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 : NumLoads(0), NumStores(0), NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
585 DT(DT), TLI(TLI), AA(AA), TheFunction(F), Induction(nullptr),
586 WidestIndTy(nullptr), HasFunNoNaNAttr(false), MaxSafeDepDistBytes(-1U) {
589 /// This enum represents the kinds of reductions that we support.
591 RK_NoReduction, ///< Not a reduction.
592 RK_IntegerAdd, ///< Sum of integers.
593 RK_IntegerMult, ///< Product of integers.
594 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
595 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
596 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
597 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
598 RK_FloatAdd, ///< Sum of floats.
599 RK_FloatMult, ///< Product of floats.
600 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
603 /// This enum represents the kinds of inductions that we support.
605 IK_NoInduction, ///< Not an induction variable.
606 IK_IntInduction, ///< Integer induction variable. Step = 1.
607 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
608 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
609 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
612 // This enum represents the kind of minmax reduction.
613 enum MinMaxReductionKind {
623 /// This struct holds information about reduction variables.
624 struct ReductionDescriptor {
625 ReductionDescriptor() : StartValue(nullptr), LoopExitInstr(nullptr),
626 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
628 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
629 MinMaxReductionKind MK)
630 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
632 // The starting value of the reduction.
633 // It does not have to be zero!
634 TrackingVH<Value> StartValue;
635 // The instruction who's value is used outside the loop.
636 Instruction *LoopExitInstr;
637 // The kind of the reduction.
639 // If this a min/max reduction the kind of reduction.
640 MinMaxReductionKind MinMaxKind;
643 /// This POD struct holds information about a potential reduction operation.
644 struct ReductionInstDesc {
645 ReductionInstDesc(bool IsRedux, Instruction *I) :
646 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
648 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
649 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
651 // Is this instruction a reduction candidate.
653 // The last instruction in a min/max pattern (select of the select(icmp())
654 // pattern), or the current reduction instruction otherwise.
655 Instruction *PatternLastInst;
656 // If this is a min/max pattern the comparison predicate.
657 MinMaxReductionKind MinMaxKind;
660 /// This struct holds information about the memory runtime legality
661 /// check that a group of pointers do not overlap.
662 struct RuntimePointerCheck {
663 RuntimePointerCheck() : Need(false) {}
665 /// Reset the state of the pointer runtime information.
672 DependencySetId.clear();
676 /// Insert a pointer and calculate the start and end SCEVs.
677 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
678 unsigned DepSetId, unsigned ASId, ValueToValueMap &Strides);
680 /// This flag indicates if we need to add the runtime check.
682 /// Holds the pointers that we need to check.
683 SmallVector<TrackingVH<Value>, 2> Pointers;
684 /// Holds the pointer value at the beginning of the loop.
685 SmallVector<const SCEV*, 2> Starts;
686 /// Holds the pointer value at the end of the loop.
687 SmallVector<const SCEV*, 2> Ends;
688 /// Holds the information if this pointer is used for writing to memory.
689 SmallVector<bool, 2> IsWritePtr;
690 /// Holds the id of the set of pointers that could be dependent because of a
691 /// shared underlying object.
692 SmallVector<unsigned, 2> DependencySetId;
693 /// Holds the id of the disjoint alias set to which this pointer belongs.
694 SmallVector<unsigned, 2> AliasSetId;
697 /// A struct for saving information about induction variables.
698 struct InductionInfo {
699 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
700 InductionInfo() : StartValue(nullptr), IK(IK_NoInduction) {}
702 TrackingVH<Value> StartValue;
707 /// ReductionList contains the reduction descriptors for all
708 /// of the reductions that were found in the loop.
709 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
711 /// InductionList saves induction variables and maps them to the
712 /// induction descriptor.
713 typedef MapVector<PHINode*, InductionInfo> InductionList;
715 /// Returns true if it is legal to vectorize this loop.
716 /// This does not mean that it is profitable to vectorize this
717 /// loop, only that it is legal to do so.
720 /// Returns the Induction variable.
721 PHINode *getInduction() { return Induction; }
723 /// Returns the reduction variables found in the loop.
724 ReductionList *getReductionVars() { return &Reductions; }
726 /// Returns the induction variables found in the loop.
727 InductionList *getInductionVars() { return &Inductions; }
729 /// Returns the widest induction type.
730 Type *getWidestInductionType() { return WidestIndTy; }
732 /// Returns True if V is an induction variable in this loop.
733 bool isInductionVariable(const Value *V);
735 /// Return true if the block BB needs to be predicated in order for the loop
736 /// to be vectorized.
737 bool blockNeedsPredication(BasicBlock *BB);
739 /// Check if this pointer is consecutive when vectorizing. This happens
740 /// when the last index of the GEP is the induction variable, or that the
741 /// pointer itself is an induction variable.
742 /// This check allows us to vectorize A[idx] into a wide load/store.
744 /// 0 - Stride is unknown or non-consecutive.
745 /// 1 - Address is consecutive.
746 /// -1 - Address is consecutive, and decreasing.
747 int isConsecutivePtr(Value *Ptr);
749 /// Returns true if the value V is uniform within the loop.
750 bool isUniform(Value *V);
752 /// Returns true if this instruction will remain scalar after vectorization.
753 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
755 /// Returns the information that we collected about runtime memory check.
756 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
758 /// This function returns the identity element (or neutral element) for
760 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
762 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
764 bool hasStride(Value *V) { return StrideSet.count(V); }
765 bool mustCheckStrides() { return !StrideSet.empty(); }
766 SmallPtrSet<Value *, 8>::iterator strides_begin() {
767 return StrideSet.begin();
769 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
772 /// Check if a single basic block loop is vectorizable.
773 /// At this point we know that this is a loop with a constant trip count
774 /// and we only need to check individual instructions.
775 bool canVectorizeInstrs();
777 /// When we vectorize loops we may change the order in which
778 /// we read and write from memory. This method checks if it is
779 /// legal to vectorize the code, considering only memory constrains.
780 /// Returns true if the loop is vectorizable
781 bool canVectorizeMemory();
783 /// Return true if we can vectorize this loop using the IF-conversion
785 bool canVectorizeWithIfConvert();
787 /// Collect the variables that need to stay uniform after vectorization.
788 void collectLoopUniforms();
790 /// Return true if all of the instructions in the block can be speculatively
791 /// executed. \p SafePtrs is a list of addresses that are known to be legal
792 /// and we know that we can read from them without segfault.
793 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
795 /// Returns True, if 'Phi' is the kind of reduction variable for type
796 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
797 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
798 /// Returns a struct describing if the instruction 'I' can be a reduction
799 /// variable of type 'Kind'. If the reduction is a min/max pattern of
800 /// select(icmp()) this function advances the instruction pointer 'I' from the
801 /// compare instruction to the select instruction and stores this pointer in
802 /// 'PatternLastInst' member of the returned struct.
803 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
804 ReductionInstDesc &Desc);
805 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
806 /// pattern corresponding to a min(X, Y) or max(X, Y).
807 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
808 ReductionInstDesc &Prev);
809 /// Returns the induction kind of Phi. This function may return NoInduction
810 /// if the PHI is not an induction variable.
811 InductionKind isInductionVariable(PHINode *Phi);
813 /// \brief Collect memory access with loop invariant strides.
815 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
817 void collectStridedAcccess(Value *LoadOrStoreInst);
819 /// Report an analysis message to assist the user in diagnosing loops that are
821 void emitAnalysis(Report &Message) {
822 DebugLoc DL = TheLoop->getStartLoc();
823 if (Instruction *I = Message.getInstr())
824 DL = I->getDebugLoc();
825 emitOptimizationRemarkAnalysis(TheFunction->getContext(), DEBUG_TYPE,
826 *TheFunction, DL, Message.str());
829 /// The loop that we evaluate.
833 /// DataLayout analysis.
834 const DataLayout *DL;
837 /// Target Library Info.
838 TargetLibraryInfo *TLI;
842 Function *TheFunction;
844 // --- vectorization state --- //
846 /// Holds the integer induction variable. This is the counter of the
849 /// Holds the reduction variables.
850 ReductionList Reductions;
851 /// Holds all of the induction variables that we found in the loop.
852 /// Notice that inductions don't need to start at zero and that induction
853 /// variables can be pointers.
854 InductionList Inductions;
855 /// Holds the widest induction type encountered.
858 /// Allowed outside users. This holds the reduction
859 /// vars which can be accessed from outside the loop.
860 SmallPtrSet<Value*, 4> AllowedExit;
861 /// This set holds the variables which are known to be uniform after
863 SmallPtrSet<Instruction*, 4> Uniforms;
864 /// We need to check that all of the pointers in this list are disjoint
866 RuntimePointerCheck PtrRtCheck;
867 /// Can we assume the absence of NaNs.
868 bool HasFunNoNaNAttr;
870 unsigned MaxSafeDepDistBytes;
872 ValueToValueMap Strides;
873 SmallPtrSet<Value *, 8> StrideSet;
876 /// LoopVectorizationCostModel - estimates the expected speedups due to
878 /// In many cases vectorization is not profitable. This can happen because of
879 /// a number of reasons. In this class we mainly attempt to predict the
880 /// expected speedup/slowdowns due to the supported instruction set. We use the
881 /// TargetTransformInfo to query the different backends for the cost of
882 /// different operations.
883 class LoopVectorizationCostModel {
885 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
886 LoopVectorizationLegality *Legal,
887 const TargetTransformInfo &TTI,
888 const DataLayout *DL, const TargetLibraryInfo *TLI,
889 AssumptionTracker *AT, const Function *F,
890 const LoopVectorizeHints *Hints)
891 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI),
892 TheFunction(F), Hints(Hints) {
893 CodeMetrics::collectEphemeralValues(L, AT, EphValues);
896 /// Information about vectorization costs
897 struct VectorizationFactor {
898 unsigned Width; // Vector width with best cost
899 unsigned Cost; // Cost of the loop with that width
901 /// \return The most profitable vectorization factor and the cost of that VF.
902 /// This method checks every power of two up to VF. If UserVF is not ZERO
903 /// then this vectorization factor will be selected if vectorization is
905 VectorizationFactor selectVectorizationFactor(bool OptForSize);
907 /// \return The size (in bits) of the widest type in the code that
908 /// needs to be vectorized. We ignore values that remain scalar such as
909 /// 64 bit loop indices.
910 unsigned getWidestType();
912 /// \return The most profitable unroll factor.
913 /// If UserUF is non-zero then this method finds the best unroll-factor
914 /// based on register pressure and other parameters.
915 /// VF and LoopCost are the selected vectorization factor and the cost of the
917 unsigned selectUnrollFactor(bool OptForSize, unsigned VF, unsigned LoopCost);
919 /// \brief A struct that represents some properties of the register usage
921 struct RegisterUsage {
922 /// Holds the number of loop invariant values that are used in the loop.
923 unsigned LoopInvariantRegs;
924 /// Holds the maximum number of concurrent live intervals in the loop.
925 unsigned MaxLocalUsers;
926 /// Holds the number of instructions in the loop.
927 unsigned NumInstructions;
930 /// \return information about the register usage of the loop.
931 RegisterUsage calculateRegisterUsage();
934 /// Returns the expected execution cost. The unit of the cost does
935 /// not matter because we use the 'cost' units to compare different
936 /// vector widths. The cost that is returned is *not* normalized by
937 /// the factor width.
938 unsigned expectedCost(unsigned VF);
940 /// Returns the execution time cost of an instruction for a given vector
941 /// width. Vector width of one means scalar.
942 unsigned getInstructionCost(Instruction *I, unsigned VF);
944 /// A helper function for converting Scalar types to vector types.
945 /// If the incoming type is void, we return void. If the VF is 1, we return
947 static Type* ToVectorTy(Type *Scalar, unsigned VF);
949 /// Returns whether the instruction is a load or store and will be a emitted
950 /// as a vector operation.
951 bool isConsecutiveLoadOrStore(Instruction *I);
953 /// Report an analysis message to assist the user in diagnosing loops that are
955 void emitAnalysis(Report &Message) {
956 DebugLoc DL = TheLoop->getStartLoc();
957 if (Instruction *I = Message.getInstr())
958 DL = I->getDebugLoc();
959 emitOptimizationRemarkAnalysis(TheFunction->getContext(), DEBUG_TYPE,
960 *TheFunction, DL, Message.str());
963 /// Values used only by @llvm.assume calls.
964 SmallPtrSet<const Value *, 32> EphValues;
966 /// The loop that we evaluate.
970 /// Loop Info analysis.
972 /// Vectorization legality.
973 LoopVectorizationLegality *Legal;
974 /// Vector target information.
975 const TargetTransformInfo &TTI;
976 /// Target data layout information.
977 const DataLayout *DL;
978 /// Target Library Info.
979 const TargetLibraryInfo *TLI;
980 const Function *TheFunction;
981 // Loop Vectorize Hint.
982 const LoopVectorizeHints *Hints;
985 /// Utility class for getting and setting loop vectorizer hints in the form
986 /// of loop metadata.
987 /// This class keeps a number of loop annotations locally (as member variables)
988 /// and can, upon request, write them back as metadata on the loop. It will
989 /// initially scan the loop for existing metadata, and will update the local
990 /// values based on information in the loop.
991 /// We cannot write all values to metadata, as the mere presence of some info,
992 /// for example 'force', means a decision has been made. So, we need to be
993 /// careful NOT to add them if the user hasn't specifically asked so.
994 class LoopVectorizeHints {
1001 /// Hint - associates name and validation with the hint value.
1004 unsigned Value; // This may have to change for non-numeric values.
1007 Hint(const char * Name, unsigned Value, HintKind Kind)
1008 : Name(Name), Value(Value), Kind(Kind) { }
1010 bool validate(unsigned Val) {
1013 return isPowerOf2_32(Val) && Val <= MaxVectorWidth;
1015 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1023 /// Vectorization width.
1025 /// Vectorization interleave factor.
1027 /// Vectorization forced
1029 /// Array to help iterating through all hints.
1030 Hint *Hints[3]; // avoiding initialisation due to MSVC2012
1032 /// Return the loop metadata prefix.
1033 static StringRef Prefix() { return "llvm.loop."; }
1037 FK_Undefined = -1, ///< Not selected.
1038 FK_Disabled = 0, ///< Forcing disabled.
1039 FK_Enabled = 1, ///< Forcing enabled.
1042 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
1043 : Width("vectorize.width", VectorizationFactor, HK_WIDTH),
1044 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1045 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1047 // FIXME: Move this up initialisation when MSVC requirement is 2013+
1049 Hints[1] = &Interleave;
1052 // Populate values with existing loop metadata.
1053 getHintsFromMetadata();
1055 // force-vector-interleave overrides DisableInterleaving.
1056 if (VectorizationInterleave.getNumOccurrences() > 0)
1057 Interleave.Value = VectorizationInterleave;
1059 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
1060 << "LV: Interleaving disabled by the pass manager\n");
1063 /// Mark the loop L as already vectorized by setting the width to 1.
1064 void setAlreadyVectorized() {
1065 Width.Value = Interleave.Value = 1;
1066 // FIXME: Change all lines below for this when we can use MSVC 2013+
1067 //writeHintsToMetadata({ Width, Unroll });
1068 std::vector<Hint> hints;
1070 hints.emplace_back(Width);
1071 hints.emplace_back(Interleave);
1072 writeHintsToMetadata(std::move(hints));
1075 /// Dumps all the hint information.
1076 std::string emitRemark() const {
1078 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1079 R << "vectorization is explicitly disabled";
1081 R << "use -Rpass-analysis=loop-vectorize for more info";
1082 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1083 R << " (Force=true";
1084 if (Width.Value != 0)
1085 R << ", Vector Width=" << Width.Value;
1086 if (Interleave.Value != 0)
1087 R << ", Interleave Count=" << Interleave.Value;
1095 unsigned getWidth() const { return Width.Value; }
1096 unsigned getInterleave() const { return Interleave.Value; }
1097 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1100 /// Find hints specified in the loop metadata and update local values.
1101 void getHintsFromMetadata() {
1102 MDNode *LoopID = TheLoop->getLoopID();
1106 // First operand should refer to the loop id itself.
1107 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1108 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1110 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1111 const MDString *S = nullptr;
1112 SmallVector<Value*, 4> Args;
1114 // The expected hint is either a MDString or a MDNode with the first
1115 // operand a MDString.
1116 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1117 if (!MD || MD->getNumOperands() == 0)
1119 S = dyn_cast<MDString>(MD->getOperand(0));
1120 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1121 Args.push_back(MD->getOperand(i));
1123 S = dyn_cast<MDString>(LoopID->getOperand(i));
1124 assert(Args.size() == 0 && "too many arguments for MDString");
1130 // Check if the hint starts with the loop metadata prefix.
1131 StringRef Name = S->getString();
1132 if (Args.size() == 1)
1133 setHint(Name, Args[0]);
1137 /// Checks string hint with one operand and set value if valid.
1138 void setHint(StringRef Name, Value *Arg) {
1139 if (!Name.startswith(Prefix()))
1141 Name = Name.substr(Prefix().size(), StringRef::npos);
1143 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
1145 unsigned Val = C->getZExtValue();
1147 for (auto H : Hints) {
1148 if (Name == H->Name) {
1149 if (H->validate(Val))
1152 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1158 /// Create a new hint from name / value pair.
1159 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1160 LLVMContext &Context = TheLoop->getHeader()->getContext();
1161 SmallVector<Value*, 2> Vals;
1162 Vals.push_back(MDString::get(Context, Name));
1163 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
1164 return MDNode::get(Context, Vals);
1167 /// Matches metadata with hint name.
1168 bool matchesHintMetadataName(MDNode *Node, std::vector<Hint> &HintTypes) {
1169 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1173 for (auto H : HintTypes)
1174 if (Name->getName().endswith(H.Name))
1179 /// Sets current hints into loop metadata, keeping other values intact.
1180 void writeHintsToMetadata(std::vector<Hint> HintTypes) {
1181 if (HintTypes.size() == 0)
1184 // Reserve the first element to LoopID (see below).
1185 SmallVector<Value*, 4> Vals(1);
1186 // If the loop already has metadata, then ignore the existing operands.
1187 MDNode *LoopID = TheLoop->getLoopID();
1189 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1190 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1191 // If node in update list, ignore old value.
1192 if (!matchesHintMetadataName(Node, HintTypes))
1193 Vals.push_back(Node);
1197 // Now, add the missing hints.
1198 for (auto H : HintTypes)
1200 createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1202 // Replace current metadata node with new one.
1203 LLVMContext &Context = TheLoop->getHeader()->getContext();
1204 MDNode *NewLoopID = MDNode::get(Context, Vals);
1205 // Set operand 0 to refer to the loop id itself.
1206 NewLoopID->replaceOperandWith(0, NewLoopID);
1208 TheLoop->setLoopID(NewLoopID);
1210 LoopID->replaceAllUsesWith(NewLoopID);
1214 /// The loop these hints belong to.
1215 const Loop *TheLoop;
1218 static void emitMissedWarning(Function *F, Loop *L,
1219 const LoopVectorizeHints &LH) {
1220 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1221 L->getStartLoc(), LH.emitRemark());
1223 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1224 if (LH.getWidth() != 1)
1225 emitLoopVectorizeWarning(
1226 F->getContext(), *F, L->getStartLoc(),
1227 "failed explicitly specified loop vectorization");
1228 else if (LH.getInterleave() != 1)
1229 emitLoopInterleaveWarning(
1230 F->getContext(), *F, L->getStartLoc(),
1231 "failed explicitly specified loop interleaving");
1235 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1237 return V.push_back(&L);
1239 for (Loop *InnerL : L)
1240 addInnerLoop(*InnerL, V);
1243 /// The LoopVectorize Pass.
