions[Phi];
1649 if (IK_PtrInduction == II.IK)
1651 else if (IK_ReversePtrInduction == II.IK)
1655 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1659 unsigned NumOperands = Gep->getNumOperands();
1660 Value *GpPtr = Gep->getPointerOperand();
1661 // If this GEP value is a consecutive pointer induction variable and all of
1662 // the indices are constant then we know it is consecutive. We can
1663 Phi = dyn_cast<PHINode>(GpPtr);
1664 if (Phi && Inductions.count(Phi)) {
1666 // Make sure that the pointer does not point to structs.
1667 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1668 if (GepPtrType->getElementType()->isAggregateType())
1671 // Make sure that all of the index operands are loop invariant.
1672 for (unsigned i = 1; i < NumOperands; ++i)
1673 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1676 InductionInfo II = Inductions[Phi];
1677 if (IK_PtrInduction == II.IK)
1679 else if (IK_ReversePtrInduction == II.IK)
1683 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1685 // Check that all of the gep indices are uniform except for our induction
1687 for (unsigned i = 0; i != NumOperands; ++i)
1688 if (i != InductionOperand &&
1689 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1692 // We can emit wide load/stores only if the last non-zero index is the
1693 // induction variable.
1694 const SCEV *Last = nullptr;
1695 if (!Strides.count(Gep))
1696 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1698 // Because of the multiplication by a stride we can have a s/zext cast.
1699 // We are going to replace this stride by 1 so the cast is safe to ignore.
1701 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1702 // %0 = trunc i64 %indvars.iv to i32
1703 // %mul = mul i32 %0, %Stride1
1704 // %idxprom = zext i32 %mul to i64 << Safe cast.
1705 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1707 Last = replaceSymbolicStrideSCEV(SE, Strides,
1708 Gep->getOperand(InductionOperand), Gep);
1709 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1711 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1715 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1716 const SCEV *Step = AR->getStepRecurrence(*SE);
1718 // The memory is consecutive because the last index is consecutive
1719 // and all other indices are loop invariant.
1722 if (Step->isAllOnesValue())
1729 bool LoopVectorizationLegality::isUniform(Value *V) {
1730 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1733 InnerLoopVectorizer::VectorParts&
1734 InnerLoopVectorizer::getVectorValue(Value *V) {
1735 assert(V != Induction && "The new induction variable should not be used.");
1736 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1738 // If we have a stride that is replaced by one, do it here.
1739 if (Legal->hasStride(V))
1740 V = ConstantInt::get(V->getType(), 1);
1742 // If we have this scalar in the map, return it.
1743 if (WidenMap.has(V))
1744 return WidenMap.get(V);
1746 // If this scalar is unknown, assume that it is a constant or that it is
1747 // loop invariant. Broadcast V and save the value for future uses.
1748 Value *B = getBroadcastInstrs(V);
1749 return WidenMap.splat(V, B);
1752 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1753 assert(Vec->getType()->isVectorTy() && "Invalid type");
1754 SmallVector<Constant*, 8> ShuffleMask;
1755 for (unsigned i = 0; i < VF; ++i)
1756 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1758 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1759 ConstantVector::get(ShuffleMask),
1763 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1764 // Attempt to issue a wide load.
1765 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1766 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1768 assert((LI || SI) && "Invalid Load/Store instruction");
1770 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1771 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1772 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1773 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1774 // An alignment of 0 means target abi alignment. We need to use the scalar's
1775 // target abi alignment in such a case.
1777 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1778 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1779 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1780 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1782 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
1783 !Legal->setMaskedOp(SI))
1784 return scalarizeInstruction(Instr, true);
1786 if (ScalarAllocatedSize != VectorElementSize)
1787 return scalarizeInstruction(Instr);
1789 // If the pointer is loop invariant or if it is non-consecutive,
1790 // scalarize the load.
1791 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1792 bool Reverse = ConsecutiveStride < 0;
1793 bool UniformLoad = LI && Legal->isUniform(Ptr);
1794 if (!ConsecutiveStride || UniformLoad)
1795 return scalarizeInstruction(Instr);
1797 Constant *Zero = Builder.getInt32(0);
1798 VectorParts &Entry = WidenMap.get(Instr);
1800 // Handle consecutive loads/stores.
1801 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1802 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1803 setDebugLocFromInst(Builder, Gep);
1804 Value *PtrOperand = Gep->getPointerOperand();
1805 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1806 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1808 // Create the new GEP with the new induction variable.
1809 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1810 Gep2->setOperand(0, FirstBasePtr);
1811 Gep2->setName("gep.indvar.base");
1812 Ptr = Builder.Insert(Gep2);
1814 setDebugLocFromInst(Builder, Gep);
1815 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1816 OrigLoop) && "Base ptr must be invariant");
1818 // The last index does not have to be the induction. It can be
1819 // consecutive and be a function of the index. For example A[I+1];
1820 unsigned NumOperands = Gep->getNumOperands();
1821 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1822 // Create the new GEP with the new induction variable.
1823 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1825 for (unsigned i = 0; i < NumOperands; ++i) {
1826 Value *GepOperand = Gep->getOperand(i);
1827 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1829 // Update last index or loop invariant instruction anchored in loop.
1830 if (i == InductionOperand ||
1831 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1832 assert((i == InductionOperand ||
1833 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1834 "Must be last index or loop invariant");
1836 VectorParts &GEPParts = getVectorValue(GepOperand);
1837 Value *Index = GEPParts[0];
1838 Index = Builder.CreateExtractElement(Index, Zero);
1839 Gep2->setOperand(i, Index);
1840 Gep2->setName("gep.indvar.idx");
1843 Ptr = Builder.Insert(Gep2);
1845 // Use the induction element ptr.
1846 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1847 setDebugLocFromInst(Builder, Ptr);
1848 VectorParts &PtrVal = getVectorValue(Ptr);
1849 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1854 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1855 "We do not allow storing to uniform addresses");
1856 setDebugLocFromInst(Builder, SI);
1857 // We don't want to update the value in the map as it might be used in
1858 // another expression. So don't use a reference type for "StoredVal".
1859 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1861 for (unsigned Part = 0; Part < UF; ++Part) {
1862 // Calculate the pointer for the specific unroll-part.
1863 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1866 // If we store to reverse consecutive memory locations then we need
1867 // to reverse the order of elements in the stored value.
1868 StoredVal[Part] = reverseVector(StoredVal[Part]);
1869 // If the address is consecutive but reversed, then the
1870 // wide store needs to start at the last vector element.
1871 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1872 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1875 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1876 DataTy->getPointerTo(AddressSpace));
1879 if (Legal->setMaskedOp(SI)) {
1881 Builder.getInt8PtrTy(PartPtr->getType()->getPointerAddressSpace());
1883 Value *I8Ptr = Builder.CreateBitCast(PartPtr, I8PtrTy);
1885 VectorParts Cond = createEdgeMask(SI->getParent()->getSinglePredecessor(),
1887 SmallVector <Value *, 8> Ops;
1888 Ops.push_back(I8Ptr);
1889 Ops.push_back(StoredVal[Part]);
1890 Ops.push_back(Builder.getInt32(Alignment));
1891 Ops.push_back(Cond[Part]);
1892 NewSI = Builder.CreateMaskedStore(Ops);
1895 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
1896 propagateMetadata(NewSI, SI);
1902 assert(LI && "Must have a load instruction");
1903 setDebugLocFromInst(Builder, LI);
1904 for (unsigned Part = 0; Part < UF; ++Part) {
1905 // Calculate the pointer for the specific unroll-part.
1906 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1909 // If the address is consecutive but reversed, then the
1910 // wide load needs to start at the last vector element.
1911 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1912 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1916 if (Legal->setMaskedOp(LI)) {
1918 Builder.getInt8PtrTy(PartPtr->getType()->getPointerAddressSpace());
1920 Value *I8Ptr = Builder.CreateBitCast(PartPtr, I8PtrTy);
1922 VectorParts SrcMask = createBlockInMask(LI->getParent());
1923 SmallVector <Value *, 8> Ops;
1924 Ops.push_back(I8Ptr);
1925 Ops.push_back(UndefValue::get(DataTy));
1926 Ops.push_back(Builder.getInt32(Alignment));
1927 Ops.push_back(SrcMask[Part]);
1928 NewLI = Builder.CreateMaskedLoad(Ops);
1931 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1932 DataTy->getPointerTo(AddressSpace));
1933 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
1935 propagateMetadata(NewLI, LI);
1936 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
1940 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1941 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1942 // Holds vector parameters or scalars, in case of uniform vals.
1943 SmallVector<VectorParts, 4> Params;
1945 setDebugLocFromInst(Builder, Instr);
1947 // Find all of the vectorized parameters.
1948 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1949 Value *SrcOp = Instr->getOperand(op);
1951 // If we are accessing the old induction variable, use the new one.
1952 if (SrcOp == OldInduction) {
1953 Params.push_back(getVectorValue(SrcOp));
1957 // Try using previously calculated values.
1958 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1960 // If the src is an instruction that appeared earlier in the basic block
1961 // then it should already be vectorized.
1962 if (SrcInst && OrigLoop->contains(SrcInst)) {
1963 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1964 // The parameter is a vector value from earlier.
1965 Params.push_back(WidenMap.get(SrcInst));
1967 // The parameter is a scalar from outside the loop. Maybe even a constant.
1968 VectorParts Scalars;
1969 Scalars.append(UF, SrcOp);
1970 Params.push_back(Scalars);
1974 assert(Params.size() == Instr->getNumOperands() &&
1975 "Invalid number of operands");
1977 // Does this instruction return a value ?
1978 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1980 Value *UndefVec = IsVoidRetTy ? nullptr :
1981 UndefValue::get(VectorType::get(Instr->getType(), VF));
1982 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1983 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1985 Instruction *InsertPt = Builder.GetInsertPoint();
1986 BasicBlock *IfBlock = Builder.GetInsertBlock();
1987 BasicBlock *CondBlock = nullptr;
1990 Loop *VectorLp = nullptr;
1991 if (IfPredicateStore) {
1992 assert(Instr->getParent()->getSinglePredecessor() &&
1993 "Only support single predecessor blocks");
1994 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1995 Instr->getParent());
1996 VectorLp = LI->getLoopFor(IfBlock);
1997 assert(VectorLp && "Must have a loop for this block");
2000 // For each vector unroll 'part':
2001 for (unsigned Part = 0; Part < UF; ++Part) {
2002 // For each scalar that we create:
2003 for (unsigned Width = 0; Width < VF; ++Width) {
2006 Value *Cmp = nullptr;
2007 if (IfPredicateStore) {
2008 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
2009 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
2010 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
2011 LoopVectorBody.push_back(CondBlock);
2012 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
2013 // Update Builder with newly created basic block.
2014 Builder.SetInsertPoint(InsertPt);
2017 Instruction *Cloned = Instr->clone();
2019 Cloned->setName(Instr->getName() + ".cloned");
2020 // Replace the operands of the cloned instructions with extracted scalars.
2021 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2022 Value *Op = Params[op][Part];
2023 // Param is a vector. Need to extract the right lane.
2024 if (Op->getType()->isVectorTy())
2025 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
2026 Cloned->setOperand(op, Op);
2029 // Place the cloned scalar in the new loop.
2030 Builder.Insert(Cloned);
2032 // If the original scalar returns a value we need to place it in a vector
2033 // so that future users will be able to use it.
2035 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
2036 Builder.getInt32(Width));
2038 if (IfPredicateStore) {
2039 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2040 LoopVectorBody.push_back(NewIfBlock);
2041 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
2042 Builder.SetInsertPoint(InsertPt);
2043 Instruction *OldBr = IfBlock->getTerminator();
2044 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2045 OldBr->eraseFromParent();
2046 IfBlock = NewIfBlock;
2052 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
2056 if (Instruction *I = dyn_cast<Instruction>(V))
2057 return I->getParent() == Loc->getParent() ? I : nullptr;
2061 std::pair<Instruction *, Instruction *>
2062 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
2063 Instruction *tnullptr = nullptr;
2064 if (!Legal->mustCheckStrides())
2065 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2067 IRBuilder<> ChkBuilder(Loc);
2070 Value *Check = nullptr;
2071 Instruction *FirstInst = nullptr;
2072 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
2073 SE = Legal->strides_end();
2075 Value *Ptr = stripIntegerCast(*SI);
2076 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
2078 // Store the first instruction we create.
2079 FirstInst = getFirstInst(FirstInst, C, Loc);
2081 Check = ChkBuilder.CreateOr(Check, C);
2086 // We have to do this trickery because the IRBuilder might fold the check to a
2087 // constant expression in which case there is no Instruction anchored in a
2089 LLVMContext &Ctx = Loc->getContext();
2090 Instruction *TheCheck =
2091 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2092 ChkBuilder.Insert(TheCheck, "stride.not.one");
2093 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2095 return std::make_pair(FirstInst, TheCheck);
2098 std::pair<Instruction *, Instruction *>
2099 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
2100 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
2101 Legal->getRuntimePointerCheck();
2103 Instruction *tnullptr = nullptr;
2104 if (!PtrRtCheck->Need)
2105 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2107 unsigned NumPointers = PtrRtCheck->Pointers.size();
2108 SmallVector<TrackingVH<Value> , 2> Starts;
2109 SmallVector<TrackingVH<Value> , 2> Ends;
2111 LLVMContext &Ctx = Loc->getContext();
2112 SCEVExpander Exp(*SE, "induction");
2113 Instruction *FirstInst = nullptr;
2115 for (unsigned i = 0; i < NumPointers; ++i) {
2116 Value *Ptr = PtrRtCheck->Pointers[i];
2117 const SCEV *Sc = SE->getSCEV(Ptr);
2119 if (SE->isLoopInvariant(Sc, OrigLoop)) {
2120 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
2122 Starts.push_back(Ptr);
2123 Ends.push_back(Ptr);
2125 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
2126 unsigned AS = Ptr->getType()->getPointerAddressSpace();
2128 // Use this type for pointer arithmetic.
2129 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
2131 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
2132 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
2133 Starts.push_back(Start);
2134 Ends.push_back(End);
2138 IRBuilder<> ChkBuilder(Loc);
2139 // Our instructions might fold to a constant.
2140 Value *MemoryRuntimeCheck = nullptr;
2141 for (unsigned i = 0; i < NumPointers; ++i) {
2142 for (unsigned j = i+1; j < NumPointers; ++j) {
2143 // No need to check if two readonly pointers intersect.
2144 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
2147 // Only need to check pointers between two different dependency sets.
2148 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
2150 // Only need to check pointers in the same alias set.
2151 if (PtrRtCheck->AliasSetId[i] != PtrRtCheck->AliasSetId[j])
2154 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
2155 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
2157 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
2158 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
2159 "Trying to bounds check pointers with different address spaces");
2161 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
2162 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
2164 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
2165 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
2166 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
2167 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
2169 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
2170 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
2171 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
2172 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
2173 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
2174 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2175 if (MemoryRuntimeCheck) {
2176 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
2178 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2180 MemoryRuntimeCheck = IsConflict;
2184 // We have to do this trickery because the IRBuilder might fold the check to a
2185 // constant expression in which case there is no Instruction anchored in a
2187 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
2188 ConstantInt::getTrue(Ctx));
2189 ChkBuilder.Insert(Check, "memcheck.conflict");
2190 FirstInst = getFirstInst(FirstInst, Check, Loc);
2191 return std::make_pair(FirstInst, Check);
2194 void InnerLoopVectorizer::createEmptyLoop() {
2196 In this function we generate a new loop. The new loop will contain
2197 the vectorized instructions while the old loop will continue to run the
2200 [ ] <-- Back-edge taken count overflow check.
2203 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2206 || [ ] <-- vector pre header.
2210 || [ ]_| <-- vector loop.
2213 | >[ ] <--- middle-block.
2216 -|- >[ ] <--- new preheader.
2220 | [ ]_| <-- old scalar loop to handle remainder.
2223 >[ ] <-- exit block.
