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
1030 /// Return the loop metadata prefix.
1031 static StringRef Prefix() { return "llvm.loop."; }
1035 FK_Undefined = -1, ///< Not selected.
1036 FK_Disabled = 0, ///< Forcing disabled.
1037 FK_Enabled = 1, ///< Forcing enabled.
1040 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
1041 : Width("vectorize.width", VectorizationFactor, HK_WIDTH),
1042 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1043 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1045 // Populate values with existing loop metadata.
1046 getHintsFromMetadata();
1048 // force-vector-interleave overrides DisableInterleaving.
1049 if (VectorizationInterleave.getNumOccurrences() > 0)
1050 Interleave.Value = VectorizationInterleave;
1052 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
1053 << "LV: Interleaving disabled by the pass manager\n");
1056 /// Mark the loop L as already vectorized by setting the width to 1.
1057 void setAlreadyVectorized() {
1058 Width.Value = Interleave.Value = 1;
1059 Hint Hints[] = {Width, Interleave};
1060 writeHintsToMetadata(Hints);
1063 /// Dumps all the hint information.
1064 std::string emitRemark() const {
1066 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1067 R << "vectorization is explicitly disabled";
1069 R << "use -Rpass-analysis=loop-vectorize for more info";
1070 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1071 R << " (Force=true";
1072 if (Width.Value != 0)
1073 R << ", Vector Width=" << Width.Value;
1074 if (Interleave.Value != 0)
1075 R << ", Interleave Count=" << Interleave.Value;
1083 unsigned getWidth() const { return Width.Value; }
1084 unsigned getInterleave() const { return Interleave.Value; }
1085 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1088 /// Find hints specified in the loop metadata and update local values.
1089 void getHintsFromMetadata() {
1090 MDNode *LoopID = TheLoop->getLoopID();
1094 // First operand should refer to the loop id itself.
1095 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1096 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1098 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1099 const MDString *S = nullptr;
1100 SmallVector<Metadata *, 4> Args;
1102 // The expected hint is either a MDString or a MDNode with the first
1103 // operand a MDString.
1104 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1105 if (!MD || MD->getNumOperands() == 0)
1107 S = dyn_cast<MDString>(MD->getOperand(0));
1108 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1109 Args.push_back(MD->getOperand(i));
1111 S = dyn_cast<MDString>(LoopID->getOperand(i));
1112 assert(Args.size() == 0 && "too many arguments for MDString");
1118 // Check if the hint starts with the loop metadata prefix.
1119 StringRef Name = S->getString();
1120 if (Args.size() == 1)
1121 setHint(Name, Args[0]);
1125 /// Checks string hint with one operand and set value if valid.
1126 void setHint(StringRef Name, Metadata *Arg) {
1127 if (!Name.startswith(Prefix()))
1129 Name = Name.substr(Prefix().size(), StringRef::npos);
1131 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1133 unsigned Val = C->getZExtValue();
1135 Hint *Hints[] = {&Width, &Interleave, &Force};
1136 for (auto H : Hints) {
1137 if (Name == H->Name) {
1138 if (H->validate(Val))
1141 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1147 /// Create a new hint from name / value pair.
1148 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1149 LLVMContext &Context = TheLoop->getHeader()->getContext();
1150 Metadata *MDs[] = {MDString::get(Context, Name),
1151 ConstantAsMetadata::get(
1152 ConstantInt::get(Type::getInt32Ty(Context), V))};
1153 return MDNode::get(Context, MDs);
1156 /// Matches metadata with hint name.
1157 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1158 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1162 for (auto H : HintTypes)
1163 if (Name->getString().endswith(H.Name))
1168 /// Sets current hints into loop metadata, keeping other values intact.
1169 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1170 if (HintTypes.size() == 0)
1173 // Reserve the first element to LoopID (see below).
1174 SmallVector<Metadata *, 4> MDs(1);
1175 // If the loop already has metadata, then ignore the existing operands.
1176 MDNode *LoopID = TheLoop->getLoopID();
1178 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1179 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1180 // If node in update list, ignore old value.
1181 if (!matchesHintMetadataName(Node, HintTypes))
1182 MDs.push_back(Node);
1186 // Now, add the missing hints.
1187 for (auto H : HintTypes)
1188 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1190 // Replace current metadata node with new one.
1191 LLVMContext &Context = TheLoop->getHeader()->getContext();
1192 MDNode *NewLoopID = MDNode::get(Context, MDs);
1193 // Set operand 0 to refer to the loop id itself.
1194 NewLoopID->replaceOperandWith(0, NewLoopID);
1196 TheLoop->setLoopID(NewLoopID);
1200 /// The loop these hints belong to.
1201 const Loop *TheLoop;
1204 static void emitMissedWarning(Function *F, Loop *L,
1205 const LoopVectorizeHints &LH) {
1206 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1207 L->getStartLoc(), LH.emitRemark());
1209 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1210 if (LH.getWidth() != 1)
1211 emitLoopVectorizeWarning(
1212 F->getContext(), *F, L->getStartLoc(),
1213 "failed explicitly specified loop vectorization");
1214 else if (LH.getInterleave() != 1)
1215 emitLoopInterleaveWarning(
1216 F->getContext(), *F, L->getStartLoc(),
1217 "failed explicitly specified loop interleaving");
1221 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1223 return V.push_back(&L);
1225 for (Loop *InnerL : L)
1226 addInnerLoop(*InnerL, V);
1229 /// The LoopVectorize Pass.
1230 struct LoopVectorize : public FunctionPass {
1231 /// Pass identification, replacement for typeid
1234 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1236 DisableUnrolling(NoUnrolling),
1237 AlwaysVectorize(AlwaysVectorize) {
1238 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1241 ScalarEvolution *SE;
1242 const DataLayout *DL;
1244 TargetTransformInfo *TTI;
1246 BlockFrequencyInfo *BFI;
1247 TargetLibraryInfo *TLI;
1249 AssumptionTracker *AT;
1250 bool DisableUnrolling;
1251 bool AlwaysVectorize;
1253 BlockFrequency ColdEntryFreq;
1255 bool runOnFunction(Function &F) override {
1256 SE = &getAnalysis<ScalarEvolution>();
1257 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1258 DL = DLP ? &DLP->getDataLayout() : nullptr;
1259 LI = &getAnalysis<LoopInfo>();
1260 TTI = &getAnalysis<TargetTransformInfo>();
1261 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1262 BFI = &getAnalysis<BlockFrequencyInfo>();
1263 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
1264 AA = &getAnalysis<AliasAnalysis>();
1265 AT = &getAnalysis<AssumptionTracker>();
1267 // Compute some weights outside of the loop over the loops. Compute this
1268 // using a BranchProbability to re-use its scaling math.
1269 const BranchProbability ColdProb(1, 5); // 20%
1270 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1272 // If the target claims to have no vector registers don't attempt
1274 if (!TTI->getNumberOfRegisters(true))
1278 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1279 << ": Missing data layout\n");
1283 // Build up a worklist of inner-loops to vectorize. This is necessary as
1284 // the act of vectorizing or partially unrolling a loop creates new loops
1285 // and can invalidate iterators across the loops.
1286 SmallVector<Loop *, 8> Worklist;
1289 addInnerLoop(*L, Worklist);
1291 LoopsAnalyzed += Worklist.size();
1293 // Now walk the identified inner loops.
1294 bool Changed = false;
1295 while (!Worklist.empty())
1296 Changed |= processLoop(Worklist.pop_back_val());
1298 // Process each loop nest in the function.
1302 bool processLoop(Loop *L) {
1303 assert(L->empty() && "Only process inner loops.");
1306 const std::string DebugLocStr = getDebugLocString(L);
1309 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1310 << L->getHeader()->getParent()->getName() << "\" from "
1311 << DebugLocStr << "\n");
1313 LoopVectorizeHints Hints(L, DisableUnrolling);
1315 DEBUG(dbgs() << "LV: Loop hints:"
1317 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1319 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1321 : "?")) << " width=" << Hints.getWidth()
1322 << " unroll=" << Hints.getInterleave() << "\n");
1324 // Function containing loop
1325 Function *F = L->getHeader()->getParent();
1327 // Looking at the diagnostic output is the only way to determine if a loop
1328 // was vectorized (other than looking at the IR or machine code), so it
1329 // is important to generate an optimization remark for each loop. Most of
1330 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1331 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1332 // less verbose reporting vectorized loops and unvectorized loops that may
1333 // benefit from vectorization, respectively.
1335 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1336 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1337 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1338 L->getStartLoc(), Hints.emitRemark());
1342 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1343 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1344 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1345 L->getStartLoc(), Hints.emitRemark());
1349 if (Hints.getWidth() == 1 && Hints.getInterleave() == 1) {
1350 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1351 emitOptimizationRemarkAnalysis(
1352 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1353 "loop not vectorized: vector width and interleave count are "
1354 "explicitly set to 1");
1358 // Check the loop for a trip count threshold:
1359 // do not vectorize loops with a tiny trip count.
1360 const unsigned TC = SE->getSmallConstantTripCount(L);
1361 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1362 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1363 << "This loop is not worth vectorizing.");
1364 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1365 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1367 DEBUG(dbgs() << "\n");
1368 emitOptimizationRemarkAnalysis(
1369 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1370 "vectorization is not beneficial and is not explicitly forced");
1375 // Check if it is legal to vectorize the loop.
1376 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, AA, F);
1377 if (!LVL.canVectorize()) {
1378 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1379 emitMissedWarning(F, L, Hints);
1383 // Use the cost model.
1384 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI, AT, F,
1387 // Check the function attributes to find out if this function should be
1388 // optimized for size.
1389 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1390 F->hasFnAttribute(Attribute::OptimizeForSize);
1392 // Compute the weighted frequency of this loop being executed and see if it
1393 // is less than 20% of the function entry baseline frequency. Note that we
1394 // always have a canonical loop here because we think we *can* vectoriez.
1395 // FIXME: This is hidden behind a flag due to pervasive problems with
1396 // exactly what block frequency models.
1397 if (LoopVectorizeWithBlockFrequency) {
1398 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1399 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1400 LoopEntryFreq < ColdEntryFreq)
1404 // Check the function attributes to see if implicit floats are allowed.a
1405 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1406 // an integer loop and the vector instructions selected are purely integer
1407 // vector instructions?
1408 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1409 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1410 "attribute is used.\n");
1411 emitOptimizationRemarkAnalysis(
1412 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1413 "loop not vectorized due to NoImplicitFloat attribute");
1414 emitMissedWarning(F, L, Hints);
1418 // Select the optimal vectorization factor.
1419 const LoopVectorizationCostModel::VectorizationFactor VF =
1420 CM.selectVectorizationFactor(OptForSize);
1422 // Select the unroll factor.
1424 CM.selectUnrollFactor(OptForSize, VF.Width, VF.Cost);
1426 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1427 << DebugLocStr << '\n');
1428 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1430 if (VF.Width == 1) {
1431 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
1434 emitOptimizationRemarkAnalysis(
1435 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1436 "not beneficial to vectorize and user disabled interleaving");
1439 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1441 // Report the unrolling decision.
1442 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1443 Twine("unrolled with interleaving factor " +
1445 " (vectorization not beneficial)"));
1447 // We decided not to vectorize, but we may want to unroll.
1449 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1450 Unroller.vectorize(&LVL);
1452 // If we decided that it is *legal* to vectorize the loop then do it.
1453 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1457 // Report the vectorization decision.
1458 emitOptimizationRemark(
1459 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1460 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1461 ", unrolling interleave factor: " + Twine(UF) + ")");
1464 // Mark the loop as already vectorized to avoid vectorizing again.
1465 Hints.setAlreadyVectorized();
1467 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1471 void getAnalysisUsage(AnalysisUsage &AU) const override {
1472 AU.addRequired<AssumptionTracker>();
1473 AU.addRequiredID(LoopSimplifyID);
1474 AU.addRequiredID(LCSSAID);
1475 AU.addRequired<BlockFrequencyInfo>();
1476 AU.addRequired<DominatorTreeWrapperPass>();
1477 AU.addRequired<LoopInfo>();
1478 AU.addRequired<ScalarEvolution>();
1479 AU.addRequired<TargetTransformInfo>();
1480 AU.addRequired<AliasAnalysis>();
1481 AU.addPreserved<LoopInfo>();
1482 AU.addPreserved<DominatorTreeWrapperPass>();
1483 AU.addPreserved<AliasAnalysis>();
1488 } // end anonymous namespace
1490 //===----------------------------------------------------------------------===//
1491 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1492 // LoopVectorizationCostModel.
1493 //===----------------------------------------------------------------------===//
1495 static Value *stripIntegerCast(Value *V) {
1496 if (CastInst *CI = dyn_cast<CastInst>(V))
1497 if (CI->getOperand(0)->getType()->isIntegerTy())
1498 return CI->getOperand(0);
1502 ///\brief Replaces the symbolic stride in a pointer SCEV expression by one.
1504 /// If \p OrigPtr is not null, use it to look up the stride value instead of
1506 static const SCEV *replaceSymbolicStrideSCEV(ScalarEvolution *SE,
1507 ValueToValueMap &PtrToStride,
1508 Value *Ptr, Value *OrigPtr = nullptr) {
1510 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
1512 // If there is an entry in the map return the SCEV of the pointer with the
1513 // symbolic stride replaced by one.
1514 ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
1515 if (SI != PtrToStride.end()) {
1516 Value *StrideVal = SI->second;
1519 StrideVal = stripIntegerCast(StrideVal);
1521 // Replace symbolic stride by one.
1522 Value *One = ConstantInt::get(StrideVal->getType(), 1);
1523 ValueToValueMap RewriteMap;
1524 RewriteMap[StrideVal] = One;
1527 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
1528 DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
1533 // Otherwise, just return the SCEV of the original pointer.
1534 return SE->getSCEV(Ptr);
1537 void LoopVectorizationLegality::RuntimePointerCheck::insert(
1538 ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
1539 unsigned ASId, ValueToValueMap &Strides) {
1540 // Get the stride replaced scev.
1541 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
1542 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1543 assert(AR && "Invalid addrec expression");
1544 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1545 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1546 Pointers.push_back(Ptr);
1547 Starts.push_back(AR->getStart());
1548 Ends.push_back(ScEnd);
1549 IsWritePtr.push_back(WritePtr);
1550 DependencySetId.push_back(DepSetId);
1551 AliasSetId.push_back(ASId);
1554 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1555 // We need to place the broadcast of invariant variables outside the loop.
1556 Instruction *Instr = dyn_cast<Instruction>(V);
1558 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1559 Instr->getParent()) != LoopVectorBody.end());
1560 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1562 // Place the code for broadcasting invariant variables in the new preheader.
1563 IRBuilder<>::InsertPointGuard Guard(Builder);
1565 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1567 // Broadcast the scalar into all locations in the vector.
1568 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1573 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1575 assert(Val->getType()->isVectorTy() && "Must be a vector");
1576 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1577 "Elem must be an integer");
1578 // Create the types.
1579 Type *ITy = Val->getType()->getScalarType();
1580 VectorType *Ty = cast<VectorType>(Val->getType());
1581 int VLen = Ty->getNumElements();
1582 SmallVector<Constant*, 8> Indices;
1584 // Create a vector of consecutive numbers from zero to VF.
1585 for (int i = 0; i < VLen; ++i) {
1586 int64_t Idx = Negate ? (-i) : i;
1587 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1590 // Add the consecutive indices to the vector value.
1591 Constant *Cv = ConstantVector::get(Indices);
1592 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1593 return Builder.CreateAdd(Val, Cv, "induction");
1596 /// \brief Find the operand of the GEP that should be checked for consecutive
1597 /// stores. This ignores trailing indices that have no effect on the final
1599 static unsigned getGEPInductionOperand(const DataLayout *DL,
1600 const GetElementPtrInst *Gep) {
1601 unsigned LastOperand = Gep->getNumOperands() - 1;
1602 unsigned GEPAllocSize = DL->getTypeAllocSize(
1603 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1605 // Walk backwards and try to peel off zeros.
1606 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1607 // Find the type we're currently indexing into.
1608 gep_type_iterator GEPTI = gep_type_begin(Gep);
1609 std::advance(GEPTI, LastOperand - 1);
1611 // If it's a type with the same allocation size as the result of the GEP we
1612 // can peel off the zero index.
1613 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1621 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1622 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1623 // Make sure that the pointer does not point to structs.
1624 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1627 // If this value is a pointer induction variable we know it is consecutive.
1628 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1629 if (Phi && Inductions.count(Phi)) {
1630 InductionInfo II = Inductions[Phi];
1631 if (IK_PtrInduction == II.IK)
1633 else if (IK_ReversePtrInduction == II.IK)
1637 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1641 unsigned NumOperands = Gep->getNumOperands();
1642 Value *GpPtr = Gep->getPointerOperand();
1643 // If this GEP value is a consecutive pointer induction variable and all of
1644 // the indices are constant then we know it is consecutive. We can
1645 Phi = dyn_cast<PHINode>(GpPtr);
1646 if (Phi && Inductions.count(Phi)) {
1648 // Make sure that the pointer does not point to structs.
1649 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1650 if (GepPtrType->getElementType()->isAggregateType())
1653 // Make sure that all of the index operands are loop invariant.
1654 for (unsigned i = 1; i < NumOperands; ++i)
1655 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1658 InductionInfo II = Inductions[Phi];
1659 if (IK_PtrInduction == II.IK)
1661 else if (IK_ReversePtrInduction == II.IK)
1665 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1667 // Check that all of the gep indices are uniform except for our induction
1669 for (unsigned i = 0; i != NumOperands; ++i)
1670 if (i != InductionOperand &&
1671 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1674 // We can emit wide load/stores only if the last non-zero index is the
1675 // induction variable.