1244 struct LoopVectorize : public FunctionPass {
1245 /// Pass identification, replacement for typeid
1248 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1250 DisableUnrolling(NoUnrolling),
1251 AlwaysVectorize(AlwaysVectorize) {
1252 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1255 ScalarEvolution *SE;
1256 const DataLayout *DL;
1258 TargetTransformInfo *TTI;
1260 BlockFrequencyInfo *BFI;
1261 TargetLibraryInfo *TLI;
1263 AssumptionTracker *AT;
1264 bool DisableUnrolling;
1265 bool AlwaysVectorize;
1267 BlockFrequency ColdEntryFreq;
1269 bool runOnFunction(Function &F) override {
1270 SE = &getAnalysis<ScalarEvolution>();
1271 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1272 DL = DLP ? &DLP->getDataLayout() : nullptr;
1273 LI = &getAnalysis<LoopInfo>();
1274 TTI = &getAnalysis<TargetTransformInfo>();
1275 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1276 BFI = &getAnalysis<BlockFrequencyInfo>();
1277 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
1278 AA = &getAnalysis<AliasAnalysis>();
1279 AT = &getAnalysis<AssumptionTracker>();
1281 // Compute some weights outside of the loop over the loops. Compute this
1282 // using a BranchProbability to re-use its scaling math.
1283 const BranchProbability ColdProb(1, 5); // 20%
1284 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1286 // If the target claims to have no vector registers don't attempt
1288 if (!TTI->getNumberOfRegisters(true))
1292 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1293 << ": Missing data layout\n");
1297 // Build up a worklist of inner-loops to vectorize. This is necessary as
1298 // the act of vectorizing or partially unrolling a loop creates new loops
1299 // and can invalidate iterators across the loops.
1300 SmallVector<Loop *, 8> Worklist;
1303 addInnerLoop(*L, Worklist);
1305 LoopsAnalyzed += Worklist.size();
1307 // Now walk the identified inner loops.
1308 bool Changed = false;
1309 while (!Worklist.empty())
1310 Changed |= processLoop(Worklist.pop_back_val());
1312 // Process each loop nest in the function.
1316 bool processLoop(Loop *L) {
1317 assert(L->empty() && "Only process inner loops.");
1320 const std::string DebugLocStr = getDebugLocString(L);
1323 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1324 << L->getHeader()->getParent()->getName() << "\" from "
1325 << DebugLocStr << "\n");
1327 LoopVectorizeHints Hints(L, DisableUnrolling);
1329 DEBUG(dbgs() << "LV: Loop hints:"
1331 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1333 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1335 : "?")) << " width=" << Hints.getWidth()
1336 << " unroll=" << Hints.getInterleave() << "\n");
1338 // Function containing loop
1339 Function *F = L->getHeader()->getParent();
1341 // Looking at the diagnostic output is the only way to determine if a loop
1342 // was vectorized (other than looking at the IR or machine code), so it
1343 // is important to generate an optimization remark for each loop. Most of
1344 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1345 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1346 // less verbose reporting vectorized loops and unvectorized loops that may
1347 // benefit from vectorization, respectively.
1349 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1350 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1351 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1352 L->getStartLoc(), Hints.emitRemark());
1356 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1357 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1358 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1359 L->getStartLoc(), Hints.emitRemark());
1363 if (Hints.getWidth() == 1 && Hints.getInterleave() == 1) {
1364 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1365 emitOptimizationRemarkAnalysis(
1366 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1367 "loop not vectorized: vector width and interleave count are "
1368 "explicitly set to 1");
1372 // Check the loop for a trip count threshold:
1373 // do not vectorize loops with a tiny trip count.
1374 const unsigned TC = SE->getSmallConstantTripCount(L);
1375 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1376 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1377 << "This loop is not worth vectorizing.");
1378 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1379 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1381 DEBUG(dbgs() << "\n");
1382 emitOptimizationRemarkAnalysis(
1383 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1384 "vectorization is not beneficial and is not explicitly forced");
1389 // Check if it is legal to vectorize the loop.
1390 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, AA, F);
1391 if (!LVL.canVectorize()) {
1392 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1393 emitMissedWarning(F, L, Hints);
1397 // Use the cost model.
1398 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI, AT, F,
1401 // Check the function attributes to find out if this function should be
1402 // optimized for size.
1403 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1404 F->hasFnAttribute(Attribute::OptimizeForSize);
1406 // Compute the weighted frequency of this loop being executed and see if it
1407 // is less than 20% of the function entry baseline frequency. Note that we
1408 // always have a canonical loop here because we think we *can* vectoriez.
1409 // FIXME: This is hidden behind a flag due to pervasive problems with
1410 // exactly what block frequency models.
1411 if (LoopVectorizeWithBlockFrequency) {
1412 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1413 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1414 LoopEntryFreq < ColdEntryFreq)
1418 // Check the function attributes to see if implicit floats are allowed.a
1419 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1420 // an integer loop and the vector instructions selected are purely integer
1421 // vector instructions?
1422 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1423 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1424 "attribute is used.\n");
1425 emitOptimizationRemarkAnalysis(
1426 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1427 "loop not vectorized due to NoImplicitFloat attribute");
1428 emitMissedWarning(F, L, Hints);
1432 // Select the optimal vectorization factor.
1433 const LoopVectorizationCostModel::VectorizationFactor VF =
1434 CM.selectVectorizationFactor(OptForSize);
1436 // Select the unroll factor.
1438 CM.selectUnrollFactor(OptForSize, VF.Width, VF.Cost);
1440 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1441 << DebugLocStr << '\n');
1442 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1444 if (VF.Width == 1) {
1445 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
1448 emitOptimizationRemarkAnalysis(
1449 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1450 "not beneficial to vectorize and user disabled interleaving");
1453 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1455 // Report the unrolling decision.
1456 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1457 Twine("unrolled with interleaving factor " +
1459 " (vectorization not beneficial)"));
1461 // We decided not to vectorize, but we may want to unroll.
1463 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1464 Unroller.vectorize(&LVL);
1466 // If we decided that it is *legal* to vectorize the loop then do it.
1467 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1471 // Report the vectorization decision.
1472 emitOptimizationRemark(
1473 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1474 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1475 ", unrolling interleave factor: " + Twine(UF) + ")");
1478 // Mark the loop as already vectorized to avoid vectorizing again.
1479 Hints.setAlreadyVectorized();
1481 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1485 void getAnalysisUsage(AnalysisUsage &AU) const override {
1486 AU.addRequired<AssumptionTracker>();
1487 AU.addRequiredID(LoopSimplifyID);
1488 AU.addRequiredID(LCSSAID);
1489 AU.addRequired<BlockFrequencyInfo>();
1490 AU.addRequired<DominatorTreeWrapperPass>();
1491 AU.addRequired<LoopInfo>();
1492 AU.addRequired<ScalarEvolution>();
1493 AU.addRequired<TargetTransformInfo>();
1494 AU.addRequired<AliasAnalysis>();
1495 AU.addPreserved<LoopInfo>();
1496 AU.addPreserved<DominatorTreeWrapperPass>();
1497 AU.addPreserved<AliasAnalysis>();
1502 } // end anonymous namespace
1504 //===----------------------------------------------------------------------===//
1505 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1506 // LoopVectorizationCostModel.
1507 //===----------------------------------------------------------------------===//
1509 static Value *stripIntegerCast(Value *V) {
1510 if (CastInst *CI = dyn_cast<CastInst>(V))
1511 if (CI->getOperand(0)->getType()->isIntegerTy())
1512 return CI->getOperand(0);
1516 ///\brief Replaces the symbolic stride in a pointer SCEV expression by one.
1518 /// If \p OrigPtr is not null, use it to look up the stride value instead of
1520 static const SCEV *replaceSymbolicStrideSCEV(ScalarEvolution *SE,
1521 ValueToValueMap &PtrToStride,
1522 Value *Ptr, Value *OrigPtr = nullptr) {
1524 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
1526 // If there is an entry in the map return the SCEV of the pointer with the
1527 // symbolic stride replaced by one.
1528 ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
1529 if (SI != PtrToStride.end()) {
1530 Value *StrideVal = SI->second;
1533 StrideVal = stripIntegerCast(StrideVal);
1535 // Replace symbolic stride by one.
1536 Value *One = ConstantInt::get(StrideVal->getType(), 1);
1537 ValueToValueMap RewriteMap;
1538 RewriteMap[StrideVal] = One;
1541 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
1542 DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
1547 // Otherwise, just return the SCEV of the original pointer.
1548 return SE->getSCEV(Ptr);
1551 void LoopVectorizationLegality::RuntimePointerCheck::insert(
1552 ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
1553 unsigned ASId, ValueToValueMap &Strides) {
1554 // Get the stride replaced scev.
1555 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
1556 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1557 assert(AR && "Invalid addrec expression");
1558 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1559 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1560 Pointers.push_back(Ptr);
1561 Starts.push_back(AR->getStart());
1562 Ends.push_back(ScEnd);
1563 IsWritePtr.push_back(WritePtr);
1564 DependencySetId.push_back(DepSetId);
1565 AliasSetId.push_back(ASId);
1568 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1569 // We need to place the broadcast of invariant variables outside the loop.
1570 Instruction *Instr = dyn_cast<Instruction>(V);
1572 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1573 Instr->getParent()) != LoopVectorBody.end());
1574 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1576 // Place the code for broadcasting invariant variables in the new preheader.
1577 IRBuilder<>::InsertPointGuard Guard(Builder);
1579 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1581 // Broadcast the scalar into all locations in the vector.
1582 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1587 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1589 assert(Val->getType()->isVectorTy() && "Must be a vector");
1590 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1591 "Elem must be an integer");
1592 // Create the types.
1593 Type *ITy = Val->getType()->getScalarType();
1594 VectorType *Ty = cast<VectorType>(Val->getType());
1595 int VLen = Ty->getNumElements();
1596 SmallVector<Constant*, 8> Indices;
1598 // Create a vector of consecutive numbers from zero to VF.
1599 for (int i = 0; i < VLen; ++i) {
1600 int64_t Idx = Negate ? (-i) : i;
1601 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1604 // Add the consecutive indices to the vector value.
1605 Constant *Cv = ConstantVector::get(Indices);
1606 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1607 return Builder.CreateAdd(Val, Cv, "induction");
1610 /// \brief Find the operand of the GEP that should be checked for consecutive
1611 /// stores. This ignores trailing indices that have no effect on the final
1613 static unsigned getGEPInductionOperand(const DataLayout *DL,
1614 const GetElementPtrInst *Gep) {
1615 unsigned LastOperand = Gep->getNumOperands() - 1;
1616 unsigned GEPAllocSize = DL->getTypeAllocSize(
1617 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1619 // Walk backwards and try to peel off zeros.
1620 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1621 // Find the type we're currently indexing into.
1622 gep_type_iterator GEPTI = gep_type_begin(Gep);
1623 std::advance(GEPTI, LastOperand - 1);
1625 // If it's a type with the same allocation size as the result of the GEP we
1626 // can peel off the zero index.
1627 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1635 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1636 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1637 // Make sure that the pointer does not point to structs.
1638 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1641 // If this value is a pointer induction variable we know it is consecutive.
1642 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1643 if (Phi && Inductions.count(Phi)) {
1644 InductionInfo II = Inductions[Phi];
1645 if (IK_PtrInduction == II.IK)
1647 else if (IK_ReversePtrInduction == II.IK)
1651 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1655 unsigned NumOperands = Gep->getNumOperands();
1656 Value *GpPtr = Gep->getPointerOperand();
1657 // If this GEP value is a consecutive pointer induction variable and all of
1658 // the indices are constant then we know it is consecutive. We can
1659 Phi = dyn_cast<PHINode>(GpPtr);
1660 if (Phi && Inductions.count(Phi)) {
1662 // Make sure that the pointer does not point to structs.
1663 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1664 if (GepPtrType->getElementType()->isAggregateType())
1667 // Make sure that all of the index operands are loop invariant.
1668 for (unsigned i = 1; i < NumOperands; ++i)
1669 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1672 InductionInfo II = Inductions[Phi];
1673 if (IK_PtrInduction == II.IK)
1675 else if (IK_ReversePtrInduction == II.IK)
1679 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1681 // Check that all of the gep indices are uniform except for our induction
1683 for (unsigned i = 0; i != NumOperands; ++i)
1684 if (i != InductionOperand &&
1685 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1688 // We can emit wide load/stores only if the last non-zero index is the
1689 // induction variable.
1690 const SCEV *Last = nullptr;
1691 if (!Strides.count(Gep))
1692 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1694 // Because of the multiplication by a stride we can have a s/zext cast.
1695 // We are going to replace this stride by 1 so the cast is safe to ignore.
1697 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1698 // %0 = trunc i64 %indvars.iv to i32
1699 // %mul = mul i32 %0, %Stride1
1700 // %idxprom = zext i32 %mul to i64 << Safe cast.
1701 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1703 Last = replaceSymbolicStrideSCEV(SE, Strides,
1704 Gep->getOperand(InductionOperand), Gep);
1705 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1707 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1711 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1712 const SCEV *Step = AR->getStepRecurrence(*SE);
1714 // The memory is consecutive because the last index is consecutive
1715 // and all other indices are loop invariant.
1718 if (Step->isAllOnesValue())
1725 bool LoopVectorizationLegality::isUniform(Value *V) {
1726 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1729 InnerLoopVectorizer::VectorParts&
1730 InnerLoopVectorizer::getVectorValue(Value *V) {
1731 assert(V != Induction && "The new induction variable should not be used.");
1732 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1734 // If we have a stride that is replaced by one, do it here.
1735 if (Legal->hasStride(V))
1736 V = ConstantInt::get(V->getType(), 1);
1738 // If we have this scalar in the map, return it.
1739 if (WidenMap.has(V))
1740 return WidenMap.get(V);
1742 // If this scalar is unknown, assume that it is a constant or that it is
1743 // loop invariant. Broadcast V and save the value for future uses.
1744 Value *B = getBroadcastInstrs(V);
1745 return WidenMap.splat(V, B);
1748 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1749 assert(Vec->getType()->isVectorTy() && "Invalid type");
1750 SmallVector<Constant*, 8> ShuffleMask;
1751 for (unsigned i = 0; i < VF; ++i)
1752 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1754 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1755 ConstantVector::get(ShuffleMask),
1759 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1760 // Attempt to issue a wide load.
1761 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1762 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1764 assert((LI || SI) && "Invalid Load/Store instruction");
1766 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1767 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1768 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1769 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1770 // An alignment of 0 means target abi alignment. We need to use the scalar's
1771 // target abi alignment in such a case.
1773 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1774 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1775 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1776 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1778 if (SI && Legal->blockNeedsPredication(SI->getParent()))
1779 return scalarizeInstruction(Instr, true);
1781 if (ScalarAllocatedSize != VectorElementSize)
1782 return scalarizeInstruction(Instr);
1784 // If the pointer is loop invariant or if it is non-consecutive,
1785 // scalarize the load.
1786 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1787 bool Reverse = ConsecutiveStride < 0;
1788 bool UniformLoad = LI && Legal->isUniform(Ptr);
1789 if (!ConsecutiveStride || UniformLoad)
1790 return scalarizeInstruction(Instr);
1792 Constant *Zero = Builder.getInt32(0);
1793 VectorParts &Entry = WidenMap.get(Instr);
1795 // Handle consecutive loads/stores.
1796 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1797 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1798 setDebugLocFromInst(Builder, Gep);
1799 Value *PtrOperand = Gep->getPointerOperand();
1800 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1801 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1803 // Create the new GEP with the new induction variable.
1804 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1805 Gep2->setOperand(0, FirstBasePtr);
1806 Gep2->setName("gep.indvar.base");
1807 Ptr = Builder.Insert(Gep2);
1809 setDebugLocFromInst(Builder, Gep);
1810 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1811 OrigLoop) && "Base ptr must be invariant");
1813 // The last index does not have to be the induction. It can be
1814 // consecutive and be a function of the index. For example A[I+1];
1815 unsigned NumOperands = Gep->getNumOperands();
1816 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1817 // Create the new GEP with the new induction variable.
1818 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1820 for (unsigned i = 0; i < NumOperands; ++i) {
1821 Value *GepOperand = Gep->getOperand(i);
1822 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1824 // Update last index or loop invariant instruction anchored in loop.
1825 if (i == InductionOperand ||
1826 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1827 assert((i == InductionOperand ||
1828 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1829 "Must be last index or loop invariant");
1831 VectorParts &GEPParts = getVectorValue(GepOperand);
1832 Value *Index = GEPParts[0];
1833 Index = Builder.CreateExtractElement(Index, Zero);
1834 Gep2->setOperand(i, Index);
1835 Gep2->setName("gep.indvar.idx");
1838 Ptr = Builder.Insert(Gep2);
1840 // Use the induction element ptr.
1841 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1842 setDebugLocFromInst(Builder, Ptr);
1843 VectorParts &PtrVal = getVectorValue(Ptr);
1844 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1849 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1850 "We do not allow storing to uniform addresses");
1851 setDebugLocFromInst(Builder, SI);
1852 // We don't want to update the value in the map as it might be used in
1853 // another expression. So don't use a reference type for "StoredVal".
1854 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1856 for (unsigned Part = 0; Part < UF; ++Part) {
1857 // Calculate the pointer for the specific unroll-part.
1858 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1861 // If we store to reverse consecutive memory locations then we need
1862 // to reverse the order of elements in the stored value.
1863 StoredVal[Part] = reverseVector(StoredVal[Part]);
1864 // If the address is consecutive but reversed, then the
1865 // wide store needs to start at the last vector element.
1866 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1867 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1870 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1871 DataTy->getPointerTo(AddressSpace));
1873 Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
1874 propagateMetadata(NewSI, SI);
1880 assert(LI && "Must have a load instruction");
1881 setDebugLocFromInst(Builder, LI);
1882 for (unsigned Part = 0; Part < UF; ++Part) {
1883 // Calculate the pointer for the specific unroll-part.
1884 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1887 // If the address is consecutive but reversed, then the
1888 // wide store needs to start at the last vector element.
1889 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1890 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1893 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1894 DataTy->getPointerTo(AddressSpace));
1895 LoadInst *NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
1896 propagateMetadata(NewLI, LI);
1897 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
1901 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1902 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1903 // Holds vector parameters or scalars, in case of uniform vals.
1904 SmallVector<VectorParts, 4> Params;
1906 setDebugLocFromInst(Builder, Instr);
1908 // Find all of the vectorized parameters.
1909 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1910 Value *SrcOp = Instr->getOperand(op);
1912 // If we are accessing the old induction variable, use the new one.
1913 if (SrcOp == OldInduction) {
1914 Params.push_back(getVectorValue(SrcOp));
1918 // Try using previously calculated values.
1919 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1921 // If the src is an instruction that appeared earlier in the basic block
1922 // then it should already be vectorized.
1923 if (SrcInst && OrigLoop->contains(SrcInst)) {
1924 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1925 // The parameter is a vector value from earlier.
1926 Params.push_back(WidenMap.get(SrcInst));
1928 // The parameter is a scalar from outside the loop. Maybe even a constant.
1929 VectorParts Scalars;
1930 Scalars.append(UF, SrcOp);
1931 Params.push_back(Scalars);
1935 assert(Params.size() == Instr->getNumOperands() &&
1936 "Invalid number of operands");
1938 // Does this instruction return a value ?