2227 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2228 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2229 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2230 assert(BypassBlock && "Invalid loop structure");
2231 assert(ExitBlock && "Must have an exit block");
2233 // Some loops have a single integer induction variable, while other loops
2234 // don't. One example is c++ iterators that often have multiple pointer
2235 // induction variables. In the code below we also support a case where we
2236 // don't have a single induction variable.
2237 OldInduction = Legal->getInduction();
2238 Type *IdxTy = Legal->getWidestInductionType();
2240 // Find the loop boundaries.
2241 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2242 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2244 // The exit count might have the type of i64 while the phi is i32. This can
2245 // happen if we have an induction variable that is sign extended before the
2246 // compare. The only way that we get a backedge taken count is that the
2247 // induction variable was signed and as such will not overflow. In such a case
2248 // truncation is legal.
2249 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2250 IdxTy->getPrimitiveSizeInBits())
2251 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2253 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2254 // Get the total trip count from the count by adding 1.
2255 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2256 SE->getConstant(BackedgeTakeCount->getType(), 1));
2258 // Expand the trip count and place the new instructions in the preheader.
2259 // Notice that the pre-header does not change, only the loop body.
2260 SCEVExpander Exp(*SE, "induction");
2262 // We need to test whether the backedge-taken count is uint##_max. Adding one
2263 // to it will cause overflow and an incorrect loop trip count in the vector
2264 // body. In case of overflow we want to directly jump to the scalar remainder
2266 Value *BackedgeCount =
2267 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2268 BypassBlock->getTerminator());
2269 if (BackedgeCount->getType()->isPointerTy())
2270 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2271 "backedge.ptrcnt.to.int",
2272 BypassBlock->getTerminator());
2273 Instruction *CheckBCOverflow =
2274 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2275 Constant::getAllOnesValue(BackedgeCount->getType()),
2276 "backedge.overflow", BypassBlock->getTerminator());
2278 // The loop index does not have to start at Zero. Find the original start
2279 // value from the induction PHI node. If we don't have an induction variable
2280 // then we know that it starts at zero.
2281 Builder.SetInsertPoint(BypassBlock->getTerminator());
2282 Value *StartIdx = ExtendedIdx = OldInduction ?
2283 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2285 ConstantInt::get(IdxTy, 0);
2287 // We need an instruction to anchor the overflow check on. StartIdx needs to
2288 // be defined before the overflow check branch. Because the scalar preheader
2289 // is going to merge the start index and so the overflow branch block needs to
2290 // contain a definition of the start index.
2291 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2292 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2293 BypassBlock->getTerminator());
2295 // Count holds the overall loop count (N).
2296 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2297 BypassBlock->getTerminator());
2299 LoopBypassBlocks.push_back(BypassBlock);
2301 // Split the single block loop into the two loop structure described above.
2302 BasicBlock *VectorPH =
2303 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2304 BasicBlock *VecBody =
2305 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2306 BasicBlock *MiddleBlock =
2307 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2308 BasicBlock *ScalarPH =
2309 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2311 // Create and register the new vector loop.
2312 Loop* Lp = new Loop();
2313 Loop *ParentLoop = OrigLoop->getParentLoop();
2315 // Insert the new loop into the loop nest and register the new basic blocks
2316 // before calling any utilities such as SCEV that require valid LoopInfo.
2318 ParentLoop->addChildLoop(Lp);
2319 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
2320 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
2321 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
2323 LI->addTopLevelLoop(Lp);
2325 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
2327 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2329 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2331 // Generate the induction variable.
2332 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2333 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2334 // The loop step is equal to the vectorization factor (num of SIMD elements)
2335 // times the unroll factor (num of SIMD instructions).
2336 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2338 // This is the IR builder that we use to add all of the logic for bypassing
2339 // the new vector loop.
2340 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2341 setDebugLocFromInst(BypassBuilder,
2342 getDebugLocFromInstOrOperands(OldInduction));
2344 // We may need to extend the index in case there is a type mismatch.
2345 // We know that the count starts at zero and does not overflow.
2346 if (Count->getType() != IdxTy) {
2347 // The exit count can be of pointer type. Convert it to the correct
2349 if (ExitCount->getType()->isPointerTy())
2350 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2352 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2355 // Add the start index to the loop count to get the new end index.
2356 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2358 // Now we need to generate the expression for N - (N % VF), which is
2359 // the part that the vectorized body will execute.
2360 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2361 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2362 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2363 "end.idx.rnd.down");
2365 // Now, compare the new count to zero. If it is zero skip the vector loop and
2366 // jump to the scalar loop.
2368 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2370 BasicBlock *LastBypassBlock = BypassBlock;
2372 // Generate code to check that the loops trip count that we computed by adding
2373 // one to the backedge-taken count will not overflow.
2375 auto PastOverflowCheck =
2376 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2377 BasicBlock *CheckBlock =
2378 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2380 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2381 LoopBypassBlocks.push_back(CheckBlock);
2382 Instruction *OldTerm = LastBypassBlock->getTerminator();
2383 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2384 OldTerm->eraseFromParent();
2385 LastBypassBlock = CheckBlock;
2388 // Generate the code to check that the strides we assumed to be one are really
2389 // one. We want the new basic block to start at the first instruction in a
2390 // sequence of instructions that form a check.
2391 Instruction *StrideCheck;
2392 Instruction *FirstCheckInst;
2393 std::tie(FirstCheckInst, StrideCheck) =
2394 addStrideCheck(LastBypassBlock->getTerminator());
2396 // Create a new block containing the stride check.
2397 BasicBlock *CheckBlock =
2398 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2400 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2401 LoopBypassBlocks.push_back(CheckBlock);
2403 // Replace the branch into the memory check block with a conditional branch
2404 // for the "few elements case".
2405 Instruction *OldTerm = LastBypassBlock->getTerminator();
2406 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2407 OldTerm->eraseFromParent();
2410 LastBypassBlock = CheckBlock;
2413 // Generate the code that checks in runtime if arrays overlap. We put the
2414 // checks into a separate block to make the more common case of few elements
2416 Instruction *MemRuntimeCheck;
2417 std::tie(FirstCheckInst, MemRuntimeCheck) =
2418 addRuntimeCheck(LastBypassBlock->getTerminator());
2419 if (MemRuntimeCheck) {
2420 // Create a new block containing the memory check.
2421 BasicBlock *CheckBlock =
2422 LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
2424 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2425 LoopBypassBlocks.push_back(CheckBlock);
2427 // Replace the branch into the memory check block with a conditional branch
2428 // for the "few elements case".
2429 Instruction *OldTerm = LastBypassBlock->getTerminator();
2430 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2431 OldTerm->eraseFromParent();
2433 Cmp = MemRuntimeCheck;
2434 LastBypassBlock = CheckBlock;
2437 LastBypassBlock->getTerminator()->eraseFromParent();
2438 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2441 // We are going to resume the execution of the scalar loop.
2442 // Go over all of the induction variables that we found and fix the
2443 // PHIs that are left in the scalar version of the loop.
2444 // The starting values of PHI nodes depend on the counter of the last
2445 // iteration in the vectorized loop.
2446 // If we come from a bypass edge then we need to start from the original
2449 // This variable saves the new starting index for the scalar loop.
2450 PHINode *ResumeIndex = nullptr;
2451 LoopVectorizationLegality::InductionList::iterator I, E;
2452 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2453 // Set builder to point to last bypass block.
2454 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2455 for (I = List->begin(), E = List->end(); I != E; ++I) {
2456 PHINode *OrigPhi = I->first;
2457 LoopVectorizationLegality::InductionInfo II = I->second;
2459 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2460 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2461 MiddleBlock->getTerminator());
2462 // We might have extended the type of the induction variable but we need a
2463 // truncated version for the scalar loop.
2464 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2465 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2466 MiddleBlock->getTerminator()) : nullptr;
2468 // Create phi nodes to merge from the backedge-taken check block.
2469 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2470 ScalarPH->getTerminator());
2471 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2473 PHINode *BCTruncResumeVal = nullptr;
2474 if (OrigPhi == OldInduction) {
2476 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2477 ScalarPH->getTerminator());
2478 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2481 Value *EndValue = nullptr;
2483 case LoopVectorizationLegality::IK_NoInduction:
2484 llvm_unreachable("Unknown induction");
2485 case LoopVectorizationLegality::IK_IntInduction: {
2486 // Handle the integer induction counter.
2487 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2489 // We have the canonical induction variable.
2490 if (OrigPhi == OldInduction) {
2491 // Create a truncated version of the resume value for the scalar loop,
2492 // we might have promoted the type to a larger width.
2494 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2495 // The new PHI merges the original incoming value, in case of a bypass,
2496 // or the value at the end of the vectorized loop.
2497 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2498 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2499 TruncResumeVal->addIncoming(EndValue, VecBody);
2501 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2503 // We know what the end value is.
2504 EndValue = IdxEndRoundDown;
2505 // We also know which PHI node holds it.
2506 ResumeIndex = ResumeVal;
2510 // Not the canonical induction variable - add the vector loop count to the
2512 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2513 II.StartValue->getType(),
2515 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
2518 case LoopVectorizationLegality::IK_ReverseIntInduction: {
2519 // Convert the CountRoundDown variable to the PHI size.
2520 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2521 II.StartValue->getType(),
2523 // Handle reverse integer induction counter.
2524 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
2527 case LoopVectorizationLegality::IK_PtrInduction: {
2528 // For pointer induction variables, calculate the offset using
2530 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
2534 case LoopVectorizationLegality::IK_ReversePtrInduction: {
2535 // The value at the end of the loop for the reverse pointer is calculated
2536 // by creating a GEP with a negative index starting from the start value.
2537 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
2538 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
2540 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
2546 // The new PHI merges the original incoming value, in case of a bypass,
2547 // or the value at the end of the vectorized loop.
2548 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2549 if (OrigPhi == OldInduction)
2550 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2552 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2554 ResumeVal->addIncoming(EndValue, VecBody);
2556 // Fix the scalar body counter (PHI node).
2557 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2559 // The old induction's phi node in the scalar body needs the truncated
2561 if (OrigPhi == OldInduction) {
2562 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2563 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2565 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2566 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2570 // If we are generating a new induction variable then we also need to
2571 // generate the code that calculates the exit value. This value is not
2572 // simply the end of the counter because we may skip the vectorized body
2573 // in case of a runtime check.
2575 assert(!ResumeIndex && "Unexpected resume value found");
2576 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2577 MiddleBlock->getTerminator());
2578 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2579 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2580 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2583 // Make sure that we found the index where scalar loop needs to continue.
2584 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2585 "Invalid resume Index");
2587 // Add a check in the middle block to see if we have completed
2588 // all of the iterations in the first vector loop.
2589 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2590 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2591 ResumeIndex, "cmp.n",
2592 MiddleBlock->getTerminator());
2594 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2595 // Remove the old terminator.
2596 MiddleBlock->getTerminator()->eraseFromParent();
2598 // Create i+1 and fill the PHINode.
2599 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2600 Induction->addIncoming(StartIdx, VectorPH);
2601 Induction->addIncoming(NextIdx, VecBody);
2602 // Create the compare.
2603 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2604 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2606 // Now we have two terminators. Remove the old one from the block.
2607 VecBody->getTerminator()->eraseFromParent();
2609 // Get ready to start creating new instructions into the vectorized body.
2610 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2613 LoopVectorPreHeader = VectorPH;
2614 LoopScalarPreHeader = ScalarPH;
2615 LoopMiddleBlock = MiddleBlock;
2616 LoopExitBlock = ExitBlock;
2617 LoopVectorBody.push_back(VecBody);
2618 LoopScalarBody = OldBasicBlock;
2620 LoopVectorizeHints Hints(Lp, true);
2621 Hints.setAlreadyVectorized();
2624 /// This function returns the identity element (or neutral element) for
2625 /// the operation K.
2627 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2632 // Adding, Xoring, Oring zero to a number does not change it.
2633 return ConstantInt::get(Tp, 0);
2634 case RK_IntegerMult:
2635 // Multiplying a number by 1 does not change it.
2636 return ConstantInt::get(Tp, 1);
2638 // AND-ing a number with an all-1 value does not change it.
2639 return ConstantInt::get(Tp, -1, true);
2641 // Multiplying a number by 1 does not change it.
2642 return ConstantFP::get(Tp, 1.0L);
2644 // Adding zero to a number does not change it.
2645 return ConstantFP::get(Tp, 0.0L);
2647 llvm_unreachable("Unknown reduction kind");
2651 /// This function translates the reduction kind to an LLVM binary operator.
2653 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2655 case LoopVectorizationLegality::RK_IntegerAdd:
2656 return Instruction::Add;
2657 case LoopVectorizationLegality::RK_IntegerMult:
2658 return Instruction::Mul;
2659 case LoopVectorizationLegality::RK_IntegerOr:
2660 return Instruction::Or;
2661 case LoopVectorizationLegality::RK_IntegerAnd:
2662 return Instruction::And;
2663 case LoopVectorizationLegality::RK_IntegerXor:
2664 return Instruction::Xor;
2665 case LoopVectorizationLegality::RK_FloatMult:
2666 return Instruction::FMul;
2667 case LoopVectorizationLegality::RK_FloatAdd:
2668 return Instruction::FAdd;
2669 case LoopVectorizationLegality::RK_IntegerMinMax:
2670 return Instruction::ICmp;
2671 case LoopVectorizationLegality::RK_FloatMinMax:
2672 return Instruction::FCmp;
2674 llvm_unreachable("Unknown reduction operation");
2678 Value *createMinMaxOp(IRBuilder<> &Builder,
2679 LoopVectorizationLegality::MinMaxReductionKind RK,
2682 CmpInst::Predicate P = CmpInst::ICMP_NE;
2685 llvm_unreachable("Unknown min/max reduction kind");
2686 case LoopVectorizationLegality::MRK_UIntMin:
2687 P = CmpInst::ICMP_ULT;
2689 case LoopVectorizationLegality::MRK_UIntMax:
2690 P = CmpInst::ICMP_UGT;
2692 case LoopVectorizationLegality::MRK_SIntMin:
2693 P = CmpInst::ICMP_SLT;
2695 case LoopVectorizationLegality::MRK_SIntMax:
2696 P = CmpInst::ICMP_SGT;
2698 case LoopVectorizationLegality::MRK_FloatMin:
2699 P = CmpInst::FCMP_OLT;
2701 case LoopVectorizationLegality::MRK_FloatMax:
2702 P = CmpInst::FCMP_OGT;
2707 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2708 RK == LoopVectorizationLegality::MRK_FloatMax)
2709 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2711 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2713 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2718 struct CSEDenseMapInfo {
2719 static bool canHandle(Instruction *I) {
2720 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2721 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2723 static inline Instruction *getEmptyKey() {
2724 return DenseMapInfo<Instruction *>::getEmptyKey();
2726 static inline Instruction *getTombstoneKey() {
2727 return DenseMapInfo<Instruction *>::getTombstoneKey();
2729 static unsigned getHashValue(Instruction *I) {
2730 assert(canHandle(I) && "Unknown instruction!");
2731 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2732 I->value_op_end()));
2734 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2735 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2736 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2738 return LHS->isIdenticalTo(RHS);
2743 /// \brief Check whether this block is a predicated block.
2744 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2745 /// = ...; " blocks. We start with one vectorized basic block. For every
2746 /// conditional block we split this vectorized block. Therefore, every second
2747 /// block will be a predicated one.
2748 static bool isPredicatedBlock(unsigned BlockNum) {
2749 return BlockNum % 2;
2752 ///\brief Perform cse of induction variable instructions.
2753 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2754 // Perform simple cse.
2755 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2756 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2757 BasicBlock *BB = BBs[i];
2758 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2759 Instruction *In = I++;
2761 if (!CSEDenseMapInfo::canHandle(In))
2764 // Check if we can replace this instruction with any of the
2765 // visited instructions.
2766 if (Instruction *V = CSEMap.lookup(In)) {
2767 In->replaceAllUsesWith(V);
2768 In->eraseFromParent();
2771 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2772 // ...;" blocks for predicated stores. Every second block is a predicated
2774 if (isPredicatedBlock(i))
2782 /// \brief Adds a 'fast' flag to floating point operations.