1676 const SCEV *Last = nullptr;
1677 if (!Strides.count(Gep))
1678 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1680 // Because of the multiplication by a stride we can have a s/zext cast.
1681 // We are going to replace this stride by 1 so the cast is safe to ignore.
1683 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1684 // %0 = trunc i64 %indvars.iv to i32
1685 // %mul = mul i32 %0, %Stride1
1686 // %idxprom = zext i32 %mul to i64 << Safe cast.
1687 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1689 Last = replaceSymbolicStrideSCEV(SE, Strides,
1690 Gep->getOperand(InductionOperand), Gep);
1691 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1693 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1697 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1698 const SCEV *Step = AR->getStepRecurrence(*SE);
1700 // The memory is consecutive because the last index is consecutive
1701 // and all other indices are loop invariant.
1704 if (Step->isAllOnesValue())
1711 bool LoopVectorizationLegality::isUniform(Value *V) {
1712 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1715 InnerLoopVectorizer::VectorParts&
1716 InnerLoopVectorizer::getVectorValue(Value *V) {
1717 assert(V != Induction && "The new induction variable should not be used.");
1718 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1720 // If we have a stride that is replaced by one, do it here.
1721 if (Legal->hasStride(V))
1722 V = ConstantInt::get(V->getType(), 1);
1724 // If we have this scalar in the map, return it.
1725 if (WidenMap.has(V))
1726 return WidenMap.get(V);
1728 // If this scalar is unknown, assume that it is a constant or that it is
1729 // loop invariant. Broadcast V and save the value for future uses.
1730 Value *B = getBroadcastInstrs(V);
1731 return WidenMap.splat(V, B);
1734 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1735 assert(Vec->getType()->isVectorTy() && "Invalid type");
1736 SmallVector<Constant*, 8> ShuffleMask;
1737 for (unsigned i = 0; i < VF; ++i)
1738 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1740 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1741 ConstantVector::get(ShuffleMask),
1745 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1746 // Attempt to issue a wide load.
1747 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1748 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1750 assert((LI || SI) && "Invalid Load/Store instruction");
1752 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1753 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1754 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1755 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1756 // An alignment of 0 means target abi alignment. We need to use the scalar's
1757 // target abi alignment in such a case.
1759 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1760 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1761 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1762 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1764 if (SI && Legal->blockNeedsPredication(SI->getParent()))
1765 return scalarizeInstruction(Instr, true);
1767 if (ScalarAllocatedSize != VectorElementSize)
1768 return scalarizeInstruction(Instr);
1770 // If the pointer is loop invariant or if it is non-consecutive,
1771 // scalarize the load.
1772 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1773 bool Reverse = ConsecutiveStride < 0;
1774 bool UniformLoad = LI && Legal->isUniform(Ptr);
1775 if (!ConsecutiveStride || UniformLoad)
1776 return scalarizeInstruction(Instr);
1778 Constant *Zero = Builder.getInt32(0);
1779 VectorParts &Entry = WidenMap.get(Instr);
1781 // Handle consecutive loads/stores.
1782 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1783 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1784 setDebugLocFromInst(Builder, Gep);
1785 Value *PtrOperand = Gep->getPointerOperand();
1786 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1787 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1789 // Create the new GEP with the new induction variable.
1790 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1791 Gep2->setOperand(0, FirstBasePtr);
1792 Gep2->setName("gep.indvar.base");
1793 Ptr = Builder.Insert(Gep2);
1795 setDebugLocFromInst(Builder, Gep);
1796 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1797 OrigLoop) && "Base ptr must be invariant");
1799 // The last index does not have to be the induction. It can be
1800 // consecutive and be a function of the index. For example A[I+1];
1801 unsigned NumOperands = Gep->getNumOperands();
1802 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1803 // Create the new GEP with the new induction variable.
1804 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1806 for (unsigned i = 0; i < NumOperands; ++i) {
1807 Value *GepOperand = Gep->getOperand(i);
1808 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1810 // Update last index or loop invariant instruction anchored in loop.
1811 if (i == InductionOperand ||
1812 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1813 assert((i == InductionOperand ||
1814 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1815 "Must be last index or loop invariant");
1817 VectorParts &GEPParts = getVectorValue(GepOperand);
1818 Value *Index = GEPParts[0];
1819 Index = Builder.CreateExtractElement(Index, Zero);
1820 Gep2->setOperand(i, Index);
1821 Gep2->setName("gep.indvar.idx");
1824 Ptr = Builder.Insert(Gep2);
1826 // Use the induction element ptr.
1827 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1828 setDebugLocFromInst(Builder, Ptr);
1829 VectorParts &PtrVal = getVectorValue(Ptr);
1830 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1835 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1836 "We do not allow storing to uniform addresses");
1837 setDebugLocFromInst(Builder, SI);
1838 // We don't want to update the value in the map as it might be used in
1839 // another expression. So don't use a reference type for "StoredVal".
1840 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1842 for (unsigned Part = 0; Part < UF; ++Part) {
1843 // Calculate the pointer for the specific unroll-part.
1844 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1847 // If we store to reverse consecutive memory locations then we need
1848 // to reverse the order of elements in the stored value.
1849 StoredVal[Part] = reverseVector(StoredVal[Part]);
1850 // If the address is consecutive but reversed, then the
1851 // wide store needs to start at the last vector element.
1852 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1853 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1856 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1857 DataTy->getPointerTo(AddressSpace));
1859 Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
1860 propagateMetadata(NewSI, SI);
1866 assert(LI && "Must have a load instruction");
1867 setDebugLocFromInst(Builder, LI);
1868 for (unsigned Part = 0; Part < UF; ++Part) {
1869 // Calculate the pointer for the specific unroll-part.
1870 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1873 // If the address is consecutive but reversed, then the
1874 // wide load needs to start at the last vector element.
1875 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1876 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1879 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1880 DataTy->getPointerTo(AddressSpace));
1881 LoadInst *NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
1882 propagateMetadata(NewLI, LI);
1883 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
1887 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1888 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1889 // Holds vector parameters or scalars, in case of uniform vals.
1890 SmallVector<VectorParts, 4> Params;
1892 setDebugLocFromInst(Builder, Instr);
1894 // Find all of the vectorized parameters.
1895 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1896 Value *SrcOp = Instr->getOperand(op);
1898 // If we are accessing the old induction variable, use the new one.
1899 if (SrcOp == OldInduction) {
1900 Params.push_back(getVectorValue(SrcOp));
1904 // Try using previously calculated values.
1905 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1907 // If the src is an instruction that appeared earlier in the basic block
1908 // then it should already be vectorized.
1909 if (SrcInst && OrigLoop->contains(SrcInst)) {
1910 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1911 // The parameter is a vector value from earlier.
1912 Params.push_back(WidenMap.get(SrcInst));
1914 // The parameter is a scalar from outside the loop. Maybe even a constant.
1915 VectorParts Scalars;
1916 Scalars.append(UF, SrcOp);
1917 Params.push_back(Scalars);
1921 assert(Params.size() == Instr->getNumOperands() &&
1922 "Invalid number of operands");
1924 // Does this instruction return a value ?
1925 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1927 Value *UndefVec = IsVoidRetTy ? nullptr :
1928 UndefValue::get(VectorType::get(Instr->getType(), VF));
1929 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1930 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1932 Instruction *InsertPt = Builder.GetInsertPoint();
1933 BasicBlock *IfBlock = Builder.GetInsertBlock();
1934 BasicBlock *CondBlock = nullptr;
1937 Loop *VectorLp = nullptr;
1938 if (IfPredicateStore) {
1939 assert(Instr->getParent()->getSinglePredecessor() &&
1940 "Only support single predecessor blocks");
1941 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1942 Instr->getParent());
1943 VectorLp = LI->getLoopFor(IfBlock);
1944 assert(VectorLp && "Must have a loop for this block");
1947 // For each vector unroll 'part':
1948 for (unsigned Part = 0; Part < UF; ++Part) {
1949 // For each scalar that we create:
1950 for (unsigned Width = 0; Width < VF; ++Width) {
1953 Value *Cmp = nullptr;
1954 if (IfPredicateStore) {
1955 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1956 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1957 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1958 LoopVectorBody.push_back(CondBlock);
1959 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
1960 // Update Builder with newly created basic block.
1961 Builder.SetInsertPoint(InsertPt);
1964 Instruction *Cloned = Instr->clone();
1966 Cloned->setName(Instr->getName() + ".cloned");
1967 // Replace the operands of the cloned instructions with extracted scalars.
1968 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1969 Value *Op = Params[op][Part];
1970 // Param is a vector. Need to extract the right lane.
1971 if (Op->getType()->isVectorTy())
1972 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1973 Cloned->setOperand(op, Op);
1976 // Place the cloned scalar in the new loop.
1977 Builder.Insert(Cloned);
1979 // If the original scalar returns a value we need to place it in a vector
1980 // so that future users will be able to use it.
1982 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1983 Builder.getInt32(Width));
1985 if (IfPredicateStore) {
1986 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1987 LoopVectorBody.push_back(NewIfBlock);
1988 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
1989 Builder.SetInsertPoint(InsertPt);
1990 Instruction *OldBr = IfBlock->getTerminator();
1991 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1992 OldBr->eraseFromParent();
1993 IfBlock = NewIfBlock;
1999 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
2003 if (Instruction *I = dyn_cast<Instruction>(V))
2004 return I->getParent() == Loc->getParent() ? I : nullptr;
2008 std::pair<Instruction *, Instruction *>
2009 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
2010 Instruction *tnullptr = nullptr;
2011 if (!Legal->mustCheckStrides())
2012 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2014 IRBuilder<> ChkBuilder(Loc);
2017 Value *Check = nullptr;
2018 Instruction *FirstInst = nullptr;
2019 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
2020 SE = Legal->strides_end();
2022 Value *Ptr = stripIntegerCast(*SI);
2023 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
2025 // Store the first instruction we create.
2026 FirstInst = getFirstInst(FirstInst, C, Loc);
2028 Check = ChkBuilder.CreateOr(Check, C);
2033 // We have to do this trickery because the IRBuilder might fold the check to a
2034 // constant expression in which case there is no Instruction anchored in a
2036 LLVMContext &Ctx = Loc->getContext();
2037 Instruction *TheCheck =
2038 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2039 ChkBuilder.Insert(TheCheck, "stride.not.one");
2040 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2042 return std::make_pair(FirstInst, TheCheck);
2045 std::pair<Instruction *, Instruction *>
2046 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
2047 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
2048 Legal->getRuntimePointerCheck();
2050 Instruction *tnullptr = nullptr;
2051 if (!PtrRtCheck->Need)
2052 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2054 unsigned NumPointers = PtrRtCheck->Pointers.size();
2055 SmallVector<TrackingVH<Value> , 2> Starts;
2056 SmallVector<TrackingVH<Value> , 2> Ends;
2058 LLVMContext &Ctx = Loc->getContext();
2059 SCEVExpander Exp(*SE, "induction");
2060 Instruction *FirstInst = nullptr;
2062 for (unsigned i = 0; i < NumPointers; ++i) {
2063 Value *Ptr = PtrRtCheck->Pointers[i];
2064 const SCEV *Sc = SE->getSCEV(Ptr);
2066 if (SE->isLoopInvariant(Sc, OrigLoop)) {
2067 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
2069 Starts.push_back(Ptr);
2070 Ends.push_back(Ptr);
2072 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
2073 unsigned AS = Ptr->getType()->getPointerAddressSpace();
2075 // Use this type for pointer arithmetic.
2076 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
2078 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
2079 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
2080 Starts.push_back(Start);
2081 Ends.push_back(End);
2085 IRBuilder<> ChkBuilder(Loc);
2086 // Our instructions might fold to a constant.
2087 Value *MemoryRuntimeCheck = nullptr;
2088 for (unsigned i = 0; i < NumPointers; ++i) {
2089 for (unsigned j = i+1; j < NumPointers; ++j) {
2090 // No need to check if two readonly pointers intersect.
2091 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
2094 // Only need to check pointers between two different dependency sets.
2095 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
2097 // Only need to check pointers in the same alias set.
2098 if (PtrRtCheck->AliasSetId[i] != PtrRtCheck->AliasSetId[j])
2101 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
2102 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
2104 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
2105 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
2106 "Trying to bounds check pointers with different address spaces");
2108 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
2109 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
2111 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
2112 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
2113 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
2114 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
2116 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
2117 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
2118 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
2119 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
2120 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
2121 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2122 if (MemoryRuntimeCheck) {
2123 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
2125 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2127 MemoryRuntimeCheck = IsConflict;
2131 // We have to do this trickery because the IRBuilder might fold the check to a
2132 // constant expression in which case there is no Instruction anchored in a
2134 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
2135 ConstantInt::getTrue(Ctx));
2136 ChkBuilder.Insert(Check, "memcheck.conflict");
2137 FirstInst = getFirstInst(FirstInst, Check, Loc);
2138 return std::make_pair(FirstInst, Check);
2141 void InnerLoopVectorizer::createEmptyLoop() {
2143 In this function we generate a new loop. The new loop will contain
2144 the vectorized instructions while the old loop will continue to run the
2147 [ ] <-- Back-edge taken count overflow check.
2150 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2153 || [ ] <-- vector pre header.
2157 || [ ]_| <-- vector loop.
2160 | >[ ] <--- middle-block.
2163 -|- >[ ] <--- new preheader.
2167 | [ ]_| <-- old scalar loop to handle remainder.
2170 >[ ] <-- exit block.
2174 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2175 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2176 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2177 assert(BypassBlock && "Invalid loop structure");
2178 assert(ExitBlock && "Must have an exit block");
2180 // Some loops have a single integer induction variable, while other loops
2181 // don't. One example is c++ iterators that often have multiple pointer
2182 // induction variables. In the code below we also support a case where we
2183 // don't have a single induction variable.
2184 OldInduction = Legal->getInduction();
2185 Type *IdxTy = Legal->getWidestInductionType();
2187 // Find the loop boundaries.
2188 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2189 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2191 // The exit count might have the type of i64 while the phi is i32. This can
2192 // happen if we have an induction variable that is sign extended before the
2193 // compare. The only way that we get a backedge taken count is that the
2194 // induction variable was signed and as such will not overflow. In such a case
2195 // truncation is legal.
2196 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2197 IdxTy->getPrimitiveSizeInBits())
2198 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2200 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2201 // Get the total trip count from the count by adding 1.
2202 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2203 SE->getConstant(BackedgeTakeCount->getType(), 1));
2205 // Expand the trip count and place the new instructions in the preheader.
2206 // Notice that the pre-header does not change, only the loop body.
2207 SCEVExpander Exp(*SE, "induction");
2209 // We need to test whether the backedge-taken count is uint##_max. Adding one
2210 // to it will cause overflow and an incorrect loop trip count in the vector
2211 // body. In case of overflow we want to directly jump to the scalar remainder
2213 Value *BackedgeCount =
2214 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2215 BypassBlock->getTerminator());
2216 if (BackedgeCount->getType()->isPointerTy())
2217 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2218 "backedge.ptrcnt.to.int",
2219 BypassBlock->getTerminator());
2220 Instruction *CheckBCOverflow =
2221 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2222 Constant::getAllOnesValue(BackedgeCount->getType()),
2223 "backedge.overflow", BypassBlock->getTerminator());
2225 // The loop index does not have to start at Zero. Find the original start
2226 // value from the induction PHI node. If we don't have an induction variable
2227 // then we know that it starts at zero.
2228 Builder.SetInsertPoint(BypassBlock->getTerminator());
2229 Value *StartIdx = ExtendedIdx = OldInduction ?
2230 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2232 ConstantInt::get(IdxTy, 0);
2234 // We need an instruction to anchor the overflow check on. StartIdx needs to
2235 // be defined before the overflow check branch. Because the scalar preheader
2236 // is going to merge the start index and so the overflow branch block needs to
2237 // contain a definition of the start index.
2238 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2239 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2240 BypassBlock->getTerminator());
2242 // Count holds the overall loop count (N).
2243 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2244 BypassBlock->getTerminator());
2246 LoopBypassBlocks.push_back(BypassBlock);
2248 // Split the single block loop into the two loop structure described above.
2249 BasicBlock *VectorPH =
2250 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2251 BasicBlock *VecBody =
2252 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2253 BasicBlock *MiddleBlock =
2254 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2255 BasicBlock *ScalarPH =
2256 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2258 // Create and register the new vector loop.
2259 Loop* Lp = new Loop();
2260 Loop *ParentLoop = OrigLoop->getParentLoop();
2262 // Insert the new loop into the loop nest and register the new basic blocks
2263 // before calling any utilities such as SCEV that require valid LoopInfo.
2265 ParentLoop->addChildLoop(Lp);
2266 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
2267 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
2268 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
2270 LI->addTopLevelLoop(Lp);
2272 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
2274 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2276 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2278 // Generate the induction variable.
2279 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2280 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2281 // The loop step is equal to the vectorization factor (num of SIMD elements)
2282 // times the unroll factor (num of SIMD instructions).
2283 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2285 // This is the IR builder that we use to add all of the logic for bypassing
2286 // the new vector loop.
2287 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2288 setDebugLocFromInst(BypassBuilder,
2289 getDebugLocFromInstOrOperands(OldInduction));
2291 // We may need to extend the index in case there is a type mismatch.