1939 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1941 Value *UndefVec = IsVoidRetTy ? nullptr :
1942 UndefValue::get(VectorType::get(Instr->getType(), VF));
1943 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1944 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1946 Instruction *InsertPt = Builder.GetInsertPoint();
1947 BasicBlock *IfBlock = Builder.GetInsertBlock();
1948 BasicBlock *CondBlock = nullptr;
1951 Loop *VectorLp = nullptr;
1952 if (IfPredicateStore) {
1953 assert(Instr->getParent()->getSinglePredecessor() &&
1954 "Only support single predecessor blocks");
1955 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1956 Instr->getParent());
1957 VectorLp = LI->getLoopFor(IfBlock);
1958 assert(VectorLp && "Must have a loop for this block");
1961 // For each vector unroll 'part':
1962 for (unsigned Part = 0; Part < UF; ++Part) {
1963 // For each scalar that we create:
1964 for (unsigned Width = 0; Width < VF; ++Width) {
1967 Value *Cmp = nullptr;
1968 if (IfPredicateStore) {
1969 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1970 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1971 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1972 LoopVectorBody.push_back(CondBlock);
1973 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
1974 // Update Builder with newly created basic block.
1975 Builder.SetInsertPoint(InsertPt);
1978 Instruction *Cloned = Instr->clone();
1980 Cloned->setName(Instr->getName() + ".cloned");
1981 // Replace the operands of the cloned instructions with extracted scalars.
1982 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1983 Value *Op = Params[op][Part];
1984 // Param is a vector. Need to extract the right lane.
1985 if (Op->getType()->isVectorTy())
1986 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1987 Cloned->setOperand(op, Op);
1990 // Place the cloned scalar in the new loop.
1991 Builder.Insert(Cloned);
1993 // If the original scalar returns a value we need to place it in a vector
1994 // so that future users will be able to use it.
1996 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1997 Builder.getInt32(Width));
1999 if (IfPredicateStore) {
2000 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2001 LoopVectorBody.push_back(NewIfBlock);
2002 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
2003 Builder.SetInsertPoint(InsertPt);
2004 Instruction *OldBr = IfBlock->getTerminator();
2005 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2006 OldBr->eraseFromParent();
2007 IfBlock = NewIfBlock;
2013 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
2017 if (Instruction *I = dyn_cast<Instruction>(V))
2018 return I->getParent() == Loc->getParent() ? I : nullptr;
2022 std::pair<Instruction *, Instruction *>
2023 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
2024 Instruction *tnullptr = nullptr;
2025 if (!Legal->mustCheckStrides())
2026 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2028 IRBuilder<> ChkBuilder(Loc);
2031 Value *Check = nullptr;
2032 Instruction *FirstInst = nullptr;
2033 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
2034 SE = Legal->strides_end();
2036 Value *Ptr = stripIntegerCast(*SI);
2037 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
2039 // Store the first instruction we create.
2040 FirstInst = getFirstInst(FirstInst, C, Loc);
2042 Check = ChkBuilder.CreateOr(Check, C);
2047 // We have to do this trickery because the IRBuilder might fold the check to a
2048 // constant expression in which case there is no Instruction anchored in a
2050 LLVMContext &Ctx = Loc->getContext();
2051 Instruction *TheCheck =
2052 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2053 ChkBuilder.Insert(TheCheck, "stride.not.one");
2054 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2056 return std::make_pair(FirstInst, TheCheck);
2059 std::pair<Instruction *, Instruction *>
2060 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
2061 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
2062 Legal->getRuntimePointerCheck();
2064 Instruction *tnullptr = nullptr;
2065 if (!PtrRtCheck->Need)
2066 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2068 unsigned NumPointers = PtrRtCheck->Pointers.size();
2069 SmallVector<TrackingVH<Value> , 2> Starts;
2070 SmallVector<TrackingVH<Value> , 2> Ends;
2072 LLVMContext &Ctx = Loc->getContext();
2073 SCEVExpander Exp(*SE, "induction");
2074 Instruction *FirstInst = nullptr;
2076 for (unsigned i = 0; i < NumPointers; ++i) {
2077 Value *Ptr = PtrRtCheck->Pointers[i];
2078 const SCEV *Sc = SE->getSCEV(Ptr);
2080 if (SE->isLoopInvariant(Sc, OrigLoop)) {
2081 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
2083 Starts.push_back(Ptr);
2084 Ends.push_back(Ptr);
2086 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
2087 unsigned AS = Ptr->getType()->getPointerAddressSpace();
2089 // Use this type for pointer arithmetic.
2090 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
2092 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
2093 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
2094 Starts.push_back(Start);
2095 Ends.push_back(End);
2099 IRBuilder<> ChkBuilder(Loc);
2100 // Our instructions might fold to a constant.
2101 Value *MemoryRuntimeCheck = nullptr;
2102 for (unsigned i = 0; i < NumPointers; ++i) {
2103 for (unsigned j = i+1; j < NumPointers; ++j) {
2104 // No need to check if two readonly pointers intersect.
2105 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
2108 // Only need to check pointers between two different dependency sets.
2109 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
2111 // Only need to check pointers in the same alias set.
2112 if (PtrRtCheck->AliasSetId[i] != PtrRtCheck->AliasSetId[j])
2115 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
2116 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
2118 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
2119 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
2120 "Trying to bounds check pointers with different address spaces");
2122 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
2123 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
2125 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
2126 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
2127 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
2128 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
2130 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
2131 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
2132 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
2133 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
2134 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
2135 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2136 if (MemoryRuntimeCheck) {
2137 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
2139 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2141 MemoryRuntimeCheck = IsConflict;
2145 // We have to do this trickery because the IRBuilder might fold the check to a
2146 // constant expression in which case there is no Instruction anchored in a
2148 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
2149 ConstantInt::getTrue(Ctx));
2150 ChkBuilder.Insert(Check, "memcheck.conflict");
2151 FirstInst = getFirstInst(FirstInst, Check, Loc);
2152 return std::make_pair(FirstInst, Check);
2155 void InnerLoopVectorizer::createEmptyLoop() {
2157 In this function we generate a new loop. The new loop will contain
2158 the vectorized instructions while the old loop will continue to run the
2161 [ ] <-- Back-edge taken count overflow check.
2164 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2167 || [ ] <-- vector pre header.
2171 || [ ]_| <-- vector loop.
2174 | >[ ] <--- middle-block.
2177 -|- >[ ] <--- new preheader.
2181 | [ ]_| <-- old scalar loop to handle remainder.
2184 >[ ] <-- exit block.
2188 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2189 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2190 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2191 assert(BypassBlock && "Invalid loop structure");
2192 assert(ExitBlock && "Must have an exit block");
2194 // Some loops have a single integer induction variable, while other loops
2195 // don't. One example is c++ iterators that often have multiple pointer
2196 // induction variables. In the code below we also support a case where we
2197 // don't have a single induction variable.
2198 OldInduction = Legal->getInduction();
2199 Type *IdxTy = Legal->getWidestInductionType();
2201 // Find the loop boundaries.
2202 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2203 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2205 // The exit count might have the type of i64 while the phi is i32. This can
2206 // happen if we have an induction variable that is sign extended before the
2207 // compare. The only way that we get a backedge taken count is that the
2208 // induction variable was signed and as such will not overflow. In such a case
2209 // truncation is legal.
2210 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2211 IdxTy->getPrimitiveSizeInBits())
2212 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2214 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2215 // Get the total trip count from the count by adding 1.
2216 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2217 SE->getConstant(BackedgeTakeCount->getType(), 1));
2219 // Expand the trip count and place the new instructions in the preheader.
2220 // Notice that the pre-header does not change, only the loop body.
2221 SCEVExpander Exp(*SE, "induction");
2223 // We need to test whether the backedge-taken count is uint##_max. Adding one
2224 // to it will cause overflow and an incorrect loop trip count in the vector
2225 // body. In case of overflow we want to directly jump to the scalar remainder
2227 Value *BackedgeCount =
2228 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2229 BypassBlock->getTerminator());
2230 if (BackedgeCount->getType()->isPointerTy())
2231 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2232 "backedge.ptrcnt.to.int",
2233 BypassBlock->getTerminator());
2234 Instruction *CheckBCOverflow =
2235 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2236 Constant::getAllOnesValue(BackedgeCount->getType()),
2237 "backedge.overflow", BypassBlock->getTerminator());
2239 // The loop index does not have to start at Zero. Find the original start
2240 // value from the induction PHI node. If we don't have an induction variable
2241 // then we know that it starts at zero.
2242 Builder.SetInsertPoint(BypassBlock->getTerminator());
2243 Value *StartIdx = ExtendedIdx = OldInduction ?
2244 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2246 ConstantInt::get(IdxTy, 0);
2248 // We need an instruction to anchor the overflow check on. StartIdx needs to
2249 // be defined before the overflow check branch. Because the scalar preheader
2250 // is going to merge the start index and so the overflow branch block needs to
2251 // contain a definition of the start index.
2252 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2253 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2254 BypassBlock->getTerminator());
2256 // Count holds the overall loop count (N).
2257 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2258 BypassBlock->getTerminator());
2260 LoopBypassBlocks.push_back(BypassBlock);
2262 // Split the single block loop into the two loop structure described above.
2263 BasicBlock *VectorPH =
2264 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2265 BasicBlock *VecBody =
2266 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2267 BasicBlock *MiddleBlock =
2268 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2269 BasicBlock *ScalarPH =
2270 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2272 // Create and register the new vector loop.
2273 Loop* Lp = new Loop();
2274 Loop *ParentLoop = OrigLoop->getParentLoop();
2276 // Insert the new loop into the loop nest and register the new basic blocks
2277 // before calling any utilities such as SCEV that require valid LoopInfo.
2279 ParentLoop->addChildLoop(Lp);
2280 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
2281 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
2282 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
2284 LI->addTopLevelLoop(Lp);
2286 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
2288 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2290 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2292 // Generate the induction variable.
2293 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2294 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2295 // The loop step is equal to the vectorization factor (num of SIMD elements)
2296 // times the unroll factor (num of SIMD instructions).
2297 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2299 // This is the IR builder that we use to add all of the logic for bypassing
2300 // the new vector loop.
2301 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2302 setDebugLocFromInst(BypassBuilder,
2303 getDebugLocFromInstOrOperands(OldInduction));
2305 // We may need to extend the index in case there is a type mismatch.
2306 // We know that the count starts at zero and does not overflow.
2307 if (Count->getType() != IdxTy) {
2308 // The exit count can be of pointer type. Convert it to the correct
2310 if (ExitCount->getType()->isPointerTy())
2311 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2313 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2316 // Add the start index to the loop count to get the new end index.
2317 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2319 // Now we need to generate the expression for N - (N % VF), which is
2320 // the part that the vectorized body will execute.
2321 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2322 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2323 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2324 "end.idx.rnd.down");
2326 // Now, compare the new count to zero. If it is zero skip the vector loop and
2327 // jump to the scalar loop.
2329 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2331 BasicBlock *LastBypassBlock = BypassBlock;
2333 // Generate code to check that the loops trip count that we computed by adding
2334 // one to the backedge-taken count will not overflow.
2336 auto PastOverflowCheck =
2337 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2338 BasicBlock *CheckBlock =
2339 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2341 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2342 LoopBypassBlocks.push_back(CheckBlock);
2343 Instruction *OldTerm = LastBypassBlock->getTerminator();
2344 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2345 OldTerm->eraseFromParent();
2346 LastBypassBlock = CheckBlock;
2349 // Generate the code to check that the strides we assumed to be one are really
2350 // one. We want the new basic block to start at the first instruction in a
2351 // sequence of instructions that form a check.
2352 Instruction *StrideCheck;
2353 Instruction *FirstCheckInst;
2354 std::tie(FirstCheckInst, StrideCheck) =
2355 addStrideCheck(LastBypassBlock->getTerminator());
2357 // Create a new block containing the stride check.
2358 BasicBlock *CheckBlock =
2359 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2361 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2362 LoopBypassBlocks.push_back(CheckBlock);
2364 // Replace the branch into the memory check block with a conditional branch
2365 // for the "few elements case".
2366 Instruction *OldTerm = LastBypassBlock->getTerminator();
2367 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2368 OldTerm->eraseFromParent();
2371 LastBypassBlock = CheckBlock;
2374 // Generate the code that checks in runtime if arrays overlap. We put the
2375 // checks into a separate block to make the more common case of few elements
2377 Instruction *MemRuntimeCheck;
2378 std::tie(FirstCheckInst, MemRuntimeCheck) =
2379 addRuntimeCheck(LastBypassBlock->getTerminator());
2380 if (MemRuntimeCheck) {
2381 // Create a new block containing the memory check.
2382 BasicBlock *CheckBlock =
2383 LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
2385 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2386 LoopBypassBlocks.push_back(CheckBlock);
2388 // Replace the branch into the memory check block with a conditional branch
2389 // for the "few elements case".
2390 Instruction *OldTerm = LastBypassBlock->getTerminator();
2391 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2392 OldTerm->eraseFromParent();
2394 Cmp = MemRuntimeCheck;
2395 LastBypassBlock = CheckBlock;
2398 LastBypassBlock->getTerminator()->eraseFromParent();
2399 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2402 // We are going to resume the execution of the scalar loop.
2403 // Go over all of the induction variables that we found and fix the
2404 // PHIs that are left in the scalar version of the loop.
2405 // The starting values of PHI nodes depend on the counter of the last
2406 // iteration in the vectorized loop.
2407 // If we come from a bypass edge then we need to start from the original
2410 // This variable saves the new starting index for the scalar loop.
2411 PHINode *ResumeIndex = nullptr;
2412 LoopVectorizationLegality::InductionList::iterator I, E;
2413 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2414 // Set builder to point to last bypass block.
2415 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2416 for (I = List->begin(), E = List->end(); I != E; ++I) {
2417 PHINode *OrigPhi = I->first;
2418 LoopVectorizationLegality::InductionInfo II = I->second;
2420 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2421 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2422 MiddleBlock->getTerminator());
2423 // We might have extended the type of the induction variable but we need a
2424 // truncated version for the scalar loop.
2425 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2426 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2427 MiddleBlock->getTerminator()) : nullptr;
2429 // Create phi nodes to merge from the backedge-taken check block.
2430 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2431 ScalarPH->getTerminator());
2432 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2434 PHINode *BCTruncResumeVal = nullptr;
2435 if (OrigPhi == OldInduction) {
2437 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2438 ScalarPH->getTerminator());
2439 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2442 Value *EndValue = nullptr;
2444 case LoopVectorizationLegality::IK_NoInduction:
2445 llvm_unreachable("Unknown induction");
2446 case LoopVectorizationLegality::IK_IntInduction: {
2447 // Handle the integer induction counter.
2448 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2450 // We have the canonical induction variable.
2451 if (OrigPhi == OldInduction) {
2452 // Create a truncated version of the resume value for the scalar loop,
2453 // we might have promoted the type to a larger width.
2455 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2456 // The new PHI merges the original incoming value, in case of a bypass,
2457 // or the value at the end of the vectorized loop.
2458 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2459 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2460 TruncResumeVal->addIncoming(EndValue, VecBody);
2462 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2464 // We know what the end value is.
2465 EndValue = IdxEndRoundDown;
2466 // We also know which PHI node holds it.
2467 ResumeIndex = ResumeVal;
2471 // Not the canonical induction variable - add the vector loop count to the
2473 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2474 II.StartValue->getType(),
2476 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
2479 case LoopVectorizationLegality::IK_ReverseIntInduction: {
2480 // Convert the CountRoundDown variable to the PHI size.
2481 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2482 II.StartValue->getType(),
2484 // Handle reverse integer induction counter.
2485 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
2488 case LoopVectorizationLegality::IK_PtrInduction: {
2489 // For pointer induction variables, calculate the offset using
2491 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
2495 case LoopVectorizationLegality::IK_ReversePtrInduction: {
2496 // The value at the end of the loop for the reverse pointer is calculated
2497 // by creating a GEP with a negative index starting from the start value.
2498 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
2499 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
2501 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
2507 // The new PHI merges the original incoming value, in case of a bypass,
2508 // or the value at the end of the vectorized loop.
2509 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2510 if (OrigPhi == OldInduction)
2511 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2513 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2515 ResumeVal->addIncoming(EndValue, VecBody);
2517 // Fix the scalar body counter (PHI node).
2518 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2520 // The old induction's phi node in the scalar body needs the truncated
2522 if (OrigPhi == OldInduction) {
2523 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2524 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2526 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2527 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2531 // If we are generating a new induction variable then we also need to
2532 // generate the code that calculates the exit value. This value is not
2533 // simply the end of the counter because we may skip the vectorized body
2534 // in case of a runtime check.
2536 assert(!ResumeIndex && "Unexpected resume value found");
2537 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2538 MiddleBlock->getTerminator());
2539 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2540 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2541 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2544 // Make sure that we found the index where scalar loop needs to continue.
2545 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2546 "Invalid resume Index");
2548 // Add a check in the middle block to see if we have completed
2549 // all of the iterations in the first vector loop.
2550 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2551 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2552 ResumeIndex, "cmp.n",
2553 MiddleBlock->getTerminator());
2555 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2556 // Remove the old terminator.
2557 MiddleBlock->getTerminator()->eraseFromParent();
2559 // Create i+1 and fill the PHINode.
2560 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2561 Induction->addIncoming(StartIdx, VectorPH);
2562 Induction->addIncoming(NextIdx, VecBody);
2563 // Create the compare.
2564 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2565 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2567 // Now we have two terminators. Remove the old one from the block.
2568 VecBody->getTerminator()->eraseFromParent();
2570 // Get ready to start creating new instructions into the vectorized body.
2571 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2574 LoopVectorPreHeader = VectorPH;
2575 LoopScalarPreHeader = ScalarPH;
2576 LoopMiddleBlock = MiddleBlock;
2577 LoopExitBlock = ExitBlock;
2578 LoopVectorBody.push_back(VecBody);
2579 LoopScalarBody = OldBasicBlock;
2581 LoopVectorizeHints Hints(Lp, true);
2582 Hints.setAlreadyVectorized();
2585 /// This function returns the identity element (or neutral element) for
2586 /// the operation K.
2588 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2593 // Adding, Xoring, Oring zero to a number does not change it.
2594 return ConstantInt::get(Tp, 0);
2595 case RK_IntegerMult:
2596 // Multiplying a number by 1 does not change it.
2597 return ConstantInt::get(Tp, 1);
2599 // AND-ing a number with an all-1 value does not change it.
2600 return ConstantInt::get(Tp, -1, true);
2602 // Multiplying a number by 1 does not change it.
2603 return ConstantFP::get(Tp, 1.0L);
2605 // Adding zero to a number does not change it.
2606 return ConstantFP::get(Tp, 0.0L);
2608 llvm_unreachable("Unknown reduction kind");
2612 /// This function translates the reduction kind to an LLVM binary operator.