2783 static Value *addFastMathFlag(Value *V) {
2784 if (isa<FPMathOperator>(V)){
2785 FastMathFlags Flags;
2786 Flags.setUnsafeAlgebra();
2787 cast<Instruction>(V)->setFastMathFlags(Flags);
2792 void InnerLoopVectorizer::vectorizeLoop() {
2793 //===------------------------------------------------===//
2795 // Notice: any optimization or new instruction that go
2796 // into the code below should be also be implemented in
2799 //===------------------------------------------------===//
2800 Constant *Zero = Builder.getInt32(0);
2802 // In order to support reduction variables we need to be able to vectorize
2803 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2804 // stages. First, we create a new vector PHI node with no incoming edges.
2805 // We use this value when we vectorize all of the instructions that use the
2806 // PHI. Next, after all of the instructions in the block are complete we
2807 // add the new incoming edges to the PHI. At this point all of the
2808 // instructions in the basic block are vectorized, so we can use them to
2809 // construct the PHI.
2810 PhiVector RdxPHIsToFix;
2812 // Scan the loop in a topological order to ensure that defs are vectorized
2814 LoopBlocksDFS DFS(OrigLoop);
2817 // Vectorize all of the blocks in the original loop.
2818 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2819 be = DFS.endRPO(); bb != be; ++bb)
2820 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2822 // At this point every instruction in the original loop is widened to
2823 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2824 // that we vectorized. The PHI nodes are currently empty because we did
2825 // not want to introduce cycles. Notice that the remaining PHI nodes
2826 // that we need to fix are reduction variables.
2828 // Create the 'reduced' values for each of the induction vars.
2829 // The reduced values are the vector values that we scalarize and combine
2830 // after the loop is finished.
2831 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2833 PHINode *RdxPhi = *it;
2834 assert(RdxPhi && "Unable to recover vectorized PHI");
2836 // Find the reduction variable descriptor.
2837 assert(Legal->getReductionVars()->count(RdxPhi) &&
2838 "Unable to find the reduction variable");
2839 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2840 (*Legal->getReductionVars())[RdxPhi];
2842 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2844 // We need to generate a reduction vector from the incoming scalar.
2845 // To do so, we need to generate the 'identity' vector and override
2846 // one of the elements with the incoming scalar reduction. We need
2847 // to do it in the vector-loop preheader.
2848 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
2850 // This is the vector-clone of the value that leaves the loop.
2851 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2852 Type *VecTy = VectorExit[0]->getType();
2854 // Find the reduction identity variable. Zero for addition, or, xor,
2855 // one for multiplication, -1 for And.
2858 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2859 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2860 // MinMax reduction have the start value as their identify.
2862 VectorStart = Identity = RdxDesc.StartValue;
2864 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2869 // Handle other reduction kinds:
2871 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2872 VecTy->getScalarType());
2875 // This vector is the Identity vector where the first element is the
2876 // incoming scalar reduction.
2877 VectorStart = RdxDesc.StartValue;
2879 Identity = ConstantVector::getSplat(VF, Iden);
2881 // This vector is the Identity vector where the first element is the
2882 // incoming scalar reduction.
2883 VectorStart = Builder.CreateInsertElement(Identity,
2884 RdxDesc.StartValue, Zero);
2888 // Fix the vector-loop phi.
2889 // We created the induction variable so we know that the
2890 // preheader is the first entry.
2891 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2893 // Reductions do not have to start at zero. They can start with
2894 // any loop invariant values.
2895 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2896 BasicBlock *Latch = OrigLoop->getLoopLatch();
2897 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2898 VectorParts &Val = getVectorValue(LoopVal);
2899 for (unsigned part = 0; part < UF; ++part) {
2900 // Make sure to add the reduction stat value only to the
2901 // first unroll part.
2902 Value *StartVal = (part == 0) ? VectorStart : Identity;
2903 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2904 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2905 LoopVectorBody.back());
2908 // Before each round, move the insertion point right between
2909 // the PHIs and the values we are going to write.
2910 // This allows us to write both PHINodes and the extractelement
2912 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2914 VectorParts RdxParts;
2915 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2916 for (unsigned part = 0; part < UF; ++part) {
2917 // This PHINode contains the vectorized reduction variable, or
2918 // the initial value vector, if we bypass the vector loop.
2919 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2920 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2921 Value *StartVal = (part == 0) ? VectorStart : Identity;
2922 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2923 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2924 NewPhi->addIncoming(RdxExitVal[part],
2925 LoopVectorBody.back());
2926 RdxParts.push_back(NewPhi);
2929 // Reduce all of the unrolled parts into a single vector.
2930 Value *ReducedPartRdx = RdxParts[0];
2931 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2932 setDebugLocFromInst(Builder, ReducedPartRdx);
2933 for (unsigned part = 1; part < UF; ++part) {
2934 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2935 // Floating point operations had to be 'fast' to enable the reduction.
2936 ReducedPartRdx = addFastMathFlag(
2937 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2938 ReducedPartRdx, "bin.rdx"));
2940 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2941 ReducedPartRdx, RdxParts[part]);
2945 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2946 // and vector ops, reducing the set of values being computed by half each
2948 assert(isPowerOf2_32(VF) &&
2949 "Reduction emission only supported for pow2 vectors!");
2950 Value *TmpVec = ReducedPartRdx;
2951 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2952 for (unsigned i = VF; i != 1; i >>= 1) {
2953 // Move the upper half of the vector to the lower half.
2954 for (unsigned j = 0; j != i/2; ++j)
2955 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2957 // Fill the rest of the mask with undef.
2958 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2959 UndefValue::get(Builder.getInt32Ty()));
2962 Builder.CreateShuffleVector(TmpVec,
2963 UndefValue::get(TmpVec->getType()),
2964 ConstantVector::get(ShuffleMask),
2967 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2968 // Floating point operations had to be 'fast' to enable the reduction.
2969 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2970 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2972 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2975 // The result is in the first element of the vector.
2976 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2977 Builder.getInt32(0));
2980 // Create a phi node that merges control-flow from the backedge-taken check
2981 // block and the middle block.
2982 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
2983 LoopScalarPreHeader->getTerminator());
2984 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
2985 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2987 // Now, we need to fix the users of the reduction variable
2988 // inside and outside of the scalar remainder loop.
2989 // We know that the loop is in LCSSA form. We need to update the
2990 // PHI nodes in the exit blocks.
2991 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2992 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2993 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2994 if (!LCSSAPhi) break;
2996 // All PHINodes need to have a single entry edge, or two if
2997 // we already fixed them.
2998 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
3000 // We found our reduction value exit-PHI. Update it with the
3001 // incoming bypass edge.
3002 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
3003 // Add an edge coming from the bypass.
3004 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
3007 }// end of the LCSSA phi scan.
3009 // Fix the scalar loop reduction variable with the incoming reduction sum
3010 // from the vector body and from the backedge value.
3011 int IncomingEdgeBlockIdx =
3012 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
3013 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
3014 // Pick the other block.
3015 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
3016 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
3017 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
3018 }// end of for each redux variable.
3022 // Remove redundant induction instructions.
3023 cse(LoopVectorBody);
3026 void InnerLoopVectorizer::fixLCSSAPHIs() {
3027 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3028 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3029 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3030 if (!LCSSAPhi) break;
3031 if (LCSSAPhi->getNumIncomingValues() == 1)
3032 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
3037 InnerLoopVectorizer::VectorParts
3038 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
3039 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
3042 // Look for cached value.
3043 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
3044 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
3045 if (ECEntryIt != MaskCache.end())
3046 return ECEntryIt->second;
3048 VectorParts SrcMask = createBlockInMask(Src);
3050 // The terminator has to be a branch inst!
3051 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
3052 assert(BI && "Unexpected terminator found");
3054 if (BI->isConditional()) {
3055 VectorParts EdgeMask = getVectorValue(BI->getCondition());
3057 if (BI->getSuccessor(0) != Dst)
3058 for (unsigned part = 0; part < UF; ++part)
3059 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
3061 for (unsigned part = 0; part < UF; ++part)
3062 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
3064 MaskCache[Edge] = EdgeMask;
3068 MaskCache[Edge] = SrcMask;
3072 InnerLoopVectorizer::VectorParts
3073 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
3074 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
3076 // Loop incoming mask is all-one.
3077 if (OrigLoop->getHeader() == BB) {
3078 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
3079 return getVectorValue(C);
3082 // This is the block mask. We OR all incoming edges, and with zero.
3083 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
3084 VectorParts BlockMask = getVectorValue(Zero);
3087 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
3088 VectorParts EM = createEdgeMask(*it, BB);
3089 for (unsigned part = 0; part < UF; ++part)
3090 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
3096 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
3097 InnerLoopVectorizer::VectorParts &Entry,
3098 unsigned UF, unsigned VF, PhiVector *PV) {
3099 PHINode* P = cast<PHINode>(PN);
3100 // Handle reduction variables:
3101 if (Legal->getReductionVars()->count(P)) {
3102 for (unsigned part = 0; part < UF; ++part) {
3103 // This is phase one of vectorizing PHIs.
3104 Type *VecTy = (VF == 1) ? PN->getType() :
3105 VectorType::get(PN->getType(), VF);
3106 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
3107 LoopVectorBody.back()-> getFirstInsertionPt());
3113 setDebugLocFromInst(Builder, P);
3114 // Check for PHI nodes that are lowered to vector selects.
3115 if (P->getParent() != OrigLoop->getHeader()) {
3116 // We know that all PHIs in non-header blocks are converted into
3117 // selects, so we don't have to worry about the insertion order and we
3118 // can just use the builder.
3119 // At this point we generate the predication tree. There may be
3120 // duplications since this is a simple recursive scan, but future
3121 // optimizations will clean it up.
3123 unsigned NumIncoming = P->getNumIncomingValues();
3125 // Generate a sequence of selects of the form:
3126 // SELECT(Mask3, In3,
3127 // SELECT(Mask2, In2,
3129 for (unsigned In = 0; In < NumIncoming; In++) {
3130 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
3132 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3134 for (unsigned part = 0; part < UF; ++part) {
3135 // We might have single edge PHIs (blocks) - use an identity
3136 // 'select' for the first PHI operand.
3138 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3141 // Select between the current value and the previous incoming edge
3142 // based on the incoming mask.
3143 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3144 Entry[part], "predphi");
3150 // This PHINode must be an induction variable.
3151 // Make sure that we know about it.
3152 assert(Legal->getInductionVars()->count(P) &&
3153 "Not an induction variable");
3155 LoopVectorizationLegality::InductionInfo II =
3156 Legal->getInductionVars()->lookup(P);
3159 case LoopVectorizationLegality::IK_NoInduction:
3160 llvm_unreachable("Unknown induction");
3161 case LoopVectorizationLegality::IK_IntInduction: {
3162 assert(P->getType() == II.StartValue->getType() && "Types must match");
3163 Type *PhiTy = P->getType();
3165 if (P == OldInduction) {
3166 // Handle the canonical induction variable. We might have had to
3168 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
3170 // Handle other induction variables that are now based on the
3172 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
3174 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
3175 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
3178 Broadcasted = getBroadcastInstrs(Broadcasted);
3179 // After broadcasting the induction variable we need to make the vector
3180 // consecutive by adding 0, 1, 2, etc.
3181 for (unsigned part = 0; part < UF; ++part)
3182 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
3185 case LoopVectorizationLegality::IK_ReverseIntInduction:
3186 case LoopVectorizationLegality::IK_PtrInduction:
3187 case LoopVectorizationLegality::IK_ReversePtrInduction:
3188 // Handle reverse integer and pointer inductions.
3189 Value *StartIdx = ExtendedIdx;
3190 // This is the normalized GEP that starts counting at zero.
3191 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
3194 // Handle the reverse integer induction variable case.
3195 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
3196 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
3197 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
3199 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
3202 // This is a new value so do not hoist it out.
3203 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
3204 // After broadcasting the induction variable we need to make the
3205 // vector consecutive by adding ... -3, -2, -1, 0.
3206 for (unsigned part = 0; part < UF; ++part)
3207 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
3212 // Handle the pointer induction variable case.
3213 assert(P->getType()->isPointerTy() && "Unexpected type.");
3215 // Is this a reverse induction ptr or a consecutive induction ptr.
3216 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
3219 // This is the vector of results. Notice that we don't generate
3220 // vector geps because scalar geps result in better code.
3221 for (unsigned part = 0; part < UF; ++part) {
3223 int EltIndex = (part) * (Reverse ? -1 : 1);
3224 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3227 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3229 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3231 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3233 Entry[part] = SclrGep;
3237 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3238 for (unsigned int i = 0; i < VF; ++i) {
3239 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
3240 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3243 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3245 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3247 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3249 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3250 Builder.getInt32(i),
3253 Entry[part] = VecVal;
3259 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3260 // For each instruction in the old loop.
3261 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3262 VectorParts &Entry = WidenMap.get(it);
3263 switch (it->getOpcode()) {
3264 case Instruction::Br:
3265 // Nothing to do for PHIs and BR, since we already took care of the
3266 // loop control flow instructions.
3268 case Instruction::PHI:{
3269 // Vectorize PHINodes.
3270 widenPHIInstruction(it, Entry, UF, VF, PV);
3274 case Instruction::Add:
3275 case Instruction::FAdd:
3276 case Instruction::Sub:
3277 case Instruction::FSub:
3278 case Instruction::Mul:
3279 case Instruction::FMul:
3280 case Instruction::UDiv:
3281 case Instruction::SDiv:
3282 case Instruction::FDiv:
3283 case Instruction::URem:
3284 case Instruction::SRem:
3285 case Instruction::FRem:
3286 case Instruction::Shl:
3287 case Instruction::LShr:
3288 case Instruction::AShr:
3289 case Instruction::And:
3290 case Instruction::Or:
3291 case Instruction::Xor: {
3292 // Just widen binops.
3293 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3294 setDebugLocFromInst(Builder, BinOp);
3295 VectorParts &A = getVectorValue(it->getOperand(0));
3296 VectorParts &B = getVectorValue(it->getOperand(1));
3298 // Use this vector value for all users of the original instruction.
3299 for (unsigned Part = 0; Part < UF; ++Part) {
3300 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3302 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3303 VecOp->copyIRFlags(BinOp);
3308 propagateMetadata(Entry, it);
3311 case Instruction::Select: {
3313 // If the selector is loop invariant we can create a select
3314 // instruction with a scalar condition. Otherwise, use vector-select.
3315 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3317 setDebugLocFromInst(Builder, it);
3319 // The condition can be loop invariant but still defined inside the
3320 // loop. This means that we can't just use the original 'cond' value.
3321 // We have to take the 'vectorized' value and pick the first lane.
3322 // Instcombine will make this a no-op.
3323 VectorParts &Cond = getVectorValue(it->getOperand(0));
3324 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3325 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3327 Value *ScalarCond = (VF == 1) ? Cond[0] :
3328 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3330 for (unsigned Part = 0; Part < UF; ++Part) {
3331 Entry[Part] = Builder.CreateSelect(
3332 InvariantCond ? ScalarCond : Cond[Part],
3337 propagateMetadata(Entry, it);
3341 case Instruction::ICmp:
3342 case Instruction::FCmp: {
3343 // Widen compares. Generate vector compares.
3344 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3345 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3346 setDebugLocFromInst(Builder, it);
3347 VectorParts &A = getVectorValue(it->getOperand(0));
3348 VectorParts &B = getVectorValue(it->getOperand(1));
3349 for (unsigned Part = 0; Part < UF; ++Part) {
3352 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3354 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3358 propagateMetadata(Entry, it);
3362 case Instruction::Store:
3363 case Instruction::Load:
3364 vectorizeMemoryInstruction(it);
3366 case Instruction::ZExt:
3367 case Instruction::SExt:
3368 case Instruction::FPToUI:
3369 case Instruction::FPToSI:
3370 case Instruction::FPExt:
3371 case Instruction::PtrToInt:
3372 case Instruction::IntToPtr:
3373 case Instruction::SIToFP:
3374 case Instruction::UIToFP:
3375 case Instruction::Trunc:
3376 case Instruction::FPTrunc:
3377 case Instruction::BitCast: {
3378 CastInst *CI = dyn_cast<CastInst>(it);
3379 setDebugLocFromInst(Builder, it);
3380 /// Optimize the special case where the source is the induction
3381 /// variable. Notice that we can only optimize the 'trunc' case
3382 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3383 /// c. other casts depend on pointer size.