2292 // We know that the count starts at zero and does not overflow.
2293 if (Count->getType() != IdxTy) {
2294 // The exit count can be of pointer type. Convert it to the correct
2296 if (ExitCount->getType()->isPointerTy())
2297 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2299 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2302 // Add the start index to the loop count to get the new end index.
2303 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2305 // Now we need to generate the expression for N - (N % VF), which is
2306 // the part that the vectorized body will execute.
2307 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2308 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2309 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2310 "end.idx.rnd.down");
2312 // Now, compare the new count to zero. If it is zero skip the vector loop and
2313 // jump to the scalar loop.
2315 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2317 BasicBlock *LastBypassBlock = BypassBlock;
2319 // Generate code to check that the loops trip count that we computed by adding
2320 // one to the backedge-taken count will not overflow.
2322 auto PastOverflowCheck =
2323 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2324 BasicBlock *CheckBlock =
2325 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2327 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2328 LoopBypassBlocks.push_back(CheckBlock);
2329 Instruction *OldTerm = LastBypassBlock->getTerminator();
2330 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2331 OldTerm->eraseFromParent();
2332 LastBypassBlock = CheckBlock;
2335 // Generate the code to check that the strides we assumed to be one are really
2336 // one. We want the new basic block to start at the first instruction in a
2337 // sequence of instructions that form a check.
2338 Instruction *StrideCheck;
2339 Instruction *FirstCheckInst;
2340 std::tie(FirstCheckInst, StrideCheck) =
2341 addStrideCheck(LastBypassBlock->getTerminator());
2343 // Create a new block containing the stride check.
2344 BasicBlock *CheckBlock =
2345 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2347 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2348 LoopBypassBlocks.push_back(CheckBlock);
2350 // Replace the branch into the memory check block with a conditional branch
2351 // for the "few elements case".
2352 Instruction *OldTerm = LastBypassBlock->getTerminator();
2353 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2354 OldTerm->eraseFromParent();
2357 LastBypassBlock = CheckBlock;
2360 // Generate the code that checks in runtime if arrays overlap. We put the
2361 // checks into a separate block to make the more common case of few elements
2363 Instruction *MemRuntimeCheck;
2364 std::tie(FirstCheckInst, MemRuntimeCheck) =
2365 addRuntimeCheck(LastBypassBlock->getTerminator());
2366 if (MemRuntimeCheck) {
2367 // Create a new block containing the memory check.
2368 BasicBlock *CheckBlock =
2369 LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
2371 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2372 LoopBypassBlocks.push_back(CheckBlock);
2374 // Replace the branch into the memory check block with a conditional branch
2375 // for the "few elements case".
2376 Instruction *OldTerm = LastBypassBlock->getTerminator();
2377 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2378 OldTerm->eraseFromParent();
2380 Cmp = MemRuntimeCheck;
2381 LastBypassBlock = CheckBlock;
2384 LastBypassBlock->getTerminator()->eraseFromParent();
2385 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2388 // We are going to resume the execution of the scalar loop.
2389 // Go over all of the induction variables that we found and fix the
2390 // PHIs that are left in the scalar version of the loop.
2391 // The starting values of PHI nodes depend on the counter of the last
2392 // iteration in the vectorized loop.
2393 // If we come from a bypass edge then we need to start from the original
2396 // This variable saves the new starting index for the scalar loop.
2397 PHINode *ResumeIndex = nullptr;
2398 LoopVectorizationLegality::InductionList::iterator I, E;
2399 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2400 // Set builder to point to last bypass block.
2401 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2402 for (I = List->begin(), E = List->end(); I != E; ++I) {
2403 PHINode *OrigPhi = I->first;
2404 LoopVectorizationLegality::InductionInfo II = I->second;
2406 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2407 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2408 MiddleBlock->getTerminator());
2409 // We might have extended the type of the induction variable but we need a
2410 // truncated version for the scalar loop.
2411 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2412 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2413 MiddleBlock->getTerminator()) : nullptr;
2415 // Create phi nodes to merge from the backedge-taken check block.
2416 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2417 ScalarPH->getTerminator());
2418 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2420 PHINode *BCTruncResumeVal = nullptr;
2421 if (OrigPhi == OldInduction) {
2423 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2424 ScalarPH->getTerminator());
2425 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2428 Value *EndValue = nullptr;
2430 case LoopVectorizationLegality::IK_NoInduction:
2431 llvm_unreachable("Unknown induction");
2432 case LoopVectorizationLegality::IK_IntInduction: {
2433 // Handle the integer induction counter.
2434 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2436 // We have the canonical induction variable.
2437 if (OrigPhi == OldInduction) {
2438 // Create a truncated version of the resume value for the scalar loop,
2439 // we might have promoted the type to a larger width.
2441 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2442 // The new PHI merges the original incoming value, in case of a bypass,
2443 // or the value at the end of the vectorized loop.
2444 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2445 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2446 TruncResumeVal->addIncoming(EndValue, VecBody);
2448 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2450 // We know what the end value is.
2451 EndValue = IdxEndRoundDown;
2452 // We also know which PHI node holds it.
2453 ResumeIndex = ResumeVal;
2457 // Not the canonical induction variable - add the vector loop count to the
2459 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2460 II.StartValue->getType(),
2462 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
2465 case LoopVectorizationLegality::IK_ReverseIntInduction: {
2466 // Convert the CountRoundDown variable to the PHI size.
2467 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2468 II.StartValue->getType(),
2470 // Handle reverse integer induction counter.
2471 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
2474 case LoopVectorizationLegality::IK_PtrInduction: {
2475 // For pointer induction variables, calculate the offset using
2477 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
2481 case LoopVectorizationLegality::IK_ReversePtrInduction: {
2482 // The value at the end of the loop for the reverse pointer is calculated
2483 // by creating a GEP with a negative index starting from the start value.
2484 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
2485 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
2487 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
2493 // The new PHI merges the original incoming value, in case of a bypass,
2494 // or the value at the end of the vectorized loop.
2495 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2496 if (OrigPhi == OldInduction)
2497 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2499 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2501 ResumeVal->addIncoming(EndValue, VecBody);
2503 // Fix the scalar body counter (PHI node).
2504 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2506 // The old induction's phi node in the scalar body needs the truncated
2508 if (OrigPhi == OldInduction) {
2509 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2510 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2512 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2513 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2517 // If we are generating a new induction variable then we also need to
2518 // generate the code that calculates the exit value. This value is not
2519 // simply the end of the counter because we may skip the vectorized body
2520 // in case of a runtime check.
2522 assert(!ResumeIndex && "Unexpected resume value found");
2523 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2524 MiddleBlock->getTerminator());
2525 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2526 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2527 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2530 // Make sure that we found the index where scalar loop needs to continue.
2531 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2532 "Invalid resume Index");
2534 // Add a check in the middle block to see if we have completed
2535 // all of the iterations in the first vector loop.
2536 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2537 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2538 ResumeIndex, "cmp.n",
2539 MiddleBlock->getTerminator());
2541 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2542 // Remove the old terminator.
2543 MiddleBlock->getTerminator()->eraseFromParent();
2545 // Create i+1 and fill the PHINode.
2546 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2547 Induction->addIncoming(StartIdx, VectorPH);
2548 Induction->addIncoming(NextIdx, VecBody);
2549 // Create the compare.
2550 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2551 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2553 // Now we have two terminators. Remove the old one from the block.
2554 VecBody->getTerminator()->eraseFromParent();
2556 // Get ready to start creating new instructions into the vectorized body.
2557 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2560 LoopVectorPreHeader = VectorPH;
2561 LoopScalarPreHeader = ScalarPH;
2562 LoopMiddleBlock = MiddleBlock;
2563 LoopExitBlock = ExitBlock;
2564 LoopVectorBody.push_back(VecBody);
2565 LoopScalarBody = OldBasicBlock;
2567 LoopVectorizeHints Hints(Lp, true);
2568 Hints.setAlreadyVectorized();
2571 /// This function returns the identity element (or neutral element) for
2572 /// the operation K.
2574 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2579 // Adding, Xoring, Oring zero to a number does not change it.
2580 return ConstantInt::get(Tp, 0);
2581 case RK_IntegerMult:
2582 // Multiplying a number by 1 does not change it.
2583 return ConstantInt::get(Tp, 1);
2585 // AND-ing a number with an all-1 value does not change it.
2586 return ConstantInt::get(Tp, -1, true);
2588 // Multiplying a number by 1 does not change it.
2589 return ConstantFP::get(Tp, 1.0L);
2591 // Adding zero to a number does not change it.
2592 return ConstantFP::get(Tp, 0.0L);
2594 llvm_unreachable("Unknown reduction kind");
2598 /// This function translates the reduction kind to an LLVM binary operator.
2600 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2602 case LoopVectorizationLegality::RK_IntegerAdd:
2603 return Instruction::Add;
2604 case LoopVectorizationLegality::RK_IntegerMult:
2605 return Instruction::Mul;
2606 case LoopVectorizationLegality::RK_IntegerOr:
2607 return Instruction::Or;
2608 case LoopVectorizationLegality::RK_IntegerAnd:
2609 return Instruction::And;
2610 case LoopVectorizationLegality::RK_IntegerXor:
2611 return Instruction::Xor;
2612 case LoopVectorizationLegality::RK_FloatMult:
2613 return Instruction::FMul;
2614 case LoopVectorizationLegality::RK_FloatAdd:
2615 return Instruction::FAdd;
2616 case LoopVectorizationLegality::RK_IntegerMinMax:
2617 return Instruction::ICmp;
2618 case LoopVectorizationLegality::RK_FloatMinMax:
2619 return Instruction::FCmp;
2621 llvm_unreachable("Unknown reduction operation");
2625 Value *createMinMaxOp(IRBuilder<> &Builder,
2626 LoopVectorizationLegality::MinMaxReductionKind RK,
2629 CmpInst::Predicate P = CmpInst::ICMP_NE;
2632 llvm_unreachable("Unknown min/max reduction kind");
2633 case LoopVectorizationLegality::MRK_UIntMin:
2634 P = CmpInst::ICMP_ULT;
2636 case LoopVectorizationLegality::MRK_UIntMax:
2637 P = CmpInst::ICMP_UGT;
2639 case LoopVectorizationLegality::MRK_SIntMin:
2640 P = CmpInst::ICMP_SLT;
2642 case LoopVectorizationLegality::MRK_SIntMax:
2643 P = CmpInst::ICMP_SGT;
2645 case LoopVectorizationLegality::MRK_FloatMin:
2646 P = CmpInst::FCMP_OLT;
2648 case LoopVectorizationLegality::MRK_FloatMax:
2649 P = CmpInst::FCMP_OGT;
2654 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2655 RK == LoopVectorizationLegality::MRK_FloatMax)
2656 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2658 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2660 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2665 struct CSEDenseMapInfo {
2666 static bool canHandle(Instruction *I) {
2667 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2668 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2670 static inline Instruction *getEmptyKey() {
2671 return DenseMapInfo<Instruction *>::getEmptyKey();
2673 static inline Instruction *getTombstoneKey() {
2674 return DenseMapInfo<Instruction *>::getTombstoneKey();
2676 static unsigned getHashValue(Instruction *I) {
2677 assert(canHandle(I) && "Unknown instruction!");
2678 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2679 I->value_op_end()));
2681 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2682 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2683 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2685 return LHS->isIdenticalTo(RHS);
2690 /// \brief Check whether this block is a predicated block.
2691 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2692 /// = ...; " blocks. We start with one vectorized basic block. For every
2693 /// conditional block we split this vectorized block. Therefore, every second
2694 /// block will be a predicated one.
2695 static bool isPredicatedBlock(unsigned BlockNum) {
2696 return BlockNum % 2;
2699 ///\brief Perform cse of induction variable instructions.
2700 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2701 // Perform simple cse.
2702 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2703 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2704 BasicBlock *BB = BBs[i];
2705 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2706 Instruction *In = I++;
2708 if (!CSEDenseMapInfo::canHandle(In))
2711 // Check if we can replace this instruction with any of the
2712 // visited instructions.
2713 if (Instruction *V = CSEMap.lookup(In)) {
2714 In->replaceAllUsesWith(V);
2715 In->eraseFromParent();
2718 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2719 // ...;" blocks for predicated stores. Every second block is a predicated
2721 if (isPredicatedBlock(i))
2729 /// \brief Adds a 'fast' flag to floating point operations.
2730 static Value *addFastMathFlag(Value *V) {
2731 if (isa<FPMathOperator>(V)){
2732 FastMathFlags Flags;
2733 Flags.setUnsafeAlgebra();
2734 cast<Instruction>(V)->setFastMathFlags(Flags);
2739 void InnerLoopVectorizer::vectorizeLoop() {
2740 //===------------------------------------------------===//
2742 // Notice: any optimization or new instruction that go
2743 // into the code below should be also be implemented in
2746 //===------------------------------------------------===//
2747 Constant *Zero = Builder.getInt32(0);
2749 // In order to support reduction variables we need to be able to vectorize
2750 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2751 // stages. First, we create a new vector PHI node with no incoming edges.
2752 // We use this value when we vectorize all of the instructions that use the
2753 // PHI. Next, after all of the instructions in the block are complete we
2754 // add the new incoming edges to the PHI. At this point all of the
2755 // instructions in the basic block are vectorized, so we can use them to
2756 // construct the PHI.
2757 PhiVector RdxPHIsToFix;
2759 // Scan the loop in a topological order to ensure that defs are vectorized
2761 LoopBlocksDFS DFS(OrigLoop);
2764 // Vectorize all of the blocks in the original loop.
2765 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2766 be = DFS.endRPO(); bb != be; ++bb)
2767 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2769 // At this point every instruction in the original loop is widened to
2770 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2771 // that we vectorized. The PHI nodes are currently empty because we did
2772 // not want to introduce cycles. Notice that the remaining PHI nodes
2773 // that we need to fix are reduction variables.
2775 // Create the 'reduced' values for each of the induction vars.
2776 // The reduced values are the vector values that we scalarize and combine
2777 // after the loop is finished.
2778 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2780 PHINode *RdxPhi = *it;
2781 assert(RdxPhi && "Unable to recover vectorized PHI");
2783 // Find the reduction variable descriptor.
2784 assert(Legal->getReductionVars()->count(RdxPhi) &&
2785 "Unable to find the reduction variable");
2786 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2787 (*Legal->getReductionVars())[RdxPhi];
2789 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2791 // We need to generate a reduction vector from the incoming scalar.
2792 // To do so, we need to generate the 'identity' vector and override
2793 // one of the elements with the incoming scalar reduction. We need
2794 // to do it in the vector-loop preheader.
2795 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
2797 // This is the vector-clone of the value that leaves the loop.
2798 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2799 Type *VecTy = VectorExit[0]->getType();
2801 // Find the reduction identity variable. Zero for addition, or, xor,
2802 // one for multiplication, -1 for And.
2805 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2806 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2807 // MinMax reduction have the start value as their identify.
2809 VectorStart = Identity = RdxDesc.StartValue;
2811 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2816 // Handle other reduction kinds:
2818 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2819 VecTy->getScalarType());
2822 // This vector is the Identity vector where the first element is the
2823 // incoming scalar reduction.
2824 VectorStart = RdxDesc.StartValue;
2826 Identity = ConstantVector::getSplat(VF, Iden);
2828 // This vector is the Identity vector where the first element is the
2829 // incoming scalar reduction.
2830 VectorStart = Builder.CreateInsertElement(Identity,
2831 RdxDesc.StartValue, Zero);
2835 // Fix the vector-loop phi.
2837 // Reductions do not have to start at zero. They can start with
2838 // any loop invariant values.
2839 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2840 BasicBlock *Latch = OrigLoop->getLoopLatch();
2841 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2842 VectorParts &Val = getVectorValue(LoopVal);
2843 for (unsigned part = 0; part < UF; ++part) {
2844 // Make sure to add the reduction stat value only to the
2845 // first unroll part.
2846 Value *StartVal = (part == 0) ? VectorStart : Identity;
2847 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
2848 LoopVectorPreHeader);
2849 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2850 LoopVectorBody.back());
2853 // Before each round, move the insertion point right between
2854 // the PHIs and the values we are going to write.
2855 // This allows us to write both PHINodes and the extractelement
2857 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2859 VectorParts RdxParts;
2860 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2861 for (unsigned part = 0; part < UF; ++part) {
2862 // This PHINode contains the vectorized reduction variable, or
2863 // the initial value vector, if we bypass the vector loop.
2864 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2865 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2866 Value *StartVal = (part == 0) ? VectorStart : Identity;
2867 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2868 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2869 NewPhi->addIncoming(RdxExitVal[part],
2870 LoopVectorBody.back());
2871 RdxParts.push_back(NewPhi);
2874 // Reduce all of the unrolled parts into a single vector.
2875 Value *ReducedPartRdx = RdxParts[0];
2876 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2877 setDebugLocFromInst(Builder, ReducedPartRdx);
2878 for (unsigned part = 1; part < UF; ++part) {
2879 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2880 // Floating point operations had to be 'fast' to enable the reduction.
2881 ReducedPartRdx = addFastMathFlag(
2882 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2883 ReducedPartRdx, "bin.rdx"));
2885 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2886 ReducedPartRdx, RdxParts[part]);
2890 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2891 // and vector ops, reducing the set of values being computed by half each
2893 assert(isPowerOf2_32(VF) &&
2894 "Reduction emission only supported for pow2 vectors!");
2895 Value *TmpVec = ReducedPartRdx;
2896 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2897 for (unsigned i = VF; i != 1; i >>= 1) {
2898 // Move the upper half of the vector to the lower half.