2614 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2616 case LoopVectorizationLegality::RK_IntegerAdd:
2617 return Instruction::Add;
2618 case LoopVectorizationLegality::RK_IntegerMult:
2619 return Instruction::Mul;
2620 case LoopVectorizationLegality::RK_IntegerOr:
2621 return Instruction::Or;
2622 case LoopVectorizationLegality::RK_IntegerAnd:
2623 return Instruction::And;
2624 case LoopVectorizationLegality::RK_IntegerXor:
2625 return Instruction::Xor;
2626 case LoopVectorizationLegality::RK_FloatMult:
2627 return Instruction::FMul;
2628 case LoopVectorizationLegality::RK_FloatAdd:
2629 return Instruction::FAdd;
2630 case LoopVectorizationLegality::RK_IntegerMinMax:
2631 return Instruction::ICmp;
2632 case LoopVectorizationLegality::RK_FloatMinMax:
2633 return Instruction::FCmp;
2635 llvm_unreachable("Unknown reduction operation");
2639 Value *createMinMaxOp(IRBuilder<> &Builder,
2640 LoopVectorizationLegality::MinMaxReductionKind RK,
2643 CmpInst::Predicate P = CmpInst::ICMP_NE;
2646 llvm_unreachable("Unknown min/max reduction kind");
2647 case LoopVectorizationLegality::MRK_UIntMin:
2648 P = CmpInst::ICMP_ULT;
2650 case LoopVectorizationLegality::MRK_UIntMax:
2651 P = CmpInst::ICMP_UGT;
2653 case LoopVectorizationLegality::MRK_SIntMin:
2654 P = CmpInst::ICMP_SLT;
2656 case LoopVectorizationLegality::MRK_SIntMax:
2657 P = CmpInst::ICMP_SGT;
2659 case LoopVectorizationLegality::MRK_FloatMin:
2660 P = CmpInst::FCMP_OLT;
2662 case LoopVectorizationLegality::MRK_FloatMax:
2663 P = CmpInst::FCMP_OGT;
2668 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2669 RK == LoopVectorizationLegality::MRK_FloatMax)
2670 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2672 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2674 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2679 struct CSEDenseMapInfo {
2680 static bool canHandle(Instruction *I) {
2681 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2682 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2684 static inline Instruction *getEmptyKey() {
2685 return DenseMapInfo<Instruction *>::getEmptyKey();
2687 static inline Instruction *getTombstoneKey() {
2688 return DenseMapInfo<Instruction *>::getTombstoneKey();
2690 static unsigned getHashValue(Instruction *I) {
2691 assert(canHandle(I) && "Unknown instruction!");
2692 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2693 I->value_op_end()));
2695 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2696 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2697 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2699 return LHS->isIdenticalTo(RHS);
2704 /// \brief Check whether this block is a predicated block.
2705 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2706 /// = ...; " blocks. We start with one vectorized basic block. For every
2707 /// conditional block we split this vectorized block. Therefore, every second
2708 /// block will be a predicated one.
2709 static bool isPredicatedBlock(unsigned BlockNum) {
2710 return BlockNum % 2;
2713 ///\brief Perform cse of induction variable instructions.
2714 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2715 // Perform simple cse.
2716 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2717 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2718 BasicBlock *BB = BBs[i];
2719 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2720 Instruction *In = I++;
2722 if (!CSEDenseMapInfo::canHandle(In))
2725 // Check if we can replace this instruction with any of the
2726 // visited instructions.
2727 if (Instruction *V = CSEMap.lookup(In)) {
2728 In->replaceAllUsesWith(V);
2729 In->eraseFromParent();
2732 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2733 // ...;" blocks for predicated stores. Every second block is a predicated
2735 if (isPredicatedBlock(i))
2743 /// \brief Adds a 'fast' flag to floating point operations.
2744 static Value *addFastMathFlag(Value *V) {
2745 if (isa<FPMathOperator>(V)){
2746 FastMathFlags Flags;
2747 Flags.setUnsafeAlgebra();
2748 cast<Instruction>(V)->setFastMathFlags(Flags);
2753 void InnerLoopVectorizer::vectorizeLoop() {
2754 //===------------------------------------------------===//
2756 // Notice: any optimization or new instruction that go
2757 // into the code below should be also be implemented in
2760 //===------------------------------------------------===//
2761 Constant *Zero = Builder.getInt32(0);
2763 // In order to support reduction variables we need to be able to vectorize
2764 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2765 // stages. First, we create a new vector PHI node with no incoming edges.
2766 // We use this value when we vectorize all of the instructions that use the
2767 // PHI. Next, after all of the instructions in the block are complete we
2768 // add the new incoming edges to the PHI. At this point all of the
2769 // instructions in the basic block are vectorized, so we can use them to
2770 // construct the PHI.
2771 PhiVector RdxPHIsToFix;
2773 // Scan the loop in a topological order to ensure that defs are vectorized
2775 LoopBlocksDFS DFS(OrigLoop);
2778 // Vectorize all of the blocks in the original loop.
2779 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2780 be = DFS.endRPO(); bb != be; ++bb)
2781 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2783 // At this point every instruction in the original loop is widened to
2784 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2785 // that we vectorized. The PHI nodes are currently empty because we did
2786 // not want to introduce cycles. Notice that the remaining PHI nodes
2787 // that we need to fix are reduction variables.
2789 // Create the 'reduced' values for each of the induction vars.
2790 // The reduced values are the vector values that we scalarize and combine
2791 // after the loop is finished.
2792 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2794 PHINode *RdxPhi = *it;
2795 assert(RdxPhi && "Unable to recover vectorized PHI");
2797 // Find the reduction variable descriptor.
2798 assert(Legal->getReductionVars()->count(RdxPhi) &&
2799 "Unable to find the reduction variable");
2800 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2801 (*Legal->getReductionVars())[RdxPhi];
2803 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2805 // We need to generate a reduction vector from the incoming scalar.
2806 // To do so, we need to generate the 'identity' vector and override
2807 // one of the elements with the incoming scalar reduction. We need
2808 // to do it in the vector-loop preheader.
2809 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
2811 // This is the vector-clone of the value that leaves the loop.
2812 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2813 Type *VecTy = VectorExit[0]->getType();
2815 // Find the reduction identity variable. Zero for addition, or, xor,
2816 // one for multiplication, -1 for And.
2819 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2820 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2821 // MinMax reduction have the start value as their identify.
2823 VectorStart = Identity = RdxDesc.StartValue;
2825 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2830 // Handle other reduction kinds:
2832 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2833 VecTy->getScalarType());
2836 // This vector is the Identity vector where the first element is the
2837 // incoming scalar reduction.
2838 VectorStart = RdxDesc.StartValue;
2840 Identity = ConstantVector::getSplat(VF, Iden);
2842 // This vector is the Identity vector where the first element is the
2843 // incoming scalar reduction.
2844 VectorStart = Builder.CreateInsertElement(Identity,
2845 RdxDesc.StartValue, Zero);
2849 // Fix the vector-loop phi.
2850 // We created the induction variable so we know that the
2851 // preheader is the first entry.
2852 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2854 // Reductions do not have to start at zero. They can start with
2855 // any loop invariant values.
2856 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2857 BasicBlock *Latch = OrigLoop->getLoopLatch();
2858 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2859 VectorParts &Val = getVectorValue(LoopVal);
2860 for (unsigned part = 0; part < UF; ++part) {
2861 // Make sure to add the reduction stat value only to the
2862 // first unroll part.
2863 Value *StartVal = (part == 0) ? VectorStart : Identity;
2864 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2865 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2866 LoopVectorBody.back());
2869 // Before each round, move the insertion point right between
2870 // the PHIs and the values we are going to write.
2871 // This allows us to write both PHINodes and the extractelement
2873 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2875 VectorParts RdxParts;
2876 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2877 for (unsigned part = 0; part < UF; ++part) {
2878 // This PHINode contains the vectorized reduction variable, or
2879 // the initial value vector, if we bypass the vector loop.
2880 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2881 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2882 Value *StartVal = (part == 0) ? VectorStart : Identity;
2883 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2884 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2885 NewPhi->addIncoming(RdxExitVal[part],
2886 LoopVectorBody.back());
2887 RdxParts.push_back(NewPhi);
2890 // Reduce all of the unrolled parts into a single vector.
2891 Value *ReducedPartRdx = RdxParts[0];
2892 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2893 setDebugLocFromInst(Builder, ReducedPartRdx);
2894 for (unsigned part = 1; part < UF; ++part) {
2895 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2896 // Floating point operations had to be 'fast' to enable the reduction.
2897 ReducedPartRdx = addFastMathFlag(
2898 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2899 ReducedPartRdx, "bin.rdx"));
2901 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2902 ReducedPartRdx, RdxParts[part]);
2906 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2907 // and vector ops, reducing the set of values being computed by half each
2909 assert(isPowerOf2_32(VF) &&
2910 "Reduction emission only supported for pow2 vectors!");
2911 Value *TmpVec = ReducedPartRdx;
2912 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2913 for (unsigned i = VF; i != 1; i >>= 1) {
2914 // Move the upper half of the vector to the lower half.
2915 for (unsigned j = 0; j != i/2; ++j)
2916 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2918 // Fill the rest of the mask with undef.
2919 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2920 UndefValue::get(Builder.getInt32Ty()));
2923 Builder.CreateShuffleVector(TmpVec,
2924 UndefValue::get(TmpVec->getType()),
2925 ConstantVector::get(ShuffleMask),
2928 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2929 // Floating point operations had to be 'fast' to enable the reduction.
2930 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2931 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2933 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2936 // The result is in the first element of the vector.
2937 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2938 Builder.getInt32(0));
2941 // Create a phi node that merges control-flow from the backedge-taken check
2942 // block and the middle block.
2943 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
2944 LoopScalarPreHeader->getTerminator());
2945 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
2946 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2948 // Now, we need to fix the users of the reduction variable
2949 // inside and outside of the scalar remainder loop.
2950 // We know that the loop is in LCSSA form. We need to update the
2951 // PHI nodes in the exit blocks.
2952 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2953 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2954 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2955 if (!LCSSAPhi) break;
2957 // All PHINodes need to have a single entry edge, or two if
2958 // we already fixed them.
2959 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2961 // We found our reduction value exit-PHI. Update it with the
2962 // incoming bypass edge.
2963 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2964 // Add an edge coming from the bypass.
2965 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2968 }// end of the LCSSA phi scan.
2970 // Fix the scalar loop reduction variable with the incoming reduction sum
2971 // from the vector body and from the backedge value.
2972 int IncomingEdgeBlockIdx =
2973 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2974 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2975 // Pick the other block.
2976 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2977 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
2978 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2979 }// end of for each redux variable.
2983 // Remove redundant induction instructions.
2984 cse(LoopVectorBody);
2987 void InnerLoopVectorizer::fixLCSSAPHIs() {
2988 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2989 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2990 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2991 if (!LCSSAPhi) break;
2992 if (LCSSAPhi->getNumIncomingValues() == 1)
2993 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2998 InnerLoopVectorizer::VectorParts
2999 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
3000 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
3003 // Look for cached value.
3004 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
3005 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
3006 if (ECEntryIt != MaskCache.end())
3007 return ECEntryIt->second;
3009 VectorParts SrcMask = createBlockInMask(Src);
3011 // The terminator has to be a branch inst!
3012 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
3013 assert(BI && "Unexpected terminator found");
3015 if (BI->isConditional()) {
3016 VectorParts EdgeMask = getVectorValue(BI->getCondition());
3018 if (BI->getSuccessor(0) != Dst)
3019 for (unsigned part = 0; part < UF; ++part)
3020 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
3022 for (unsigned part = 0; part < UF; ++part)
3023 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
3025 MaskCache[Edge] = EdgeMask;
3029 MaskCache[Edge] = SrcMask;
3033 InnerLoopVectorizer::VectorParts
3034 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
3035 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
3037 // Loop incoming mask is all-one.
3038 if (OrigLoop->getHeader() == BB) {
3039 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
3040 return getVectorValue(C);
3043 // This is the block mask. We OR all incoming edges, and with zero.
3044 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
3045 VectorParts BlockMask = getVectorValue(Zero);
3048 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
3049 VectorParts EM = createEdgeMask(*it, BB);
3050 for (unsigned part = 0; part < UF; ++part)
3051 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
3057 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
3058 InnerLoopVectorizer::VectorParts &Entry,
3059 unsigned UF, unsigned VF, PhiVector *PV) {
3060 PHINode* P = cast<PHINode>(PN);
3061 // Handle reduction variables:
3062 if (Legal->getReductionVars()->count(P)) {
3063 for (unsigned part = 0; part < UF; ++part) {
3064 // This is phase one of vectorizing PHIs.
3065 Type *VecTy = (VF == 1) ? PN->getType() :
3066 VectorType::get(PN->getType(), VF);
3067 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
3068 LoopVectorBody.back()-> getFirstInsertionPt());
3074 setDebugLocFromInst(Builder, P);
3075 // Check for PHI nodes that are lowered to vector selects.
3076 if (P->getParent() != OrigLoop->getHeader()) {
3077 // We know that all PHIs in non-header blocks are converted into
3078 // selects, so we don't have to worry about the insertion order and we
3079 // can just use the builder.
3080 // At this point we generate the predication tree. There may be
3081 // duplications since this is a simple recursive scan, but future
3082 // optimizations will clean it up.
3084 unsigned NumIncoming = P->getNumIncomingValues();
3086 // Generate a sequence of selects of the form:
3087 // SELECT(Mask3, In3,
3088 // SELECT(Mask2, In2,
3090 for (unsigned In = 0; In < NumIncoming; In++) {
3091 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
3093 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3095 for (unsigned part = 0; part < UF; ++part) {
3096 // We might have single edge PHIs (blocks) - use an identity
3097 // 'select' for the first PHI operand.
3099 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3102 // Select between the current value and the previous incoming edge
3103 // based on the incoming mask.
3104 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3105 Entry[part], "predphi");
3111 // This PHINode must be an induction variable.
3112 // Make sure that we know about it.
3113 assert(Legal->getInductionVars()->count(P) &&
3114 "Not an induction variable");
3116 LoopVectorizationLegality::InductionInfo II =
3117 Legal->getInductionVars()->lookup(P);
3120 case LoopVectorizationLegality::IK_NoInduction:
3121 llvm_unreachable("Unknown induction");
3122 case LoopVectorizationLegality::IK_IntInduction: {
3123 assert(P->getType() == II.StartValue->getType() && "Types must match");
3124 Type *PhiTy = P->getType();
3126 if (P == OldInduction) {
3127 // Handle the canonical induction variable. We might have had to
3129 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
3131 // Handle other induction variables that are now based on the
3133 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
3135 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
3136 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
3139 Broadcasted = getBroadcastInstrs(Broadcasted);
3140 // After broadcasting the induction variable we need to make the vector
3141 // consecutive by adding 0, 1, 2, etc.
3142 for (unsigned part = 0; part < UF; ++part)
3143 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
3146 case LoopVectorizationLegality::IK_ReverseIntInduction:
3147 case LoopVectorizationLegality::IK_PtrInduction:
3148 case LoopVectorizationLegality::IK_ReversePtrInduction:
3149 // Handle reverse integer and pointer inductions.
3150 Value *StartIdx = ExtendedIdx;
3151 // This is the normalized GEP that starts counting at zero.
3152 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
3155 // Handle the reverse integer induction variable case.
3156 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
3157 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
3158 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
3160 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
3163 // This is a new value so do not hoist it out.
3164 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
3165 // After broadcasting the induction variable we need to make the
3166 // vector consecutive by adding ... -3, -2, -1, 0.
3167 for (unsigned part = 0; part < UF; ++part)
3168 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
3173 // Handle the pointer induction variable case.
3174 assert(P->getType()->isPointerTy() && "Unexpected type.");
3176 // Is this a reverse induction ptr or a consecutive induction ptr.
3177 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
3180 // This is the vector of results. Notice that we don't generate
3181 // vector geps because scalar geps result in better code.
3182 for (unsigned part = 0; part < UF; ++part) {
3184 int EltIndex = (part) * (Reverse ? -1 : 1);
3185 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3188 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3190 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3192 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3194 Entry[part] = SclrGep;
3198 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3199 for (unsigned int i = 0; i < VF; ++i) {
3200 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
3201 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3204 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3206 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3208 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3210 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3211 Builder.getInt32(i),
3214 Entry[part] = VecVal;
3220 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3221 // For each instruction in the old loop.
3222 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3223 VectorParts &Entry = WidenMap.get(it);
3224 switch (it->getOpcode()) {
3225 case Instruction::Br:
3226 // Nothing to do for PHIs and BR, since we already took care of the
3227 // loop control flow instructions.
3229 case Instruction::PHI:{
3230 // Vectorize PHINodes.
3231 widenPHIInstruction(it, Entry, UF, VF, PV);
3235 case Instruction::Add:
3236 case Instruction::FAdd:
3237 case Instruction::Sub:
3238 case Instruction::FSub:
3239 case Instruction::Mul:
3240 case Instruction::FMul:
3241 case Instruction::UDiv:
3242 case Instruction::SDiv:
3243 case Instruction::FDiv:
3244 case Instruction::URem:
3245 case Instruction::SRem:
3246 case Instruction::FRem:
3247 case Instruction::Shl:
3248 case Instruction::LShr:
3249 case Instruction::AShr:
3250 case Instruction::And:
3251 case Instruction::Or:
3252 case Instruction::Xor: {
3253 // Just widen binops.
3254 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3255 setDebugLocFromInst(Builder, BinOp);
3256 VectorParts &A = getVectorValue(it->getOperand(0));
3257 VectorParts &B = getVectorValue(it->getOperand(1));
3259 // Use this vector value for all users of the original instruction.
3260 for (unsigned Part = 0; Part < UF; ++Part) {
3261 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3263 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3264 VecOp->copyIRFlags(BinOp);
3269 propagateMetadata(Entry, it);
3272 case Instruction::Select: {
3274 // If the selector is loop invariant we can create a select
3275 // instruction with a scalar condition. Otherwise, use vector-select.
3276 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3278 setDebugLocFromInst(Builder, it);
3280 // The condition can be loop invariant but still defined inside the
3281 // loop. This means that we can't just use the original 'cond' value.
3282 // We have to take the 'vectorized' value and pick the first lane.
3283 // Instcombine will make this a no-op.
3284 VectorParts &Cond = getVectorValue(it->getOperand(0));
3285 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3286 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3288 Value *ScalarCond = (VF == 1) ? Cond[0] :
3289 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3291 for (unsigned Part = 0; Part < UF; ++Part) {
3292 Entry[Part] = Builder.CreateSelect(
3293 InvariantCond ? ScalarCond : Cond[Part],
3298 propagateMetadata(Entry, it);
3302 case Instruction::ICmp:
3303 case Instruction::FCmp: {
3304 // Widen compares. Generate vector compares.
3305 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3306 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3307 setDebugLocFromInst(Builder, it);
3308 VectorParts &A = getVectorValue(it->getOperand(0));
3309 VectorParts &B = getVectorValue(it->getOperand(1));
3310 for (unsigned Part = 0; Part < UF; ++Part) {
3313 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3315 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3319 propagateMetadata(Entry, it);
3323 case Instruction::Store:
3324 case Instruction::Load:
3325 vectorizeMemoryInstruction(it);
3327 case Instruction::ZExt:
3328 case Instruction::SExt:
3329 case Instruction::FPToUI:
3330 case Instruction::FPToSI:
3331 case Instruction::FPExt:
3332 case Instruction::PtrToInt:
3333 case Instruction::IntToPtr:
3334 case Instruction::SIToFP:
3335 case Instruction::UIToFP:
3336 case Instruction::Trunc:
3337 case Instruction::FPTrunc:
3338 case Instruction::BitCast: {
3339 CastInst *CI = dyn_cast<CastInst>(it);
3340 setDebugLocFromInst(Builder, it);
3341 /// Optimize the special case where the source is the induction
3342 /// variable. Notice that we can only optimize the 'trunc' case
3343 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3344 /// c. other casts depend on pointer size.
3345 if (CI->getOperand(0) == OldInduction &&
3346 it->getOpcode() == Instruction::Trunc) {
3347 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3349 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3350 for (unsigned Part = 0; Part < UF; ++Part)
3351 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
3352 propagateMetadata(Entry, it);
3355 /// Vectorize casts.