3384 if (CI->getOperand(0) == OldInduction &&
3385 it->getOpcode() == Instruction::Trunc) {
3386 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3388 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3389 for (unsigned Part = 0; Part < UF; ++Part)
3390 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
3391 propagateMetadata(Entry, it);
3394 /// Vectorize casts.
3395 Type *DestTy = (VF == 1) ? CI->getType() :
3396 VectorType::get(CI->getType(), VF);
3398 VectorParts &A = getVectorValue(it->getOperand(0));
3399 for (unsigned Part = 0; Part < UF; ++Part)
3400 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3401 propagateMetadata(Entry, it);
3405 case Instruction::Call: {
3406 // Ignore dbg intrinsics.
3407 if (isa<DbgInfoIntrinsic>(it))
3409 setDebugLocFromInst(Builder, it);
3411 Module *M = BB->getParent()->getParent();
3412 CallInst *CI = cast<CallInst>(it);
3413 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3414 assert(ID && "Not an intrinsic call!");
3416 case Intrinsic::assume:
3417 case Intrinsic::lifetime_end:
3418 case Intrinsic::lifetime_start:
3419 scalarizeInstruction(it);
3422 bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
3423 for (unsigned Part = 0; Part < UF; ++Part) {
3424 SmallVector<Value *, 4> Args;
3425 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3426 if (HasScalarOpd && i == 1) {
3427 Args.push_back(CI->getArgOperand(i));
3430 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3431 Args.push_back(Arg[Part]);
3433 Type *Tys[] = {CI->getType()};
3435 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3437 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3438 Entry[Part] = Builder.CreateCall(F, Args);
3441 propagateMetadata(Entry, it);
3448 // All other instructions are unsupported. Scalarize them.
3449 scalarizeInstruction(it);
3452 }// end of for_each instr.
3455 void InnerLoopVectorizer::updateAnalysis() {
3456 // Forget the original basic block.
3457 SE->forgetLoop(OrigLoop);
3459 // Update the dominator tree information.
3460 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3461 "Entry does not dominate exit.");
3463 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3464 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3465 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3467 // Due to if predication of stores we might create a sequence of "if(pred)
3468 // a[i] = ...; " blocks.
3469 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3471 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3472 else if (isPredicatedBlock(i)) {
3473 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3475 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3479 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3480 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3481 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3482 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3484 DEBUG(DT->verifyDomTree());
3487 /// \brief Check whether it is safe to if-convert this phi node.
3489 /// Phi nodes with constant expressions that can trap are not safe to if
3491 static bool canIfConvertPHINodes(BasicBlock *BB) {
3492 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3493 PHINode *Phi = dyn_cast<PHINode>(I);
3496 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3497 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3504 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3505 if (!EnableIfConversion) {
3506 emitAnalysis(Report() << "if-conversion is disabled");
3510 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3512 // A list of pointers that we can safely read and write to.
3513 SmallPtrSet<Value *, 8> SafePointes;
3515 // Collect safe addresses.
3516 for (Loop::block_iterator BI = TheLoop->block_begin(),
3517 BE = TheLoop->block_end(); BI != BE; ++BI) {
3518 BasicBlock *BB = *BI;
3520 if (blockNeedsPredication(BB))
3523 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3524 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3525 SafePointes.insert(LI->getPointerOperand());
3526 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3527 SafePointes.insert(SI->getPointerOperand());
3531 // Collect the blocks that need predication.
3532 BasicBlock *Header = TheLoop->getHeader();
3533 for (Loop::block_iterator BI = TheLoop->block_begin(),
3534 BE = TheLoop->block_end(); BI != BE; ++BI) {
3535 BasicBlock *BB = *BI;
3537 // We don't support switch statements inside loops.
3538 if (!isa<BranchInst>(BB->getTerminator())) {
3539 emitAnalysis(Report(BB->getTerminator())
3540 << "loop contains a switch statement");
3544 // We must be able to predicate all blocks that need to be predicated.
3545 if (blockNeedsPredication(BB)) {
3546 if (!blockCanBePredicated(BB, SafePointes)) {
3547 emitAnalysis(Report(BB->getTerminator())
3548 << "control flow cannot be substituted for a select");
3551 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3552 emitAnalysis(Report(BB->getTerminator())
3553 << "control flow cannot be substituted for a select");
3558 // We can if-convert this loop.
3562 bool LoopVectorizationLegality::canVectorize() {
3563 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3564 // be canonicalized.
3565 if (!TheLoop->getLoopPreheader()) {
3567 Report() << "loop control flow is not understood by vectorizer");
3571 // We can only vectorize innermost loops.
3572 if (TheLoop->getSubLoopsVector().size()) {
3573 emitAnalysis(Report() << "loop is not the innermost loop");
3577 // We must have a single backedge.
3578 if (TheLoop->getNumBackEdges() != 1) {
3580 Report() << "loop control flow is not understood by vectorizer");
3584 // We must have a single exiting block.
3585 if (!TheLoop->getExitingBlock()) {
3587 Report() << "loop control flow is not understood by vectorizer");
3591 // We need to have a loop header.
3592 DEBUG(dbgs() << "LV: Found a loop: " <<
3593 TheLoop->getHeader()->getName() << '\n');
3595 // Check if we can if-convert non-single-bb loops.
3596 unsigned NumBlocks = TheLoop->getNumBlocks();
3597 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3598 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3602 // ScalarEvolution needs to be able to find the exit count.
3603 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3604 if (ExitCount == SE->getCouldNotCompute()) {
3605 emitAnalysis(Report() << "could not determine number of loop iterations");
3606 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3610 // Check if we can vectorize the instructions and CFG in this loop.
3611 if (!canVectorizeInstrs()) {
3612 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3616 // Go over each instruction and look at memory deps.
3617 if (!canVectorizeMemory()) {
3618 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3622 // Collect all of the variables that remain uniform after vectorization.
3623 collectLoopUniforms();
3625 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3626 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
3629 // Okay! We can vectorize. At this point we don't have any other mem analysis
3630 // which may limit our maximum vectorization factor, so just return true with
3635 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3636 if (Ty->isPointerTy())
3637 return DL.getIntPtrType(Ty);
3639 // It is possible that char's or short's overflow when we ask for the loop's
3640 // trip count, work around this by changing the type size.
3641 if (Ty->getScalarSizeInBits() < 32)
3642 return Type::getInt32Ty(Ty->getContext());
3647 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3648 Ty0 = convertPointerToIntegerType(DL, Ty0);
3649 Ty1 = convertPointerToIntegerType(DL, Ty1);
3650 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3655 /// \brief Check that the instruction has outside loop users and is not an
3656 /// identified reduction variable.
3657 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3658 SmallPtrSetImpl<Value *> &Reductions) {
3659 // Reduction instructions are allowed to have exit users. All other
3660 // instructions must not have external users.
3661 if (!Reductions.count(Inst))
3662 //Check that all of the users of the loop are inside the BB.
3663 for (User *U : Inst->users()) {
3664 Instruction *UI = cast<Instruction>(U);
3665 // This user may be a reduction exit value.
3666 if (!TheLoop->contains(UI)) {
3667 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3674 bool LoopVectorizationLegality::canVectorizeInstrs() {
3675 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3676 BasicBlock *Header = TheLoop->getHeader();
3678 // Look for the attribute signaling the absence of NaNs.
3679 Function &F = *Header->getParent();
3680 if (F.hasFnAttribute("no-nans-fp-math"))
3681 HasFunNoNaNAttr = F.getAttributes().getAttribute(
3682 AttributeSet::FunctionIndex,
3683 "no-nans-fp-math").getValueAsString() == "true";
3685 // For each block in the loop.
3686 for (Loop::block_iterator bb = TheLoop->block_begin(),
3687 be = TheLoop->block_end(); bb != be; ++bb) {
3689 // Scan the instructions in the block and look for hazards.
3690 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3693 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3694 Type *PhiTy = Phi->getType();
3695 // Check that this PHI type is allowed.
3696 if (!PhiTy->isIntegerTy() &&
3697 !PhiTy->isFloatingPointTy() &&
3698 !PhiTy->isPointerTy()) {
3699 emitAnalysis(Report(it)
3700 << "loop control flow is not understood by vectorizer");
3701 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3705 // If this PHINode is not in the header block, then we know that we
3706 // can convert it to select during if-conversion. No need to check if
3707 // the PHIs in this block are induction or reduction variables.
3708 if (*bb != Header) {
3709 // Check that this instruction has no outside users or is an
3710 // identified reduction value with an outside user.
3711 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3713 emitAnalysis(Report(it) << "value could not be identified as "
3714 "an induction or reduction variable");
3718 // We only allow if-converted PHIs with more than two incoming values.
3719 if (Phi->getNumIncomingValues() != 2) {
3720 emitAnalysis(Report(it)
3721 << "control flow not understood by vectorizer");
3722 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3726 // This is the value coming from the preheader.
3727 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3728 // Check if this is an induction variable.
3729 InductionKind IK = isInductionVariable(Phi);
3731 if (IK_NoInduction != IK) {
3732 // Get the widest type.
3734 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3736 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3738 // Int inductions are special because we only allow one IV.
3739 if (IK == IK_IntInduction) {
3740 // Use the phi node with the widest type as induction. Use the last
3741 // one if there are multiple (no good reason for doing this other
3742 // than it is expedient).
3743 if (!Induction || PhiTy == WidestIndTy)
3747 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3748 Inductions[Phi] = InductionInfo(StartValue, IK);
3750 // Until we explicitly handle the case of an induction variable with
3751 // an outside loop user we have to give up vectorizing this loop.
3752 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3753 emitAnalysis(Report(it) << "use of induction value outside of the "
3754 "loop is not handled by vectorizer");
3761 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3762 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3765 if (AddReductionVar(Phi, RK_IntegerMult)) {
3766 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3769 if (AddReductionVar(Phi, RK_IntegerOr)) {
3770 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3773 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3774 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3777 if (AddReductionVar(Phi, RK_IntegerXor)) {
3778 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3781 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3782 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3785 if (AddReductionVar(Phi, RK_FloatMult)) {
3786 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3789 if (AddReductionVar(Phi, RK_FloatAdd)) {
3790 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3793 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3794 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3799 emitAnalysis(Report(it) << "value that could not be identified as "
3800 "reduction is used outside the loop");
3801 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3803 }// end of PHI handling
3805 // We still don't handle functions. However, we can ignore dbg intrinsic
3806 // calls and we do handle certain intrinsic and libm functions.
3807 CallInst *CI = dyn_cast<CallInst>(it);
3808 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3809 emitAnalysis(Report(it) << "call instruction cannot be vectorized");
3810 DEBUG(dbgs() << "LV: Found a call site.\n");
3814 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
3815 // second argument is the same (i.e. loop invariant)
3817 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
3818 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
3819 emitAnalysis(Report(it)
3820 << "intrinsic instruction cannot be vectorized");
3821 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
3826 // Check that the instruction return type is vectorizable.
3827 // Also, we can't vectorize extractelement instructions.
3828 if ((!VectorType::isValidElementType(it->getType()) &&
3829 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3830 emitAnalysis(Report(it)
3831 << "instruction return type cannot be vectorized");
3832 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3836 // Check that the stored type is vectorizable.
3837 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3838 Type *T = ST->getValueOperand()->getType();
3839 if (!VectorType::isValidElementType(T)) {
3840 emitAnalysis(Report(ST) << "store instruction cannot be vectorized");
3843 if (EnableMemAccessVersioning)
3844 collectStridedAcccess(ST);
3847 if (EnableMemAccessVersioning)
3848 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3849 collectStridedAcccess(LI);
3851 // Reduction instructions are allowed to have exit users.
3852 // All other instructions must not have external users.
3853 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3854 emitAnalysis(Report(it) << "value cannot be used outside the loop");
3863 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3864 if (Inductions.empty()) {
3865 emitAnalysis(Report()
3866 << "loop induction variable could not be identified");
3874 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3875 /// return the induction operand of the gep pointer.
3876 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3877 const DataLayout *DL, Loop *Lp) {
3878 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3882 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3884 // Check that all of the gep indices are uniform except for our induction
3886 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3887 if (i != InductionOperand &&
3888 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3890 return GEP->getOperand(InductionOperand);
3893 ///\brief Look for a cast use of the passed value.
3894 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3895 Value *UniqueCast = nullptr;
3896 for (User *U : Ptr->users()) {
3897 CastInst *CI = dyn_cast<CastInst>(U);
3898 if (CI && CI->getType() == Ty) {
3908 ///\brief Get the stride of a pointer access in a loop.
3909 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3910 /// pointer to the Value, or null otherwise.
3911 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3912 const DataLayout *DL, Loop *Lp) {
3913 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3914 if (!PtrTy || PtrTy->isAggregateType())
3917 // Try to remove a gep instruction to make the pointer (actually index at this
3918 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3919 // pointer, otherwise, we are analyzing the index.
3920 Value *OrigPtr = Ptr;
3922 // The size of the pointer access.
3923 int64_t PtrAccessSize = 1;
3925 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3926 const SCEV *V = SE->getSCEV(Ptr);
3930 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3931 V = C->getOperand();
3933 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3937 V = S->getStepRecurrence(*SE);
3941 // Strip off the size of access multiplication if we are still analyzing the
3943 if (OrigPtr == Ptr) {
3944 DL->getTypeAllocSize(PtrTy->getElementType());
3945 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3946 if (M->getOperand(0)->getSCEVType() != scConstant)
3949 const APInt &APStepVal =
3950 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3952 // Huge step value - give up.
3953 if (APStepVal.getBitWidth() > 64)
3956 int64_t StepVal = APStepVal.getSExtValue();
3957 if (PtrAccessSize != StepVal)
3959 V = M->getOperand(1);
3964 Type *StripedOffRecurrenceCast = nullptr;
3965 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3966 StripedOffRecurrenceCast = C->getType();
3967 V = C->getOperand();
3970 // Look for the loop invariant symbolic value.
3971 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3975 Value *Stride = U->getValue();
3976 if (!Lp->isLoopInvariant(Stride))
3979 // If we have stripped off the recurrence cast we have to make sure that we
3980 // return the value that is used in this loop so that we can replace it later.
3981 if (StripedOffRecurrenceCast)
3982 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3987 void LoopVectorizationLegality::collectStridedAcccess(Value *MemAccess) {
3988 Value *Ptr = nullptr;
3989 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3990 Ptr = LI->getPointerOperand();
3991 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3992 Ptr = SI->getPointerOperand();
3996 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
4000 DEBUG(dbgs() << "LV: Found a strided access that we can version");
4001 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
4002 Strides[Ptr] = Stride;
4003 StrideSet.insert(Stride);
4006 void LoopVectorizationLegality::collectLoopUniforms() {
4007 // We now know that the loop is vectorizable!
4008 // Collect variables that will remain uniform after vectorization.
4009 std::vector<Value*> Worklist;
4010 BasicBlock *Latch = TheLoop->getLoopLatch();
4012 // Start with the conditional branch and walk up the block.
4013 Worklist.push_back(Latch->getTerminator()->getOperand(0));
4015 // Also add all consecutive pointer values; these values will be uniform
4016 // after vectorization (and subsequent cleanup) and, until revectorization is
4017 // supported, all dependencies must also be uniform.
4018 for (Loop::block_iterator B = TheLoop->block_begin(),
4019 BE = TheLoop->block_end(); B != BE; ++B)
4020 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
4022 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
4023 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4025 while (Worklist.size()) {
4026 Instruction *I = dyn_cast<Instruction>(Worklist.back());
4027 Worklist.pop_back();
4029 // Look at instructions inside this loop.
4030 // Stop when reaching PHI nodes.
4031 // TODO: we need to follow values all over the loop, not only in this block.