2899 for (unsigned j = 0; j != i/2; ++j)
2900 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2902 // Fill the rest of the mask with undef.
2903 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2904 UndefValue::get(Builder.getInt32Ty()));
2907 Builder.CreateShuffleVector(TmpVec,
2908 UndefValue::get(TmpVec->getType()),
2909 ConstantVector::get(ShuffleMask),
2912 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2913 // Floating point operations had to be 'fast' to enable the reduction.
2914 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2915 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2917 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2920 // The result is in the first element of the vector.
2921 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2922 Builder.getInt32(0));
2925 // Create a phi node that merges control-flow from the backedge-taken check
2926 // block and the middle block.
2927 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
2928 LoopScalarPreHeader->getTerminator());
2929 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
2930 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2932 // Now, we need to fix the users of the reduction variable
2933 // inside and outside of the scalar remainder loop.
2934 // We know that the loop is in LCSSA form. We need to update the
2935 // PHI nodes in the exit blocks.
2936 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2937 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2938 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2939 if (!LCSSAPhi) break;
2941 // All PHINodes need to have a single entry edge, or two if
2942 // we already fixed them.
2943 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2945 // We found our reduction value exit-PHI. Update it with the
2946 // incoming bypass edge.
2947 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2948 // Add an edge coming from the bypass.
2949 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2952 }// end of the LCSSA phi scan.
2954 // Fix the scalar loop reduction variable with the incoming reduction sum
2955 // from the vector body and from the backedge value.
2956 int IncomingEdgeBlockIdx =
2957 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2958 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2959 // Pick the other block.
2960 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2961 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
2962 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2963 }// end of for each redux variable.
2967 // Remove redundant induction instructions.
2968 cse(LoopVectorBody);
2971 void InnerLoopVectorizer::fixLCSSAPHIs() {
2972 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2973 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2974 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2975 if (!LCSSAPhi) break;
2976 if (LCSSAPhi->getNumIncomingValues() == 1)
2977 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2982 InnerLoopVectorizer::VectorParts
2983 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2984 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2987 // Look for cached value.
2988 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2989 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2990 if (ECEntryIt != MaskCache.end())
2991 return ECEntryIt->second;
2993 VectorParts SrcMask = createBlockInMask(Src);
2995 // The terminator has to be a branch inst!
2996 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2997 assert(BI && "Unexpected terminator found");
2999 if (BI->isConditional()) {
3000 VectorParts EdgeMask = getVectorValue(BI->getCondition());
3002 if (BI->getSuccessor(0) != Dst)
3003 for (unsigned part = 0; part < UF; ++part)
3004 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
3006 for (unsigned part = 0; part < UF; ++part)
3007 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
3009 MaskCache[Edge] = EdgeMask;
3013 MaskCache[Edge] = SrcMask;
3017 InnerLoopVectorizer::VectorParts
3018 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
3019 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
3021 // Loop incoming mask is all-one.
3022 if (OrigLoop->getHeader() == BB) {
3023 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
3024 return getVectorValue(C);
3027 // This is the block mask. We OR all incoming edges, and with zero.
3028 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
3029 VectorParts BlockMask = getVectorValue(Zero);
3032 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
3033 VectorParts EM = createEdgeMask(*it, BB);
3034 for (unsigned part = 0; part < UF; ++part)
3035 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
3041 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
3042 InnerLoopVectorizer::VectorParts &Entry,
3043 unsigned UF, unsigned VF, PhiVector *PV) {
3044 PHINode* P = cast<PHINode>(PN);
3045 // Handle reduction variables:
3046 if (Legal->getReductionVars()->count(P)) {
3047 for (unsigned part = 0; part < UF; ++part) {
3048 // This is phase one of vectorizing PHIs.
3049 Type *VecTy = (VF == 1) ? PN->getType() :
3050 VectorType::get(PN->getType(), VF);
3051 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
3052 LoopVectorBody.back()-> getFirstInsertionPt());
3058 setDebugLocFromInst(Builder, P);
3059 // Check for PHI nodes that are lowered to vector selects.
3060 if (P->getParent() != OrigLoop->getHeader()) {
3061 // We know that all PHIs in non-header blocks are converted into
3062 // selects, so we don't have to worry about the insertion order and we
3063 // can just use the builder.
3064 // At this point we generate the predication tree. There may be
3065 // duplications since this is a simple recursive scan, but future
3066 // optimizations will clean it up.
3068 unsigned NumIncoming = P->getNumIncomingValues();
3070 // Generate a sequence of selects of the form:
3071 // SELECT(Mask3, In3,
3072 // SELECT(Mask2, In2,
3074 for (unsigned In = 0; In < NumIncoming; In++) {
3075 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
3077 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3079 for (unsigned part = 0; part < UF; ++part) {
3080 // We might have single edge PHIs (blocks) - use an identity
3081 // 'select' for the first PHI operand.
3083 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3086 // Select between the current value and the previous incoming edge
3087 // based on the incoming mask.
3088 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3089 Entry[part], "predphi");
3095 // This PHINode must be an induction variable.
3096 // Make sure that we know about it.
3097 assert(Legal->getInductionVars()->count(P) &&
3098 "Not an induction variable");
3100 LoopVectorizationLegality::InductionInfo II =
3101 Legal->getInductionVars()->lookup(P);
3104 case LoopVectorizationLegality::IK_NoInduction:
3105 llvm_unreachable("Unknown induction");
3106 case LoopVectorizationLegality::IK_IntInduction: {
3107 assert(P->getType() == II.StartValue->getType() && "Types must match");
3108 Type *PhiTy = P->getType();
3110 if (P == OldInduction) {
3111 // Handle the canonical induction variable. We might have had to
3113 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
3115 // Handle other induction variables that are now based on the
3117 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
3119 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
3120 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
3123 Broadcasted = getBroadcastInstrs(Broadcasted);
3124 // After broadcasting the induction variable we need to make the vector
3125 // consecutive by adding 0, 1, 2, etc.
3126 for (unsigned part = 0; part < UF; ++part)
3127 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
3130 case LoopVectorizationLegality::IK_ReverseIntInduction:
3131 case LoopVectorizationLegality::IK_PtrInduction:
3132 case LoopVectorizationLegality::IK_ReversePtrInduction:
3133 // Handle reverse integer and pointer inductions.
3134 Value *StartIdx = ExtendedIdx;
3135 // This is the normalized GEP that starts counting at zero.
3136 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
3139 // Handle the reverse integer induction variable case.
3140 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
3141 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
3142 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
3144 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
3147 // This is a new value so do not hoist it out.
3148 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
3149 // After broadcasting the induction variable we need to make the
3150 // vector consecutive by adding ... -3, -2, -1, 0.
3151 for (unsigned part = 0; part < UF; ++part)
3152 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
3157 // Handle the pointer induction variable case.
3158 assert(P->getType()->isPointerTy() && "Unexpected type.");
3160 // Is this a reverse induction ptr or a consecutive induction ptr.
3161 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
3164 // This is the vector of results. Notice that we don't generate
3165 // vector geps because scalar geps result in better code.
3166 for (unsigned part = 0; part < UF; ++part) {
3168 int EltIndex = (part) * (Reverse ? -1 : 1);
3169 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3172 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3174 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3176 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3178 Entry[part] = SclrGep;
3182 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3183 for (unsigned int i = 0; i < VF; ++i) {
3184 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
3185 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3188 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3190 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3192 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3194 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3195 Builder.getInt32(i),
3198 Entry[part] = VecVal;
3204 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3205 // For each instruction in the old loop.
3206 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3207 VectorParts &Entry = WidenMap.get(it);
3208 switch (it->getOpcode()) {
3209 case Instruction::Br:
3210 // Nothing to do for PHIs and BR, since we already took care of the
3211 // loop control flow instructions.
3213 case Instruction::PHI:{
3214 // Vectorize PHINodes.
3215 widenPHIInstruction(it, Entry, UF, VF, PV);
3219 case Instruction::Add:
3220 case Instruction::FAdd:
3221 case Instruction::Sub:
3222 case Instruction::FSub:
3223 case Instruction::Mul:
3224 case Instruction::FMul:
3225 case Instruction::UDiv:
3226 case Instruction::SDiv:
3227 case Instruction::FDiv:
3228 case Instruction::URem:
3229 case Instruction::SRem:
3230 case Instruction::FRem:
3231 case Instruction::Shl:
3232 case Instruction::LShr:
3233 case Instruction::AShr:
3234 case Instruction::And:
3235 case Instruction::Or:
3236 case Instruction::Xor: {
3237 // Just widen binops.
3238 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3239 setDebugLocFromInst(Builder, BinOp);
3240 VectorParts &A = getVectorValue(it->getOperand(0));
3241 VectorParts &B = getVectorValue(it->getOperand(1));
3243 // Use this vector value for all users of the original instruction.
3244 for (unsigned Part = 0; Part < UF; ++Part) {
3245 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3247 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3248 VecOp->copyIRFlags(BinOp);
3253 propagateMetadata(Entry, it);
3256 case Instruction::Select: {
3258 // If the selector is loop invariant we can create a select
3259 // instruction with a scalar condition. Otherwise, use vector-select.
3260 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3262 setDebugLocFromInst(Builder, it);
3264 // The condition can be loop invariant but still defined inside the
3265 // loop. This means that we can't just use the original 'cond' value.
3266 // We have to take the 'vectorized' value and pick the first lane.
3267 // Instcombine will make this a no-op.
3268 VectorParts &Cond = getVectorValue(it->getOperand(0));
3269 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3270 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3272 Value *ScalarCond = (VF == 1) ? Cond[0] :
3273 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3275 for (unsigned Part = 0; Part < UF; ++Part) {
3276 Entry[Part] = Builder.CreateSelect(
3277 InvariantCond ? ScalarCond : Cond[Part],
3282 propagateMetadata(Entry, it);
3286 case Instruction::ICmp:
3287 case Instruction::FCmp: {
3288 // Widen compares. Generate vector compares.
3289 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3290 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3291 setDebugLocFromInst(Builder, it);
3292 VectorParts &A = getVectorValue(it->getOperand(0));
3293 VectorParts &B = getVectorValue(it->getOperand(1));
3294 for (unsigned Part = 0; Part < UF; ++Part) {
3297 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3299 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3303 propagateMetadata(Entry, it);
3307 case Instruction::Store:
3308 case Instruction::Load:
3309 vectorizeMemoryInstruction(it);
3311 case Instruction::ZExt:
3312 case Instruction::SExt:
3313 case Instruction::FPToUI:
3314 case Instruction::FPToSI:
3315 case Instruction::FPExt:
3316 case Instruction::PtrToInt:
3317 case Instruction::IntToPtr:
3318 case Instruction::SIToFP:
3319 case Instruction::UIToFP:
3320 case Instruction::Trunc:
3321 case Instruction::FPTrunc:
3322 case Instruction::BitCast: {
3323 CastInst *CI = dyn_cast<CastInst>(it);
3324 setDebugLocFromInst(Builder, it);
3325 /// Optimize the special case where the source is the induction
3326 /// variable. Notice that we can only optimize the 'trunc' case
3327 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3328 /// c. other casts depend on pointer size.
3329 if (CI->getOperand(0) == OldInduction &&
3330 it->getOpcode() == Instruction::Trunc) {
3331 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3333 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3334 for (unsigned Part = 0; Part < UF; ++Part)
3335 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
3336 propagateMetadata(Entry, it);
3339 /// Vectorize casts.
3340 Type *DestTy = (VF == 1) ? CI->getType() :
3341 VectorType::get(CI->getType(), VF);
3343 VectorParts &A = getVectorValue(it->getOperand(0));
3344 for (unsigned Part = 0; Part < UF; ++Part)
3345 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3346 propagateMetadata(Entry, it);
3350 case Instruction::Call: {
3351 // Ignore dbg intrinsics.
3352 if (isa<DbgInfoIntrinsic>(it))
3354 setDebugLocFromInst(Builder, it);
3356 Module *M = BB->getParent()->getParent();
3357 CallInst *CI = cast<CallInst>(it);
3358 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3359 assert(ID && "Not an intrinsic call!");
3361 case Intrinsic::assume:
3362 case Intrinsic::lifetime_end:
3363 case Intrinsic::lifetime_start:
3364 scalarizeInstruction(it);
3367 bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
3368 for (unsigned Part = 0; Part < UF; ++Part) {
3369 SmallVector<Value *, 4> Args;
3370 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3371 if (HasScalarOpd && i == 1) {
3372 Args.push_back(CI->getArgOperand(i));
3375 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3376 Args.push_back(Arg[Part]);
3378 Type *Tys[] = {CI->getType()};
3380 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3382 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3383 Entry[Part] = Builder.CreateCall(F, Args);
3386 propagateMetadata(Entry, it);
3393 // All other instructions are unsupported. Scalarize them.
3394 scalarizeInstruction(it);
3397 }// end of for_each instr.
3400 void InnerLoopVectorizer::updateAnalysis() {
3401 // Forget the original basic block.
3402 SE->forgetLoop(OrigLoop);
3404 // Update the dominator tree information.
3405 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3406 "Entry does not dominate exit.");
3408 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3409 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3410 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3412 // Due to if predication of stores we might create a sequence of "if(pred)
3413 // a[i] = ...; " blocks.
3414 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3416 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3417 else if (isPredicatedBlock(i)) {
3418 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3420 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3424 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3425 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3426 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3427 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3429 DEBUG(DT->verifyDomTree());
3432 /// \brief Check whether it is safe to if-convert this phi node.
3434 /// Phi nodes with constant expressions that can trap are not safe to if
3436 static bool canIfConvertPHINodes(BasicBlock *BB) {
3437 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3438 PHINode *Phi = dyn_cast<PHINode>(I);
3441 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3442 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3449 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3450 if (!EnableIfConversion) {
3451 emitAnalysis(Report() << "if-conversion is disabled");
3455 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3457 // A list of pointers that we can safely read and write to.
3458 SmallPtrSet<Value *, 8> SafePointes;
3460 // Collect safe addresses.
3461 for (Loop::block_iterator BI = TheLoop->block_begin(),
3462 BE = TheLoop->block_end(); BI != BE; ++BI) {
3463 BasicBlock *BB = *BI;
3465 if (blockNeedsPredication(BB))
3468 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3469 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3470 SafePointes.insert(LI->getPointerOperand());
3471 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3472 SafePointes.insert(SI->getPointerOperand());
3476 // Collect the blocks that need predication.
3477 BasicBlock *Header = TheLoop->getHeader();
3478 for (Loop::block_iterator BI = TheLoop->block_begin(),
3479 BE = TheLoop->block_end(); BI != BE; ++BI) {
3480 BasicBlock *BB = *BI;
3482 // We don't support switch statements inside loops.
3483 if (!isa<BranchInst>(BB->getTerminator())) {
3484 emitAnalysis(Report(BB->getTerminator())
3485 << "loop contains a switch statement");
3489 // We must be able to predicate all blocks that need to be predicated.
3490 if (blockNeedsPredication(BB)) {
3491 if (!blockCanBePredicated(BB, SafePointes)) {
3492 emitAnalysis(Report(BB->getTerminator())
3493 << "control flow cannot be substituted for a select");
3496 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3497 emitAnalysis(Report(BB->getTerminator())
3498 << "control flow cannot be substituted for a select");
3503 // We can if-convert this loop.
3507 bool LoopVectorizationLegality::canVectorize() {
3508 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3509 // be canonicalized.
3510 if (!TheLoop->getLoopPreheader()) {
3512 Report() << "loop control flow is not understood by vectorizer");
3516 // We can only vectorize innermost loops.
3517 if (TheLoop->getSubLoopsVector().size()) {
3518 emitAnalysis(Report() << "loop is not the innermost loop");
3522 // We must have a single backedge.
3523 if (TheLoop->getNumBackEdges() != 1) {
3525 Report() << "loop control flow is not understood by vectorizer");
3529 // We must have a single exiting block.
3530 if (!TheLoop->getExitingBlock()) {
3532 Report() << "loop control flow is not understood by vectorizer");
3536 // We only handle bottom-tested loops, i.e. loop in which the condition is
3537 // checked at the end of each iteration. With that we can assume that all
3538 // instructions in the loop are executed the same number of times.
3539 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
3541 Report() << "loop control flow is not understood by vectorizer");
3545 // We need to have a loop header.
3546 DEBUG(dbgs() << "LV: Found a loop: " <<
3547 TheLoop->getHeader()->getName() << '\n');
3549 // Check if we can if-convert non-single-bb loops.
3550 unsigned NumBlocks = TheLoop->getNumBlocks();
3551 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3552 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3556 // ScalarEvolution needs to be able to find the exit count.
3557 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3558 if (ExitCount == SE->getCouldNotCompute()) {
3559 emitAnalysis(Report() << "could not determine number of loop iterations");
3560 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3564 // Check if we can vectorize the instructions and CFG in this loop.
3565 if (!canVectorizeInstrs()) {
3566 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3570 // Go over each instruction and look at memory deps.
3571 if (!canVectorizeMemory()) {
3572 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3576 // Collect all of the variables that remain uniform after vectorization.
3577 collectLoopUniforms();
3579 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3580 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
3583 // Okay! We can vectorize. At this point we don't have any other mem analysis
3584 // which may limit our maximum vectorization factor, so just return true with
3589 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3590 if (Ty->isPointerTy())
3591 return DL.getIntPtrType(Ty);
3593 // It is possible that char's or short's overflow when we ask for the loop's
3594 // trip count, work around this by changing the type size.