3356 Type *DestTy = (VF == 1) ? CI->getType() :
3357 VectorType::get(CI->getType(), VF);
3359 VectorParts &A = getVectorValue(it->getOperand(0));
3360 for (unsigned Part = 0; Part < UF; ++Part)
3361 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3362 propagateMetadata(Entry, it);
3366 case Instruction::Call: {
3367 // Ignore dbg intrinsics.
3368 if (isa<DbgInfoIntrinsic>(it))
3370 setDebugLocFromInst(Builder, it);
3372 Module *M = BB->getParent()->getParent();
3373 CallInst *CI = cast<CallInst>(it);
3374 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3375 assert(ID && "Not an intrinsic call!");
3377 case Intrinsic::assume:
3378 case Intrinsic::lifetime_end:
3379 case Intrinsic::lifetime_start:
3380 scalarizeInstruction(it);
3383 bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
3384 for (unsigned Part = 0; Part < UF; ++Part) {
3385 SmallVector<Value *, 4> Args;
3386 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3387 if (HasScalarOpd && i == 1) {
3388 Args.push_back(CI->getArgOperand(i));
3391 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3392 Args.push_back(Arg[Part]);
3394 Type *Tys[] = {CI->getType()};
3396 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3398 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3399 Entry[Part] = Builder.CreateCall(F, Args);
3402 propagateMetadata(Entry, it);
3409 // All other instructions are unsupported. Scalarize them.
3410 scalarizeInstruction(it);
3413 }// end of for_each instr.
3416 void InnerLoopVectorizer::updateAnalysis() {
3417 // Forget the original basic block.
3418 SE->forgetLoop(OrigLoop);
3420 // Update the dominator tree information.
3421 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3422 "Entry does not dominate exit.");
3424 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3425 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3426 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3428 // Due to if predication of stores we might create a sequence of "if(pred)
3429 // a[i] = ...; " blocks.
3430 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3432 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3433 else if (isPredicatedBlock(i)) {
3434 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3436 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3440 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3441 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3442 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3443 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
3445 DEBUG(DT->verifyDomTree());
3448 /// \brief Check whether it is safe to if-convert this phi node.
3450 /// Phi nodes with constant expressions that can trap are not safe to if
3452 static bool canIfConvertPHINodes(BasicBlock *BB) {
3453 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3454 PHINode *Phi = dyn_cast<PHINode>(I);
3457 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3458 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3465 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3466 if (!EnableIfConversion) {
3467 emitAnalysis(Report() << "if-conversion is disabled");
3471 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3473 // A list of pointers that we can safely read and write to.
3474 SmallPtrSet<Value *, 8> SafePointes;
3476 // Collect safe addresses.
3477 for (Loop::block_iterator BI = TheLoop->block_begin(),
3478 BE = TheLoop->block_end(); BI != BE; ++BI) {
3479 BasicBlock *BB = *BI;
3481 if (blockNeedsPredication(BB))
3484 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3485 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3486 SafePointes.insert(LI->getPointerOperand());
3487 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3488 SafePointes.insert(SI->getPointerOperand());
3492 // Collect the blocks that need predication.
3493 BasicBlock *Header = TheLoop->getHeader();
3494 for (Loop::block_iterator BI = TheLoop->block_begin(),
3495 BE = TheLoop->block_end(); BI != BE; ++BI) {
3496 BasicBlock *BB = *BI;
3498 // We don't support switch statements inside loops.
3499 if (!isa<BranchInst>(BB->getTerminator())) {
3500 emitAnalysis(Report(BB->getTerminator())
3501 << "loop contains a switch statement");
3505 // We must be able to predicate all blocks that need to be predicated.
3506 if (blockNeedsPredication(BB)) {
3507 if (!blockCanBePredicated(BB, SafePointes)) {
3508 emitAnalysis(Report(BB->getTerminator())
3509 << "control flow cannot be substituted for a select");
3512 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3513 emitAnalysis(Report(BB->getTerminator())
3514 << "control flow cannot be substituted for a select");
3519 // We can if-convert this loop.
3523 bool LoopVectorizationLegality::canVectorize() {
3524 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3525 // be canonicalized.
3526 if (!TheLoop->getLoopPreheader()) {
3528 Report() << "loop control flow is not understood by vectorizer");
3532 // We can only vectorize innermost loops.
3533 if (TheLoop->getSubLoopsVector().size()) {
3534 emitAnalysis(Report() << "loop is not the innermost loop");
3538 // We must have a single backedge.
3539 if (TheLoop->getNumBackEdges() != 1) {
3541 Report() << "loop control flow is not understood by vectorizer");
3545 // We must have a single exiting block.
3546 if (!TheLoop->getExitingBlock()) {
3548 Report() << "loop control flow is not understood by vectorizer");
3552 // We need to have a loop header.
3553 DEBUG(dbgs() << "LV: Found a loop: " <<
3554 TheLoop->getHeader()->getName() << '\n');
3556 // Check if we can if-convert non-single-bb loops.
3557 unsigned NumBlocks = TheLoop->getNumBlocks();
3558 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3559 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3563 // ScalarEvolution needs to be able to find the exit count.
3564 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3565 if (ExitCount == SE->getCouldNotCompute()) {
3566 emitAnalysis(Report() << "could not determine number of loop iterations");
3567 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3571 // Check if we can vectorize the instructions and CFG in this loop.
3572 if (!canVectorizeInstrs()) {
3573 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3577 // Go over each instruction and look at memory deps.
3578 if (!canVectorizeMemory()) {
3579 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3583 // Collect all of the variables that remain uniform after vectorization.
3584 collectLoopUniforms();
3586 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3587 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
3590 // Okay! We can vectorize. At this point we don't have any other mem analysis
3591 // which may limit our maximum vectorization factor, so just return true with
3596 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3597 if (Ty->isPointerTy())
3598 return DL.getIntPtrType(Ty);
3600 // It is possible that char's or short's overflow when we ask for the loop's
3601 // trip count, work around this by changing the type size.
3602 if (Ty->getScalarSizeInBits() < 32)
3603 return Type::getInt32Ty(Ty->getContext());
3608 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3609 Ty0 = convertPointerToIntegerType(DL, Ty0);
3610 Ty1 = convertPointerToIntegerType(DL, Ty1);
3611 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3616 /// \brief Check that the instruction has outside loop users and is not an
3617 /// identified reduction variable.
3618 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3619 SmallPtrSetImpl<Value *> &Reductions) {
3620 // Reduction instructions are allowed to have exit users. All other
3621 // instructions must not have external users.
3622 if (!Reductions.count(Inst))
3623 //Check that all of the users of the loop are inside the BB.
3624 for (User *U : Inst->users()) {
3625 Instruction *UI = cast<Instruction>(U);
3626 // This user may be a reduction exit value.
3627 if (!TheLoop->contains(UI)) {
3628 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3635 bool LoopVectorizationLegality::canVectorizeInstrs() {
3636 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3637 BasicBlock *Header = TheLoop->getHeader();
3639 // Look for the attribute signaling the absence of NaNs.
3640 Function &F = *Header->getParent();
3641 if (F.hasFnAttribute("no-nans-fp-math"))
3642 HasFunNoNaNAttr = F.getAttributes().getAttribute(
3643 AttributeSet::FunctionIndex,
3644 "no-nans-fp-math").getValueAsString() == "true";
3646 // For each block in the loop.
3647 for (Loop::block_iterator bb = TheLoop->block_begin(),
3648 be = TheLoop->block_end(); bb != be; ++bb) {
3650 // Scan the instructions in the block and look for hazards.
3651 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3654 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3655 Type *PhiTy = Phi->getType();
3656 // Check that this PHI type is allowed.
3657 if (!PhiTy->isIntegerTy() &&
3658 !PhiTy->isFloatingPointTy() &&
3659 !PhiTy->isPointerTy()) {
3660 emitAnalysis(Report(it)
3661 << "loop control flow is not understood by vectorizer");
3662 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3666 // If this PHINode is not in the header block, then we know that we
3667 // can convert it to select during if-conversion. No need to check if
3668 // the PHIs in this block are induction or reduction variables.
3669 if (*bb != Header) {
3670 // Check that this instruction has no outside users or is an
3671 // identified reduction value with an outside user.
3672 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3674 emitAnalysis(Report(it) << "value could not be identified as "
3675 "an induction or reduction variable");
3679 // We only allow if-converted PHIs with more than two incoming values.
3680 if (Phi->getNumIncomingValues() != 2) {
3681 emitAnalysis(Report(it)
3682 << "control flow not understood by vectorizer");
3683 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3687 // This is the value coming from the preheader.
3688 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3689 // Check if this is an induction variable.
3690 InductionKind IK = isInductionVariable(Phi);
3692 if (IK_NoInduction != IK) {
3693 // Get the widest type.
3695 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3697 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3699 // Int inductions are special because we only allow one IV.
3700 if (IK == IK_IntInduction) {
3701 // Use the phi node with the widest type as induction. Use the last
3702 // one if there are multiple (no good reason for doing this other
3703 // than it is expedient).
3704 if (!Induction || PhiTy == WidestIndTy)
3708 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3709 Inductions[Phi] = InductionInfo(StartValue, IK);
3711 // Until we explicitly handle the case of an induction variable with
3712 // an outside loop user we have to give up vectorizing this loop.
3713 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3714 emitAnalysis(Report(it) << "use of induction value outside of the "
3715 "loop is not handled by vectorizer");
3722 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3723 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3726 if (AddReductionVar(Phi, RK_IntegerMult)) {
3727 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3730 if (AddReductionVar(Phi, RK_IntegerOr)) {
3731 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3734 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3735 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3738 if (AddReductionVar(Phi, RK_IntegerXor)) {
3739 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3742 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3743 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3746 if (AddReductionVar(Phi, RK_FloatMult)) {
3747 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3750 if (AddReductionVar(Phi, RK_FloatAdd)) {
3751 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3754 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3755 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3760 emitAnalysis(Report(it) << "value that could not be identified as "
3761 "reduction is used outside the loop");
3762 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3764 }// end of PHI handling
3766 // We still don't handle functions. However, we can ignore dbg intrinsic
3767 // calls and we do handle certain intrinsic and libm functions.
3768 CallInst *CI = dyn_cast<CallInst>(it);
3769 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3770 emitAnalysis(Report(it) << "call instruction cannot be vectorized");
3771 DEBUG(dbgs() << "LV: Found a call site.\n");
3775 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
3776 // second argument is the same (i.e. loop invariant)
3778 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
3779 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
3780 emitAnalysis(Report(it)
3781 << "intrinsic instruction cannot be vectorized");
3782 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
3787 // Check that the instruction return type is vectorizable.
3788 // Also, we can't vectorize extractelement instructions.
3789 if ((!VectorType::isValidElementType(it->getType()) &&
3790 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3791 emitAnalysis(Report(it)
3792 << "instruction return type cannot be vectorized");
3793 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3797 // Check that the stored type is vectorizable.
3798 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3799 Type *T = ST->getValueOperand()->getType();
3800 if (!VectorType::isValidElementType(T)) {
3801 emitAnalysis(Report(ST) << "store instruction cannot be vectorized");
3804 if (EnableMemAccessVersioning)
3805 collectStridedAcccess(ST);
3808 if (EnableMemAccessVersioning)
3809 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3810 collectStridedAcccess(LI);
3812 // Reduction instructions are allowed to have exit users.
3813 // All other instructions must not have external users.
3814 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3815 emitAnalysis(Report(it) << "value cannot be used outside the loop");
3824 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3825 if (Inductions.empty()) {
3826 emitAnalysis(Report()
3827 << "loop induction variable could not be identified");
3835 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3836 /// return the induction operand of the gep pointer.
3837 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3838 const DataLayout *DL, Loop *Lp) {
3839 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3843 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3845 // Check that all of the gep indices are uniform except for our induction
3847 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3848 if (i != InductionOperand &&
3849 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3851 return GEP->getOperand(InductionOperand);
3854 ///\brief Look for a cast use of the passed value.
3855 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3856 Value *UniqueCast = nullptr;
3857 for (User *U : Ptr->users()) {
3858 CastInst *CI = dyn_cast<CastInst>(U);
3859 if (CI && CI->getType() == Ty) {
3869 ///\brief Get the stride of a pointer access in a loop.
3870 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3871 /// pointer to the Value, or null otherwise.
3872 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3873 const DataLayout *DL, Loop *Lp) {
3874 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3875 if (!PtrTy || PtrTy->isAggregateType())
3878 // Try to remove a gep instruction to make the pointer (actually index at this
3879 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3880 // pointer, otherwise, we are analyzing the index.
3881 Value *OrigPtr = Ptr;
3883 // The size of the pointer access.
3884 int64_t PtrAccessSize = 1;
3886 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3887 const SCEV *V = SE->getSCEV(Ptr);
3891 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3892 V = C->getOperand();
3894 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3898 V = S->getStepRecurrence(*SE);
3902 // Strip off the size of access multiplication if we are still analyzing the
3904 if (OrigPtr == Ptr) {
3905 DL->getTypeAllocSize(PtrTy->getElementType());
3906 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3907 if (M->getOperand(0)->getSCEVType() != scConstant)
3910 const APInt &APStepVal =
3911 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3913 // Huge step value - give up.
3914 if (APStepVal.getBitWidth() > 64)
3917 int64_t StepVal = APStepVal.getSExtValue();
3918 if (PtrAccessSize != StepVal)
3920 V = M->getOperand(1);
3925 Type *StripedOffRecurrenceCast = nullptr;
3926 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3927 StripedOffRecurrenceCast = C->getType();
3928 V = C->getOperand();
3931 // Look for the loop invariant symbolic value.
3932 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3936 Value *Stride = U->getValue();
3937 if (!Lp->isLoopInvariant(Stride))
3940 // If we have stripped off the recurrence cast we have to make sure that we
3941 // return the value that is used in this loop so that we can replace it later.
3942 if (StripedOffRecurrenceCast)
3943 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3948 void LoopVectorizationLegality::collectStridedAcccess(Value *MemAccess) {
3949 Value *Ptr = nullptr;
3950 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3951 Ptr = LI->getPointerOperand();
3952 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3953 Ptr = SI->getPointerOperand();
3957 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3961 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3962 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3963 Strides[Ptr] = Stride;
3964 StrideSet.insert(Stride);
3967 void LoopVectorizationLegality::collectLoopUniforms() {
3968 // We now know that the loop is vectorizable!
3969 // Collect variables that will remain uniform after vectorization.
3970 std::vector<Value*> Worklist;
3971 BasicBlock *Latch = TheLoop->getLoopLatch();
3973 // Start with the conditional branch and walk up the block.
3974 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3976 // Also add all consecutive pointer values; these values will be uniform
3977 // after vectorization (and subsequent cleanup) and, until revectorization is
3978 // supported, all dependencies must also be uniform.
3979 for (Loop::block_iterator B = TheLoop->block_begin(),
3980 BE = TheLoop->block_end(); B != BE; ++B)
3981 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
3983 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
3984 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3986 while (Worklist.size()) {
3987 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3988 Worklist.pop_back();
3990 // Look at instructions inside this loop.
3991 // Stop when reaching PHI nodes.
3992 // TODO: we need to follow values all over the loop, not only in this block.
3993 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3996 // This is a known uniform.
3999 // Insert all operands.
4000 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4005 /// \brief Analyses memory accesses in a loop.
4007 /// Checks whether run time pointer checks are needed and builds sets for data
4008 /// dependence checking.
4009 class AccessAnalysis {
4011 /// \brief Read or write access location.
4012 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
4013 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
4015 /// \brief Set of potential dependent memory accesses.
4016 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
4018 AccessAnalysis(const DataLayout *Dl, AliasAnalysis *AA, DepCandidates &DA) :
4019 DL(Dl), AST(*AA), DepCands(DA), IsRTCheckNeeded(false) {}
4021 /// \brief Register a load and whether it is only read from.
4022 void addLoad(AliasAnalysis::Location &Loc, bool IsReadOnly) {
4023 Value *Ptr = const_cast<Value*>(Loc.Ptr);
4024 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
4025 Accesses.insert(MemAccessInfo(Ptr, false));
4027 ReadOnlyPtr.insert(Ptr);
4030 /// \brief Register a store.
4031 void addStore(AliasAnalysis::Location &Loc) {
4032 Value *Ptr = const_cast<Value*>(Loc.Ptr);
4033 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
4034 Accesses.insert(MemAccessInfo(Ptr, true));
4037 /// \brief Check whether we can check the pointers at runtime for
4038 /// non-intersection.
4039 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
4040 unsigned &NumComparisons, ScalarEvolution *SE,
4041 Loop *TheLoop, ValueToValueMap &Strides,
4042 bool ShouldCheckStride = false);
4044 /// \brief Goes over all memory accesses, checks whether a RT check is needed
4045 /// and builds sets of dependent accesses.
4046 void buildDependenceSets() {
4047 processMemAccesses();
4050 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
4052 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
4053 void resetDepChecks() { CheckDeps.clear(); }
4055 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
4058 typedef SetVector<MemAccessInfo> PtrAccessSet;
4060 /// \brief Go over all memory access and check whether runtime pointer checks
4061 /// are needed /// and build sets of dependency check candidates.
4062 void processMemAccesses();
4064 /// Set of all accesses.
4065 PtrAccessSet Accesses;
4067 /// Set of accesses that need a further dependence check.
4068 MemAccessInfoSet CheckDeps;
4070 /// Set of pointers that are read only.
4071 SmallPtrSet<Value*, 16> ReadOnlyPtr;
4073 const DataLayout *DL;
4075 /// An alias set tracker to partition the access set by underlying object and
4076 //intrinsic property (such as TBAA metadata).
4077 AliasSetTracker AST;
4079 /// Sets of potentially dependent accesses - members of one set share an
4080 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
4081 /// dependence check.
4082 DepCandidates &DepCands;
4084 bool IsRTCheckNeeded;
4087 } // end anonymous namespace
4089 /// \brief Check whether a pointer can participate in a runtime bounds check.
4090 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
4092 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
4093 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4097 return AR->isAffine();
4100 /// \brief Check the stride of the pointer and ensure that it does not wrap in
4101 /// the address space.
4102 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4103 const Loop *Lp, ValueToValueMap &StridesMap);
4105 bool AccessAnalysis::canCheckPtrAtRT(
4106 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
4107 unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
4108 ValueToValueMap &StridesMap, bool ShouldCheckStride) {
4109 // Find pointers with computable bounds. We are going to use this information
4110 // to place a runtime bound check.
4111 bool CanDoRT = true;
4113 bool IsDepCheckNeeded = isDependencyCheckNeeded();
4116 // We assign a consecutive id to access from different alias sets.
4117 // Accesses between different groups doesn't need to be checked.
4119 for (auto &AS : AST) {
4120 unsigned NumReadPtrChecks = 0;
4121 unsigned NumWritePtrChecks = 0;
4123 // We assign consecutive id to access from different dependence sets.
4124 // Accesses within the same set don't need a runtime check.
4125 unsigned RunningDepId = 1;
4126 DenseMap<Value *, unsigned> DepSetId;
4129 Value *Ptr = A.getValue();
4130 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
4131 MemAccessInfo Access(Ptr, IsWrite);
4134 ++NumWritePtrChecks;
4138 if (hasComputableBounds(SE, StridesMap, Ptr) &&
4139 // When we run after a failing dependency check we have to make sure we
4140 // don't have wrapping pointers.
4141 (!ShouldCheckStride ||
4142 isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
4143 // The id of the dependence set.
4146 if (IsDepCheckNeeded) {
4147 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
4148 unsigned &LeaderId = DepSetId[Leader];
4150 LeaderId = RunningDepId++;
4153 // Each access has its own dependence set.