4032 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
4035 // This is a known uniform.
4038 // Insert all operands.
4039 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4044 /// \brief Analyses memory accesses in a loop.
4046 /// Checks whether run time pointer checks are needed and builds sets for data
4047 /// dependence checking.
4048 class AccessAnalysis {
4050 /// \brief Read or write access location.
4051 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
4052 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
4054 /// \brief Set of potential dependent memory accesses.
4055 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
4057 AccessAnalysis(const DataLayout *Dl, AliasAnalysis *AA, DepCandidates &DA) :
4058 DL(Dl), AST(*AA), DepCands(DA), IsRTCheckNeeded(false) {}
4060 /// \brief Register a load and whether it is only read from.
4061 void addLoad(AliasAnalysis::Location &Loc, bool IsReadOnly) {
4062 Value *Ptr = const_cast<Value*>(Loc.Ptr);
4063 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
4064 Accesses.insert(MemAccessInfo(Ptr, false));
4066 ReadOnlyPtr.insert(Ptr);
4069 /// \brief Register a store.
4070 void addStore(AliasAnalysis::Location &Loc) {
4071 Value *Ptr = const_cast<Value*>(Loc.Ptr);
4072 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
4073 Accesses.insert(MemAccessInfo(Ptr, true));
4076 /// \brief Check whether we can check the pointers at runtime for
4077 /// non-intersection.
4078 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
4079 unsigned &NumComparisons, ScalarEvolution *SE,
4080 Loop *TheLoop, ValueToValueMap &Strides,
4081 bool ShouldCheckStride = false);
4083 /// \brief Goes over all memory accesses, checks whether a RT check is needed
4084 /// and builds sets of dependent accesses.
4085 void buildDependenceSets() {
4086 processMemAccesses();
4089 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
4091 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
4092 void resetDepChecks() { CheckDeps.clear(); }
4094 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
4097 typedef SetVector<MemAccessInfo> PtrAccessSet;
4099 /// \brief Go over all memory access and check whether runtime pointer checks
4100 /// are needed /// and build sets of dependency check candidates.
4101 void processMemAccesses();
4103 /// Set of all accesses.
4104 PtrAccessSet Accesses;
4106 /// Set of accesses that need a further dependence check.
4107 MemAccessInfoSet CheckDeps;
4109 /// Set of pointers that are read only.
4110 SmallPtrSet<Value*, 16> ReadOnlyPtr;
4112 const DataLayout *DL;
4114 /// An alias set tracker to partition the access set by underlying object and
4115 //intrinsic property (such as TBAA metadata).
4116 AliasSetTracker AST;
4118 /// Sets of potentially dependent accesses - members of one set share an
4119 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
4120 /// dependence check.
4121 DepCandidates &DepCands;
4123 bool IsRTCheckNeeded;
4126 } // end anonymous namespace
4128 /// \brief Check whether a pointer can participate in a runtime bounds check.
4129 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
4131 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
4132 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4136 return AR->isAffine();
4139 /// \brief Check the stride of the pointer and ensure that it does not wrap in
4140 /// the address space.
4141 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4142 const Loop *Lp, ValueToValueMap &StridesMap);
4144 bool AccessAnalysis::canCheckPtrAtRT(
4145 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
4146 unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
4147 ValueToValueMap &StridesMap, bool ShouldCheckStride) {
4148 // Find pointers with computable bounds. We are going to use this information
4149 // to place a runtime bound check.
4150 bool CanDoRT = true;
4152 bool IsDepCheckNeeded = isDependencyCheckNeeded();
4155 // We assign a consecutive id to access from different alias sets.
4156 // Accesses between different groups doesn't need to be checked.
4158 for (auto &AS : AST) {
4159 unsigned NumReadPtrChecks = 0;
4160 unsigned NumWritePtrChecks = 0;
4162 // We assign consecutive id to access from different dependence sets.
4163 // Accesses within the same set don't need a runtime check.
4164 unsigned RunningDepId = 1;
4165 DenseMap<Value *, unsigned> DepSetId;
4168 Value *Ptr = A.getValue();
4169 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
4170 MemAccessInfo Access(Ptr, IsWrite);
4173 ++NumWritePtrChecks;
4177 if (hasComputableBounds(SE, StridesMap, Ptr) &&
4178 // When we run after a failing dependency check we have to make sure we
4179 // don't have wrapping pointers.
4180 (!ShouldCheckStride ||
4181 isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
4182 // The id of the dependence set.
4185 if (IsDepCheckNeeded) {
4186 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
4187 unsigned &LeaderId = DepSetId[Leader];
4189 LeaderId = RunningDepId++;
4192 // Each access has its own dependence set.
4193 DepId = RunningDepId++;
4195 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
4197 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
4203 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
4204 NumComparisons += 0; // Only one dependence set.
4206 NumComparisons += (NumWritePtrChecks * (NumReadPtrChecks +
4207 NumWritePtrChecks - 1));
4213 // If the pointers that we would use for the bounds comparison have different
4214 // address spaces, assume the values aren't directly comparable, so we can't
4215 // use them for the runtime check. We also have to assume they could
4216 // overlap. In the future there should be metadata for whether address spaces
4218 unsigned NumPointers = RtCheck.Pointers.size();
4219 for (unsigned i = 0; i < NumPointers; ++i) {
4220 for (unsigned j = i + 1; j < NumPointers; ++j) {
4221 // Only need to check pointers between two different dependency sets.
4222 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
4224 // Only need to check pointers in the same alias set.
4225 if (RtCheck.AliasSetId[i] != RtCheck.AliasSetId[j])
4228 Value *PtrI = RtCheck.Pointers[i];
4229 Value *PtrJ = RtCheck.Pointers[j];
4231 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
4232 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
4234 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
4235 " different address spaces\n");
4244 void AccessAnalysis::processMemAccesses() {
4245 // We process the set twice: first we process read-write pointers, last we
4246 // process read-only pointers. This allows us to skip dependence tests for
4247 // read-only pointers.
4249 DEBUG(dbgs() << "LV: Processing memory accesses...\n");
4250 DEBUG(dbgs() << " AST: "; AST.dump());
4251 DEBUG(dbgs() << "LV: Accesses:\n");
4253 for (auto A : Accesses)
4254 dbgs() << "\t" << *A.getPointer() << " (" <<
4255 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
4256 "read-only" : "read")) << ")\n";
4259 // The AliasSetTracker has nicely partitioned our pointers by metadata
4260 // compatibility and potential for underlying-object overlap. As a result, we
4261 // only need to check for potential pointer dependencies within each alias
4263 for (auto &AS : AST) {
4264 // Note that both the alias-set tracker and the alias sets themselves used
4265 // linked lists internally and so the iteration order here is deterministic
4266 // (matching the original instruction order within each set).
4268 bool SetHasWrite = false;
4270 // Map of pointers to last access encountered.
4271 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
4272 UnderlyingObjToAccessMap ObjToLastAccess;
4274 // Set of access to check after all writes have been processed.
4275 PtrAccessSet DeferredAccesses;
4277 // Iterate over each alias set twice, once to process read/write pointers,
4278 // and then to process read-only pointers.
4279 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
4280 bool UseDeferred = SetIteration > 0;
4281 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
4284 Value *Ptr = A.getValue();
4285 bool IsWrite = S.count(MemAccessInfo(Ptr, true));
4287 // If we're using the deferred access set, then it contains only reads.
4288 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
4289 if (UseDeferred && !IsReadOnlyPtr)
4291 // Otherwise, the pointer must be in the PtrAccessSet, either as a read
4293 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
4294 S.count(MemAccessInfo(Ptr, false))) &&
4295 "Alias-set pointer not in the access set?");
4297 MemAccessInfo Access(Ptr, IsWrite);
4298 DepCands.insert(Access);
4300 // Memorize read-only pointers for later processing and skip them in the
4301 // first round (they need to be checked after we have seen all write
4302 // pointers). Note: we also mark pointer that are not consecutive as
4303 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need
4304 // the second check for "!IsWrite".
4305 if (!UseDeferred && IsReadOnlyPtr) {
4306 DeferredAccesses.insert(Access);
4310 // If this is a write - check other reads and writes for conflicts. If
4311 // this is a read only check other writes for conflicts (but only if
4312 // there is no other write to the ptr - this is an optimization to
4313 // catch "a[i] = a[i] + " without having to do a dependence check).
4314 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
4315 CheckDeps.insert(Access);
4316 IsRTCheckNeeded = true;
4322 // Create sets of pointers connected by a shared alias set and
4323 // underlying object.
4324 typedef SmallVector<Value *, 16> ValueVector;
4325 ValueVector TempObjects;
4326 GetUnderlyingObjects(Ptr, TempObjects, DL);
4327 for (Value *UnderlyingObj : TempObjects) {
4328 UnderlyingObjToAccessMap::iterator Prev =
4329 ObjToLastAccess.find(UnderlyingObj);
4330 if (Prev != ObjToLastAccess.end())
4331 DepCands.unionSets(Access, Prev->second);
4333 ObjToLastAccess[UnderlyingObj] = Access;
4341 /// \brief Checks memory dependences among accesses to the same underlying
4342 /// object to determine whether there vectorization is legal or not (and at
4343 /// which vectorization factor).
4345 /// This class works under the assumption that we already checked that memory
4346 /// locations with different underlying pointers are "must-not alias".
4347 /// We use the ScalarEvolution framework to symbolically evalutate access
4348 /// functions pairs. Since we currently don't restructure the loop we can rely
4349 /// on the program order of memory accesses to determine their safety.
4350 /// At the moment we will only deem accesses as safe for:
4351 /// * A negative constant distance assuming program order.
4353 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
4354 /// a[i] = tmp; y = a[i];
4356 /// The latter case is safe because later checks guarantuee that there can't
4357 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
4358 /// the same variable: a header phi can only be an induction or a reduction, a
4359 /// reduction can't have a memory sink, an induction can't have a memory
4360 /// source). This is important and must not be violated (or we have to
4361 /// resort to checking for cycles through memory).
4363 /// * A positive constant distance assuming program order that is bigger
4364 /// than the biggest memory access.
4366 /// tmp = a[i] OR b[i] = x
4367 /// a[i+2] = tmp y = b[i+2];
4369 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
4371 /// * Zero distances and all accesses have the same size.
4373 class MemoryDepChecker {
4375 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
4376 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
4378 MemoryDepChecker(ScalarEvolution *Se, const DataLayout *Dl, const Loop *L)
4379 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
4380 ShouldRetryWithRuntimeCheck(false) {}
4382 /// \brief Register the location (instructions are given increasing numbers)
4383 /// of a write access.
4384 void addAccess(StoreInst *SI) {
4385 Value *Ptr = SI->getPointerOperand();
4386 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
4387 InstMap.push_back(SI);
4391 /// \brief Register the location (instructions are given increasing numbers)
4392 /// of a write access.
4393 void addAccess(LoadInst *LI) {
4394 Value *Ptr = LI->getPointerOperand();
4395 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
4396 InstMap.push_back(LI);
4400 /// \brief Check whether the dependencies between the accesses are safe.
4402 /// Only checks sets with elements in \p CheckDeps.
4403 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4404 MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
4406 /// \brief The maximum number of bytes of a vector register we can vectorize
4407 /// the accesses safely with.
4408 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
4410 /// \brief In same cases when the dependency check fails we can still
4411 /// vectorize the loop with a dynamic array access check.
4412 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
4415 ScalarEvolution *SE;
4416 const DataLayout *DL;
4417 const Loop *InnermostLoop;
4419 /// \brief Maps access locations (ptr, read/write) to program order.
4420 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
4422 /// \brief Memory access instructions in program order.
4423 SmallVector<Instruction *, 16> InstMap;
4425 /// \brief The program order index to be used for the next instruction.
4428 // We can access this many bytes in parallel safely.
4429 unsigned MaxSafeDepDistBytes;
4431 /// \brief If we see a non-constant dependence distance we can still try to
4432 /// vectorize this loop with runtime checks.
4433 bool ShouldRetryWithRuntimeCheck;
4435 /// \brief Check whether there is a plausible dependence between the two
4438 /// Access \p A must happen before \p B in program order. The two indices
4439 /// identify the index into the program order map.
4441 /// This function checks whether there is a plausible dependence (or the
4442 /// absence of such can't be proved) between the two accesses. If there is a
4443 /// plausible dependence but the dependence distance is bigger than one
4444 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
4445 /// distance is smaller than any other distance encountered so far).
4446 /// Otherwise, this function returns true signaling a possible dependence.
4447 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
4448 const MemAccessInfo &B, unsigned BIdx,
4449 ValueToValueMap &Strides);
4451 /// \brief Check whether the data dependence could prevent store-load
4453 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
4456 } // end anonymous namespace
4458 static bool isInBoundsGep(Value *Ptr) {
4459 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
4460 return GEP->isInBounds();
4464 /// \brief Check whether the access through \p Ptr has a constant stride.
4465 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4466 const Loop *Lp, ValueToValueMap &StridesMap) {
4467 const Type *Ty = Ptr->getType();
4468 assert(Ty->isPointerTy() && "Unexpected non-ptr");
4470 // Make sure that the pointer does not point to aggregate types.
4471 const PointerType *PtrTy = cast<PointerType>(Ty);
4472 if (PtrTy->getElementType()->isAggregateType()) {
4473 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
4478 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
4480 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4482 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
4483 << *Ptr << " SCEV: " << *PtrScev << "\n");
4487 // The accesss function must stride over the innermost loop.
4488 if (Lp != AR->getLoop()) {
4489 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
4490 *Ptr << " SCEV: " << *PtrScev << "\n");
4493 // The address calculation must not wrap. Otherwise, a dependence could be
4495 // An inbounds getelementptr that is a AddRec with a unit stride
4496 // cannot wrap per definition. The unit stride requirement is checked later.
4497 // An getelementptr without an inbounds attribute and unit stride would have
4498 // to access the pointer value "0" which is undefined behavior in address
4499 // space 0, therefore we can also vectorize this case.
4500 bool IsInBoundsGEP = isInBoundsGep(Ptr);
4501 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
4502 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
4503 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
4504 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
4505 << *Ptr << " SCEV: " << *PtrScev << "\n");
4509 // Check the step is constant.
4510 const SCEV *Step = AR->getStepRecurrence(*SE);
4512 // Calculate the pointer stride and check if it is consecutive.
4513 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4515 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
4516 " SCEV: " << *PtrScev << "\n");
4520 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
4521 const APInt &APStepVal = C->getValue()->getValue();
4523 // Huge step value - give up.
4524 if (APStepVal.getBitWidth() > 64)
4527 int64_t StepVal = APStepVal.getSExtValue();
4530 int64_t Stride = StepVal / Size;
4531 int64_t Rem = StepVal % Size;
4535 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
4536 // know we can't "wrap around the address space". In case of address space
4537 // zero we know that this won't happen without triggering undefined behavior.
4538 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
4539 Stride != 1 && Stride != -1)
4545 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
4546 unsigned TypeByteSize) {
4547 // If loads occur at a distance that is not a multiple of a feasible vector
4548 // factor store-load forwarding does not take place.
4549 // Positive dependences might cause troubles because vectorizing them might
4550 // prevent store-load forwarding making vectorized code run a lot slower.
4551 // a[i] = a[i-3] ^ a[i-8];
4552 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
4553 // hence on your typical architecture store-load forwarding does not take
4554 // place. Vectorizing in such cases does not make sense.
4555 // Store-load forwarding distance.
4556 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
4557 // Maximum vector factor.
4558 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
4559 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
4560 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
4562 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
4564 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
4565 MaxVFWithoutSLForwardIssues = (vf >>=1);
4570 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
4571 DEBUG(dbgs() << "LV: Distance " << Distance <<
4572 " that could cause a store-load forwarding conflict\n");
4576 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
4577 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
4578 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
4582 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
4583 const MemAccessInfo &B, unsigned BIdx,
4584 ValueToValueMap &Strides) {
4585 assert (AIdx < BIdx && "Must pass arguments in program order");
4587 Value *APtr = A.getPointer();
4588 Value *BPtr = B.getPointer();
4589 bool AIsWrite = A.getInt();
4590 bool BIsWrite = B.getInt();
4592 // Two reads are independent.