3595 if (Ty->getScalarSizeInBits() < 32)
3596 return Type::getInt32Ty(Ty->getContext());
3601 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3602 Ty0 = convertPointerToIntegerType(DL, Ty0);
3603 Ty1 = convertPointerToIntegerType(DL, Ty1);
3604 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3609 /// \brief Check that the instruction has outside loop users and is not an
3610 /// identified reduction variable.
3611 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3612 SmallPtrSetImpl<Value *> &Reductions) {
3613 // Reduction instructions are allowed to have exit users. All other
3614 // instructions must not have external users.
3615 if (!Reductions.count(Inst))
3616 //Check that all of the users of the loop are inside the BB.
3617 for (User *U : Inst->users()) {
3618 Instruction *UI = cast<Instruction>(U);
3619 // This user may be a reduction exit value.
3620 if (!TheLoop->contains(UI)) {
3621 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3628 bool LoopVectorizationLegality::canVectorizeInstrs() {
3629 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3630 BasicBlock *Header = TheLoop->getHeader();
3632 // Look for the attribute signaling the absence of NaNs.
3633 Function &F = *Header->getParent();
3634 if (F.hasFnAttribute("no-nans-fp-math"))
3635 HasFunNoNaNAttr = F.getAttributes().getAttribute(
3636 AttributeSet::FunctionIndex,
3637 "no-nans-fp-math").getValueAsString() == "true";
3639 // For each block in the loop.
3640 for (Loop::block_iterator bb = TheLoop->block_begin(),
3641 be = TheLoop->block_end(); bb != be; ++bb) {
3643 // Scan the instructions in the block and look for hazards.
3644 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3647 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3648 Type *PhiTy = Phi->getType();
3649 // Check that this PHI type is allowed.
3650 if (!PhiTy->isIntegerTy() &&
3651 !PhiTy->isFloatingPointTy() &&
3652 !PhiTy->isPointerTy()) {
3653 emitAnalysis(Report(it)
3654 << "loop control flow is not understood by vectorizer");
3655 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3659 // If this PHINode is not in the header block, then we know that we
3660 // can convert it to select during if-conversion. No need to check if
3661 // the PHIs in this block are induction or reduction variables.
3662 if (*bb != Header) {
3663 // Check that this instruction has no outside users or is an
3664 // identified reduction value with an outside user.
3665 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3667 emitAnalysis(Report(it) << "value could not be identified as "
3668 "an induction or reduction variable");
3672 // We only allow if-converted PHIs with more than two incoming values.
3673 if (Phi->getNumIncomingValues() != 2) {
3674 emitAnalysis(Report(it)
3675 << "control flow not understood by vectorizer");
3676 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3680 // This is the value coming from the preheader.
3681 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3682 // Check if this is an induction variable.
3683 InductionKind IK = isInductionVariable(Phi);
3685 if (IK_NoInduction != IK) {
3686 // Get the widest type.
3688 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3690 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3692 // Int inductions are special because we only allow one IV.
3693 if (IK == IK_IntInduction) {
3694 // Use the phi node with the widest type as induction. Use the last
3695 // one if there are multiple (no good reason for doing this other
3696 // than it is expedient).
3697 if (!Induction || PhiTy == WidestIndTy)
3701 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3702 Inductions[Phi] = InductionInfo(StartValue, IK);
3704 // Until we explicitly handle the case of an induction variable with
3705 // an outside loop user we have to give up vectorizing this loop.
3706 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3707 emitAnalysis(Report(it) << "use of induction value outside of the "
3708 "loop is not handled by vectorizer");
3715 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3716 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3719 if (AddReductionVar(Phi, RK_IntegerMult)) {
3720 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3723 if (AddReductionVar(Phi, RK_IntegerOr)) {
3724 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3727 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3728 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3731 if (AddReductionVar(Phi, RK_IntegerXor)) {
3732 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3735 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3736 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3739 if (AddReductionVar(Phi, RK_FloatMult)) {
3740 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3743 if (AddReductionVar(Phi, RK_FloatAdd)) {
3744 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3747 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3748 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3753 emitAnalysis(Report(it) << "value that could not be identified as "
3754 "reduction is used outside the loop");
3755 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3757 }// end of PHI handling
3759 // We still don't handle functions. However, we can ignore dbg intrinsic
3760 // calls and we do handle certain intrinsic and libm functions.
3761 CallInst *CI = dyn_cast<CallInst>(it);
3762 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3763 emitAnalysis(Report(it) << "call instruction cannot be vectorized");
3764 DEBUG(dbgs() << "LV: Found a call site.\n");
3768 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
3769 // second argument is the same (i.e. loop invariant)
3771 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
3772 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
3773 emitAnalysis(Report(it)
3774 << "intrinsic instruction cannot be vectorized");
3775 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
3780 // Check that the instruction return type is vectorizable.
3781 // Also, we can't vectorize extractelement instructions.
3782 if ((!VectorType::isValidElementType(it->getType()) &&
3783 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3784 emitAnalysis(Report(it)
3785 << "instruction return type cannot be vectorized");
3786 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3790 // Check that the stored type is vectorizable.
3791 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3792 Type *T = ST->getValueOperand()->getType();
3793 if (!VectorType::isValidElementType(T)) {
3794 emitAnalysis(Report(ST) << "store instruction cannot be vectorized");
3797 if (EnableMemAccessVersioning)
3798 collectStridedAcccess(ST);
3801 if (EnableMemAccessVersioning)
3802 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3803 collectStridedAcccess(LI);
3805 // Reduction instructions are allowed to have exit users.
3806 // All other instructions must not have external users.
3807 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3808 emitAnalysis(Report(it) << "value cannot be used outside the loop");
3817 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3818 if (Inductions.empty()) {
3819 emitAnalysis(Report()
3820 << "loop induction variable could not be identified");
3828 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3829 /// return the induction operand of the gep pointer.
3830 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3831 const DataLayout *DL, Loop *Lp) {
3832 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3836 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3838 // Check that all of the gep indices are uniform except for our induction
3840 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3841 if (i != InductionOperand &&
3842 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3844 return GEP->getOperand(InductionOperand);
3847 ///\brief Look for a cast use of the passed value.
3848 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3849 Value *UniqueCast = nullptr;
3850 for (User *U : Ptr->users()) {
3851 CastInst *CI = dyn_cast<CastInst>(U);
3852 if (CI && CI->getType() == Ty) {
3862 ///\brief Get the stride of a pointer access in a loop.
3863 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3864 /// pointer to the Value, or null otherwise.
3865 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3866 const DataLayout *DL, Loop *Lp) {
3867 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3868 if (!PtrTy || PtrTy->isAggregateType())
3871 // Try to remove a gep instruction to make the pointer (actually index at this
3872 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3873 // pointer, otherwise, we are analyzing the index.
3874 Value *OrigPtr = Ptr;
3876 // The size of the pointer access.
3877 int64_t PtrAccessSize = 1;
3879 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3880 const SCEV *V = SE->getSCEV(Ptr);
3884 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3885 V = C->getOperand();
3887 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3891 V = S->getStepRecurrence(*SE);
3895 // Strip off the size of access multiplication if we are still analyzing the
3897 if (OrigPtr == Ptr) {
3898 DL->getTypeAllocSize(PtrTy->getElementType());
3899 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3900 if (M->getOperand(0)->getSCEVType() != scConstant)
3903 const APInt &APStepVal =
3904 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3906 // Huge step value - give up.
3907 if (APStepVal.getBitWidth() > 64)
3910 int64_t StepVal = APStepVal.getSExtValue();
3911 if (PtrAccessSize != StepVal)
3913 V = M->getOperand(1);
3918 Type *StripedOffRecurrenceCast = nullptr;
3919 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3920 StripedOffRecurrenceCast = C->getType();
3921 V = C->getOperand();
3924 // Look for the loop invariant symbolic value.
3925 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3929 Value *Stride = U->getValue();
3930 if (!Lp->isLoopInvariant(Stride))
3933 // If we have stripped off the recurrence cast we have to make sure that we
3934 // return the value that is used in this loop so that we can replace it later.
3935 if (StripedOffRecurrenceCast)
3936 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3941 void LoopVectorizationLegality::collectStridedAcccess(Value *MemAccess) {
3942 Value *Ptr = nullptr;
3943 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3944 Ptr = LI->getPointerOperand();
3945 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3946 Ptr = SI->getPointerOperand();
3950 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3954 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3955 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3956 Strides[Ptr] = Stride;
3957 StrideSet.insert(Stride);
3960 void LoopVectorizationLegality::collectLoopUniforms() {
3961 // We now know that the loop is vectorizable!
3962 // Collect variables that will remain uniform after vectorization.
3963 std::vector<Value*> Worklist;
3964 BasicBlock *Latch = TheLoop->getLoopLatch();
3966 // Start with the conditional branch and walk up the block.
3967 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3969 // Also add all consecutive pointer values; these values will be uniform
3970 // after vectorization (and subsequent cleanup) and, until revectorization is
3971 // supported, all dependencies must also be uniform.
3972 for (Loop::block_iterator B = TheLoop->block_begin(),
3973 BE = TheLoop->block_end(); B != BE; ++B)
3974 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
3976 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
3977 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3979 while (Worklist.size()) {
3980 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3981 Worklist.pop_back();
3983 // Look at instructions inside this loop.
3984 // Stop when reaching PHI nodes.
3985 // TODO: we need to follow values all over the loop, not only in this block.
3986 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3989 // This is a known uniform.
3992 // Insert all operands.
3993 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3998 /// \brief Analyses memory accesses in a loop.
4000 /// Checks whether run time pointer checks are needed and builds sets for data
4001 /// dependence checking.
4002 class AccessAnalysis {
4004 /// \brief Read or write access location.
4005 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
4006 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
4008 /// \brief Set of potential dependent memory accesses.
4009 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
4011 AccessAnalysis(const DataLayout *Dl, AliasAnalysis *AA, DepCandidates &DA) :
4012 DL(Dl), AST(*AA), DepCands(DA), IsRTCheckNeeded(false) {}
4014 /// \brief Register a load and whether it is only read from.
4015 void addLoad(AliasAnalysis::Location &Loc, bool IsReadOnly) {
4016 Value *Ptr = const_cast<Value*>(Loc.Ptr);
4017 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
4018 Accesses.insert(MemAccessInfo(Ptr, false));
4020 ReadOnlyPtr.insert(Ptr);
4023 /// \brief Register a store.
4024 void addStore(AliasAnalysis::Location &Loc) {
4025 Value *Ptr = const_cast<Value*>(Loc.Ptr);
4026 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
4027 Accesses.insert(MemAccessInfo(Ptr, true));
4030 /// \brief Check whether we can check the pointers at runtime for
4031 /// non-intersection.
4032 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
4033 unsigned &NumComparisons, ScalarEvolution *SE,
4034 Loop *TheLoop, ValueToValueMap &Strides,
4035 bool ShouldCheckStride = false);
4037 /// \brief Goes over all memory accesses, checks whether a RT check is needed
4038 /// and builds sets of dependent accesses.
4039 void buildDependenceSets() {
4040 processMemAccesses();
4043 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
4045 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
4046 void resetDepChecks() { CheckDeps.clear(); }
4048 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
4051 typedef SetVector<MemAccessInfo> PtrAccessSet;
4053 /// \brief Go over all memory access and check whether runtime pointer checks
4054 /// are needed /// and build sets of dependency check candidates.
4055 void processMemAccesses();
4057 /// Set of all accesses.
4058 PtrAccessSet Accesses;
4060 /// Set of accesses that need a further dependence check.
4061 MemAccessInfoSet CheckDeps;
4063 /// Set of pointers that are read only.
4064 SmallPtrSet<Value*, 16> ReadOnlyPtr;
4066 const DataLayout *DL;
4068 /// An alias set tracker to partition the access set by underlying object and
4069 //intrinsic property (such as TBAA metadata).
4070 AliasSetTracker AST;
4072 /// Sets of potentially dependent accesses - members of one set share an
4073 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
4074 /// dependence check.
4075 DepCandidates &DepCands;
4077 bool IsRTCheckNeeded;
4080 } // end anonymous namespace
4082 /// \brief Check whether a pointer can participate in a runtime bounds check.
4083 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
4085 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
4086 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4090 return AR->isAffine();
4093 /// \brief Check the stride of the pointer and ensure that it does not wrap in
4094 /// the address space.
4095 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4096 const Loop *Lp, ValueToValueMap &StridesMap);
4098 bool AccessAnalysis::canCheckPtrAtRT(
4099 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
4100 unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
4101 ValueToValueMap &StridesMap, bool ShouldCheckStride) {
4102 // Find pointers with computable bounds. We are going to use this information
4103 // to place a runtime bound check.
4104 bool CanDoRT = true;
4106 bool IsDepCheckNeeded = isDependencyCheckNeeded();
4109 // We assign a consecutive id to access from different alias sets.
4110 // Accesses between different groups doesn't need to be checked.
4112 for (auto &AS : AST) {
4113 unsigned NumReadPtrChecks = 0;
4114 unsigned NumWritePtrChecks = 0;
4116 // We assign consecutive id to access from different dependence sets.
4117 // Accesses within the same set don't need a runtime check.
4118 unsigned RunningDepId = 1;
4119 DenseMap<Value *, unsigned> DepSetId;
4122 Value *Ptr = A.getValue();
4123 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
4124 MemAccessInfo Access(Ptr, IsWrite);
4127 ++NumWritePtrChecks;
4131 if (hasComputableBounds(SE, StridesMap, Ptr) &&
4132 // When we run after a failing dependency check we have to make sure we
4133 // don't have wrapping pointers.
4134 (!ShouldCheckStride ||
4135 isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
4136 // The id of the dependence set.
4139 if (IsDepCheckNeeded) {
4140 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
4141 unsigned &LeaderId = DepSetId[Leader];
4143 LeaderId = RunningDepId++;
4146 // Each access has its own dependence set.
4147 DepId = RunningDepId++;
4149 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
4151 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
4157 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
4158 NumComparisons += 0; // Only one dependence set.
4160 NumComparisons += (NumWritePtrChecks * (NumReadPtrChecks +
4161 NumWritePtrChecks - 1));
4167 // If the pointers that we would use for the bounds comparison have different
4168 // address spaces, assume the values aren't directly comparable, so we can't
4169 // use them for the runtime check. We also have to assume they could
4170 // overlap. In the future there should be metadata for whether address spaces
4172 unsigned NumPointers = RtCheck.Pointers.size();
4173 for (unsigned i = 0; i < NumPointers; ++i) {
4174 for (unsigned j = i + 1; j < NumPointers; ++j) {
4175 // Only need to check pointers between two different dependency sets.
4176 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
4178 // Only need to check pointers in the same alias set.
4179 if (RtCheck.AliasSetId[i] != RtCheck.AliasSetId[j])
4182 Value *PtrI = RtCheck.Pointers[i];
4183 Value *PtrJ = RtCheck.Pointers[j];
4185 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
4186 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
4188 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
4189 " different address spaces\n");
4198 void AccessAnalysis::processMemAccesses() {
4199 // We process the set twice: first we process read-write pointers, last we
4200 // process read-only pointers. This allows us to skip dependence tests for
4201 // read-only pointers.
4203 DEBUG(dbgs() << "LV: Processing memory accesses...\n");
4204 DEBUG(dbgs() << " AST: "; AST.dump());
4205 DEBUG(dbgs() << "LV: Accesses:\n");
4207 for (auto A : Accesses)
4208 dbgs() << "\t" << *A.getPointer() << " (" <<
4209 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
4210 "read-only" : "read")) << ")\n";
4213 // The AliasSetTracker has nicely partitioned our pointers by metadata
4214 // compatibility and potential for underlying-object overlap. As a result, we
4215 // only need to check for potential pointer dependencies within each alias
4217 for (auto &AS : AST) {
4218 // Note that both the alias-set tracker and the alias sets themselves used
4219 // linked lists internally and so the iteration order here is deterministic
4220 // (matching the original instruction order within each set).
4222 bool SetHasWrite = false;
4224 // Map of pointers to last access encountered.
4225 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
4226 UnderlyingObjToAccessMap ObjToLastAccess;
4228 // Set of access to check after all writes have been processed.
4229 PtrAccessSet DeferredAccesses;
4231 // Iterate over each alias set twice, once to process read/write pointers,
4232 // and then to process read-only pointers.
4233 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
4234 bool UseDeferred = SetIteration > 0;
4235 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
4238 Value *Ptr = A.getValue();
4239 bool IsWrite = S.count(MemAccessInfo(Ptr, true));
4241 // If we're using the deferred access set, then it contains only reads.
4242 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
4243 if (UseDeferred && !IsReadOnlyPtr)
4245 // Otherwise, the pointer must be in the PtrAccessSet, either as a read
4247 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
4248 S.count(MemAccessInfo(Ptr, false))) &&
4249 "Alias-set pointer not in the access set?");
4251 MemAccessInfo Access(Ptr, IsWrite);
4252 DepCands.insert(Access);
4254 // Memorize read-only pointers for later processing and skip them in the
4255 // first round (they need to be checked after we have seen all write
4256 // pointers). Note: we also mark pointer that are not consecutive as
4257 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need
4258 // the second check for "!IsWrite".