4154 DepId = RunningDepId++;
4156 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
4158 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
4164 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
4165 NumComparisons += 0; // Only one dependence set.
4167 NumComparisons += (NumWritePtrChecks * (NumReadPtrChecks +
4168 NumWritePtrChecks - 1));
4174 // If the pointers that we would use for the bounds comparison have different
4175 // address spaces, assume the values aren't directly comparable, so we can't
4176 // use them for the runtime check. We also have to assume they could
4177 // overlap. In the future there should be metadata for whether address spaces
4179 unsigned NumPointers = RtCheck.Pointers.size();
4180 for (unsigned i = 0; i < NumPointers; ++i) {
4181 for (unsigned j = i + 1; j < NumPointers; ++j) {
4182 // Only need to check pointers between two different dependency sets.
4183 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
4185 // Only need to check pointers in the same alias set.
4186 if (RtCheck.AliasSetId[i] != RtCheck.AliasSetId[j])
4189 Value *PtrI = RtCheck.Pointers[i];
4190 Value *PtrJ = RtCheck.Pointers[j];
4192 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
4193 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
4195 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
4196 " different address spaces\n");
4205 void AccessAnalysis::processMemAccesses() {
4206 // We process the set twice: first we process read-write pointers, last we
4207 // process read-only pointers. This allows us to skip dependence tests for
4208 // read-only pointers.
4210 DEBUG(dbgs() << "LV: Processing memory accesses...\n");
4211 DEBUG(dbgs() << " AST: "; AST.dump());
4212 DEBUG(dbgs() << "LV: Accesses:\n");
4214 for (auto A : Accesses)
4215 dbgs() << "\t" << *A.getPointer() << " (" <<
4216 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
4217 "read-only" : "read")) << ")\n";
4220 // The AliasSetTracker has nicely partitioned our pointers by metadata
4221 // compatibility and potential for underlying-object overlap. As a result, we
4222 // only need to check for potential pointer dependencies within each alias
4224 for (auto &AS : AST) {
4225 // Note that both the alias-set tracker and the alias sets themselves used
4226 // linked lists internally and so the iteration order here is deterministic
4227 // (matching the original instruction order within each set).
4229 bool SetHasWrite = false;
4231 // Map of pointers to last access encountered.
4232 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
4233 UnderlyingObjToAccessMap ObjToLastAccess;
4235 // Set of access to check after all writes have been processed.
4236 PtrAccessSet DeferredAccesses;
4238 // Iterate over each alias set twice, once to process read/write pointers,
4239 // and then to process read-only pointers.
4240 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
4241 bool UseDeferred = SetIteration > 0;
4242 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
4245 Value *Ptr = A.getValue();
4246 bool IsWrite = S.count(MemAccessInfo(Ptr, true));
4248 // If we're using the deferred access set, then it contains only reads.
4249 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
4250 if (UseDeferred && !IsReadOnlyPtr)
4252 // Otherwise, the pointer must be in the PtrAccessSet, either as a read
4254 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
4255 S.count(MemAccessInfo(Ptr, false))) &&
4256 "Alias-set pointer not in the access set?");
4258 MemAccessInfo Access(Ptr, IsWrite);
4259 DepCands.insert(Access);
4261 // Memorize read-only pointers for later processing and skip them in the
4262 // first round (they need to be checked after we have seen all write
4263 // pointers). Note: we also mark pointer that are not consecutive as
4264 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need
4265 // the second check for "!IsWrite".
4266 if (!UseDeferred && IsReadOnlyPtr) {
4267 DeferredAccesses.insert(Access);
4271 // If this is a write - check other reads and writes for conflicts. If
4272 // this is a read only check other writes for conflicts (but only if
4273 // there is no other write to the ptr - this is an optimization to
4274 // catch "a[i] = a[i] + " without having to do a dependence check).
4275 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
4276 CheckDeps.insert(Access);
4277 IsRTCheckNeeded = true;
4283 // Create sets of pointers connected by a shared alias set and
4284 // underlying object.
4285 typedef SmallVector<Value*, 16> ValueVector;
4286 ValueVector TempObjects;
4287 GetUnderlyingObjects(Ptr, TempObjects, DL);
4288 for (Value *UnderlyingObj : TempObjects) {
4289 UnderlyingObjToAccessMap::iterator Prev =
4290 ObjToLastAccess.find(UnderlyingObj);
4291 if (Prev != ObjToLastAccess.end())
4292 DepCands.unionSets(Access, Prev->second);
4294 ObjToLastAccess[UnderlyingObj] = Access;
4302 /// \brief Checks memory dependences among accesses to the same underlying
4303 /// object to determine whether there vectorization is legal or not (and at
4304 /// which vectorization factor).
4306 /// This class works under the assumption that we already checked that memory
4307 /// locations with different underlying pointers are "must-not alias".
4308 /// We use the ScalarEvolution framework to symbolically evalutate access
4309 /// functions pairs. Since we currently don't restructure the loop we can rely
4310 /// on the program order of memory accesses to determine their safety.
4311 /// At the moment we will only deem accesses as safe for:
4312 /// * A negative constant distance assuming program order.
4314 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
4315 /// a[i] = tmp; y = a[i];
4317 /// The latter case is safe because later checks guarantuee that there can't
4318 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
4319 /// the same variable: a header phi can only be an induction or a reduction, a
4320 /// reduction can't have a memory sink, an induction can't have a memory
4321 /// source). This is important and must not be violated (or we have to
4322 /// resort to checking for cycles through memory).
4324 /// * A positive constant distance assuming program order that is bigger
4325 /// than the biggest memory access.
4327 /// tmp = a[i] OR b[i] = x
4328 /// a[i+2] = tmp y = b[i+2];
4330 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
4332 /// * Zero distances and all accesses have the same size.
4334 class MemoryDepChecker {
4336 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
4337 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
4339 MemoryDepChecker(ScalarEvolution *Se, const DataLayout *Dl, const Loop *L)
4340 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
4341 ShouldRetryWithRuntimeCheck(false) {}
4343 /// \brief Register the location (instructions are given increasing numbers)
4344 /// of a write access.
4345 void addAccess(StoreInst *SI) {
4346 Value *Ptr = SI->getPointerOperand();
4347 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
4348 InstMap.push_back(SI);
4352 /// \brief Register the location (instructions are given increasing numbers)
4353 /// of a write access.
4354 void addAccess(LoadInst *LI) {
4355 Value *Ptr = LI->getPointerOperand();
4356 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
4357 InstMap.push_back(LI);
4361 /// \brief Check whether the dependencies between the accesses are safe.
4363 /// Only checks sets with elements in \p CheckDeps.
4364 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4365 MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
4367 /// \brief The maximum number of bytes of a vector register we can vectorize
4368 /// the accesses safely with.
4369 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
4371 /// \brief In same cases when the dependency check fails we can still
4372 /// vectorize the loop with a dynamic array access check.
4373 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
4376 ScalarEvolution *SE;
4377 const DataLayout *DL;
4378 const Loop *InnermostLoop;
4380 /// \brief Maps access locations (ptr, read/write) to program order.
4381 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
4383 /// \brief Memory access instructions in program order.
4384 SmallVector<Instruction *, 16> InstMap;
4386 /// \brief The program order index to be used for the next instruction.
4389 // We can access this many bytes in parallel safely.
4390 unsigned MaxSafeDepDistBytes;
4392 /// \brief If we see a non-constant dependence distance we can still try to
4393 /// vectorize this loop with runtime checks.
4394 bool ShouldRetryWithRuntimeCheck;
4396 /// \brief Check whether there is a plausible dependence between the two
4399 /// Access \p A must happen before \p B in program order. The two indices
4400 /// identify the index into the program order map.
4402 /// This function checks whether there is a plausible dependence (or the
4403 /// absence of such can't be proved) between the two accesses. If there is a
4404 /// plausible dependence but the dependence distance is bigger than one
4405 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
4406 /// distance is smaller than any other distance encountered so far).
4407 /// Otherwise, this function returns true signaling a possible dependence.
4408 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
4409 const MemAccessInfo &B, unsigned BIdx,
4410 ValueToValueMap &Strides);
4412 /// \brief Check whether the data dependence could prevent store-load
4414 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
4417 } // end anonymous namespace
4419 static bool isInBoundsGep(Value *Ptr) {
4420 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
4421 return GEP->isInBounds();
4425 /// \brief Check whether the access through \p Ptr has a constant stride.
4426 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4427 const Loop *Lp, ValueToValueMap &StridesMap) {
4428 const Type *Ty = Ptr->getType();
4429 assert(Ty->isPointerTy() && "Unexpected non-ptr");
4431 // Make sure that the pointer does not point to aggregate types.
4432 const PointerType *PtrTy = cast<PointerType>(Ty);
4433 if (PtrTy->getElementType()->isAggregateType()) {
4434 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
4439 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
4441 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4443 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
4444 << *Ptr << " SCEV: " << *PtrScev << "\n");
4448 // The accesss function must stride over the innermost loop.
4449 if (Lp != AR->getLoop()) {
4450 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
4451 *Ptr << " SCEV: " << *PtrScev << "\n");
4454 // The address calculation must not wrap. Otherwise, a dependence could be
4456 // An inbounds getelementptr that is a AddRec with a unit stride
4457 // cannot wrap per definition. The unit stride requirement is checked later.
4458 // An getelementptr without an inbounds attribute and unit stride would have
4459 // to access the pointer value "0" which is undefined behavior in address
4460 // space 0, therefore we can also vectorize this case.
4461 bool IsInBoundsGEP = isInBoundsGep(Ptr);
4462 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
4463 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
4464 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
4465 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
4466 << *Ptr << " SCEV: " << *PtrScev << "\n");
4470 // Check the step is constant.
4471 const SCEV *Step = AR->getStepRecurrence(*SE);
4473 // Calculate the pointer stride and check if it is consecutive.
4474 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4476 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
4477 " SCEV: " << *PtrScev << "\n");
4481 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
4482 const APInt &APStepVal = C->getValue()->getValue();
4484 // Huge step value - give up.
4485 if (APStepVal.getBitWidth() > 64)
4488 int64_t StepVal = APStepVal.getSExtValue();
4491 int64_t Stride = StepVal / Size;
4492 int64_t Rem = StepVal % Size;
4496 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
4497 // know we can't "wrap around the address space". In case of address space
4498 // zero we know that this won't happen without triggering undefined behavior.
4499 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
4500 Stride != 1 && Stride != -1)
4506 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
4507 unsigned TypeByteSize) {
4508 // If loads occur at a distance that is not a multiple of a feasible vector
4509 // factor store-load forwarding does not take place.
4510 // Positive dependences might cause troubles because vectorizing them might
4511 // prevent store-load forwarding making vectorized code run a lot slower.
4512 // a[i] = a[i-3] ^ a[i-8];
4513 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
4514 // hence on your typical architecture store-load forwarding does not take
4515 // place. Vectorizing in such cases does not make sense.
4516 // Store-load forwarding distance.
4517 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
4518 // Maximum vector factor.
4519 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
4520 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
4521 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
4523 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
4525 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
4526 MaxVFWithoutSLForwardIssues = (vf >>=1);
4531 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
4532 DEBUG(dbgs() << "LV: Distance " << Distance <<
4533 " that could cause a store-load forwarding conflict\n");
4537 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
4538 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
4539 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
4543 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
4544 const MemAccessInfo &B, unsigned BIdx,
4545 ValueToValueMap &Strides) {
4546 assert (AIdx < BIdx && "Must pass arguments in program order");
4548 Value *APtr = A.getPointer();
4549 Value *BPtr = B.getPointer();
4550 bool AIsWrite = A.getInt();
4551 bool BIsWrite = B.getInt();
4553 // Two reads are independent.
4554 if (!AIsWrite && !BIsWrite)
4557 // We cannot check pointers in different address spaces.
4558 if (APtr->getType()->getPointerAddressSpace() !=
4559 BPtr->getType()->getPointerAddressSpace())
4562 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
4563 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
4565 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
4566 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
4568 const SCEV *Src = AScev;
4569 const SCEV *Sink = BScev;
4571 // If the induction step is negative we have to invert source and sink of the
4573 if (StrideAPtr < 0) {
4576 std::swap(APtr, BPtr);
4577 std::swap(Src, Sink);
4578 std::swap(AIsWrite, BIsWrite);
4579 std::swap(AIdx, BIdx);
4580 std::swap(StrideAPtr, StrideBPtr);
4583 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
4585 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
4586 << "(Induction step: " << StrideAPtr << ")\n");
4587 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
4588 << *InstMap[BIdx] << ": " << *Dist << "\n");
4590 // Need consecutive accesses. We don't want to vectorize
4591 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
4592 // the address space.
4593 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
4594 DEBUG(dbgs() << "Non-consecutive pointer access\n");
4598 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
4600 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
4601 ShouldRetryWithRuntimeCheck = true;
4605 Type *ATy = APtr->getType()->getPointerElementType();
4606 Type *BTy = BPtr->getType()->getPointerElementType();
4607 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
4609 // Negative distances are not plausible dependencies.
4610 const APInt &Val = C->getValue()->getValue();
4611 if (Val.isNegative()) {
4612 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
4613 if (IsTrueDataDependence &&
4614 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
4618 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
4622 // Write to the same location with the same size.
4623 // Could be improved to assert type sizes are the same (i32 == float, etc).
4627 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
4631 assert(Val.isStrictlyPositive() && "Expect a positive value");
4633 // Positive distance bigger than max vectorization factor.
4636 "LV: ReadWrite-Write positive dependency with different types\n");
4640 unsigned Distance = (unsigned) Val.getZExtValue();
4642 // Bail out early if passed-in parameters make vectorization not feasible.
4643 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
4644 unsigned ForcedUnroll = VectorizationInterleave ? VectorizationInterleave : 1;
4646 // The distance must be bigger than the size needed for a vectorized version
4647 // of the operation and the size of the vectorized operation must not be
4648 // bigger than the currrent maximum size.
4649 if (Distance < 2*TypeByteSize ||
4650 2*TypeByteSize > MaxSafeDepDistBytes ||
4651 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
4652 DEBUG(dbgs() << "LV: Failure because of Positive distance "
4653 << Val.getSExtValue() << '\n');
4657 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
4658 Distance : MaxSafeDepDistBytes;
4660 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
4661 if (IsTrueDataDependence &&
4662 couldPreventStoreLoadForward(Distance, TypeByteSize))
4665 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
4666 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
4671 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4672 MemAccessInfoSet &CheckDeps,
4673 ValueToValueMap &Strides) {
4675 MaxSafeDepDistBytes = -1U;
4676 while (!CheckDeps.empty()) {
4677 MemAccessInfo CurAccess = *CheckDeps.begin();
4679 // Get the relevant memory access set.
4680 EquivalenceClasses<MemAccessInfo>::iterator I =
4681 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
4683 // Check accesses within this set.
4684 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
4685 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
4687 // Check every access pair.
4689 CheckDeps.erase(*AI);
4690 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
4692 // Check every accessing instruction pair in program order.
4693 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
4694 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
4695 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
4696 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
4697 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
4699 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
4710 bool LoopVectorizationLegality::canVectorizeMemory() {
4712 typedef SmallVector<Value*, 16> ValueVector;
4713 typedef SmallPtrSet<Value*, 16> ValueSet;
4715 // Holds the Load and Store *instructions*.
4719 // Holds all the different accesses in the loop.
4720 unsigned NumReads = 0;
4721 unsigned NumReadWrites = 0;
4723 PtrRtCheck.Pointers.clear();
4724 PtrRtCheck.Need = false;
4726 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
4727 MemoryDepChecker DepChecker(SE, DL, TheLoop);
4730 for (Loop::block_iterator bb = TheLoop->block_begin(),
4731 be = TheLoop->block_end(); bb != be; ++bb) {
4733 // Scan the BB and collect legal loads and stores.
4734 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4737 // If this is a load, save it. If this instruction can read from memory
4738 // but is not a load, then we quit. Notice that we don't handle function
4739 // calls that read or write.
4740 if (it->mayReadFromMemory()) {
4741 // Many math library functions read the rounding mode. We will only
4742 // vectorize a loop if it contains known function calls that don't set
4743 // the flag. Therefore, it is safe to ignore this read from memory.
4744 CallInst *Call = dyn_cast<CallInst>(it);
4745 if (Call && getIntrinsicIDForCall(Call, TLI))
4748 LoadInst *Ld = dyn_cast<LoadInst>(it);
4749 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
4750 emitAnalysis(Report(Ld)
4751 << "read with atomic ordering or volatile read");
4752 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
4756 Loads.push_back(Ld);
4757 DepChecker.addAccess(Ld);
4761 // Save 'store' instructions. Abort if other instructions write to memory.
4762 if (it->mayWriteToMemory()) {
4763 StoreInst *St = dyn_cast<StoreInst>(it);
4765 emitAnalysis(Report(it) << "instruction cannot be vectorized");
4768 if (!St->isSimple() && !IsAnnotatedParallel) {
4769 emitAnalysis(Report(St)
4770 << "write with atomic ordering or volatile write");
4771 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
4775 Stores.push_back(St);
4776 DepChecker.addAccess(St);
4781 // Now we have two lists that hold the loads and the stores.
4782 // Next, we find the pointers that they use.
4784 // Check if we see any stores. If there are no stores, then we don't
4785 // care if the pointers are *restrict*.
4786 if (!Stores.size()) {
4787 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
4791 AccessAnalysis::DepCandidates DependentAccesses;
4792 AccessAnalysis Accesses(DL, AA, DependentAccesses);
4794 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
4795 // multiple times on the same object. If the ptr is accessed twice, once
4796 // for read and once for write, it will only appear once (on the write
4797 // list). This is okay, since we are going to check for conflicts between
4798 // writes and between reads and writes, but not between reads and reads.
4801 ValueVector::iterator I, IE;
4802 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
4803 StoreInst *ST = cast<StoreInst>(*I);
4804 Value* Ptr = ST->getPointerOperand();
4806 if (isUniform(Ptr)) {
4809 << "write to a loop invariant address could not be vectorized");
4810 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4814 // If we did *not* see this pointer before, insert it to the read-write
4815 // list. At this phase it is only a 'write' list.
4816 if (Seen.insert(Ptr)) {
4819 AliasAnalysis::Location Loc = AA->getLocation(ST);
4820 // The TBAA metadata could have a control dependency on the predication
4821 // condition, so we cannot rely on it when determining whether or not we
4822 // need runtime pointer checks.
4823 if (blockNeedsPredication(ST->getParent()))
4824 Loc.AATags.TBAA = nullptr;
4826 Accesses.addStore(Loc);
4830 if (IsAnnotatedParallel) {
4832 << "LV: A loop annotated parallel, ignore memory dependency "
4837 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
4838 LoadInst *LD = cast<LoadInst>(*I);
4839 Value* Ptr = LD->getPointerOperand();
4840 // If we did *not* see this pointer before, insert it to the
4841 // read list. If we *did* see it before, then it is already in
4842 // the read-write list. This allows us to vectorize expressions
4843 // such as A[i] += x; Because the address of A[i] is a read-write
4844 // pointer. This only works if the index of A[i] is consecutive.
4845 // If the address of i is unknown (for example A[B[i]]) then we may
4846 // read a few words, modify, and write a few words, and some of the
4847 // words may be written to the same address.
4848 bool IsReadOnlyPtr = false;
4849 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
4851 IsReadOnlyPtr = true;
4854 AliasAnalysis::Location Loc = AA->getLocation(LD);
4855 // The TBAA metadata could have a control dependency on the predication
4856 // condition, so we cannot rely on it when determining whether or not we
4857 // need runtime pointer checks.