4593 if (!AIsWrite && !BIsWrite)
4596 // We cannot check pointers in different address spaces.
4597 if (APtr->getType()->getPointerAddressSpace() !=
4598 BPtr->getType()->getPointerAddressSpace())
4601 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
4602 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
4604 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
4605 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
4607 const SCEV *Src = AScev;
4608 const SCEV *Sink = BScev;
4610 // If the induction step is negative we have to invert source and sink of the
4612 if (StrideAPtr < 0) {
4615 std::swap(APtr, BPtr);
4616 std::swap(Src, Sink);
4617 std::swap(AIsWrite, BIsWrite);
4618 std::swap(AIdx, BIdx);
4619 std::swap(StrideAPtr, StrideBPtr);
4622 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
4624 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
4625 << "(Induction step: " << StrideAPtr << ")\n");
4626 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
4627 << *InstMap[BIdx] << ": " << *Dist << "\n");
4629 // Need consecutive accesses. We don't want to vectorize
4630 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
4631 // the address space.
4632 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
4633 DEBUG(dbgs() << "Non-consecutive pointer access\n");
4637 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
4639 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
4640 ShouldRetryWithRuntimeCheck = true;
4644 Type *ATy = APtr->getType()->getPointerElementType();
4645 Type *BTy = BPtr->getType()->getPointerElementType();
4646 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
4648 // Negative distances are not plausible dependencies.
4649 const APInt &Val = C->getValue()->getValue();
4650 if (Val.isNegative()) {
4651 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
4652 if (IsTrueDataDependence &&
4653 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
4657 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
4661 // Write to the same location with the same size.
4662 // Could be improved to assert type sizes are the same (i32 == float, etc).
4666 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
4670 assert(Val.isStrictlyPositive() && "Expect a positive value");
4672 // Positive distance bigger than max vectorization factor.
4675 "LV: ReadWrite-Write positive dependency with different types\n");
4679 unsigned Distance = (unsigned) Val.getZExtValue();
4681 // Bail out early if passed-in parameters make vectorization not feasible.
4682 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
4683 unsigned ForcedUnroll = VectorizationInterleave ? VectorizationInterleave : 1;
4685 // The distance must be bigger than the size needed for a vectorized version
4686 // of the operation and the size of the vectorized operation must not be
4687 // bigger than the currrent maximum size.
4688 if (Distance < 2*TypeByteSize ||
4689 2*TypeByteSize > MaxSafeDepDistBytes ||
4690 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
4691 DEBUG(dbgs() << "LV: Failure because of Positive distance "
4692 << Val.getSExtValue() << '\n');
4696 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
4697 Distance : MaxSafeDepDistBytes;
4699 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
4700 if (IsTrueDataDependence &&
4701 couldPreventStoreLoadForward(Distance, TypeByteSize))
4704 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
4705 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
4710 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4711 MemAccessInfoSet &CheckDeps,
4712 ValueToValueMap &Strides) {
4714 MaxSafeDepDistBytes = -1U;
4715 while (!CheckDeps.empty()) {
4716 MemAccessInfo CurAccess = *CheckDeps.begin();
4718 // Get the relevant memory access set.
4719 EquivalenceClasses<MemAccessInfo>::iterator I =
4720 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
4722 // Check accesses within this set.
4723 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
4724 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
4726 // Check every access pair.
4728 CheckDeps.erase(*AI);
4729 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
4731 // Check every accessing instruction pair in program order.
4732 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
4733 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
4734 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
4735 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
4736 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
4738 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
4749 bool LoopVectorizationLegality::canVectorizeMemory() {
4751 typedef SmallVector<Value*, 16> ValueVector;
4752 typedef SmallPtrSet<Value*, 16> ValueSet;
4754 // Holds the Load and Store *instructions*.
4758 // Holds all the different accesses in the loop.
4759 unsigned NumReads = 0;
4760 unsigned NumReadWrites = 0;
4762 PtrRtCheck.Pointers.clear();
4763 PtrRtCheck.Need = false;
4765 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
4766 MemoryDepChecker DepChecker(SE, DL, TheLoop);
4769 for (Loop::block_iterator bb = TheLoop->block_begin(),
4770 be = TheLoop->block_end(); bb != be; ++bb) {
4772 // Scan the BB and collect legal loads and stores.
4773 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4776 // If this is a load, save it. If this instruction can read from memory
4777 // but is not a load, then we quit. Notice that we don't handle function
4778 // calls that read or write.
4779 if (it->mayReadFromMemory()) {
4780 // Many math library functions read the rounding mode. We will only
4781 // vectorize a loop if it contains known function calls that don't set
4782 // the flag. Therefore, it is safe to ignore this read from memory.
4783 CallInst *Call = dyn_cast<CallInst>(it);
4784 if (Call && getIntrinsicIDForCall(Call, TLI))
4787 LoadInst *Ld = dyn_cast<LoadInst>(it);
4788 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
4789 emitAnalysis(Report(Ld)
4790 << "read with atomic ordering or volatile read");
4791 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
4795 Loads.push_back(Ld);
4796 DepChecker.addAccess(Ld);
4800 // Save 'store' instructions. Abort if other instructions write to memory.
4801 if (it->mayWriteToMemory()) {
4802 StoreInst *St = dyn_cast<StoreInst>(it);
4804 emitAnalysis(Report(it) << "instruction cannot be vectorized");
4807 if (!St->isSimple() && !IsAnnotatedParallel) {
4808 emitAnalysis(Report(St)
4809 << "write with atomic ordering or volatile write");
4810 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
4814 Stores.push_back(St);
4815 DepChecker.addAccess(St);
4820 // Now we have two lists that hold the loads and the stores.
4821 // Next, we find the pointers that they use.
4823 // Check if we see any stores. If there are no stores, then we don't
4824 // care if the pointers are *restrict*.
4825 if (!Stores.size()) {
4826 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
4830 AccessAnalysis::DepCandidates DependentAccesses;
4831 AccessAnalysis Accesses(DL, AA, DependentAccesses);
4833 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
4834 // multiple times on the same object. If the ptr is accessed twice, once
4835 // for read and once for write, it will only appear once (on the write
4836 // list). This is okay, since we are going to check for conflicts between
4837 // writes and between reads and writes, but not between reads and reads.
4840 ValueVector::iterator I, IE;
4841 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
4842 StoreInst *ST = cast<StoreInst>(*I);
4843 Value* Ptr = ST->getPointerOperand();
4845 if (isUniform(Ptr)) {
4848 << "write to a loop invariant address could not be vectorized");
4849 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4853 // If we did *not* see this pointer before, insert it to the read-write
4854 // list. At this phase it is only a 'write' list.
4855 if (Seen.insert(Ptr).second) {
4858 AliasAnalysis::Location Loc = AA->getLocation(ST);
4859 // The TBAA metadata could have a control dependency on the predication
4860 // condition, so we cannot rely on it when determining whether or not we
4861 // need runtime pointer checks.
4862 if (blockNeedsPredication(ST->getParent()))
4863 Loc.AATags.TBAA = nullptr;
4865 Accesses.addStore(Loc);
4869 if (IsAnnotatedParallel) {
4871 << "LV: A loop annotated parallel, ignore memory dependency "
4876 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
4877 LoadInst *LD = cast<LoadInst>(*I);
4878 Value* Ptr = LD->getPointerOperand();
4879 // If we did *not* see this pointer before, insert it to the
4880 // read list. If we *did* see it before, then it is already in
4881 // the read-write list. This allows us to vectorize expressions
4882 // such as A[i] += x; Because the address of A[i] is a read-write
4883 // pointer. This only works if the index of A[i] is consecutive.
4884 // If the address of i is unknown (for example A[B[i]]) then we may
4885 // read a few words, modify, and write a few words, and some of the
4886 // words may be written to the same address.
4887 bool IsReadOnlyPtr = false;
4888 if (Seen.insert(Ptr).second ||
4889 !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
4891 IsReadOnlyPtr = true;
4894 AliasAnalysis::Location Loc = AA->getLocation(LD);
4895 // The TBAA metadata could have a control dependency on the predication
4896 // condition, so we cannot rely on it when determining whether or not we
4897 // need runtime pointer checks.
4898 if (blockNeedsPredication(LD->getParent()))
4899 Loc.AATags.TBAA = nullptr;
4901 Accesses.addLoad(Loc, IsReadOnlyPtr);
4904 // If we write (or read-write) to a single destination and there are no
4905 // other reads in this loop then is it safe to vectorize.
4906 if (NumReadWrites == 1 && NumReads == 0) {
4907 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
4911 // Build dependence sets and check whether we need a runtime pointer bounds
4913 Accesses.buildDependenceSets();
4914 bool NeedRTCheck = Accesses.isRTCheckNeeded();
4916 // Find pointers with computable bounds. We are going to use this information
4917 // to place a runtime bound check.
4918 unsigned NumComparisons = 0;
4919 bool CanDoRT = false;
4921 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
4924 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
4925 " pointer comparisons.\n");
4927 // If we only have one set of dependences to check pointers among we don't
4928 // need a runtime check.
4929 if (NumComparisons == 0 && NeedRTCheck)
4930 NeedRTCheck = false;
4932 // Check that we did not collect too many pointers or found an unsizeable
4934 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4940 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
4943 if (NeedRTCheck && !CanDoRT) {
4944 emitAnalysis(Report() << "cannot identify array bounds");
4945 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
4946 "the array bounds.\n");
4951 PtrRtCheck.Need = NeedRTCheck;
4953 bool CanVecMem = true;
4954 if (Accesses.isDependencyCheckNeeded()) {
4955 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
4956 CanVecMem = DepChecker.areDepsSafe(
4957 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
4958 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
4960 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
4961 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
4964 // Clear the dependency checks. We assume they are not needed.
4965 Accesses.resetDepChecks();
4968 PtrRtCheck.Need = true;
4970 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
4971 TheLoop, Strides, true);
4972 // Check that we did not collect too many pointers or found an unsizeable
4974 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4975 if (!CanDoRT && NumComparisons > 0)
4976 emitAnalysis(Report()
4977 << "cannot check memory dependencies at runtime");
4979 emitAnalysis(Report()
4980 << NumComparisons << " exceeds limit of "
4981 << RuntimeMemoryCheckThreshold
4982 << " dependent memory operations checked at runtime");
4983 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
4993 emitAnalysis(Report() << "unsafe dependent memory operations in loop");
4995 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
4996 " need a runtime memory check.\n");
5001 static bool hasMultipleUsesOf(Instruction *I,
5002 SmallPtrSetImpl<Instruction *> &Insts) {
5003 unsigned NumUses = 0;
5004 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
5005 if (Insts.count(dyn_cast<Instruction>(*Use)))
5014 static bool areAllUsesIn(Instruction *I, SmallPtrSetImpl<Instruction *> &Set) {
5015 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
5016 if (!Set.count(dyn_cast<Instruction>(*Use)))
5021 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
5022 ReductionKind Kind) {
5023 if (Phi->getNumIncomingValues() != 2)
5026 // Reduction variables are only found in the loop header block.
5027 if (Phi->getParent() != TheLoop->getHeader())
5030 // Obtain the reduction start value from the value that comes from the loop
5032 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
5034 // ExitInstruction is the single value which is used outside the loop.
5035 // We only allow for a single reduction value to be used outside the loop.
5036 // This includes users of the reduction, variables (which form a cycle
5037 // which ends in the phi node).
5038 Instruction *ExitInstruction = nullptr;
5039 // Indicates that we found a reduction operation in our scan.
5040 bool FoundReduxOp = false;
5042 // We start with the PHI node and scan for all of the users of this
5043 // instruction. All users must be instructions that can be used as reduction
5044 // variables (such as ADD). We must have a single out-of-block user. The cycle
5045 // must include the original PHI.
5046 bool FoundStartPHI = false;
5048 // To recognize min/max patterns formed by a icmp select sequence, we store
5049 // the number of instruction we saw from the recognized min/max pattern,
5050 // to make sure we only see exactly the two instructions.
5051 unsigned NumCmpSelectPatternInst = 0;
5052 ReductionInstDesc ReduxDesc(false, nullptr);
5054 SmallPtrSet<Instruction *, 8> VisitedInsts;
5055 SmallVector<Instruction *, 8> Worklist;
5056 Worklist.push_back(Phi);
5057 VisitedInsts.insert(Phi);
5059 // A value in the reduction can be used:
5060 // - By the reduction:
5061 // - Reduction operation:
5062 // - One use of reduction value (safe).
5063 // - Multiple use of reduction value (not safe).
5065 // - All uses of the PHI must be the reduction (safe).
5066 // - Otherwise, not safe.
5067 // - By one instruction outside of the loop (safe).
5068 // - By further instructions outside of the loop (not safe).
5069 // - By an instruction that is not part of the reduction (not safe).
5071 // * An instruction type other than PHI or the reduction operation.
5072 // * A PHI in the header other than the initial PHI.
5073 while (!Worklist.empty()) {
5074 Instruction *Cur = Worklist.back();
5075 Worklist.pop_back();
5078 // If the instruction has no users then this is a broken chain and can't be
5079 // a reduction variable.
5080 if (Cur->use_empty())
5083 bool IsAPhi = isa<PHINode>(Cur);
5085 // A header PHI use other than the original PHI.
5086 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
5089 // Reductions of instructions such as Div, and Sub is only possible if the
5090 // LHS is the reduction variable.
5091 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
5092 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
5093 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
5096 // Any reduction instruction must be of one of the allowed kinds.
5097 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
5098 if (!ReduxDesc.IsReduction)
5101 // A reduction operation must only have one use of the reduction value.
5102 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
5103 hasMultipleUsesOf(Cur, VisitedInsts))
5106 // All inputs to a PHI node must be a reduction value.
5107 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
5110 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
5111 isa<SelectInst>(Cur)))
5112 ++NumCmpSelectPatternInst;
5113 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
5114 isa<SelectInst>(Cur)))
5115 ++NumCmpSelectPatternInst;
5117 // Check whether we found a reduction operator.
5118 FoundReduxOp |= !IsAPhi;
5120 // Process users of current instruction. Push non-PHI nodes after PHI nodes
5121 // onto the stack. This way we are going to have seen all inputs to PHI
5122 // nodes once we get to them.
5123 SmallVector<Instruction *, 8> NonPHIs;
5124 SmallVector<Instruction *, 8> PHIs;
5125 for (User *U : Cur->users()) {
5126 Instruction *UI = cast<Instruction>(U);
5128 // Check if we found the exit user.
5129 BasicBlock *Parent = UI->getParent();
5130 if (!TheLoop->contains(Parent)) {
5131 // Exit if you find multiple outside users or if the header phi node is
5132 // being used. In this case the user uses the value of the previous
5133 // iteration, in which case we would loose "VF-1" iterations of the
5134 // reduction operation if we vectorize.
5135 if (ExitInstruction != nullptr || Cur == Phi)
5138 // The instruction used by an outside user must be the last instruction
5139 // before we feed back to the reduction phi. Otherwise, we loose VF-1
5140 // operations on the value.
5141 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
5144 ExitInstruction = Cur;
5148 // Process instructions only once (termination). Each reduction cycle
5149 // value must only be used once, except by phi nodes and min/max
5150 // reductions which are represented as a cmp followed by a select.
5151 ReductionInstDesc IgnoredVal(false, nullptr);
5152 if (VisitedInsts.insert(UI).second) {
5153 if (isa<PHINode>(UI))
5156 NonPHIs.push_back(UI);
5157 } else if (!isa<PHINode>(UI) &&
5158 ((!isa<FCmpInst>(UI) &&
5159 !isa<ICmpInst>(UI) &&
5160 !isa<SelectInst>(UI)) ||
5161 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
5164 // Remember that we completed the cycle.
5166 FoundStartPHI = true;
5168 Worklist.append(PHIs.begin(), PHIs.end());
5169 Worklist.append(NonPHIs.begin(), NonPHIs.end());
5172 // This means we have seen one but not the other instruction of the
5173 // pattern or more than just a select and cmp.