4259 if (!UseDeferred && IsReadOnlyPtr) {
4260 DeferredAccesses.insert(Access);
4264 // If this is a write - check other reads and writes for conflicts. If
4265 // this is a read only check other writes for conflicts (but only if
4266 // there is no other write to the ptr - this is an optimization to
4267 // catch "a[i] = a[i] + " without having to do a dependence check).
4268 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
4269 CheckDeps.insert(Access);
4270 IsRTCheckNeeded = true;
4276 // Create sets of pointers connected by a shared alias set and
4277 // underlying object.
4278 typedef SmallVector<Value *, 16> ValueVector;
4279 ValueVector TempObjects;
4280 GetUnderlyingObjects(Ptr, TempObjects, DL);
4281 for (Value *UnderlyingObj : TempObjects) {
4282 UnderlyingObjToAccessMap::iterator Prev =
4283 ObjToLastAccess.find(UnderlyingObj);
4284 if (Prev != ObjToLastAccess.end())
4285 DepCands.unionSets(Access, Prev->second);
4287 ObjToLastAccess[UnderlyingObj] = Access;
4295 /// \brief Checks memory dependences among accesses to the same underlying
4296 /// object to determine whether there vectorization is legal or not (and at
4297 /// which vectorization factor).
4299 /// This class works under the assumption that we already checked that memory
4300 /// locations with different underlying pointers are "must-not alias".
4301 /// We use the ScalarEvolution framework to symbolically evalutate access
4302 /// functions pairs. Since we currently don't restructure the loop we can rely
4303 /// on the program order of memory accesses to determine their safety.
4304 /// At the moment we will only deem accesses as safe for:
4305 /// * A negative constant distance assuming program order.
4307 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
4308 /// a[i] = tmp; y = a[i];
4310 /// The latter case is safe because later checks guarantuee that there can't
4311 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
4312 /// the same variable: a header phi can only be an induction or a reduction, a
4313 /// reduction can't have a memory sink, an induction can't have a memory
4314 /// source). This is important and must not be violated (or we have to
4315 /// resort to checking for cycles through memory).
4317 /// * A positive constant distance assuming program order that is bigger
4318 /// than the biggest memory access.
4320 /// tmp = a[i] OR b[i] = x
4321 /// a[i+2] = tmp y = b[i+2];
4323 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
4325 /// * Zero distances and all accesses have the same size.
4327 class MemoryDepChecker {
4329 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
4330 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
4332 MemoryDepChecker(ScalarEvolution *Se, const DataLayout *Dl, const Loop *L)
4333 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
4334 ShouldRetryWithRuntimeCheck(false) {}
4336 /// \brief Register the location (instructions are given increasing numbers)
4337 /// of a write access.
4338 void addAccess(StoreInst *SI) {
4339 Value *Ptr = SI->getPointerOperand();
4340 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
4341 InstMap.push_back(SI);
4345 /// \brief Register the location (instructions are given increasing numbers)
4346 /// of a write access.
4347 void addAccess(LoadInst *LI) {
4348 Value *Ptr = LI->getPointerOperand();
4349 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
4350 InstMap.push_back(LI);
4354 /// \brief Check whether the dependencies between the accesses are safe.
4356 /// Only checks sets with elements in \p CheckDeps.
4357 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4358 MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
4360 /// \brief The maximum number of bytes of a vector register we can vectorize
4361 /// the accesses safely with.
4362 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
4364 /// \brief In same cases when the dependency check fails we can still
4365 /// vectorize the loop with a dynamic array access check.
4366 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
4369 ScalarEvolution *SE;
4370 const DataLayout *DL;
4371 const Loop *InnermostLoop;
4373 /// \brief Maps access locations (ptr, read/write) to program order.
4374 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
4376 /// \brief Memory access instructions in program order.
4377 SmallVector<Instruction *, 16> InstMap;
4379 /// \brief The program order index to be used for the next instruction.
4382 // We can access this many bytes in parallel safely.
4383 unsigned MaxSafeDepDistBytes;
4385 /// \brief If we see a non-constant dependence distance we can still try to
4386 /// vectorize this loop with runtime checks.
4387 bool ShouldRetryWithRuntimeCheck;
4389 /// \brief Check whether there is a plausible dependence between the two
4392 /// Access \p A must happen before \p B in program order. The two indices
4393 /// identify the index into the program order map.
4395 /// This function checks whether there is a plausible dependence (or the
4396 /// absence of such can't be proved) between the two accesses. If there is a
4397 /// plausible dependence but the dependence distance is bigger than one
4398 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
4399 /// distance is smaller than any other distance encountered so far).
4400 /// Otherwise, this function returns true signaling a possible dependence.
4401 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
4402 const MemAccessInfo &B, unsigned BIdx,
4403 ValueToValueMap &Strides);
4405 /// \brief Check whether the data dependence could prevent store-load
4407 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
4410 } // end anonymous namespace
4412 static bool isInBoundsGep(Value *Ptr) {
4413 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
4414 return GEP->isInBounds();
4418 /// \brief Check whether the access through \p Ptr has a constant stride.
4419 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4420 const Loop *Lp, ValueToValueMap &StridesMap) {
4421 const Type *Ty = Ptr->getType();
4422 assert(Ty->isPointerTy() && "Unexpected non-ptr");
4424 // Make sure that the pointer does not point to aggregate types.
4425 const PointerType *PtrTy = cast<PointerType>(Ty);
4426 if (PtrTy->getElementType()->isAggregateType()) {
4427 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
4432 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
4434 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4436 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
4437 << *Ptr << " SCEV: " << *PtrScev << "\n");
4441 // The accesss function must stride over the innermost loop.
4442 if (Lp != AR->getLoop()) {
4443 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
4444 *Ptr << " SCEV: " << *PtrScev << "\n");
4447 // The address calculation must not wrap. Otherwise, a dependence could be
4449 // An inbounds getelementptr that is a AddRec with a unit stride
4450 // cannot wrap per definition. The unit stride requirement is checked later.
4451 // An getelementptr without an inbounds attribute and unit stride would have
4452 // to access the pointer value "0" which is undefined behavior in address
4453 // space 0, therefore we can also vectorize this case.
4454 bool IsInBoundsGEP = isInBoundsGep(Ptr);
4455 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
4456 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
4457 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
4458 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
4459 << *Ptr << " SCEV: " << *PtrScev << "\n");
4463 // Check the step is constant.
4464 const SCEV *Step = AR->getStepRecurrence(*SE);
4466 // Calculate the pointer stride and check if it is consecutive.
4467 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4469 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
4470 " SCEV: " << *PtrScev << "\n");
4474 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
4475 const APInt &APStepVal = C->getValue()->getValue();
4477 // Huge step value - give up.
4478 if (APStepVal.getBitWidth() > 64)
4481 int64_t StepVal = APStepVal.getSExtValue();
4484 int64_t Stride = StepVal / Size;
4485 int64_t Rem = StepVal % Size;
4489 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
4490 // know we can't "wrap around the address space". In case of address space
4491 // zero we know that this won't happen without triggering undefined behavior.
4492 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
4493 Stride != 1 && Stride != -1)
4499 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
4500 unsigned TypeByteSize) {
4501 // If loads occur at a distance that is not a multiple of a feasible vector
4502 // factor store-load forwarding does not take place.
4503 // Positive dependences might cause troubles because vectorizing them might
4504 // prevent store-load forwarding making vectorized code run a lot slower.
4505 // a[i] = a[i-3] ^ a[i-8];
4506 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
4507 // hence on your typical architecture store-load forwarding does not take
4508 // place. Vectorizing in such cases does not make sense.
4509 // Store-load forwarding distance.
4510 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
4511 // Maximum vector factor.
4512 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
4513 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
4514 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
4516 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
4518 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
4519 MaxVFWithoutSLForwardIssues = (vf >>=1);
4524 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
4525 DEBUG(dbgs() << "LV: Distance " << Distance <<
4526 " that could cause a store-load forwarding conflict\n");
4530 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
4531 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
4532 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
4536 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
4537 const MemAccessInfo &B, unsigned BIdx,
4538 ValueToValueMap &Strides) {
4539 assert (AIdx < BIdx && "Must pass arguments in program order");
4541 Value *APtr = A.getPointer();
4542 Value *BPtr = B.getPointer();
4543 bool AIsWrite = A.getInt();
4544 bool BIsWrite = B.getInt();
4546 // Two reads are independent.
4547 if (!AIsWrite && !BIsWrite)
4550 // We cannot check pointers in different address spaces.
4551 if (APtr->getType()->getPointerAddressSpace() !=
4552 BPtr->getType()->getPointerAddressSpace())
4555 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
4556 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
4558 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
4559 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
4561 const SCEV *Src = AScev;
4562 const SCEV *Sink = BScev;
4564 // If the induction step is negative we have to invert source and sink of the
4566 if (StrideAPtr < 0) {
4569 std::swap(APtr, BPtr);
4570 std::swap(Src, Sink);
4571 std::swap(AIsWrite, BIsWrite);
4572 std::swap(AIdx, BIdx);
4573 std::swap(StrideAPtr, StrideBPtr);
4576 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
4578 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
4579 << "(Induction step: " << StrideAPtr << ")\n");
4580 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
4581 << *InstMap[BIdx] << ": " << *Dist << "\n");
4583 // Need consecutive accesses. We don't want to vectorize
4584 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
4585 // the address space.
4586 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
4587 DEBUG(dbgs() << "Non-consecutive pointer access\n");
4591 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
4593 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
4594 ShouldRetryWithRuntimeCheck = true;
4598 Type *ATy = APtr->getType()->getPointerElementType();
4599 Type *BTy = BPtr->getType()->getPointerElementType();
4600 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
4602 // Negative distances are not plausible dependencies.
4603 const APInt &Val = C->getValue()->getValue();
4604 if (Val.isNegative()) {
4605 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
4606 if (IsTrueDataDependence &&
4607 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
4611 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
4615 // Write to the same location with the same size.
4616 // Could be improved to assert type sizes are the same (i32 == float, etc).
4620 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
4624 assert(Val.isStrictlyPositive() && "Expect a positive value");
4626 // Positive distance bigger than max vectorization factor.
4629 "LV: ReadWrite-Write positive dependency with different types\n");
4633 unsigned Distance = (unsigned) Val.getZExtValue();
4635 // Bail out early if passed-in parameters make vectorization not feasible.
4636 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
4637 unsigned ForcedUnroll = VectorizationInterleave ? VectorizationInterleave : 1;
4639 // The distance must be bigger than the size needed for a vectorized version
4640 // of the operation and the size of the vectorized operation must not be
4641 // bigger than the currrent maximum size.
4642 if (Distance < 2*TypeByteSize ||
4643 2*TypeByteSize > MaxSafeDepDistBytes ||
4644 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
4645 DEBUG(dbgs() << "LV: Failure because of Positive distance "
4646 << Val.getSExtValue() << '\n');
4650 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
4651 Distance : MaxSafeDepDistBytes;
4653 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
4654 if (IsTrueDataDependence &&
4655 couldPreventStoreLoadForward(Distance, TypeByteSize))
4658 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
4659 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
4664 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4665 MemAccessInfoSet &CheckDeps,
4666 ValueToValueMap &Strides) {
4668 MaxSafeDepDistBytes = -1U;
4669 while (!CheckDeps.empty()) {
4670 MemAccessInfo CurAccess = *CheckDeps.begin();
4672 // Get the relevant memory access set.
4673 EquivalenceClasses<MemAccessInfo>::iterator I =
4674 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
4676 // Check accesses within this set.
4677 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
4678 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
4680 // Check every access pair.
4682 CheckDeps.erase(*AI);
4683 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
4685 // Check every accessing instruction pair in program order.
4686 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
4687 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
4688 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
4689 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
4690 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
4692 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
4703 bool LoopVectorizationLegality::canVectorizeMemory() {
4705 typedef SmallVector<Value*, 16> ValueVector;
4706 typedef SmallPtrSet<Value*, 16> ValueSet;
4708 // Holds the Load and Store *instructions*.
4712 // Holds all the different accesses in the loop.
4713 unsigned NumReads = 0;
4714 unsigned NumReadWrites = 0;
4716 PtrRtCheck.Pointers.clear();
4717 PtrRtCheck.Need = false;
4719 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
4720 MemoryDepChecker DepChecker(SE, DL, TheLoop);
4723 for (Loop::block_iterator bb = TheLoop->block_begin(),
4724 be = TheLoop->block_end(); bb != be; ++bb) {
4726 // Scan the BB and collect legal loads and stores.
4727 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4730 // If this is a load, save it. If this instruction can read from memory
4731 // but is not a load, then we quit. Notice that we don't handle function
4732 // calls that read or write.
4733 if (it->mayReadFromMemory()) {
4734 // Many math library functions read the rounding mode. We will only
4735 // vectorize a loop if it contains known function calls that don't set
4736 // the flag. Therefore, it is safe to ignore this read from memory.
4737 CallInst *Call = dyn_cast<CallInst>(it);
4738 if (Call && getIntrinsicIDForCall(Call, TLI))
4741 LoadInst *Ld = dyn_cast<LoadInst>(it);
4742 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
4743 emitAnalysis(Report(Ld)
4744 << "read with atomic ordering or volatile read");
4745 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
4749 Loads.push_back(Ld);
4750 DepChecker.addAccess(Ld);
4754 // Save 'store' instructions. Abort if other instructions write to memory.
4755 if (it->mayWriteToMemory()) {
4756 StoreInst *St = dyn_cast<StoreInst>(it);
4758 emitAnalysis(Report(it) << "instruction cannot be vectorized");
4761 if (!St->isSimple() && !IsAnnotatedParallel) {
4762 emitAnalysis(Report(St)
4763 << "write with atomic ordering or volatile write");
4764 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
4768 Stores.push_back(St);
4769 DepChecker.addAccess(St);
4774 // Now we have two lists that hold the loads and the stores.
4775 // Next, we find the pointers that they use.
4777 // Check if we see any stores. If there are no stores, then we don't
4778 // care if the pointers are *restrict*.
4779 if (!Stores.size()) {
4780 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
4784 AccessAnalysis::DepCandidates DependentAccesses;
4785 AccessAnalysis Accesses(DL, AA, DependentAccesses);
4787 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
4788 // multiple times on the same object. If the ptr is accessed twice, once
4789 // for read and once for write, it will only appear once (on the write
4790 // list). This is okay, since we are going to check for conflicts between
4791 // writes and between reads and writes, but not between reads and reads.
4794 ValueVector::iterator I, IE;
4795 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
4796 StoreInst *ST = cast<StoreInst>(*I);
4797 Value* Ptr = ST->getPointerOperand();
4799 if (isUniform(Ptr)) {
4802 << "write to a loop invariant address could not be vectorized");
4803 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4807 // If we did *not* see this pointer before, insert it to the read-write
4808 // list. At this phase it is only a 'write' list.
4809 if (Seen.insert(Ptr).second) {
4812 AliasAnalysis::Location Loc = AA->getLocation(ST);
4813 // The TBAA metadata could have a control dependency on the predication
4814 // condition, so we cannot rely on it when determining whether or not we
4815 // need runtime pointer checks.
4816 if (blockNeedsPredication(ST->getParent()))
4817 Loc.AATags.TBAA = nullptr;
4819 Accesses.addStore(Loc);
4823 if (IsAnnotatedParallel) {
4825 << "LV: A loop annotated parallel, ignore memory dependency "
4830 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
4831 LoadInst *LD = cast<LoadInst>(*I);
4832 Value* Ptr = LD->getPointerOperand();
4833 // If we did *not* see this pointer before, insert it to the
4834 // read list. If we *did* see it before, then it is already in
4835 // the read-write list. This allows us to vectorize expressions
4836 // such as A[i] += x; Because the address of A[i] is a read-write
4837 // pointer. This only works if the index of A[i] is consecutive.
4838 // If the address of i is unknown (for example A[B[i]]) then we may
4839 // read a few words, modify, and write a few words, and some of the
4840 // words may be written to the same address.
4841 bool IsReadOnlyPtr = false;
4842 if (Seen.insert(Ptr).second ||
4843 !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
4845 IsReadOnlyPtr = true;
4848 AliasAnalysis::Location Loc = AA->getLocation(LD);
4849 // The TBAA metadata could have a control dependency on the predication
4850 // condition, so we cannot rely on it when determining whether or not we
4851 // need runtime pointer checks.
4852 if (blockNeedsPredication(LD->getParent()))
4853 Loc.AATags.TBAA = nullptr;
4855 Accesses.addLoad(Loc, IsReadOnlyPtr);
4858 // If we write (or read-write) to a single destination and there are no
4859 // other reads in this loop then is it safe to vectorize.
4860 if (NumReadWrites == 1 && NumReads == 0) {
4861 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
4865 // Build dependence sets and check whether we need a runtime pointer bounds
4867 Accesses.buildDependenceSets();
4868 bool NeedRTCheck = Accesses.isRTCheckNeeded();
4870 // Find pointers with computable bounds. We are going to use this information
4871 // to place a runtime bound check.
4872 unsigned NumComparisons = 0;
4873 bool CanDoRT = false;
4875 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
4878 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
4879 " pointer comparisons.\n");
4881 // If we only have one set of dependences to check pointers among we don't
4882 // need a runtime check.
4883 if (NumComparisons == 0 && NeedRTCheck)
4884 NeedRTCheck = false;
4886 // Check that we did not collect too many pointers or found an unsizeable
4888 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4894 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
4897 if (NeedRTCheck && !CanDoRT) {
4898 emitAnalysis(Report() << "cannot identify array bounds");
4899 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
4900 "the array bounds.\n");
4905 PtrRtCheck.Need = NeedRTCheck;
4907 bool CanVecMem = true;
4908 if (Accesses.isDependencyCheckNeeded()) {
4909 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
4910 CanVecMem = DepChecker.areDepsSafe(
4911 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
4912 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
4914 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
4915 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
4918 // Clear the dependency checks. We assume they are not needed.