4858 if (blockNeedsPredication(LD->getParent()))
4859 Loc.AATags.TBAA = nullptr;
4861 Accesses.addLoad(Loc, IsReadOnlyPtr);
4864 // If we write (or read-write) to a single destination and there are no
4865 // other reads in this loop then is it safe to vectorize.
4866 if (NumReadWrites == 1 && NumReads == 0) {
4867 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
4871 // Build dependence sets and check whether we need a runtime pointer bounds
4873 Accesses.buildDependenceSets();
4874 bool NeedRTCheck = Accesses.isRTCheckNeeded();
4876 // Find pointers with computable bounds. We are going to use this information
4877 // to place a runtime bound check.
4878 unsigned NumComparisons = 0;
4879 bool CanDoRT = false;
4881 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
4884 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
4885 " pointer comparisons.\n");
4887 // If we only have one set of dependences to check pointers among we don't
4888 // need a runtime check.
4889 if (NumComparisons == 0 && NeedRTCheck)
4890 NeedRTCheck = false;
4892 // Check that we did not collect too many pointers or found an unsizeable
4894 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4900 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
4903 if (NeedRTCheck && !CanDoRT) {
4904 emitAnalysis(Report() << "cannot identify array bounds");
4905 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
4906 "the array bounds.\n");
4911 PtrRtCheck.Need = NeedRTCheck;
4913 bool CanVecMem = true;
4914 if (Accesses.isDependencyCheckNeeded()) {
4915 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
4916 CanVecMem = DepChecker.areDepsSafe(
4917 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
4918 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
4920 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
4921 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
4924 // Clear the dependency checks. We assume they are not needed.
4925 Accesses.resetDepChecks();
4928 PtrRtCheck.Need = true;
4930 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
4931 TheLoop, Strides, true);
4932 // Check that we did not collect too many pointers or found an unsizeable
4934 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4935 if (!CanDoRT && NumComparisons > 0)
4936 emitAnalysis(Report()
4937 << "cannot check memory dependencies at runtime");
4939 emitAnalysis(Report()
4940 << NumComparisons << " exceeds limit of "
4941 << RuntimeMemoryCheckThreshold
4942 << " dependent memory operations checked at runtime");
4943 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
4953 emitAnalysis(Report() << "unsafe dependent memory operations in loop");
4955 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
4956 " need a runtime memory check.\n");
4961 static bool hasMultipleUsesOf(Instruction *I,
4962 SmallPtrSetImpl<Instruction *> &Insts) {
4963 unsigned NumUses = 0;
4964 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
4965 if (Insts.count(dyn_cast<Instruction>(*Use)))
4974 static bool areAllUsesIn(Instruction *I, SmallPtrSetImpl<Instruction *> &Set) {
4975 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
4976 if (!Set.count(dyn_cast<Instruction>(*Use)))
4981 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
4982 ReductionKind Kind) {
4983 if (Phi->getNumIncomingValues() != 2)
4986 // Reduction variables are only found in the loop header block.
4987 if (Phi->getParent() != TheLoop->getHeader())
4990 // Obtain the reduction start value from the value that comes from the loop
4992 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
4994 // ExitInstruction is the single value which is used outside the loop.
4995 // We only allow for a single reduction value to be used outside the loop.
4996 // This includes users of the reduction, variables (which form a cycle
4997 // which ends in the phi node).
4998 Instruction *ExitInstruction = nullptr;
4999 // Indicates that we found a reduction operation in our scan.
5000 bool FoundReduxOp = false;
5002 // We start with the PHI node and scan for all of the users of this
5003 // instruction. All users must be instructions that can be used as reduction
5004 // variables (such as ADD). We must have a single out-of-block user. The cycle
5005 // must include the original PHI.
5006 bool FoundStartPHI = false;
5008 // To recognize min/max patterns formed by a icmp select sequence, we store
5009 // the number of instruction we saw from the recognized min/max pattern,
5010 // to make sure we only see exactly the two instructions.
5011 unsigned NumCmpSelectPatternInst = 0;
5012 ReductionInstDesc ReduxDesc(false, nullptr);
5014 SmallPtrSet<Instruction *, 8> VisitedInsts;
5015 SmallVector<Instruction *, 8> Worklist;
5016 Worklist.push_back(Phi);
5017 VisitedInsts.insert(Phi);
5019 // A value in the reduction can be used:
5020 // - By the reduction:
5021 // - Reduction operation:
5022 // - One use of reduction value (safe).
5023 // - Multiple use of reduction value (not safe).
5025 // - All uses of the PHI must be the reduction (safe).
5026 // - Otherwise, not safe.
5027 // - By one instruction outside of the loop (safe).
5028 // - By further instructions outside of the loop (not safe).
5029 // - By an instruction that is not part of the reduction (not safe).
5031 // * An instruction type other than PHI or the reduction operation.
5032 // * A PHI in the header other than the initial PHI.
5033 while (!Worklist.empty()) {
5034 Instruction *Cur = Worklist.back();
5035 Worklist.pop_back();
5038 // If the instruction has no users then this is a broken chain and can't be
5039 // a reduction variable.
5040 if (Cur->use_empty())
5043 bool IsAPhi = isa<PHINode>(Cur);
5045 // A header PHI use other than the original PHI.
5046 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
5049 // Reductions of instructions such as Div, and Sub is only possible if the
5050 // LHS is the reduction variable.
5051 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
5052 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
5053 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
5056 // Any reduction instruction must be of one of the allowed kinds.
5057 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
5058 if (!ReduxDesc.IsReduction)
5061 // A reduction operation must only have one use of the reduction value.
5062 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
5063 hasMultipleUsesOf(Cur, VisitedInsts))
5066 // All inputs to a PHI node must be a reduction value.
5067 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
5070 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
5071 isa<SelectInst>(Cur)))
5072 ++NumCmpSelectPatternInst;
5073 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
5074 isa<SelectInst>(Cur)))
5075 ++NumCmpSelectPatternInst;
5077 // Check whether we found a reduction operator.
5078 FoundReduxOp |= !IsAPhi;
5080 // Process users of current instruction. Push non-PHI nodes after PHI nodes
5081 // onto the stack. This way we are going to have seen all inputs to PHI
5082 // nodes once we get to them.
5083 SmallVector<Instruction *, 8> NonPHIs;
5084 SmallVector<Instruction *, 8> PHIs;
5085 for (User *U : Cur->users()) {
5086 Instruction *UI = cast<Instruction>(U);
5088 // Check if we found the exit user.
5089 BasicBlock *Parent = UI->getParent();
5090 if (!TheLoop->contains(Parent)) {
5091 // Exit if you find multiple outside users or if the header phi node is
5092 // being used. In this case the user uses the value of the previous
5093 // iteration, in which case we would loose "VF-1" iterations of the
5094 // reduction operation if we vectorize.
5095 if (ExitInstruction != nullptr || Cur == Phi)
5098 // The instruction used by an outside user must be the last instruction
5099 // before we feed back to the reduction phi. Otherwise, we loose VF-1
5100 // operations on the value.
5101 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
5104 ExitInstruction = Cur;
5108 // Process instructions only once (termination). Each reduction cycle
5109 // value must only be used once, except by phi nodes and min/max
5110 // reductions which are represented as a cmp followed by a select.
5111 ReductionInstDesc IgnoredVal(false, nullptr);
5112 if (VisitedInsts.insert(UI)) {
5113 if (isa<PHINode>(UI))
5116 NonPHIs.push_back(UI);
5117 } else if (!isa<PHINode>(UI) &&
5118 ((!isa<FCmpInst>(UI) &&
5119 !isa<ICmpInst>(UI) &&
5120 !isa<SelectInst>(UI)) ||
5121 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
5124 // Remember that we completed the cycle.
5126 FoundStartPHI = true;
5128 Worklist.append(PHIs.begin(), PHIs.end());
5129 Worklist.append(NonPHIs.begin(), NonPHIs.end());
5132 // This means we have seen one but not the other instruction of the
5133 // pattern or more than just a select and cmp.
5134 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
5135 NumCmpSelectPatternInst != 2)
5138 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
5141 // We found a reduction var if we have reached the original phi node and we
5142 // only have a single instruction with out-of-loop users.
5144 // This instruction is allowed to have out-of-loop users.
5145 AllowedExit.insert(ExitInstruction);
5147 // Save the description of this reduction variable.
5148 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
5149 ReduxDesc.MinMaxKind);
5150 Reductions[Phi] = RD;
5151 // We've ended the cycle. This is a reduction variable if we have an
5152 // outside user and it has a binary op.
5157 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
5158 /// pattern corresponding to a min(X, Y) or max(X, Y).
5159 LoopVectorizationLegality::ReductionInstDesc
5160 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
5161 ReductionInstDesc &Prev) {
5163 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
5164 "Expect a select instruction");
5165 Instruction *Cmp = nullptr;
5166 SelectInst *Select = nullptr;
5168 // We must handle the select(cmp()) as a single instruction. Advance to the
5170 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
5171 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
5172 return ReductionInstDesc(false, I);
5173 return ReductionInstDesc(Select, Prev.MinMaxKind);
5176 // Only handle single use cases for now.
5177 if (!(Select = dyn_cast<SelectInst>(I)))
5178 return ReductionInstDesc(false, I);
5179 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
5180 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
5181 return ReductionInstDesc(false, I);
5182 if (!Cmp->hasOneUse())
5183 return ReductionInstDesc(false, I);
5188 // Look for a min/max pattern.
5189 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5190 return ReductionInstDesc(Select, MRK_UIntMin);
5191 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5192 return ReductionInstDesc(Select, MRK_UIntMax);
5193 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5194 return ReductionInstDesc(Select, MRK_SIntMax);
5195 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5196 return ReductionInstDesc(Select, MRK_SIntMin);
5197 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5198 return ReductionInstDesc(Select, MRK_FloatMin);
5199 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5200 return ReductionInstDesc(Select, MRK_FloatMax);
5201 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5202 return ReductionInstDesc(Select, MRK_FloatMin);
5203 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5204 return ReductionInstDesc(Select, MRK_FloatMax);
5206 return ReductionInstDesc(false, I);
5209 LoopVectorizationLegality::ReductionInstDesc
5210 LoopVectorizationLegality::isReductionInstr(Instruction *I,
5212 ReductionInstDesc &Prev) {
5213 bool FP = I->getType()->isFloatingPointTy();
5214 bool FastMath = FP && I->hasUnsafeAlgebra();
5215 switch (I->getOpcode()) {
5217 return ReductionInstDesc(false, I);
5218 case Instruction::PHI:
5219 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
5220 Kind != RK_FloatMinMax))
5221 return ReductionInstDesc(false, I);
5222 return ReductionInstDesc(I, Prev.MinMaxKind);
5223 case Instruction::Sub:
5224 case Instruction::Add:
5225 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
5226 case Instruction::Mul:
5227 return ReductionInstDesc(Kind == RK_IntegerMult, I);
5228 case Instruction::And:
5229 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
5230 case Instruction::Or:
5231 return ReductionInstDesc(Kind == RK_IntegerOr, I);
5232 case Instruction::Xor:
5233 return ReductionInstDesc(Kind == RK_IntegerXor, I);
5234 case Instruction::FMul:
5235 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
5236 case Instruction::FSub:
5237 case Instruction::FAdd:
5238 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
5239 case Instruction::FCmp:
5240 case Instruction::ICmp:
5241 case Instruction::Select:
5242 if (Kind != RK_IntegerMinMax &&
5243 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
5244 return ReductionInstDesc(false, I);
5245 return isMinMaxSelectCmpPattern(I, Prev);
5249 LoopVectorizationLegality::InductionKind
5250 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
5251 Type *PhiTy = Phi->getType();
5252 // We only handle integer and pointer inductions variables.
5253 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
5254 return IK_NoInduction;
5256 // Check that the PHI is consecutive.
5257 const SCEV *PhiScev = SE->getSCEV(Phi);
5258 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
5260 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
5261 return IK_NoInduction;
5263 const SCEV *Step = AR->getStepRecurrence(*SE);
5265 // Integer inductions need to have a stride of one.
5266 if (PhiTy->isIntegerTy()) {
5268 return IK_IntInduction;
5269 if (Step->isAllOnesValue())
5270 return IK_ReverseIntInduction;
5271 return IK_NoInduction;
5274 // Calculate the pointer stride and check if it is consecutive.
5275 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5277 return IK_NoInduction;
5279 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
5280 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
5281 if (C->getValue()->equalsInt(Size))
5282 return IK_PtrInduction;
5283 else if (C->getValue()->equalsInt(0 - Size))
5284 return IK_ReversePtrInduction;
5286 return IK_NoInduction;
5289 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
5290 Value *In0 = const_cast<Value*>(V);
5291 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
5295 return Inductions.count(PN);
5298 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
5299 assert(TheLoop->contains(BB) && "Unknown block used");
5301 // Blocks that do not dominate the latch need predication.
5302 BasicBlock* Latch = TheLoop->getLoopLatch();
5303 return !DT->dominates(BB, Latch);
5306 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
5307 SmallPtrSetImpl<Value *> &SafePtrs) {
5308 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5309 // We might be able to hoist the load.
5310 if (it->mayReadFromMemory()) {
5311 LoadInst *LI = dyn_cast<LoadInst>(it);
5312 if (!LI || !SafePtrs.count(LI->getPointerOperand()))
5316 // We don't predicate stores at the moment.
5317 if (it->mayWriteToMemory()) {
5318 StoreInst *SI = dyn_cast<StoreInst>(it);
5319 // We only support predication of stores in basic blocks with one
5321 if (!SI || ++NumPredStores > NumberOfStoresToPredicate ||
5322 !SafePtrs.count(SI->getPointerOperand()) ||
5323 !SI->getParent()->getSinglePredecessor())
5329 // Check that we don't have a constant expression that can trap as operand.
5330 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
5332 if (Constant *C = dyn_cast<Constant>(*OI))
5337 // The instructions below can trap.
5338 switch (it->getOpcode()) {
5340 case Instruction::UDiv:
5341 case Instruction::SDiv:
5342 case Instruction::URem:
5343 case Instruction::SRem:
5351 LoopVectorizationCostModel::VectorizationFactor
5352 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
5353 // Width 1 means no vectorize
5354 VectorizationFactor Factor = { 1U, 0U };
5355 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
5356 emitAnalysis(Report() << "runtime pointer checks needed. Enable vectorization of this loop with '#pragma clang loop vectorize(enable)' when compiling with -Os");
5357 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
5361 if (!EnableCondStoresVectorization && Legal->NumPredStores) {
5362 emitAnalysis(Report() << "store that is conditionally executed prevents vectorization");
5363 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
5367 // Find the trip count.
5368 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
5369 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
5371 unsigned WidestType = getWidestType();
5372 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
5373 unsigned MaxSafeDepDist = -1U;
5374 if (Legal->getMaxSafeDepDistBytes() != -1U)
5375 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
5376 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
5377 WidestRegister : MaxSafeDepDist);
5378 unsigned MaxVectorSize = WidestRegister / WidestType;
5379 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
5380 DEBUG(dbgs() << "LV: The Widest register is: "
5381 << WidestRegister << " bits.\n");
5383 if (MaxVectorSize == 0) {
5384 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
5388 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
5389 " into one vector!");
5391 unsigned VF = MaxVectorSize;
5393 // If we optimize the program for size, avoid creating the tail loop.
5395 // If we are unable to calculate the trip count then don't try to vectorize.
5397 emitAnalysis(Report() << "unable to calculate the loop count due to complex control flow");
5398 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5402 // Find the maximum SIMD width that can fit within the trip count.
5403 VF = TC % MaxVectorSize;
5408 // If the trip count that we found modulo the vectorization factor is not
5409 // zero then we require a tail.
5411 emitAnalysis(Report() << "cannot optimize for size and vectorize at the same time. Enable vectorization of this loop with '#pragma clang loop vectorize(enable)' when compiling with -Os");
5412 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5417 int UserVF = Hints->getWidth();
5419 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
5420 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
5422 Factor.Width = UserVF;
5426 float Cost = expectedCost(1);
5428 const float ScalarCost = Cost;
5431 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
5433 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
5434 // Ignore scalar width, because the user explicitly wants vectorization.
5435 if (ForceVectorization && VF > 1) {
5437 Cost = expectedCost(Width) / (float)Width;
5440 for (unsigned i=2; i <= VF; i*=2) {
5441 // Notice that the vector loop needs to be executed less times, so
5442 // we need to divide the cost of the vector loops by the width of
5443 // the vector elements.
5444 float VectorCost = expectedCost(i) / (float)i;
5445 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
5446 (int)VectorCost << ".\n");
5447 if (VectorCost < Cost) {
5453 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
5454 << "LV: Vectorization seems to be not beneficial, "
5455 << "but was forced by a user.\n");
5456 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
5457 Factor.Width = Width;
5458 Factor.Cost = Width * Cost;
5462 unsigned LoopVectorizationCostModel::getWidestType() {
5463 unsigned MaxWidth = 8;
5466 for (Loop::block_iterator bb = TheLoop->block_begin(),
5467 be = TheLoop->block_end(); bb != be; ++bb) {
5468 BasicBlock *BB = *bb;
5470 // For each instruction in the loop.
5471 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5472 Type *T = it->getType();
5474 // Ignore ephemeral values.
5475 if (EphValues.count(it))
5478 // Only examine Loads, Stores and PHINodes.
5479 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
5482 // Examine PHI nodes that are reduction variables.
5483 if (PHINode *PN = dyn_cast<PHINode>(it))
5484 if (!Legal->getReductionVars()->count(PN))
5487 // Examine the stored values.
5488 if (StoreInst *ST = dyn_cast<StoreInst>(it))
5489 T = ST->getValueOperand()->getType();
5491 // Ignore loaded pointer types and stored pointer types that are not
5492 // consecutive. However, we do want to take consecutive stores/loads of
5493 // pointer vectors into account.
5494 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
5497 MaxWidth = std::max(MaxWidth,
5498 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
5506 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
5508 unsigned LoopCost) {
5510 // -- The unroll heuristics --
5511 // We unroll the loop in order to expose ILP and reduce the loop overhead.
5512 // There are many micro-architectural considerations that we can't predict
5513 // at this level. For example, frontend pressure (on decode or fetch) due to
5514 // code size, or the number and capabilities of the execution ports.
5516 // We use the following heuristics to select the unroll factor:
5517 // 1. If the code has reductions, then we unroll in order to break the cross
5518 // iteration dependency.
5519 // 2. If the loop is really small, then we unroll in order to reduce the loop
5521 // 3. We don't unroll if we think that we will spill registers to memory due
5522 // to the increased register pressure.
5524 // Use the user preference, unless 'auto' is selected.
5525 int UserUF = Hints->getInterleave();
5529 // When we optimize for size, we don't unroll.
5533 // We used the distance for the unroll factor.
5534 if (Legal->getMaxSafeDepDistBytes() != -1U)
5537 // Do not unroll loops with a relatively small trip count.
5538 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
5539 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
5542 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
5543 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
5547 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
5548 TargetNumRegisters = ForceTargetNumScalarRegs;
5550 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
5551 TargetNumRegisters = ForceTargetNumVectorRegs;
5554 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
5555 // We divide by these constants so assume that we have at least one
5556 // instruction that uses at least one register.
5557 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
5558 R.NumInstructions = std::max(R.NumInstructions, 1U);
5560 // We calculate the unroll factor using the following formula.
5561 // Subtract the number of loop invariants from the number of available
5562 // registers. These registers are used by all of the unrolled instances.
5563 // Next, divide the remaining registers by the number of registers that is
5564 // required by the loop, in order to estimate how many parallel instances
5565 // fit without causing spills. All of this is rounded down if necessary to be
5566 // a power of two. We want power of two unroll factors to simplify any
5567 // addressing operations or alignment considerations.