5174 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
5175 NumCmpSelectPatternInst != 2)
5178 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
5181 // We found a reduction var if we have reached the original phi node and we
5182 // only have a single instruction with out-of-loop users.
5184 // This instruction is allowed to have out-of-loop users.
5185 AllowedExit.insert(ExitInstruction);
5187 // Save the description of this reduction variable.
5188 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
5189 ReduxDesc.MinMaxKind);
5190 Reductions[Phi] = RD;
5191 // We've ended the cycle. This is a reduction variable if we have an
5192 // outside user and it has a binary op.
5197 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
5198 /// pattern corresponding to a min(X, Y) or max(X, Y).
5199 LoopVectorizationLegality::ReductionInstDesc
5200 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
5201 ReductionInstDesc &Prev) {
5203 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
5204 "Expect a select instruction");
5205 Instruction *Cmp = nullptr;
5206 SelectInst *Select = nullptr;
5208 // We must handle the select(cmp()) as a single instruction. Advance to the
5210 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
5211 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
5212 return ReductionInstDesc(false, I);
5213 return ReductionInstDesc(Select, Prev.MinMaxKind);
5216 // Only handle single use cases for now.
5217 if (!(Select = dyn_cast<SelectInst>(I)))
5218 return ReductionInstDesc(false, I);
5219 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
5220 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
5221 return ReductionInstDesc(false, I);
5222 if (!Cmp->hasOneUse())
5223 return ReductionInstDesc(false, I);
5228 // Look for a min/max pattern.
5229 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5230 return ReductionInstDesc(Select, MRK_UIntMin);
5231 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5232 return ReductionInstDesc(Select, MRK_UIntMax);
5233 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5234 return ReductionInstDesc(Select, MRK_SIntMax);
5235 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5236 return ReductionInstDesc(Select, MRK_SIntMin);
5237 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5238 return ReductionInstDesc(Select, MRK_FloatMin);
5239 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5240 return ReductionInstDesc(Select, MRK_FloatMax);
5241 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5242 return ReductionInstDesc(Select, MRK_FloatMin);
5243 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5244 return ReductionInstDesc(Select, MRK_FloatMax);
5246 return ReductionInstDesc(false, I);
5249 LoopVectorizationLegality::ReductionInstDesc
5250 LoopVectorizationLegality::isReductionInstr(Instruction *I,
5252 ReductionInstDesc &Prev) {
5253 bool FP = I->getType()->isFloatingPointTy();
5254 bool FastMath = FP && I->hasUnsafeAlgebra();
5255 switch (I->getOpcode()) {
5257 return ReductionInstDesc(false, I);
5258 case Instruction::PHI:
5259 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
5260 Kind != RK_FloatMinMax))
5261 return ReductionInstDesc(false, I);
5262 return ReductionInstDesc(I, Prev.MinMaxKind);
5263 case Instruction::Sub:
5264 case Instruction::Add:
5265 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
5266 case Instruction::Mul:
5267 return ReductionInstDesc(Kind == RK_IntegerMult, I);
5268 case Instruction::And:
5269 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
5270 case Instruction::Or:
5271 return ReductionInstDesc(Kind == RK_IntegerOr, I);
5272 case Instruction::Xor:
5273 return ReductionInstDesc(Kind == RK_IntegerXor, I);
5274 case Instruction::FMul:
5275 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
5276 case Instruction::FSub:
5277 case Instruction::FAdd:
5278 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
5279 case Instruction::FCmp:
5280 case Instruction::ICmp:
5281 case Instruction::Select:
5282 if (Kind != RK_IntegerMinMax &&
5283 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
5284 return ReductionInstDesc(false, I);
5285 return isMinMaxSelectCmpPattern(I, Prev);
5289 LoopVectorizationLegality::InductionKind
5290 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
5291 Type *PhiTy = Phi->getType();
5292 // We only handle integer and pointer inductions variables.
5293 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
5294 return IK_NoInduction;
5296 // Check that the PHI is consecutive.
5297 const SCEV *PhiScev = SE->getSCEV(Phi);
5298 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
5300 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
5301 return IK_NoInduction;
5303 const SCEV *Step = AR->getStepRecurrence(*SE);
5305 // Integer inductions need to have a stride of one.
5306 if (PhiTy->isIntegerTy()) {
5308 return IK_IntInduction;
5309 if (Step->isAllOnesValue())
5310 return IK_ReverseIntInduction;
5311 return IK_NoInduction;
5314 // Calculate the pointer stride and check if it is consecutive.
5315 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5317 return IK_NoInduction;
5319 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
5320 Type *PointerElementType = PhiTy->getPointerElementType();
5321 // The pointer stride cannot be determined if the pointer element type is not
5323 if (!PointerElementType->isSized())
5324 return IK_NoInduction;
5326 uint64_t Size = DL->getTypeAllocSize(PointerElementType);
5327 if (C->getValue()->equalsInt(Size))
5328 return IK_PtrInduction;
5329 else if (C->getValue()->equalsInt(0 - Size))
5330 return IK_ReversePtrInduction;
5332 return IK_NoInduction;
5335 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
5336 Value *In0 = const_cast<Value*>(V);
5337 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
5341 return Inductions.count(PN);
5344 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
5345 assert(TheLoop->contains(BB) && "Unknown block used");
5347 // Blocks that do not dominate the latch need predication.
5348 BasicBlock* Latch = TheLoop->getLoopLatch();
5349 return !DT->dominates(BB, Latch);
5352 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
5353 SmallPtrSetImpl<Value *> &SafePtrs) {
5354 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5355 // We might be able to hoist the load.
5356 if (it->mayReadFromMemory()) {
5357 LoadInst *LI = dyn_cast<LoadInst>(it);
5360 if (!SafePtrs.count(LI->getPointerOperand())) {
5361 if (canPredicateLoad(LI->getType(), LI->getPointerOperand())) {
5362 MaskedOp.insert(LI);
5369 // We don't predicate stores at the moment.
5370 if (it->mayWriteToMemory()) {
5371 StoreInst *SI = dyn_cast<StoreInst>(it);
5372 // We only support predication of stores in basic blocks with one
5377 if (++NumPredStores > NumberOfStoresToPredicate ||
5378 !SafePtrs.count(SI->getPointerOperand()) ||
5379 !SI->getParent()->getSinglePredecessor()) {
5380 if (canPredicateStore(SI->getValueOperand()->getType(),
5381 SI->getPointerOperand())) {
5382 MaskedOp.insert(SI);
5392 // Check that we don't have a constant expression that can trap as operand.
5393 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
5395 if (Constant *C = dyn_cast<Constant>(*OI))
5400 // The instructions below can trap.
5401 switch (it->getOpcode()) {
5403 case Instruction::UDiv:
5404 case Instruction::SDiv:
5405 case Instruction::URem:
5406 case Instruction::SRem:
5414 LoopVectorizationCostModel::VectorizationFactor
5415 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
5416 // Width 1 means no vectorize
5417 VectorizationFactor Factor = { 1U, 0U };
5418 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
5419 emitAnalysis(Report() << "runtime pointer checks needed. Enable vectorization of this loop with '#pragma clang loop vectorize(enable)' when compiling with -Os");
5420 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
5424 if (!EnableCondStoresVectorization && Legal->NumPredStores) {
5425 emitAnalysis(Report() << "store that is conditionally executed prevents vectorization");
5426 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
5430 // Find the trip count.
5431 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
5432 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
5434 unsigned WidestType = getWidestType();
5435 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
5436 unsigned MaxSafeDepDist = -1U;
5437 if (Legal->getMaxSafeDepDistBytes() != -1U)
5438 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
5439 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
5440 WidestRegister : MaxSafeDepDist);
5441 unsigned MaxVectorSize = WidestRegister / WidestType;
5442 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
5443 DEBUG(dbgs() << "LV: The Widest register is: "
5444 << WidestRegister << " bits.\n");
5446 if (MaxVectorSize == 0) {
5447 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
5451 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
5452 " into one vector!");
5454 unsigned VF = MaxVectorSize;
5456 // If we optimize the program for size, avoid creating the tail loop.
5458 // If we are unable to calculate the trip count then don't try to vectorize.
5460 emitAnalysis(Report() << "unable to calculate the loop count due to complex control flow");
5461 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5465 // Find the maximum SIMD width that can fit within the trip count.
5466 VF = TC % MaxVectorSize;
5471 // If the trip count that we found modulo the vectorization factor is not
5472 // zero then we require a tail.
5474 emitAnalysis(Report() << "cannot optimize for size and vectorize at the "
5475 "same time. Enable vectorization of this loop "
5476 "with '#pragma clang loop vectorize(enable)' "
5477 "when compiling with -Os");
5478 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5483 int UserVF = Hints->getWidth();
5485 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
5486 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
5488 Factor.Width = UserVF;
5492 float Cost = expectedCost(1);
5494 const float ScalarCost = Cost;
5497 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
5499 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
5500 // Ignore scalar width, because the user explicitly wants vectorization.
5501 if (ForceVectorization && VF > 1) {
5503 Cost = expectedCost(Width) / (float)Width;
5506 for (unsigned i=2; i <= VF; i*=2) {
5507 // Notice that the vector loop needs to be executed less times, so
5508 // we need to divide the cost of the vector loops by the width of
5509 // the vector elements.
5510 float VectorCost = expectedCost(i) / (float)i;
5511 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
5512 VectorCost << ".\n");
5513 if (VectorCost < Cost) {
5519 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
5520 << "LV: Vectorization seems to be not beneficial, "
5521 << "but was forced by a user.\n");
5522 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
5523 Factor.Width = Width;
5524 Factor.Cost = Width * Cost;
5528 unsigned LoopVectorizationCostModel::getWidestType() {
5529 unsigned MaxWidth = 8;
5532 for (Loop::block_iterator bb = TheLoop->block_begin(),
5533 be = TheLoop->block_end(); bb != be; ++bb) {
5534 BasicBlock *BB = *bb;
5536 // For each instruction in the loop.
5537 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5538 Type *T = it->getType();
5540 // Ignore ephemeral values.
5541 if (EphValues.count(it))
5544 // Only examine Loads, Stores and PHINodes.
5545 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
5548 // Examine PHI nodes that are reduction variables.
5549 if (PHINode *PN = dyn_cast<PHINode>(it))
5550 if (!Legal->getReductionVars()->count(PN))
5553 // Examine the stored values.
5554 if (StoreInst *ST = dyn_cast<StoreInst>(it))
5555 T = ST->getValueOperand()->getType();
5557 // Ignore loaded pointer types and stored pointer types that are not
5558 // consecutive. However, we do want to take consecutive stores/loads of
5559 // pointer vectors into account.
5560 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
5563 MaxWidth = std::max(MaxWidth,
5564 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
5572 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
5574 unsigned LoopCost) {
5576 // -- The unroll heuristics --
5577 // We unroll the loop in order to expose ILP and reduce the loop overhead.
5578 // There are many micro-architectural considerations that we can't predict
5579 // at this level. For example, frontend pressure (on decode or fetch) due to
5580 // code size, or the number and capabilities of the execution ports.
5582 // We use the following heuristics to select the unroll factor:
5583 // 1. If the code has reductions, then we unroll in order to break the cross
5584 // iteration dependency.
5585 // 2. If the loop is really small, then we unroll in order to reduce the loop
5587 // 3. We don't unroll if we think that we will spill registers to memory due
5588 // to the increased register pressure.
5590 // Use the user preference, unless 'auto' is selected.
5591 int UserUF = Hints->getInterleave();
5595 // When we optimize for size, we don't unroll.
5599 // We used the distance for the unroll factor.
5600 if (Legal->getMaxSafeDepDistBytes() != -1U)
5603 // Do not unroll loops with a relatively small trip count.
5604 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
5605 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
5608 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
5609 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
5613 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
5614 TargetNumRegisters = ForceTargetNumScalarRegs;
5616 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
5617 TargetNumRegisters = ForceTargetNumVectorRegs;
5620 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
5621 // We divide by these constants so assume that we have at least one
5622 // instruction that uses at least one register.
5623 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
5624 R.NumInstructions = std::max(R.NumInstructions, 1U);
5626 // We calculate the unroll factor using the following formula.
5627 // Subtract the number of loop invariants from the number of available
5628 // registers. These registers are used by all of the unrolled instances.
5629 // Next, divide the remaining registers by the number of registers that is
5630 // required by the loop, in order to estimate how many parallel instances
5631 // fit without causing spills. All of this is rounded down if necessary to be
5632 // a power of two. We want power of two unroll factors to simplify any
5633 // addressing operations or alignment considerations.
5634 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
5637 // Don't count the induction variable as unrolled.
5638 if (EnableIndVarRegisterHeur)
5639 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
5640 std::max(1U, (R.MaxLocalUsers - 1)));
5642 // Clamp the unroll factor ranges to reasonable factors.
5643 unsigned MaxInterleaveSize = TTI.getMaxInterleaveFactor();
5645 // Check if the user has overridden the unroll max.
5647 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
5648 MaxInterleaveSize = ForceTargetMaxScalarInterleaveFactor;
5650 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
5651 MaxInterleaveSize = ForceTargetMaxVectorInterleaveFactor;
5654 // If we did not calculate the cost for VF (because the user selected the VF)
5655 // then we calculate the cost of VF here.
5657 LoopCost = expectedCost(VF);
5659 // Clamp the calculated UF to be between the 1 and the max unroll factor
5660 // that the target allows.
5661 if (UF > MaxInterleaveSize)
5662 UF = MaxInterleaveSize;
5666 // Unroll if we vectorized this loop and there is a reduction that could
5667 // benefit from unrolling.
5668 if (VF > 1 && Legal->getReductionVars()->size()) {
5669 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
5673 // Note that if we've already vectorized the loop we will have done the
5674 // runtime check and so unrolling won't require further checks.
5675 bool UnrollingRequiresRuntimePointerCheck =
5676 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
5678 // We want to unroll small loops in order to reduce the loop overhead and
5679 // potentially expose ILP opportunities.
5680 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
5681 if (!UnrollingRequiresRuntimePointerCheck &&
5682 LoopCost < SmallLoopCost) {
5683 // We assume that the cost overhead is 1 and we use the cost model
5684 // to estimate the cost of the loop and unroll until the cost of the
5685 // loop overhead is about 5% of the cost of the loop.
5686 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
5688 // Unroll until store/load ports (estimated by max unroll factor) are
5690 unsigned StoresUF = UF / (Legal->NumStores ? Legal->NumStores : 1);
5691 unsigned LoadsUF = UF / (Legal->NumLoads ? Legal->NumLoads : 1);
5693 // If we have a scalar reduction (vector reductions are already dealt with
5694 // by this point), we can increase the critical path length if the loop
5695 // we're unrolling is inside another loop. Limit, by default to 2, so the
5696 // critical path only gets increased by one reduction operation.
5697 if (Legal->getReductionVars()->size() &&
5698 TheLoop->getLoopDepth() > 1) {
5699 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionUF);
5700 SmallUF = std::min(SmallUF, F);
5701 StoresUF = std::min(StoresUF, F);
5702 LoadsUF = std::min(LoadsUF, F);
5705 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
5706 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
5707 return std::max(StoresUF, LoadsUF);
5710 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
5714 DEBUG(dbgs() << "LV: Not Unrolling.\n");
5718 LoopVectorizationCostModel::RegisterUsage
5719 LoopVectorizationCostModel::calculateRegisterUsage() {
5720 // This function calculates the register usage by measuring the highest number
5721 // of values that are alive at a single location. Obviously, this is a very
5722 // rough estimation. We scan the loop in a topological order in order and
5723 // assign a number to each instruction. We use RPO to ensure that defs are
5724 // met before their users. We assume that each instruction that has in-loop
5725 // users starts an interval. We record every time that an in-loop value is
5726 // used, so we have a list of the first and last occurrences of each
5727 // instruction. Next, we transpose this data structure into a multi map that
5728 // holds the list of intervals that *end* at a specific location. This multi
5729 // map allows us to perform a linear search. We scan the instructions linearly
5730 // and record each time that a new interval starts, by placing it in a set.