4919 Accesses.resetDepChecks();
4922 PtrRtCheck.Need = true;
4924 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
4925 TheLoop, Strides, true);
4926 // Check that we did not collect too many pointers or found an unsizeable
4928 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4929 if (!CanDoRT && NumComparisons > 0)
4930 emitAnalysis(Report()
4931 << "cannot check memory dependencies at runtime");
4933 emitAnalysis(Report()
4934 << NumComparisons << " exceeds limit of "
4935 << RuntimeMemoryCheckThreshold
4936 << " dependent memory operations checked at runtime");
4937 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
4947 emitAnalysis(Report() << "unsafe dependent memory operations in loop");
4949 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
4950 " need a runtime memory check.\n");
4955 static bool hasMultipleUsesOf(Instruction *I,
4956 SmallPtrSetImpl<Instruction *> &Insts) {
4957 unsigned NumUses = 0;
4958 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
4959 if (Insts.count(dyn_cast<Instruction>(*Use)))
4968 static bool areAllUsesIn(Instruction *I, SmallPtrSetImpl<Instruction *> &Set) {
4969 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
4970 if (!Set.count(dyn_cast<Instruction>(*Use)))
4975 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
4976 ReductionKind Kind) {
4977 if (Phi->getNumIncomingValues() != 2)
4980 // Reduction variables are only found in the loop header block.
4981 if (Phi->getParent() != TheLoop->getHeader())
4984 // Obtain the reduction start value from the value that comes from the loop
4986 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
4988 // ExitInstruction is the single value which is used outside the loop.
4989 // We only allow for a single reduction value to be used outside the loop.
4990 // This includes users of the reduction, variables (which form a cycle
4991 // which ends in the phi node).
4992 Instruction *ExitInstruction = nullptr;
4993 // Indicates that we found a reduction operation in our scan.
4994 bool FoundReduxOp = false;
4996 // We start with the PHI node and scan for all of the users of this
4997 // instruction. All users must be instructions that can be used as reduction
4998 // variables (such as ADD). We must have a single out-of-block user. The cycle
4999 // must include the original PHI.
5000 bool FoundStartPHI = false;
5002 // To recognize min/max patterns formed by a icmp select sequence, we store
5003 // the number of instruction we saw from the recognized min/max pattern,
5004 // to make sure we only see exactly the two instructions.
5005 unsigned NumCmpSelectPatternInst = 0;
5006 ReductionInstDesc ReduxDesc(false, nullptr);
5008 SmallPtrSet<Instruction *, 8> VisitedInsts;
5009 SmallVector<Instruction *, 8> Worklist;
5010 Worklist.push_back(Phi);
5011 VisitedInsts.insert(Phi);
5013 // A value in the reduction can be used:
5014 // - By the reduction:
5015 // - Reduction operation:
5016 // - One use of reduction value (safe).
5017 // - Multiple use of reduction value (not safe).
5019 // - All uses of the PHI must be the reduction (safe).
5020 // - Otherwise, not safe.
5021 // - By one instruction outside of the loop (safe).
5022 // - By further instructions outside of the loop (not safe).
5023 // - By an instruction that is not part of the reduction (not safe).
5025 // * An instruction type other than PHI or the reduction operation.
5026 // * A PHI in the header other than the initial PHI.
5027 while (!Worklist.empty()) {
5028 Instruction *Cur = Worklist.back();
5029 Worklist.pop_back();
5032 // If the instruction has no users then this is a broken chain and can't be
5033 // a reduction variable.
5034 if (Cur->use_empty())
5037 bool IsAPhi = isa<PHINode>(Cur);
5039 // A header PHI use other than the original PHI.
5040 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
5043 // Reductions of instructions such as Div, and Sub is only possible if the
5044 // LHS is the reduction variable.
5045 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
5046 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
5047 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
5050 // Any reduction instruction must be of one of the allowed kinds.
5051 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
5052 if (!ReduxDesc.IsReduction)
5055 // A reduction operation must only have one use of the reduction value.
5056 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
5057 hasMultipleUsesOf(Cur, VisitedInsts))
5060 // All inputs to a PHI node must be a reduction value.
5061 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
5064 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
5065 isa<SelectInst>(Cur)))
5066 ++NumCmpSelectPatternInst;
5067 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
5068 isa<SelectInst>(Cur)))
5069 ++NumCmpSelectPatternInst;
5071 // Check whether we found a reduction operator.
5072 FoundReduxOp |= !IsAPhi;
5074 // Process users of current instruction. Push non-PHI nodes after PHI nodes
5075 // onto the stack. This way we are going to have seen all inputs to PHI
5076 // nodes once we get to them.
5077 SmallVector<Instruction *, 8> NonPHIs;
5078 SmallVector<Instruction *, 8> PHIs;
5079 for (User *U : Cur->users()) {
5080 Instruction *UI = cast<Instruction>(U);
5082 // Check if we found the exit user.
5083 BasicBlock *Parent = UI->getParent();
5084 if (!TheLoop->contains(Parent)) {
5085 // Exit if you find multiple outside users or if the header phi node is
5086 // being used. In this case the user uses the value of the previous
5087 // iteration, in which case we would loose "VF-1" iterations of the
5088 // reduction operation if we vectorize.
5089 if (ExitInstruction != nullptr || Cur == Phi)
5092 // The instruction used by an outside user must be the last instruction
5093 // before we feed back to the reduction phi. Otherwise, we loose VF-1
5094 // operations on the value.
5095 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
5098 ExitInstruction = Cur;
5102 // Process instructions only once (termination). Each reduction cycle
5103 // value must only be used once, except by phi nodes and min/max
5104 // reductions which are represented as a cmp followed by a select.
5105 ReductionInstDesc IgnoredVal(false, nullptr);
5106 if (VisitedInsts.insert(UI).second) {
5107 if (isa<PHINode>(UI))
5110 NonPHIs.push_back(UI);
5111 } else if (!isa<PHINode>(UI) &&
5112 ((!isa<FCmpInst>(UI) &&
5113 !isa<ICmpInst>(UI) &&
5114 !isa<SelectInst>(UI)) ||
5115 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
5118 // Remember that we completed the cycle.
5120 FoundStartPHI = true;
5122 Worklist.append(PHIs.begin(), PHIs.end());
5123 Worklist.append(NonPHIs.begin(), NonPHIs.end());
5126 // This means we have seen one but not the other instruction of the
5127 // pattern or more than just a select and cmp.
5128 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
5129 NumCmpSelectPatternInst != 2)
5132 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
5135 // We found a reduction var if we have reached the original phi node and we
5136 // only have a single instruction with out-of-loop users.
5138 // This instruction is allowed to have out-of-loop users.
5139 AllowedExit.insert(ExitInstruction);
5141 // Save the description of this reduction variable.
5142 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
5143 ReduxDesc.MinMaxKind);
5144 Reductions[Phi] = RD;
5145 // We've ended the cycle. This is a reduction variable if we have an
5146 // outside user and it has a binary op.
5151 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
5152 /// pattern corresponding to a min(X, Y) or max(X, Y).
5153 LoopVectorizationLegality::ReductionInstDesc
5154 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
5155 ReductionInstDesc &Prev) {
5157 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
5158 "Expect a select instruction");
5159 Instruction *Cmp = nullptr;
5160 SelectInst *Select = nullptr;
5162 // We must handle the select(cmp()) as a single instruction. Advance to the
5164 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
5165 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
5166 return ReductionInstDesc(false, I);
5167 return ReductionInstDesc(Select, Prev.MinMaxKind);
5170 // Only handle single use cases for now.
5171 if (!(Select = dyn_cast<SelectInst>(I)))
5172 return ReductionInstDesc(false, I);
5173 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
5174 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
5175 return ReductionInstDesc(false, I);
5176 if (!Cmp->hasOneUse())
5177 return ReductionInstDesc(false, I);
5182 // Look for a min/max pattern.
5183 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5184 return ReductionInstDesc(Select, MRK_UIntMin);
5185 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5186 return ReductionInstDesc(Select, MRK_UIntMax);
5187 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5188 return ReductionInstDesc(Select, MRK_SIntMax);
5189 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5190 return ReductionInstDesc(Select, MRK_SIntMin);
5191 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5192 return ReductionInstDesc(Select, MRK_FloatMin);
5193 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5194 return ReductionInstDesc(Select, MRK_FloatMax);
5195 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5196 return ReductionInstDesc(Select, MRK_FloatMin);
5197 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5198 return ReductionInstDesc(Select, MRK_FloatMax);
5200 return ReductionInstDesc(false, I);
5203 LoopVectorizationLegality::ReductionInstDesc
5204 LoopVectorizationLegality::isReductionInstr(Instruction *I,
5206 ReductionInstDesc &Prev) {
5207 bool FP = I->getType()->isFloatingPointTy();
5208 bool FastMath = FP && I->hasUnsafeAlgebra();
5209 switch (I->getOpcode()) {
5211 return ReductionInstDesc(false, I);
5212 case Instruction::PHI:
5213 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
5214 Kind != RK_FloatMinMax))
5215 return ReductionInstDesc(false, I);
5216 return ReductionInstDesc(I, Prev.MinMaxKind);
5217 case Instruction::Sub:
5218 case Instruction::Add:
5219 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
5220 case Instruction::Mul:
5221 return ReductionInstDesc(Kind == RK_IntegerMult, I);
5222 case Instruction::And:
5223 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
5224 case Instruction::Or:
5225 return ReductionInstDesc(Kind == RK_IntegerOr, I);
5226 case Instruction::Xor:
5227 return ReductionInstDesc(Kind == RK_IntegerXor, I);
5228 case Instruction::FMul:
5229 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
5230 case Instruction::FSub:
5231 case Instruction::FAdd:
5232 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
5233 case Instruction::FCmp:
5234 case Instruction::ICmp:
5235 case Instruction::Select:
5236 if (Kind != RK_IntegerMinMax &&
5237 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
5238 return ReductionInstDesc(false, I);
5239 return isMinMaxSelectCmpPattern(I, Prev);
5243 LoopVectorizationLegality::InductionKind
5244 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
5245 Type *PhiTy = Phi->getType();
5246 // We only handle integer and pointer inductions variables.
5247 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
5248 return IK_NoInduction;
5250 // Check that the PHI is consecutive.
5251 const SCEV *PhiScev = SE->getSCEV(Phi);
5252 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
5254 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
5255 return IK_NoInduction;
5257 const SCEV *Step = AR->getStepRecurrence(*SE);
5259 // Integer inductions need to have a stride of one.
5260 if (PhiTy->isIntegerTy()) {
5262 return IK_IntInduction;
5263 if (Step->isAllOnesValue())
5264 return IK_ReverseIntInduction;
5265 return IK_NoInduction;
5268 // Calculate the pointer stride and check if it is consecutive.
5269 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5271 return IK_NoInduction;
5273 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
5274 Type *PointerElementType = PhiTy->getPointerElementType();
5275 // The pointer stride cannot be determined if the pointer element type is not
5277 if (!PointerElementType->isSized())
5278 return IK_NoInduction;
5280 uint64_t Size = DL->getTypeAllocSize(PointerElementType);
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 <= 64 && "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 "
5412 "same time. Enable vectorization of this loop "
5413 "with '#pragma clang loop vectorize(enable)' "
5414 "when compiling with -Os");
5415 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5420 int UserVF = Hints->getWidth();
5422 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
5423 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
5425 Factor.Width = UserVF;
5429 float Cost = expectedCost(1);
5431 const float ScalarCost = Cost;
5434 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
5436 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
5437 // Ignore scalar width, because the user explicitly wants vectorization.
5438 if (ForceVectorization && VF > 1) {
5440 Cost = expectedCost(Width) / (float)Width;
5443 for (unsigned i=2; i <= VF; i*=2) {
5444 // Notice that the vector loop needs to be executed less times, so
5445 // we need to divide the cost of the vector loops by the width of
5446 // the vector elements.
5447 float VectorCost = expectedCost(i) / (float)i;
5448 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
5449 (int)VectorCost << ".\n");
5450 if (VectorCost < Cost) {
5456 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
5457 << "LV: Vectorization seems to be not beneficial, "
5458 << "but was forced by a user.\n");
5459 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
5460 Factor.Width = Width;
5461 Factor.Cost = Width * Cost;
5465 unsigned LoopVectorizationCostModel::getWidestType() {
5466 unsigned MaxWidth = 8;
5469 for (Loop::block_iterator bb = TheLoop->block_begin(),
5470 be = TheLoop->block_end(); bb != be; ++bb) {
5471 BasicBlock *BB = *bb;
5473 // For each instruction in the loop.
5474 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5475 Type *T = it->getType();
5477 // Ignore ephemeral values.
5478 if (EphValues.count(it))
5481 // Only examine Loads, Stores and PHINodes.
5482 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
5485 // Examine PHI nodes that are reduction variables.
5486 if (PHINode *PN = dyn_cast<PHINode>(it))
5487 if (!Legal->getReductionVars()->count(PN))
5490 // Examine the stored values.
5491 if (StoreInst *ST = dyn_cast<StoreInst>(it))
5492 T = ST->getValueOperand()->getType();
5494 // Ignore loaded pointer types and stored pointer types that are not
5495 // consecutive. However, we do want to take consecutive stores/loads of
5496 // pointer vectors into account.
5497 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
5500 MaxWidth = std::max(MaxWidth,
5501 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
5509 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
5511 unsigned LoopCost) {
5513 // -- The unroll heuristics --
5514 // We unroll the loop in order to expose ILP and reduce the loop overhead.
5515 // There are many micro-architectural considerations that we can't predict
5516 // at this level. For example, frontend pressure (on decode or fetch) due to
5517 // code size, or the number and capabilities of the execution ports.
5519 // We use the following heuristics to select the unroll factor:
5520 // 1. If the code has reductions, then we unroll in order to break the cross
5521 // iteration dependency.
5522 // 2. If the loop is really small, then we unroll in order to reduce the loop
5524 // 3. We don't unroll if we think that we will spill registers to memory due
5525 // to the increased register pressure.
5527 // Use the user preference, unless 'auto' is selected.
5528 int UserUF = Hints->getInterleave();
5532 // When we optimize for size, we don't unroll.
5536 // We used the distance for the unroll factor.
5537 if (Legal->getMaxSafeDepDistBytes() != -1U)
5540 // Do not unroll loops with a relatively small trip count.
5541 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
5542 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
5545 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
5546 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
5550 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
5551 TargetNumRegisters = ForceTargetNumScalarRegs;
5553 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
5554 TargetNumRegisters = ForceTargetNumVectorRegs;
5557 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
5558 // We divide by these constants so assume that we have at least one
5559 // instruction that uses at least one register.
5560 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
5561 R.NumInstructions = std::max(R.NumInstructions, 1U);
5563 // We calculate the unroll factor using the following formula.
5564 // Subtract the number of loop invariants from the number of available
5565 // registers. These registers are used by all of the unrolled instances.
5566 // Next, divide the remaining registers by the number of registers that is
5567 // required by the loop, in order to estimate how many parallel instances
5568 // fit without causing spills. All of this is rounded down if necessary to be
5569 // a power of two. We want power of two unroll factors to simplify any
5570 // addressing operations or alignment considerations.
5571 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
5574 // Don't count the induction variable as unrolled.
5575 if (EnableIndVarRegisterHeur)
5576 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
5577 std::max(1U, (R.MaxLocalUsers - 1)));
5579 // Clamp the unroll factor ranges to reasonable factors.
5580 unsigned MaxInterleaveSize = TTI.getMaxInterleaveFactor();
5582 // Check if the user has overridden the unroll max.
5584 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
5585 MaxInterleaveSize = ForceTargetMaxScalarInterleaveFactor;
5587 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
5588 MaxInterleaveSize = ForceTargetMaxVectorInterleaveFactor;
5591 // If we did not calculate the cost for VF (because the user selected the VF)
5592 // then we calculate the cost of VF here.
5594 LoopCost = expectedCost(VF);
5596 // Clamp the calculated UF to be between the 1 and the max unroll factor
5597 // that the target allows.
5598 if (UF > MaxInterleaveSize)
5599 UF = MaxInterleaveSize;
5603 // Unroll if we vectorized this loop and there is a reduction that could
5604 // benefit from unrolling.
5605 if (VF > 1 && Legal->getReductionVars()->size()) {
5606 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
5610 // Note that if we've already vectorized the loop we will have done the
5611 // runtime check and so unrolling won't require further checks.
5612 bool UnrollingRequiresRuntimePointerCheck =
5613 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
5615 // We want to unroll small loops in order to reduce the loop overhead and
5616 // potentially expose ILP opportunities.
5617 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
5618 if (!UnrollingRequiresRuntimePointerCheck &&
5619 LoopCost < SmallLoopCost) {
5620 // We assume that the cost overhead is 1 and we use the cost model
5621 // to estimate the cost of the loop and unroll until the cost of the
5622 // loop overhead is about 5% of the cost of the loop.
5623 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
5625 // Unroll until store/load ports (estimated by max unroll factor) are
5627 unsigned StoresUF = UF / (Legal->NumStores ? Legal->NumStores : 1);
5628 unsigned LoadsUF = UF / (Legal->NumLoads ? Legal->NumLoads : 1);
5630 // If we have a scalar reduction (vector reductions are already dealt with
5631 // by this point), we can increase the critical path length if the loop
5632 // we're unrolling is inside another loop. Limit, by default to 2, so the
5633 // critical path only gets increased by one reduction operation.