5568 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
5571 // Don't count the induction variable as unrolled.
5572 if (EnableIndVarRegisterHeur)
5573 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
5574 std::max(1U, (R.MaxLocalUsers - 1)));
5576 // Clamp the unroll factor ranges to reasonable factors.
5577 unsigned MaxInterleaveSize = TTI.getMaxInterleaveFactor();
5579 // Check if the user has overridden the unroll max.
5581 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
5582 MaxInterleaveSize = ForceTargetMaxScalarInterleaveFactor;
5584 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
5585 MaxInterleaveSize = ForceTargetMaxVectorInterleaveFactor;
5588 // If we did not calculate the cost for VF (because the user selected the VF)
5589 // then we calculate the cost of VF here.
5591 LoopCost = expectedCost(VF);
5593 // Clamp the calculated UF to be between the 1 and the max unroll factor
5594 // that the target allows.
5595 if (UF > MaxInterleaveSize)
5596 UF = MaxInterleaveSize;
5600 // Unroll if we vectorized this loop and there is a reduction that could
5601 // benefit from unrolling.
5602 if (VF > 1 && Legal->getReductionVars()->size()) {
5603 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
5607 // Note that if we've already vectorized the loop we will have done the
5608 // runtime check and so unrolling won't require further checks.
5609 bool UnrollingRequiresRuntimePointerCheck =
5610 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
5612 // We want to unroll small loops in order to reduce the loop overhead and
5613 // potentially expose ILP opportunities.
5614 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
5615 if (!UnrollingRequiresRuntimePointerCheck &&
5616 LoopCost < SmallLoopCost) {
5617 // We assume that the cost overhead is 1 and we use the cost model
5618 // to estimate the cost of the loop and unroll until the cost of the
5619 // loop overhead is about 5% of the cost of the loop.
5620 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
5622 // Unroll until store/load ports (estimated by max unroll factor) are
5624 unsigned StoresUF = UF / (Legal->NumStores ? Legal->NumStores : 1);
5625 unsigned LoadsUF = UF / (Legal->NumLoads ? Legal->NumLoads : 1);
5627 // If we have a scalar reduction (vector reductions are already dealt with
5628 // by this point), we can increase the critical path length if the loop
5629 // we're unrolling is inside another loop. Limit, by default to 2, so the
5630 // critical path only gets increased by one reduction operation.
5631 if (Legal->getReductionVars()->size() &&
5632 TheLoop->getLoopDepth() > 1) {
5633 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionUF);
5634 SmallUF = std::min(SmallUF, F);
5635 StoresUF = std::min(StoresUF, F);
5636 LoadsUF = std::min(LoadsUF, F);
5639 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
5640 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
5641 return std::max(StoresUF, LoadsUF);
5644 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
5648 DEBUG(dbgs() << "LV: Not Unrolling.\n");
5652 LoopVectorizationCostModel::RegisterUsage
5653 LoopVectorizationCostModel::calculateRegisterUsage() {
5654 // This function calculates the register usage by measuring the highest number
5655 // of values that are alive at a single location. Obviously, this is a very
5656 // rough estimation. We scan the loop in a topological order in order and
5657 // assign a number to each instruction. We use RPO to ensure that defs are
5658 // met before their users. We assume that each instruction that has in-loop
5659 // users starts an interval. We record every time that an in-loop value is
5660 // used, so we have a list of the first and last occurrences of each
5661 // instruction. Next, we transpose this data structure into a multi map that
5662 // holds the list of intervals that *end* at a specific location. This multi
5663 // map allows us to perform a linear search. We scan the instructions linearly
5664 // and record each time that a new interval starts, by placing it in a set.
5665 // If we find this value in the multi-map then we remove it from the set.
5666 // The max register usage is the maximum size of the set.
5667 // We also search for instructions that are defined outside the loop, but are
5668 // used inside the loop. We need this number separately from the max-interval
5669 // usage number because when we unroll, loop-invariant values do not take
5671 LoopBlocksDFS DFS(TheLoop);
5675 R.NumInstructions = 0;
5677 // Each 'key' in the map opens a new interval. The values
5678 // of the map are the index of the 'last seen' usage of the
5679 // instruction that is the key.
5680 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5681 // Maps instruction to its index.
5682 DenseMap<unsigned, Instruction*> IdxToInstr;
5683 // Marks the end of each interval.
5684 IntervalMap EndPoint;
5685 // Saves the list of instruction indices that are used in the loop.
5686 SmallSet<Instruction*, 8> Ends;
5687 // Saves the list of values that are used in the loop but are
5688 // defined outside the loop, such as arguments and constants.
5689 SmallPtrSet<Value*, 8> LoopInvariants;
5692 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5693 be = DFS.endRPO(); bb != be; ++bb) {
5694 R.NumInstructions += (*bb)->size();
5695 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
5697 Instruction *I = it;
5698 IdxToInstr[Index++] = I;
5700 // Save the end location of each USE.
5701 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
5702 Value *U = I->getOperand(i);
5703 Instruction *Instr = dyn_cast<Instruction>(U);
5705 // Ignore non-instruction values such as arguments, constants, etc.
5706 if (!Instr) continue;
5708 // If this instruction is outside the loop then record it and continue.
5709 if (!TheLoop->contains(Instr)) {
5710 LoopInvariants.insert(Instr);
5714 // Overwrite previous end points.
5715 EndPoint[Instr] = Index;
5721 // Saves the list of intervals that end with the index in 'key'.
5722 typedef SmallVector<Instruction*, 2> InstrList;
5723 DenseMap<unsigned, InstrList> TransposeEnds;
5725 // Transpose the EndPoints to a list of values that end at each index.
5726 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5728 TransposeEnds[it->second].push_back(it->first);
5730 SmallSet<Instruction*, 8> OpenIntervals;
5731 unsigned MaxUsage = 0;
5734 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5735 for (unsigned int i = 0; i < Index; ++i) {
5736 Instruction *I = IdxToInstr[i];
5737 // Ignore instructions that are never used within the loop.
5738 if (!Ends.count(I)) continue;
5740 // Ignore ephemeral values.
5741 if (EphValues.count(I))
5744 // Remove all of the instructions that end at this location.
5745 InstrList &List = TransposeEnds[i];
5746 for (unsigned int j=0, e = List.size(); j < e; ++j)
5747 OpenIntervals.erase(List[j]);
5749 // Count the number of live interals.
5750 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5752 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5753 OpenIntervals.size() << '\n');
5755 // Add the current instruction to the list of open intervals.
5756 OpenIntervals.insert(I);
5759 unsigned Invariant = LoopInvariants.size();
5760 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5761 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5762 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5764 R.LoopInvariantRegs = Invariant;
5765 R.MaxLocalUsers = MaxUsage;
5769 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5773 for (Loop::block_iterator bb = TheLoop->block_begin(),
5774 be = TheLoop->block_end(); bb != be; ++bb) {
5775 unsigned BlockCost = 0;
5776 BasicBlock *BB = *bb;
5778 // For each instruction in the old loop.
5779 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5780 // Skip dbg intrinsics.
5781 if (isa<DbgInfoIntrinsic>(it))
5784 // Ignore ephemeral values.
5785 if (EphValues.count(it))
5788 unsigned C = getInstructionCost(it, VF);
5790 // Check if we should override the cost.
5791 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5792 C = ForceTargetInstructionCost;
5795 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5796 VF << " For instruction: " << *it << '\n');
5799 // We assume that if-converted blocks have a 50% chance of being executed.
5800 // When the code is scalar then some of the blocks are avoided due to CF.
5801 // When the code is vectorized we execute all code paths.
5802 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5811 /// \brief Check whether the address computation for a non-consecutive memory
5812 /// access looks like an unlikely candidate for being merged into the indexing
5815 /// We look for a GEP which has one index that is an induction variable and all
5816 /// other indices are loop invariant. If the stride of this access is also
5817 /// within a small bound we decide that this address computation can likely be
5818 /// merged into the addressing mode.
5819 /// In all other cases, we identify the address computation as complex.
5820 static bool isLikelyComplexAddressComputation(Value *Ptr,
5821 LoopVectorizationLegality *Legal,
5822 ScalarEvolution *SE,
5823 const Loop *TheLoop) {
5824 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5828 // We are looking for a gep with all loop invariant indices except for one
5829 // which should be an induction variable.
5830 unsigned NumOperands = Gep->getNumOperands();
5831 for (unsigned i = 1; i < NumOperands; ++i) {
5832 Value *Opd = Gep->getOperand(i);
5833 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5834 !Legal->isInductionVariable(Opd))
5838 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5839 // can likely be merged into the address computation.
5840 unsigned MaxMergeDistance = 64;
5842 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5846 // Check the step is constant.
5847 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5848 // Calculate the pointer stride and check if it is consecutive.
5849 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5853 const APInt &APStepVal = C->getValue()->getValue();
5855 // Huge step value - give up.
5856 if (APStepVal.getBitWidth() > 64)
5859 int64_t StepVal = APStepVal.getSExtValue();
5861 return StepVal > MaxMergeDistance;
5864 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5865 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5871 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5872 // If we know that this instruction will remain uniform, check the cost of
5873 // the scalar version.
5874 if (Legal->isUniformAfterVectorization(I))
5877 Type *RetTy = I->getType();
5878 Type *VectorTy = ToVectorTy(RetTy, VF);
5880 // TODO: We need to estimate the cost of intrinsic calls.
5881 switch (I->getOpcode()) {
5882 case Instruction::GetElementPtr:
5883 // We mark this instruction as zero-cost because the cost of GEPs in
5884 // vectorized code depends on whether the corresponding memory instruction
5885 // is scalarized or not. Therefore, we handle GEPs with the memory
5886 // instruction cost.
5888 case Instruction::Br: {
5889 return TTI.getCFInstrCost(I->getOpcode());
5891 case Instruction::PHI:
5892 //TODO: IF-converted IFs become selects.
5894 case Instruction::Add:
5895 case Instruction::FAdd:
5896 case Instruction::Sub:
5897 case Instruction::FSub:
5898 case Instruction::Mul:
5899 case Instruction::FMul:
5900 case Instruction::UDiv:
5901 case Instruction::SDiv:
5902 case Instruction::FDiv:
5903 case Instruction::URem:
5904 case Instruction::SRem:
5905 case Instruction::FRem:
5906 case Instruction::Shl:
5907 case Instruction::LShr:
5908 case Instruction::AShr:
5909 case Instruction::And:
5910 case Instruction::Or:
5911 case Instruction::Xor: {
5912 // Since we will replace the stride by 1 the multiplication should go away.
5913 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5915 // Certain instructions can be cheaper to vectorize if they have a constant
5916 // second vector operand. One example of this are shifts on x86.
5917 TargetTransformInfo::OperandValueKind Op1VK =
5918 TargetTransformInfo::OK_AnyValue;
5919 TargetTransformInfo::OperandValueKind Op2VK =
5920 TargetTransformInfo::OK_AnyValue;
5921 TargetTransformInfo::OperandValueProperties Op1VP =
5922 TargetTransformInfo::OP_None;
5923 TargetTransformInfo::OperandValueProperties Op2VP =
5924 TargetTransformInfo::OP_None;
5925 Value *Op2 = I->getOperand(1);
5927 // Check for a splat of a constant or for a non uniform vector of constants.
5928 if (isa<ConstantInt>(Op2)) {
5929 ConstantInt *CInt = cast<ConstantInt>(Op2);
5930 if (CInt && CInt->getValue().isPowerOf2())
5931 Op2VP = TargetTransformInfo::OP_PowerOf2;
5932 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5933 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5934 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5935 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
5937 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
5938 if (CInt && CInt->getValue().isPowerOf2())
5939 Op2VP = TargetTransformInfo::OP_PowerOf2;
5940 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5944 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
5947 case Instruction::Select: {
5948 SelectInst *SI = cast<SelectInst>(I);
5949 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5950 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5951 Type *CondTy = SI->getCondition()->getType();
5953 CondTy = VectorType::get(CondTy, VF);
5955 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5957 case Instruction::ICmp:
5958 case Instruction::FCmp: {
5959 Type *ValTy = I->getOperand(0)->getType();
5960 VectorTy = ToVectorTy(ValTy, VF);
5961 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5963 case Instruction::Store:
5964 case Instruction::Load: {
5965 StoreInst *SI = dyn_cast<StoreInst>(I);
5966 LoadInst *LI = dyn_cast<LoadInst>(I);
5967 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5969 VectorTy = ToVectorTy(ValTy, VF);
5971 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5972 unsigned AS = SI ? SI->getPointerAddressSpace() :
5973 LI->getPointerAddressSpace();
5974 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5975 // We add the cost of address computation here instead of with the gep
5976 // instruction because only here we know whether the operation is
5979 return TTI.getAddressComputationCost(VectorTy) +
5980 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5982 // Scalarized loads/stores.
5983 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5984 bool Reverse = ConsecutiveStride < 0;
5985 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
5986 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
5987 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5988 bool IsComplexComputation =
5989 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5991 // The cost of extracting from the value vector and pointer vector.
5992 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5993 for (unsigned i = 0; i < VF; ++i) {
5994 // The cost of extracting the pointer operand.
5995 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5996 // In case of STORE, the cost of ExtractElement from the vector.
5997 // In case of LOAD, the cost of InsertElement into the returned
5999 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
6000 Instruction::InsertElement,
6004 // The cost of the scalar loads/stores.
6005 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
6006 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
6011 // Wide load/stores.
6012 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
6013 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
6016 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
6020 case Instruction::ZExt:
6021 case Instruction::SExt:
6022 case Instruction::FPToUI:
6023 case Instruction::FPToSI:
6024 case Instruction::FPExt:
6025 case Instruction::PtrToInt:
6026 case Instruction::IntToPtr:
6027 case Instruction::SIToFP:
6028 case Instruction::UIToFP:
6029 case Instruction::Trunc:
6030 case Instruction::FPTrunc:
6031 case Instruction::BitCast: {
6032 // We optimize the truncation of induction variable.
6033 // The cost of these is the same as the scalar operation.
6034 if (I->getOpcode() == Instruction::Trunc &&
6035 Legal->isInductionVariable(I->getOperand(0)))
6036 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
6037 I->getOperand(0)->getType());
6039 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
6040 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
6042 case Instruction::Call: {
6043 CallInst *CI = cast<CallInst>(I);
6044 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
6045 assert(ID && "Not an intrinsic call!");
6046 Type *RetTy = ToVectorTy(CI->getType(), VF);
6047 SmallVector<Type*, 4> Tys;
6048 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
6049 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
6050 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
6053 // We are scalarizing the instruction. Return the cost of the scalar
6054 // instruction, plus the cost of insert and extract into vector
6055 // elements, times the vector width.
6058 if (!RetTy->isVoidTy() && VF != 1) {
6059 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
6061 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
6064 // The cost of inserting the results plus extracting each one of the
6066 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
6069 // The cost of executing VF copies of the scalar instruction. This opcode
6070 // is unknown. Assume that it is the same as 'mul'.
6071 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
6077 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
6078 if (Scalar->isVoidTy() || VF == 1)
6080 return VectorType::get(Scalar, VF);
6083 char LoopVectorize::ID = 0;
6084 static const char lv_name[] = "Loop Vectorization";
6085 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
6086 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
6087 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
6088 INITIALIZE_PASS_DEPENDENCY(AssumptionTracker)
6089 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
6090 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
6091 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
6092 INITIALIZE_PASS_DEPENDENCY(LCSSA)
6093 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
6094 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
6095 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
6098 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
6099 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
6103 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
6104 // Check for a store.
6105 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
6106 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
6108 // Check for a load.
6109 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
6110 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
6116 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
6117 bool IfPredicateStore) {
6118 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
6119 // Holds vector parameters or scalars, in case of uniform vals.
6120 SmallVector<VectorParts, 4> Params;
6122 setDebugLocFromInst(Builder, Instr);
6124 // Find all of the vectorized parameters.
6125 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
6126 Value *SrcOp = Instr->getOperand(op);
6128 // If we are accessing the old induction variable, use the new one.
6129 if (SrcOp == OldInduction) {
6130 Params.push_back(getVectorValue(SrcOp));
6134 // Try using previously calculated values.
6135 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
6137 // If the src is an instruction that appeared earlier in the basic block
6138 // then it should already be vectorized.
6139 if (SrcInst && OrigLoop->contains(SrcInst)) {
6140 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
6141 // The parameter is a vector value from earlier.
6142 Params.push_back(WidenMap.get(SrcInst));
6144 // The parameter is a scalar from outside the loop. Maybe even a constant.
6145 VectorParts Scalars;
6146 Scalars.append(UF, SrcOp);
6147 Params.push_back(Scalars);
6151 assert(Params.size() == Instr->getNumOperands() &&
6152 "Invalid number of operands");
6154 // Does this instruction return a value ?
6155 bool IsVoidRetTy = Instr->getType()->isVoidTy();
6157 Value *UndefVec = IsVoidRetTy ? nullptr :
6158 UndefValue::get(Instr->getType());
6159 // Create a new entry in the WidenMap and initialize it to Undef or Null.
6160 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
6162 Instruction *InsertPt = Builder.GetInsertPoint();
6163 BasicBlock *IfBlock = Builder.GetInsertBlock();
6164 BasicBlock *CondBlock = nullptr;
6167 Loop *VectorLp = nullptr;
6168 if (IfPredicateStore) {
6169 assert(Instr->getParent()->getSinglePredecessor() &&
6170 "Only support single predecessor blocks");
6171 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
6172 Instr->getParent());
6173 VectorLp = LI->getLoopFor(IfBlock);
6174 assert(VectorLp && "Must have a loop for this block");
6177 // For each vector unroll 'part':
6178 for (unsigned Part = 0; Part < UF; ++Part) {
6179 // For each scalar that we create:
6181 // Start an "if (pred) a[i] = ..." block.
6182 Value *Cmp = nullptr;
6183 if (IfPredicateStore) {
6184 if (Cond[Part]->getType()->isVectorTy())
6186 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
6187 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
6188 ConstantInt::get(Cond[Part]->getType(), 1));
6189 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
6190 LoopVectorBody.push_back(CondBlock);
6191 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
6192 // Update Builder with newly created basic block.
6193 Builder.SetInsertPoint(InsertPt);
6196 Instruction *Cloned = Instr->clone();
6198 Cloned->setName(Instr->getName() + ".cloned");
6199 // Replace the operands of the cloned instructions with extracted scalars.
6200 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
6201 Value *Op = Params[op][Part];
6202 Cloned->setOperand(op, Op);
6205 // Place the cloned scalar in the new loop.
6206 Builder.Insert(Cloned);
6208 // If the original scalar returns a value we need to place it in a vector
6209 // so that future users will be able to use it.
6211 VecResults[Part] = Cloned;
6214 if (IfPredicateStore) {
6215 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
6216 LoopVectorBody.push_back(NewIfBlock);
6217 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
6218 Builder.SetInsertPoint(InsertPt);
6219 Instruction *OldBr = IfBlock->getTerminator();
6220 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
6221 OldBr->eraseFromParent();
6222 IfBlock = NewIfBlock;
6227 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
6228 StoreInst *SI = dyn_cast<StoreInst>(Instr);
6229 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
6231 return scalarizeInstruction(Instr, IfPredicateStore);
6234 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
6238 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
6242 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
6244 // When unrolling and the VF is 1, we only need to add a simple scalar.
6245 Type *ITy = Val->getType();
6246 assert(!ITy->isVectorTy() && "Val must be a scalar");
6247 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
6248 return Builder.CreateAdd(Val, C, "induction");