5731 // If we find this value in the multi-map then we remove it from the set.
5732 // The max register usage is the maximum size of the set.
5733 // We also search for instructions that are defined outside the loop, but are
5734 // used inside the loop. We need this number separately from the max-interval
5735 // usage number because when we unroll, loop-invariant values do not take
5737 LoopBlocksDFS DFS(TheLoop);
5741 R.NumInstructions = 0;
5743 // Each 'key' in the map opens a new interval. The values
5744 // of the map are the index of the 'last seen' usage of the
5745 // instruction that is the key.
5746 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5747 // Maps instruction to its index.
5748 DenseMap<unsigned, Instruction*> IdxToInstr;
5749 // Marks the end of each interval.
5750 IntervalMap EndPoint;
5751 // Saves the list of instruction indices that are used in the loop.
5752 SmallSet<Instruction*, 8> Ends;
5753 // Saves the list of values that are used in the loop but are
5754 // defined outside the loop, such as arguments and constants.
5755 SmallPtrSet<Value*, 8> LoopInvariants;
5758 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5759 be = DFS.endRPO(); bb != be; ++bb) {
5760 R.NumInstructions += (*bb)->size();
5761 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
5763 Instruction *I = it;
5764 IdxToInstr[Index++] = I;
5766 // Save the end location of each USE.
5767 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
5768 Value *U = I->getOperand(i);
5769 Instruction *Instr = dyn_cast<Instruction>(U);
5771 // Ignore non-instruction values such as arguments, constants, etc.
5772 if (!Instr) continue;
5774 // If this instruction is outside the loop then record it and continue.
5775 if (!TheLoop->contains(Instr)) {
5776 LoopInvariants.insert(Instr);
5780 // Overwrite previous end points.
5781 EndPoint[Instr] = Index;
5787 // Saves the list of intervals that end with the index in 'key'.
5788 typedef SmallVector<Instruction*, 2> InstrList;
5789 DenseMap<unsigned, InstrList> TransposeEnds;
5791 // Transpose the EndPoints to a list of values that end at each index.
5792 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5794 TransposeEnds[it->second].push_back(it->first);
5796 SmallSet<Instruction*, 8> OpenIntervals;
5797 unsigned MaxUsage = 0;
5800 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5801 for (unsigned int i = 0; i < Index; ++i) {
5802 Instruction *I = IdxToInstr[i];
5803 // Ignore instructions that are never used within the loop.
5804 if (!Ends.count(I)) continue;
5806 // Ignore ephemeral values.
5807 if (EphValues.count(I))
5810 // Remove all of the instructions that end at this location.
5811 InstrList &List = TransposeEnds[i];
5812 for (unsigned int j=0, e = List.size(); j < e; ++j)
5813 OpenIntervals.erase(List[j]);
5815 // Count the number of live interals.
5816 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5818 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5819 OpenIntervals.size() << '\n');
5821 // Add the current instruction to the list of open intervals.
5822 OpenIntervals.insert(I);
5825 unsigned Invariant = LoopInvariants.size();
5826 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5827 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5828 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5830 R.LoopInvariantRegs = Invariant;
5831 R.MaxLocalUsers = MaxUsage;
5835 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5839 for (Loop::block_iterator bb = TheLoop->block_begin(),
5840 be = TheLoop->block_end(); bb != be; ++bb) {
5841 unsigned BlockCost = 0;
5842 BasicBlock *BB = *bb;
5844 // For each instruction in the old loop.
5845 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5846 // Skip dbg intrinsics.
5847 if (isa<DbgInfoIntrinsic>(it))
5850 // Ignore ephemeral values.
5851 if (EphValues.count(it))
5854 unsigned C = getInstructionCost(it, VF);
5856 // Check if we should override the cost.
5857 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5858 C = ForceTargetInstructionCost;
5861 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5862 VF << " For instruction: " << *it << '\n');
5865 // We assume that if-converted blocks have a 50% chance of being executed.
5866 // When the code is scalar then some of the blocks are avoided due to CF.
5867 // When the code is vectorized we execute all code paths.
5868 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5877 /// \brief Check whether the address computation for a non-consecutive memory
5878 /// access looks like an unlikely candidate for being merged into the indexing
5881 /// We look for a GEP which has one index that is an induction variable and all
5882 /// other indices are loop invariant. If the stride of this access is also
5883 /// within a small bound we decide that this address computation can likely be
5884 /// merged into the addressing mode.
5885 /// In all other cases, we identify the address computation as complex.
5886 static bool isLikelyComplexAddressComputation(Value *Ptr,
5887 LoopVectorizationLegality *Legal,
5888 ScalarEvolution *SE,
5889 const Loop *TheLoop) {
5890 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5894 // We are looking for a gep with all loop invariant indices except for one
5895 // which should be an induction variable.
5896 unsigned NumOperands = Gep->getNumOperands();
5897 for (unsigned i = 1; i < NumOperands; ++i) {
5898 Value *Opd = Gep->getOperand(i);
5899 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5900 !Legal->isInductionVariable(Opd))
5904 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5905 // can likely be merged into the address computation.
5906 unsigned MaxMergeDistance = 64;
5908 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5912 // Check the step is constant.
5913 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5914 // Calculate the pointer stride and check if it is consecutive.
5915 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5919 const APInt &APStepVal = C->getValue()->getValue();
5921 // Huge step value - give up.
5922 if (APStepVal.getBitWidth() > 64)
5925 int64_t StepVal = APStepVal.getSExtValue();
5927 return StepVal > MaxMergeDistance;
5930 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5931 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5937 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5938 // If we know that this instruction will remain uniform, check the cost of
5939 // the scalar version.
5940 if (Legal->isUniformAfterVectorization(I))
5943 Type *RetTy = I->getType();
5944 Type *VectorTy = ToVectorTy(RetTy, VF);
5946 // TODO: We need to estimate the cost of intrinsic calls.
5947 switch (I->getOpcode()) {
5948 case Instruction::GetElementPtr:
5949 // We mark this instruction as zero-cost because the cost of GEPs in
5950 // vectorized code depends on whether the corresponding memory instruction
5951 // is scalarized or not. Therefore, we handle GEPs with the memory
5952 // instruction cost.
5954 case Instruction::Br: {
5955 return TTI.getCFInstrCost(I->getOpcode());
5957 case Instruction::PHI:
5958 //TODO: IF-converted IFs become selects.
5960 case Instruction::Add:
5961 case Instruction::FAdd:
5962 case Instruction::Sub:
5963 case Instruction::FSub:
5964 case Instruction::Mul:
5965 case Instruction::FMul:
5966 case Instruction::UDiv:
5967 case Instruction::SDiv:
5968 case Instruction::FDiv:
5969 case Instruction::URem:
5970 case Instruction::SRem:
5971 case Instruction::FRem:
5972 case Instruction::Shl:
5973 case Instruction::LShr:
5974 case Instruction::AShr:
5975 case Instruction::And:
5976 case Instruction::Or:
5977 case Instruction::Xor: {
5978 // Since we will replace the stride by 1 the multiplication should go away.
5979 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5981 // Certain instructions can be cheaper to vectorize if they have a constant
5982 // second vector operand. One example of this are shifts on x86.
5983 TargetTransformInfo::OperandValueKind Op1VK =
5984 TargetTransformInfo::OK_AnyValue;
5985 TargetTransformInfo::OperandValueKind Op2VK =
5986 TargetTransformInfo::OK_AnyValue;
5987 TargetTransformInfo::OperandValueProperties Op1VP =
5988 TargetTransformInfo::OP_None;
5989 TargetTransformInfo::OperandValueProperties Op2VP =
5990 TargetTransformInfo::OP_None;
5991 Value *Op2 = I->getOperand(1);
5993 // Check for a splat of a constant or for a non uniform vector of constants.
5994 if (isa<ConstantInt>(Op2)) {
5995 ConstantInt *CInt = cast<ConstantInt>(Op2);
5996 if (CInt && CInt->getValue().isPowerOf2())
5997 Op2VP = TargetTransformInfo::OP_PowerOf2;
5998 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5999 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
6000 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
6001 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
6003 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
6004 if (CInt && CInt->getValue().isPowerOf2())
6005 Op2VP = TargetTransformInfo::OP_PowerOf2;
6006 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
6010 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
6013 case Instruction::Select: {
6014 SelectInst *SI = cast<SelectInst>(I);
6015 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
6016 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
6017 Type *CondTy = SI->getCondition()->getType();
6019 CondTy = VectorType::get(CondTy, VF);
6021 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
6023 case Instruction::ICmp:
6024 case Instruction::FCmp: {
6025 Type *ValTy = I->getOperand(0)->getType();
6026 VectorTy = ToVectorTy(ValTy, VF);
6027 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
6029 case Instruction::Store:
6030 case Instruction::Load: {
6031 StoreInst *SI = dyn_cast<StoreInst>(I);
6032 LoadInst *LI = dyn_cast<LoadInst>(I);
6033 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
6035 VectorTy = ToVectorTy(ValTy, VF);
6037 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
6038 unsigned AS = SI ? SI->getPointerAddressSpace() :
6039 LI->getPointerAddressSpace();
6040 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
6041 // We add the cost of address computation here instead of with the gep
6042 // instruction because only here we know whether the operation is
6045 return TTI.getAddressComputationCost(VectorTy) +
6046 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
6048 // Scalarized loads/stores.
6049 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
6050 bool Reverse = ConsecutiveStride < 0;
6051 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
6052 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
6053 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
6054 bool IsComplexComputation =
6055 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
6057 // The cost of extracting from the value vector and pointer vector.
6058 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
6059 for (unsigned i = 0; i < VF; ++i) {
6060 // The cost of extracting the pointer operand.
6061 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
6062 // In case of STORE, the cost of ExtractElement from the vector.
6063 // In case of LOAD, the cost of InsertElement into the returned
6065 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
6066 Instruction::InsertElement,
6070 // The cost of the scalar loads/stores.
6071 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
6072 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
6077 // Wide load/stores.
6078 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
6079 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
6082 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
6086 case Instruction::ZExt:
6087 case Instruction::SExt:
6088 case Instruction::FPToUI:
6089 case Instruction::FPToSI:
6090 case Instruction::FPExt:
6091 case Instruction::PtrToInt:
6092 case Instruction::IntToPtr:
6093 case Instruction::SIToFP:
6094 case Instruction::UIToFP:
6095 case Instruction::Trunc:
6096 case Instruction::FPTrunc:
6097 case Instruction::BitCast: {
6098 // We optimize the truncation of induction variable.
6099 // The cost of these is the same as the scalar operation.
6100 if (I->getOpcode() == Instruction::Trunc &&
6101 Legal->isInductionVariable(I->getOperand(0)))
6102 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
6103 I->getOperand(0)->getType());
6105 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
6106 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
6108 case Instruction::Call: {
6109 CallInst *CI = cast<CallInst>(I);
6110 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
6111 assert(ID && "Not an intrinsic call!");
6112 Type *RetTy = ToVectorTy(CI->getType(), VF);
6113 SmallVector<Type*, 4> Tys;
6114 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
6115 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
6116 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
6119 // We are scalarizing the instruction. Return the cost of the scalar
6120 // instruction, plus the cost of insert and extract into vector
6121 // elements, times the vector width.
6124 if (!RetTy->isVoidTy() && VF != 1) {
6125 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
6127 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
6130 // The cost of inserting the results plus extracting each one of the
6132 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
6135 // The cost of executing VF copies of the scalar instruction. This opcode
6136 // is unknown. Assume that it is the same as 'mul'.
6137 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
6143 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
6144 if (Scalar->isVoidTy() || VF == 1)
6146 return VectorType::get(Scalar, VF);
6149 char LoopVectorize::ID = 0;
6150 static const char lv_name[] = "Loop Vectorization";
6151 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
6152 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
6153 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
6154 INITIALIZE_PASS_DEPENDENCY(AssumptionTracker)
6155 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
6156 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
6157 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
6158 INITIALIZE_PASS_DEPENDENCY(LCSSA)
6159 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
6160 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
6161 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
6164 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
6165 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
6169 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
6170 // Check for a store.
6171 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
6172 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
6174 // Check for a load.
6175 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
6176 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
6182 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
6183 bool IfPredicateStore) {
6184 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
6185 // Holds vector parameters or scalars, in case of uniform vals.
6186 SmallVector<VectorParts, 4> Params;
6188 setDebugLocFromInst(Builder, Instr);
6190 // Find all of the vectorized parameters.
6191 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
6192 Value *SrcOp = Instr->getOperand(op);
6194 // If we are accessing the old induction variable, use the new one.
6195 if (SrcOp == OldInduction) {
6196 Params.push_back(getVectorValue(SrcOp));
6200 // Try using previously calculated values.
6201 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
6203 // If the src is an instruction that appeared earlier in the basic block
6204 // then it should already be vectorized.
6205 if (SrcInst && OrigLoop->contains(SrcInst)) {
6206 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
6207 // The parameter is a vector value from earlier.
6208 Params.push_back(WidenMap.get(SrcInst));
6210 // The parameter is a scalar from outside the loop. Maybe even a constant.
6211 VectorParts Scalars;
6212 Scalars.append(UF, SrcOp);
6213 Params.push_back(Scalars);
6217 assert(Params.size() == Instr->getNumOperands() &&
6218 "Invalid number of operands");
6220 // Does this instruction return a value ?
6221 bool IsVoidRetTy = Instr->getType()->isVoidTy();
6223 Value *UndefVec = IsVoidRetTy ? nullptr :
6224 UndefValue::get(Instr->getType());
6225 // Create a new entry in the WidenMap and initialize it to Undef or Null.
6226 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
6228 Instruction *InsertPt = Builder.GetInsertPoint();
6229 BasicBlock *IfBlock = Builder.GetInsertBlock();
6230 BasicBlock *CondBlock = nullptr;
6233 Loop *VectorLp = nullptr;
6234 if (IfPredicateStore) {
6235 assert(Instr->getParent()->getSinglePredecessor() &&
6236 "Only support single predecessor blocks");
6237 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
6238 Instr->getParent());
6239 VectorLp = LI->getLoopFor(IfBlock);
6240 assert(VectorLp && "Must have a loop for this block");
6243 // For each vector unroll 'part':
6244 for (unsigned Part = 0; Part < UF; ++Part) {
6245 // For each scalar that we create:
6247 // Start an "if (pred) a[i] = ..." block.
6248 Value *Cmp = nullptr;
6249 if (IfPredicateStore) {
6250 if (Cond[Part]->getType()->isVectorTy())
6252 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
6253 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
6254 ConstantInt::get(Cond[Part]->getType(), 1));
6255 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
6256 LoopVectorBody.push_back(CondBlock);
6257 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
6258 // Update Builder with newly created basic block.
6259 Builder.SetInsertPoint(InsertPt);
6262 Instruction *Cloned = Instr->clone();
6264 Cloned->setName(Instr->getName() + ".cloned");
6265 // Replace the operands of the cloned instructions with extracted scalars.
6266 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
6267 Value *Op = Params[op][Part];
6268 Cloned->setOperand(op, Op);
6271 // Place the cloned scalar in the new loop.
6272 Builder.Insert(Cloned);
6274 // If the original scalar returns a value we need to place it in a vector
6275 // so that future users will be able to use it.
6277 VecResults[Part] = Cloned;
6280 if (IfPredicateStore) {
6281 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
6282 LoopVectorBody.push_back(NewIfBlock);
6283 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
6284 Builder.SetInsertPoint(InsertPt);
6285 Instruction *OldBr = IfBlock->getTerminator();
6286 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
6287 OldBr->eraseFromParent();
6288 IfBlock = NewIfBlock;
6293 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
6294 StoreInst *SI = dyn_cast<StoreInst>(Instr);
6295 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
6297 return scalarizeInstruction(Instr, IfPredicateStore);
6300 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
6304 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
6308 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
6310 // When unrolling and the VF is 1, we only need to add a simple scalar.
6311 Type *ITy = Val->getType();
6312 assert(!ITy->isVectorTy() && "Val must be a scalar");
6313 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
6314 return Builder.CreateAdd(Val, C, "induction");