5634 if (Legal->getReductionVars()->size() &&
5635 TheLoop->getLoopDepth() > 1) {
5636 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionUF);
5637 SmallUF = std::min(SmallUF, F);
5638 StoresUF = std::min(StoresUF, F);
5639 LoadsUF = std::min(LoadsUF, F);
5642 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
5643 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
5644 return std::max(StoresUF, LoadsUF);
5647 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
5651 DEBUG(dbgs() << "LV: Not Unrolling.\n");
5655 LoopVectorizationCostModel::RegisterUsage
5656 LoopVectorizationCostModel::calculateRegisterUsage() {
5657 // This function calculates the register usage by measuring the highest number
5658 // of values that are alive at a single location. Obviously, this is a very
5659 // rough estimation. We scan the loop in a topological order in order and
5660 // assign a number to each instruction. We use RPO to ensure that defs are
5661 // met before their users. We assume that each instruction that has in-loop
5662 // users starts an interval. We record every time that an in-loop value is
5663 // used, so we have a list of the first and last occurrences of each
5664 // instruction. Next, we transpose this data structure into a multi map that
5665 // holds the list of intervals that *end* at a specific location. This multi
5666 // map allows us to perform a linear search. We scan the instructions linearly
5667 // and record each time that a new interval starts, by placing it in a set.
5668 // If we find this value in the multi-map then we remove it from the set.
5669 // The max register usage is the maximum size of the set.
5670 // We also search for instructions that are defined outside the loop, but are
5671 // used inside the loop. We need this number separately from the max-interval
5672 // usage number because when we unroll, loop-invariant values do not take
5674 LoopBlocksDFS DFS(TheLoop);
5678 R.NumInstructions = 0;
5680 // Each 'key' in the map opens a new interval. The values
5681 // of the map are the index of the 'last seen' usage of the
5682 // instruction that is the key.
5683 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5684 // Maps instruction to its index.
5685 DenseMap<unsigned, Instruction*> IdxToInstr;
5686 // Marks the end of each interval.
5687 IntervalMap EndPoint;
5688 // Saves the list of instruction indices that are used in the loop.
5689 SmallSet<Instruction*, 8> Ends;
5690 // Saves the list of values that are used in the loop but are
5691 // defined outside the loop, such as arguments and constants.
5692 SmallPtrSet<Value*, 8> LoopInvariants;
5695 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5696 be = DFS.endRPO(); bb != be; ++bb) {
5697 R.NumInstructions += (*bb)->size();
5698 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
5700 Instruction *I = it;
5701 IdxToInstr[Index++] = I;
5703 // Save the end location of each USE.
5704 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
5705 Value *U = I->getOperand(i);
5706 Instruction *Instr = dyn_cast<Instruction>(U);
5708 // Ignore non-instruction values such as arguments, constants, etc.
5709 if (!Instr) continue;
5711 // If this instruction is outside the loop then record it and continue.
5712 if (!TheLoop->contains(Instr)) {
5713 LoopInvariants.insert(Instr);
5717 // Overwrite previous end points.
5718 EndPoint[Instr] = Index;
5724 // Saves the list of intervals that end with the index in 'key'.
5725 typedef SmallVector<Instruction*, 2> InstrList;
5726 DenseMap<unsigned, InstrList> TransposeEnds;
5728 // Transpose the EndPoints to a list of values that end at each index.
5729 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5731 TransposeEnds[it->second].push_back(it->first);
5733 SmallSet<Instruction*, 8> OpenIntervals;
5734 unsigned MaxUsage = 0;
5737 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5738 for (unsigned int i = 0; i < Index; ++i) {
5739 Instruction *I = IdxToInstr[i];
5740 // Ignore instructions that are never used within the loop.
5741 if (!Ends.count(I)) continue;
5743 // Ignore ephemeral values.
5744 if (EphValues.count(I))
5747 // Remove all of the instructions that end at this location.
5748 InstrList &List = TransposeEnds[i];
5749 for (unsigned int j=0, e = List.size(); j < e; ++j)
5750 OpenIntervals.erase(List[j]);
5752 // Count the number of live interals.
5753 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5755 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5756 OpenIntervals.size() << '\n');
5758 // Add the current instruction to the list of open intervals.
5759 OpenIntervals.insert(I);
5762 unsigned Invariant = LoopInvariants.size();
5763 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5764 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5765 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5767 R.LoopInvariantRegs = Invariant;
5768 R.MaxLocalUsers = MaxUsage;
5772 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5776 for (Loop::block_iterator bb = TheLoop->block_begin(),
5777 be = TheLoop->block_end(); bb != be; ++bb) {
5778 unsigned BlockCost = 0;
5779 BasicBlock *BB = *bb;
5781 // For each instruction in the old loop.
5782 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5783 // Skip dbg intrinsics.
5784 if (isa<DbgInfoIntrinsic>(it))
5787 // Ignore ephemeral values.
5788 if (EphValues.count(it))
5791 unsigned C = getInstructionCost(it, VF);
5793 // Check if we should override the cost.
5794 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5795 C = ForceTargetInstructionCost;
5798 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5799 VF << " For instruction: " << *it << '\n');
5802 // We assume that if-converted blocks have a 50% chance of being executed.
5803 // When the code is scalar then some of the blocks are avoided due to CF.
5804 // When the code is vectorized we execute all code paths.
5805 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5814 /// \brief Check whether the address computation for a non-consecutive memory
5815 /// access looks like an unlikely candidate for being merged into the indexing
5818 /// We look for a GEP which has one index that is an induction variable and all
5819 /// other indices are loop invariant. If the stride of this access is also
5820 /// within a small bound we decide that this address computation can likely be
5821 /// merged into the addressing mode.
5822 /// In all other cases, we identify the address computation as complex.
5823 static bool isLikelyComplexAddressComputation(Value *Ptr,
5824 LoopVectorizationLegality *Legal,
5825 ScalarEvolution *SE,
5826 const Loop *TheLoop) {
5827 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5831 // We are looking for a gep with all loop invariant indices except for one
5832 // which should be an induction variable.
5833 unsigned NumOperands = Gep->getNumOperands();
5834 for (unsigned i = 1; i < NumOperands; ++i) {
5835 Value *Opd = Gep->getOperand(i);
5836 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5837 !Legal->isInductionVariable(Opd))
5841 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5842 // can likely be merged into the address computation.
5843 unsigned MaxMergeDistance = 64;
5845 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5849 // Check the step is constant.
5850 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5851 // Calculate the pointer stride and check if it is consecutive.
5852 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5856 const APInt &APStepVal = C->getValue()->getValue();
5858 // Huge step value - give up.
5859 if (APStepVal.getBitWidth() > 64)
5862 int64_t StepVal = APStepVal.getSExtValue();
5864 return StepVal > MaxMergeDistance;
5867 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5868 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5874 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5875 // If we know that this instruction will remain uniform, check the cost of
5876 // the scalar version.
5877 if (Legal->isUniformAfterVectorization(I))
5880 Type *RetTy = I->getType();
5881 Type *VectorTy = ToVectorTy(RetTy, VF);
5883 // TODO: We need to estimate the cost of intrinsic calls.
5884 switch (I->getOpcode()) {
5885 case Instruction::GetElementPtr:
5886 // We mark this instruction as zero-cost because the cost of GEPs in
5887 // vectorized code depends on whether the corresponding memory instruction
5888 // is scalarized or not. Therefore, we handle GEPs with the memory
5889 // instruction cost.
5891 case Instruction::Br: {
5892 return TTI.getCFInstrCost(I->getOpcode());
5894 case Instruction::PHI:
5895 //TODO: IF-converted IFs become selects.
5897 case Instruction::Add:
5898 case Instruction::FAdd:
5899 case Instruction::Sub:
5900 case Instruction::FSub:
5901 case Instruction::Mul:
5902 case Instruction::FMul:
5903 case Instruction::UDiv:
5904 case Instruction::SDiv:
5905 case Instruction::FDiv:
5906 case Instruction::URem:
5907 case Instruction::SRem:
5908 case Instruction::FRem:
5909 case Instruction::Shl:
5910 case Instruction::LShr:
5911 case Instruction::AShr:
5912 case Instruction::And:
5913 case Instruction::Or:
5914 case Instruction::Xor: {
5915 // Since we will replace the stride by 1 the multiplication should go away.
5916 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5918 // Certain instructions can be cheaper to vectorize if they have a constant
5919 // second vector operand. One example of this are shifts on x86.
5920 TargetTransformInfo::OperandValueKind Op1VK =
5921 TargetTransformInfo::OK_AnyValue;
5922 TargetTransformInfo::OperandValueKind Op2VK =
5923 TargetTransformInfo::OK_AnyValue;
5924 TargetTransformInfo::OperandValueProperties Op1VP =
5925 TargetTransformInfo::OP_None;
5926 TargetTransformInfo::OperandValueProperties Op2VP =
5927 TargetTransformInfo::OP_None;
5928 Value *Op2 = I->getOperand(1);
5930 // Check for a splat of a constant or for a non uniform vector of constants.
5931 if (isa<ConstantInt>(Op2)) {
5932 ConstantInt *CInt = cast<ConstantInt>(Op2);
5933 if (CInt && CInt->getValue().isPowerOf2())
5934 Op2VP = TargetTransformInfo::OP_PowerOf2;
5935 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5936 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5937 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5938 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
5940 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
5941 if (CInt && CInt->getValue().isPowerOf2())
5942 Op2VP = TargetTransformInfo::OP_PowerOf2;
5943 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5947 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
5950 case Instruction::Select: {
5951 SelectInst *SI = cast<SelectInst>(I);
5952 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5953 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5954 Type *CondTy = SI->getCondition()->getType();
5956 CondTy = VectorType::get(CondTy, VF);
5958 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5960 case Instruction::ICmp:
5961 case Instruction::FCmp: {
5962 Type *ValTy = I->getOperand(0)->getType();
5963 VectorTy = ToVectorTy(ValTy, VF);
5964 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5966 case Instruction::Store:
5967 case Instruction::Load: {
5968 StoreInst *SI = dyn_cast<StoreInst>(I);
5969 LoadInst *LI = dyn_cast<LoadInst>(I);
5970 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5972 VectorTy = ToVectorTy(ValTy, VF);
5974 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5975 unsigned AS = SI ? SI->getPointerAddressSpace() :
5976 LI->getPointerAddressSpace();
5977 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5978 // We add the cost of address computation here instead of with the gep
5979 // instruction because only here we know whether the operation is
5982 return TTI.getAddressComputationCost(VectorTy) +
5983 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5985 // Scalarized loads/stores.
5986 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5987 bool Reverse = ConsecutiveStride < 0;
5988 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
5989 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
5990 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5991 bool IsComplexComputation =
5992 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5994 // The cost of extracting from the value vector and pointer vector.
5995 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5996 for (unsigned i = 0; i < VF; ++i) {
5997 // The cost of extracting the pointer operand.
5998 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5999 // In case of STORE, the cost of ExtractElement from the vector.
6000 // In case of LOAD, the cost of InsertElement into the returned
6002 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
6003 Instruction::InsertElement,
6007 // The cost of the scalar loads/stores.
6008 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
6009 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
6014 // Wide load/stores.
6015 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
6016 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
6019 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
6023 case Instruction::ZExt:
6024 case Instruction::SExt:
6025 case Instruction::FPToUI:
6026 case Instruction::FPToSI:
6027 case Instruction::FPExt:
6028 case Instruction::PtrToInt:
6029 case Instruction::IntToPtr:
6030 case Instruction::SIToFP:
6031 case Instruction::UIToFP:
6032 case Instruction::Trunc:
6033 case Instruction::FPTrunc:
6034 case Instruction::BitCast: {
6035 // We optimize the truncation of induction variable.
6036 // The cost of these is the same as the scalar operation.
6037 if (I->getOpcode() == Instruction::Trunc &&
6038 Legal->isInductionVariable(I->getOperand(0)))
6039 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
6040 I->getOperand(0)->getType());
6042 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
6043 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
6045 case Instruction::Call: {
6046 CallInst *CI = cast<CallInst>(I);
6047 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
6048 assert(ID && "Not an intrinsic call!");
6049 Type *RetTy = ToVectorTy(CI->getType(), VF);
6050 SmallVector<Type*, 4> Tys;
6051 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
6052 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
6053 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
6056 // We are scalarizing the instruction. Return the cost of the scalar
6057 // instruction, plus the cost of insert and extract into vector
6058 // elements, times the vector width.
6061 if (!RetTy->isVoidTy() && VF != 1) {
6062 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
6064 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
6067 // The cost of inserting the results plus extracting each one of the
6069 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
6072 // The cost of executing VF copies of the scalar instruction. This opcode
6073 // is unknown. Assume that it is the same as 'mul'.
6074 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
6080 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
6081 if (Scalar->isVoidTy() || VF == 1)
6083 return VectorType::get(Scalar, VF);
6086 char LoopVectorize::ID = 0;
6087 static const char lv_name[] = "Loop Vectorization";
6088 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
6089 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
6090 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
6091 INITIALIZE_PASS_DEPENDENCY(AssumptionTracker)
6092 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
6093 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
6094 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
6095 INITIALIZE_PASS_DEPENDENCY(LCSSA)
6096 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
6097 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
6098 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
6101 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
6102 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
6106 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
6107 // Check for a store.
6108 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
6109 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
6111 // Check for a load.
6112 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
6113 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
6119 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
6120 bool IfPredicateStore) {
6121 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
6122 // Holds vector parameters or scalars, in case of uniform vals.
6123 SmallVector<VectorParts, 4> Params;
6125 setDebugLocFromInst(Builder, Instr);
6127 // Find all of the vectorized parameters.
6128 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
6129 Value *SrcOp = Instr->getOperand(op);
6131 // If we are accessing the old induction variable, use the new one.
6132 if (SrcOp == OldInduction) {
6133 Params.push_back(getVectorValue(SrcOp));
6137 // Try using previously calculated values.
6138 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
6140 // If the src is an instruction that appeared earlier in the basic block
6141 // then it should already be vectorized.
6142 if (SrcInst && OrigLoop->contains(SrcInst)) {
6143 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
6144 // The parameter is a vector value from earlier.
6145 Params.push_back(WidenMap.get(SrcInst));
6147 // The parameter is a scalar from outside the loop. Maybe even a constant.
6148 VectorParts Scalars;
6149 Scalars.append(UF, SrcOp);
6150 Params.push_back(Scalars);
6154 assert(Params.size() == Instr->getNumOperands() &&
6155 "Invalid number of operands");
6157 // Does this instruction return a value ?
6158 bool IsVoidRetTy = Instr->getType()->isVoidTy();
6160 Value *UndefVec = IsVoidRetTy ? nullptr :
6161 UndefValue::get(Instr->getType());
6162 // Create a new entry in the WidenMap and initialize it to Undef or Null.
6163 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
6165 Instruction *InsertPt = Builder.GetInsertPoint();
6166 BasicBlock *IfBlock = Builder.GetInsertBlock();
6167 BasicBlock *CondBlock = nullptr;
6170 Loop *VectorLp = nullptr;
6171 if (IfPredicateStore) {
6172 assert(Instr->getParent()->getSinglePredecessor() &&
6173 "Only support single predecessor blocks");
6174 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
6175 Instr->getParent());
6176 VectorLp = LI->getLoopFor(IfBlock);
6177 assert(VectorLp && "Must have a loop for this block");
6180 // For each vector unroll 'part':
6181 for (unsigned Part = 0; Part < UF; ++Part) {
6182 // For each scalar that we create:
6184 // Start an "if (pred) a[i] = ..." block.
6185 Value *Cmp = nullptr;
6186 if (IfPredicateStore) {
6187 if (Cond[Part]->getType()->isVectorTy())
6189 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
6190 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
6191 ConstantInt::get(Cond[Part]->getType(), 1));
6192 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
6193 LoopVectorBody.push_back(CondBlock);
6194 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
6195 // Update Builder with newly created basic block.
6196 Builder.SetInsertPoint(InsertPt);
6199 Instruction *Cloned = Instr->clone();
6201 Cloned->setName(Instr->getName() + ".cloned");
6202 // Replace the operands of the cloned instructions with extracted scalars.
6203 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
6204 Value *Op = Params[op][Part];
6205 Cloned->setOperand(op, Op);
6208 // Place the cloned scalar in the new loop.
6209 Builder.Insert(Cloned);
6211 // If the original scalar returns a value we need to place it in a vector
6212 // so that future users will be able to use it.
6214 VecResults[Part] = Cloned;
6217 if (IfPredicateStore) {
6218 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
6219 LoopVectorBody.push_back(NewIfBlock);
6220 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
6221 Builder.SetInsertPoint(InsertPt);
6222 Instruction *OldBr = IfBlock->getTerminator();
6223 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
6224 OldBr->eraseFromParent();
6225 IfBlock = NewIfBlock;
6230 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
6231 StoreInst *SI = dyn_cast<StoreInst>(Instr);
6232 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
6234 return scalarizeInstruction(Instr, IfPredicateStore);
6237 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
6241 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
6245 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
6247 // When unrolling and the VF is 1, we only need to add a simple scalar.
6248 Type *ITy = Val->getType();
6249 assert(!ITy->isVectorTy() && "Val must be a scalar");
6250 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
6251 return Builder.CreateAdd(Val, C, "induction");